Voicetechhub Feed

AI Cybersecurity vs. Security for AI vs. Red Teaming

Sun, 11 Jan 2026 21:31:25 GMT

What is AI Cybersecurity and what is Security for AI?

AI cybersecurity is the use of artificial intelligence to protect IT systems, networks, and data from cyber threats.

detect attacks and anomalies in logs and traffic,
identify phishing, malware, and fraud,
prioritize and respond to security incidents faster than humans can.

Overview: AI for Cybersecurity vs. Security for AI (Visualization: gamma.app)

What is Red Teaming?

Red teaming is part of AI security and safety testing: it uses targeted attack scenarios to systematically identify weaknesses in AI systems and derive protective measures from them.

There are many ways to attach AI systems. We will discuss them in the next section.

Important Red Teaming Definitions Business People Need to Know

Prompt Injection & Jailbreaks – How users can trick an AI into ignoring its rules and doing things it should not do
• RAG Poisoning – How false or manipulated documents are injected into a system’s knowledge base to produce wrong answers
• Data Exfiltration – How sensitive data can be leaked through an AI system without being noticed

• PII Leakage – When an AI exposes personal data that should be protected
• Tool & Function-Call Abuse – When an AI is manipulated into executing connected tools or systems in dangerous ways
• Insecure Output Handling – When AI responses are used without being checked for risks such as code, commands, or sensitive content
• Long-Context Contamination – A security and quality issue where earlier or hidden inputs in long contexts push the model into unsafe, incorrect, or manipulable behavior
• Cost Abuse & Denial of Service (DoS) – When an AI system is flooded with requests to drive up costs or make it unavailable

• Hallucination Induction – Techniques that deliberately cause an AI to generate convincing but false information

The AI Security Checklist - How can I Mitigate the Risks of AI Attacks?

If you want to mitigate the risks you'll need to do several actions.

We put together all key points in our AI Security Checklist for you as a guideline:

Do you need a Sparring Partner for AI Security and Red Teaming?

Which AI Presentation Tools are Suitable for Enterprises?

ninhabicht@gmail.com (Nina Habicht) — Tue, 22 Jul 2025 23:40:47 GMT

Which Tools are Best for AI Powerpoint Creation?

We recently wrote about "What are the Best AI Powerpoint Tools" . The AI race in the fields of creative work and AI marketing has evolved significantly since then. That’s why we’re providing an updated overview.

In addition to Gamma.app, there are now many other tools available to help you create stunning, management-level presentations . However, some limitations still remain and it's important to be aware of them.

Check out our helpful comparison table, based on insights from over 6,500 courses we’ve delivered at VOICETECHHUB . Also explore our Corporate Training on AI Management Presentations .

What PowerPoint AI Tools are Suitable for Enterprises?

Most tools are based in the US and therefore process your data abroad. Some platforms – such as Gamma.app have SOC and ISO certifications, but they still handle valuable corporate insights.

Microsoft Copilot is not yet on par with PowerPoint-specific tools like Gamma.app or Canva. In particular, the user experience still needs improvement.

To stay fully on the safe side, always carefully consider what data you choose to share. As the saying goes: If it’s free, you’re likely paying with your data!

AI Presentation Overview

Looking for enterprise-ready AI presentation software? Whether you're a corporate team in the EU, a startup scaling globally, or a consultant seeking structure and speed? Here is our overview.

Microsoft Copilot for PowerPoint
Perfectly integrated into Microsoft 365, Copilot is built for enterprise environments. It's GDPR-compliant , processes data within secure ecosystems, and seamlessly transforms Word or PDF documents into structured slide decks — ideal for large organizations and regulated industries.

Gamma
A standout for data privacy and compliance , Gamma offers ISO27001 and SOC2-certified infrastructure. It’s ideal for companies in Europe or privacy-sensitive sectors that need interactive, AI-driven presentations without compromising data integrity.

Both Copilot and Gamma lead the field in enterprise-level data protection and compliance — making them strong choices for EU-based firms or businesses handling confidential data.

Beautiful.ai
Smart templates, brand-compliant layouts, and automated slide design make Beautiful.ai a great choice for teams needing visually polished business presentations and fast.

Canva Magic Studio
Canva’s Magic Design and Brand Kit features are ideal for creative teams producing social-media-friendly or branded decks with minimal manual effort.

Plus AI
At just $15/month, Plus AI delivers seamless integration with PowerPoint and Google Slides , plus robust charting and template tools. It is great for startups and SMBs.

GenSpark
Ideal for consultants and business storytellers, GenSpark turns raw notes into structured, professional presentations. Its freemium model makes it accessible while offering real impact for daily business use .

AI Powerpoint Tools in a Nutshell - Security and Enterprise-Grade Readiness. Prices may change, consult the provider directly. This table has been automatically translated by our Secure dreamleap AI Suite Translator Agent.

What are the Limiations of AI Presentation Tools?

Limitations of current AI Powerpoint Tools.

This table has been automatically translated by our Secure dreamleap AI Suite Translator Agent .

How to Create Stunning AI-Generated PowerPoint Presentations

Depending on whether you are a little agency, and end user or want to implement powerpoint automations in your company, we recommend different tools.

Register to Get your Personal Step-by-Step Guide based on your PPT AI Challenges.

Are These Reviews Paid?

No – our reviews are independent and unpaid. We regularly test and use a wide range of tools ourselves to provide a neutral, experience-based, and holistic perspective on what's out there.

We value transparency and objectivity above all.

Need tailored advice? Feel free to reach out, we're happy to help .

Which Tool for What? State of AI Tools 2025

ninhabicht@gmail.com (Nina Habicht) — Thu, 08 May 2025 11:27:32 GMT

Should I use several AI tools or stick to one platform?

That's a question I often hear from clients.

�� ? �� .

Ask yourself: What problem are you trying to solve?

Our guideline to be successful with your AI tool journey

1. Start by exploring a few major large language model platforms (ChatGPT, Gemini, Claude, etc.).

- Gemini -> Amazing multimodality, images

- ChatGPT -> Swiss Knife for AI, great for coding, logical and analytical tasks.

- Claude -> Psychological, enhanced writing and strong with coding

2. Once you’ve defined your use case, commit to one main tool and consider upgrading to a paid version for the full experience. Still continue experimenting with specialised tools for certain tasks, so you learn, get ideas and can depriorize certain use cases.

3. Most importantly, invest in learning prompt engineering and focus on solving real problems that deliver value for you or your business and your clients. Sometimes, you don’t even need AI!

Conclusion

Depending on whether you are a user or want to implement AI in your enterprise evaluation of tools differ. Usually enterprise-ready software platforms consider more aspects such as data privacy, security, flexibility, ethical aspects and skill & organizational options within your company.

Contact us if you need more in-depth advice on tools, platforms and successful AI transformation.

How to Use AI Chatbots in Your Business – and Stay Compliant

ninhabicht@gmail.com (Nina Habicht) — Tue, 29 Apr 2025 08:25:12 GMT

AI-powered chatbots, whether developed in-house or deployed through trusted platforms, are revolutionizing customer service, knowledge access, and internal communication.
However, alongside these opportunities come new legal obligations: data protection , transparency , and EU AI Act compliance must be addressed carefully.

This article covers:

Where AI chatbots bring business value
What compliance risks you must manage
How to implement AI chatbots successfully and securely

Why Chatbots Matter in 2025

Typical Use Cases for AI Chatbots

Effective chatbot deployment starts with clear use cases. Whether as customer assistant, HR bot or knowledge manager – options are vast.

Customer service : FAQs, troubleshooting
Lead generation : Engaging website visitors
HR & recruiting : Applications, interview scheduling
Internal support : Policy lookup, onboarding help
Sales & e-commerce : Product advice and returns

Business Benefits of Customized AI Chatbots

Custom AI chatbots offer more than automation – they deliver tangible value by streamlining operations and increasing user satisfaction.

✅ Scalable automation of communication
✅ Faster information for employees and customers
✅ Better service with fewer resources
✅ Innovation through tailored tech solutions

Legal & Compliance Aspects

Compliance is essential for sustainable chatbot use. Here's what to keep in mind:

a) Transparency

Let users know they’re interacting with a company-managed chatbot.
→ Tip: Add a welcome message (“You’re communicating with an AI assistant developed by our organization.”)

b) Data Minimization

Collect only necessary data and avoid saving sensitive info unnecessarily.

c) Vendor Due Diligence

Prefer EU/Swiss-hosted platforms or internal hosting
Ensure GDPR/Swiss DPA compliance
Clarify data processing in contracts

Checklist:

Is data use clearly defined?
Can logs be reviewed and deleted?
Are ISO 27001 standards supported?

d) EU AI Act Risk Classification

For high-risk systems (e.g. recruitment, finance), specific documentation and risk assessments are mandatory.

�� Need help creating internal documentation, AI risk assessments, or role definitions?
We support companies with customized AI Usage Guidelines – structured, ready-to-use, and EU AI Act–compliant.

Implementation Roadmap

Conclusion

Custom AI chatbots are powerful – but only if implemented responsibly.
Companies with strong AI governance & compliance processes will be more trusted, more efficient, and future-ready.

For technical implementation and governance, we work closely with companies - and, if required, with specialized partners such as LEXR for the legal review and safeguarding of your AI processes.

Reasoning vs. Deep Research vs. RAG

ninhabicht@gmail.com (Nina Habicht) — Tue, 25 Mar 2025 19:04:15 GMT

What's RAG?

The goal is for the language model is not to draw on its own knowledge (from the model), but for information to be enriched in the prompt. This is usually your own data you provide to the model (PDFs, systems).

RAG explained with vector databases.

What is Reasoning?

Reasoning models are specialized in drawing logical conclusions, understanding arguments, and evaluating relationships.

Example: Deep Seek R1 or OpenAI Model o1 (e.g. for very complex logical tasks), o3-mini high (e.g. programming) and o3-mini (e.g. for smaller complex task).

How does Deep Research work?

Deep Research is specialized in conducting internet research and compiling as much data and insights as possible in a structured manner. It follows links and opens files on websites, extracts data. Keep in mind that it can take several minutes due to many steps it is executing.

Examples:

Deep Research in OpenAI
Deep Research in Perplexity based on Deep Seek Models

Ideal for:

Extensive Research Tasks
Competitive Analysis
Strategic Analysis
Analyst Work

Who should use Deep Research?

Medical Researchers
University Employees
Marketing Researchers
Market Researchers
Investment Managers and Venture Capitalists
Financial Institutions

Deep Research Explained

Conclusion:

Each method brings a unique set of capabilities to the table, and trying a few out can help you determine the one that aligns best with your tasks. Keep in mind that reasoning models are not always necessary and need many ressources - especially if you do not need to execute logical tasks such as programming. Instead for writing texts and simple summaries use GPT-4o (or alternative models without reasoning components) instead.

Do you want to create your secure Deep Research for your company?

AI Video Creation: The Ultimate Guide to Runway, Luma AI, Haiper.ai, and Hailuo AI

ninhabicht@gmail.com (Nina Habicht) — Sun, 16 Feb 2025 21:32:52 GMT

Why AI Video creation tools are game-changers?

The rise of AI-powered video creation platforms has democratized content creation, enabling individuals and small businesses to produce professional-grade videos quickly and affordably.

These tools come with several advantages:

Speed: AI accelerates the video production process, reducing editing time significantly.
Creativity: AI-powered effects, 3D modeling, and animations help creators achieve complex visuals without technical expertise.
Cost-Effectiveness: AI tools reduce the need for expensive software and hardware, making high-quality video production accessible.

With that in mind, let’s dive into a comparison of the top four AI video creation platforms.

Hailuo video creator tools

1. Runway - The Professional AI Video Suite

AI-Runway is a popular AI-powered video editing platform that offers a suite of tools to create, edit, and enhance videos. It’s particularly favored for its ease of use and flexibility. Runway has gained traction for enabling users to remove backgrounds, create motion graphics, and integrate advanced effects into their videos - all without the need for a green screen or extensive manual editing.

Key Features

Green Screen Tool: Runway’s standout feature is its automatic background removal tool, which eliminates the need for a traditional green screen.
Text-to-Image and Text-to-Video: With Runway, you can create visual assets from text prompts, making it easier to build scenes from scratch.
Inpainting: This feature allows users to modify specific parts of their video frame-by-frame, giving high precision in editing.
Collaboration Tools: Built with teams in mind, Runway offers tools to share projects and collaborate in real time.

Ideal for: Runway is ideal for content creators, marketers, and YouTubers who want a straightforward tool to create professional-looking videos quickly.

2. Luma AI: Revolutionizing 3D with Photorealistic Modeling

Luma AI specializes in 3D modeling and rendering , making it a favorite among creators focused on animation, virtual reality, and augmented reality. Luma AI’s technology allows users to create 3D models from images or videos, generating photorealistic assets that can be used across various media. Its NeRF (Neural Radiance Fields) technology is particularly notable, enabling high-quality 3D assets with relatively little input.

Key Features

NeRF 3D Modeling : Turn 2D photos or video clips into interactive 3D models.
Photo to 3D Model Conversion : Generate 3D models using just a few photos, making it accessible for users without specialized cameras.
AR/VR Compatibility : Export models that are ready to use in augmented and virtual reality applications.
Integrated Workflows
Luma AI integrates with major 3D platforms like Blender and Unreal Engine, streamlining the workflow for professionals.

Ideal for: Luma AI is best suited for professionals in gaming, VR/AR development, and film production who need high-quality 3D models and assets.

3. Haiper AI: Simplified AI for Creative Video Ads

Haiper.ai offers AI-assisted tools for creating short, engaging video ads . This platform emphasizes speed and ease, providing templates and a straightforward editor to produce marketing videos without a steep learning curve. Haiper.ai’s templates are designed to be engaging, making it an ideal choice for businesses looking to create eye-catching promotional content quickly.

Like its "bigger brother" runway it is a entire suite to create movies. Quality-wise - at least in the beginning - it was not comparable to expert narrow AI movie tools or runway.

Key Features

Pre-Made Templates: Choose from a variety of templates designed for different industries and marketing goals.
AI-Powered Suggestions: Haiper.ai can generate suggestions for video edits and effects, allowing users to enhance their content with minimal effort.
Text-to-Video Capabilities: This feature allows you to input text prompts, and the platform will generate a video around the content, including relevant visuals.
Marketing Analytics: Access insights and analytics to measure the performance of your videos.

Ideal for: Haiper.ai is perfect for small business owners, social media managers, and digital marketers who need quick, effective videos for ad campaigns.

4. Hailuo AI: The Artistic AI for Video Content

Hailuo AI is a creative-focused AI platform aimed at helping artists and creators design visually compelling video content. With a range of AI-powered tools, Hailuo AI supports everything from animation to artistic filters, enabling unique and stylized video effects that are particularly suited to social media and short-form content and especially animations.

The small animations are higher in quality than with haiper. Easily upload a midjourney image (have a look at our practical image creation blog posts ).

Key Features

Style Transfer Effects: Apply artistic filters and effects to transform ordinary footage into stylized video art.
Animation Support: Hailuo AI offers tools to animate text and images, making it ideal for creating visual storytelling videos.
Social Media Optimization: Generate content that’s automatically optimized for social media formats.
AI-Powered Editing: Includes tools to automatically adjust color grading, lighting, and other aspects for a polished look.

Ideal for: Hailuo AI is tailored for artists, influencers, and social media creators looking to enhance their visual storytelling with unique effects and animations.

Conclusion: Choosing the Right AI Video Creation Tool

The right AI video creation tool for you will depend on your project goals and your familiarity with video editing and 3D modeling.

Here’s a quick rundown:

Runway is perfect if you need a versatile video editor with AI-driven tools.
Luma AI is ideal if you’re focused on creating realistic 3D models for AR or VR.
Haiper.ai is best for marketers looking for fast and efficient ad creation.
Hailuo AI is suited for artists seeking creative filters and animations for social media.

Each platform brings a unique set of capabilities to the table, and trying a few out can help you determine the one that aligns best with your vision. With the power of AI, creating engaging video content is now easier than ever!

Do you want to deep-dive into AI video creation?

We can provide you enablement options for your Marketing, Martech and Branding teams.

What are the Best AI Powerpoint Tools

ninhabicht@gmail.com (Nina Habicht) — Sun, 16 Feb 2025 21:32:48 GMT

What are useful AI Powerpoint Tools?

While working with several tools we identified beautiful ai and gamma app as great options to become creative and save time with powerpoint generation tasks.

While with beautiful.ai the focus lies on uploading context and creating the powerpoint output as well as ideas for presentations, gamma.app focuses on several user needs, such as website creation, card generations and crafting powerpoints (so a slightly broader option of output formats are available compared to beautiful.ai).

Powerpoints can be easily generated with beautiful ai and gamma app.

What are the drawbacks of AI powerpoint tools?

The problem with AI powerpoint tools is that they tend to be too "generative". As a user you want to slightly change existing slides or refill documents with your information, instead you need to "repair" the generated powerpoint. This in the end takes a lot of time.

However, AI-powered Powerpoint creator tools are excellent for individuals who need to generate presentation ideas or develop storylines with visual inspiration. Their true strength lies in enhancing the creative process rather than simply populating slides with pre-existing data.

Do you want to deep-dive into AI powerpoint creation?

We can provide you enablement options for your Marketing, Martech and Branding teams.

Image Generator and AI Branding Tool Review for Marketing and Brand Managers

ninhabicht@gmail.com (Nina Habicht) — Sun, 24 Nov 2024 22:28:32 GMT

What to Choose: Midjourney, Flux, Getty Images or Adobe Firefly?

We compared several tools already in our all-you-need-to-know session to learn several midjourney prompt techniques .

Today, we want to deep-dive in a more extended overview on creative image generation tools. For that we compared several tools regarding dog images. Most importantly, it's crucial to understand that not every tool is suitable for every application. Some tools are better suited for animals, while others are more optimized for environments, product placements, and similar tasks.

We always used the same prompt: "generate a close lookup of a dog looking happy into the camera fotorealistic sunlight".

Image generated with "imagen" ( Google Deepmind ) vs. Image generated with Midjourney (right). Both show high-resolution and details that are very much outperforming DALL E from OpenAI.

DALL E (OpenAI) Image of a dog right looks like an illustration (painting) while Adobe Firefly provides realistic, however, overly flashy image.

Pros and Cons of Image Creators

We analyzed and used several tools during our Midjourney and image creation trainings and workshops . These are our key learnings:

Midjourney

Strengths

- Ideal for stylized and aesthetic images, such as stock images or mood boards.

- Fast and appealing image generation.

- Works well for creative, artistic projects.

Weaknesses

- Struggles with complex scenes involving multiple subjects and text.

- Text rendering is often inaccurate.

- Usage is only possible with all subscription packages.

FLUX

Strengths

- Offers maximum control for realistic scenarios, such as street photography and product photos.

- High precision in text and brushstrokes.

- Excels in text-based images and provides multiple optimization options since the latest release in November.

Weaknesses

- The 'FLUX Fast' model delivers lower quality in complex scenes compared to other variants.

RunDiffusion

Strengths

- Outstanding for realistic portraits and street photography.

- Produces natural and professional-looking images.

Weaknesses

- Struggles with complex scenes or images involving multiple subjects.

- Lacks detail accuracy in intricate designs.

Firefly (Adobe)

Strengths

- Delivers good results for abstract art and simple scenes.

- Ideal for creative and non-realistic projects.

- Includes stock images that are legally safe for use.

Weaknesses

- Faces and details often appear unnatural.

- Performs poorly on portraits and complex scenes.

Stable Diffusion

Strengths

- Generates decent results for simple, realistic scenes.

- Suitable for less complex tasks and straightforward designs.

Weaknesses

- Faces challenges with complex scenes, proportions, and text generation.

- Details can often lack precision and realism.

Freepik

Strengths

- Performs exceptionally well for product photos and realistic scenes.

- Particularly effective for upscaling images.

- Well-suited for stock image creation.

Weaknesses

- Sometimes produces overdone effects or heavily edited-looking visuals.

Ideogram

Strengths

- Excellent for black-and-white photography and stylized art.

- Shines in abstract projects and creative designs.

Weaknesses

- Struggles with portraits, scenes involving multiple subjects, and complex color palettes.

Imagen (Google)

Strengths

- Top-tier performance in generating highly detailed and realistic pictures.

- Produces visually accurate and high-quality results.

Weaknesses

- Commercial usage is currently unclear.

Bria

Strengths

- Allows for commercial usage, making it ideal for professional projects.

- Well-suited for creating customized visuals.

Weaknesses

- Limited features for creating highly complex designs compared to other tools.

Source: Image Generator Research and Reviews by VOICETECHHUB summarized and visualized with the help of ChatGPT.

Do you need support with Image Creation Skills?

We can provide you enablement options for your Marketing, Martech and Branding teams.

Optimizing your Website for AI: How to get found by ChatGPT

ninhabicht@gmail.com (Nina Habicht) — Sat, 24 Aug 2024 23:07:58 GMT

In today's digital landscape, it's no longer sufficient to optimize solely for traditional search engines. With the rise of AI tools like ChatGPT and Perplexity AI (Check our blog post How does Perplexity work? ) Search Engine Optimization is shifting towards the need to be provide "AI compatible content" . It's crucial to rethink your SEO strategies. Here are some surprising tips and tricks that will impress even the most seasoned SEO experts.

What do Users Tell on ChatGPT vs. Google Search?

How Swiss People Search with AI Tools

In Switzerland, one third are using ChatGPT for search reasons. The latest report from Nau Media AG reveals surprising facts from Switzerland:

- 2/3 have not yet used ChatGPT

- 75% use Google for searches

- 25% use ChatGPT for searches

Critical research should always be done, regardless of the tool used.

But what does AI mean for Company Inbound Marketing?

We see why companies should be present in AI tools and write their content to be "AI-compatible."

Perplexity AI should be added as a relevant tool for AI Research & Search Tools as well as "GPTs" from OpenAI where companies can build their own company or task-specific AI (Check our articles about Are G PTs relevant for my AI Strategy? Or Comprehensive Guide to MyGPTs and Alternatives . Here, you can see the sources of articles and directly display graphics. Moreover, websites will soon be directly created using AI, which means companies could potentially have advantages if they are represented within these AI Tools.

Worldwide studies show happy users with AI Search

A study comparing Google Search vs. ChatGPT found that: Users spent less time on tasks when using ChatGPT compared to Google Search ( [2307.01135] ChatGPT vs. Google: A Comparative Study of Search Performance and User Experience (arxiv.org) )
ChatGPT's responses were perceived as having higher information quality ( [2307.01135] ChatGPT vs. Google: A Comparative Study of Search Performance and User Experience (arxiv.org) )
Users reported better experiences with ChatGPT in terms of usefulness, enjoyment, and satisfaction ( [2307.01135] ChatGPT vs. Google: A Comparative Study of Search Performance and User Experience (arxiv.org)
Both tools had similar levels of user trust .

Statistics 2024 - Google vs. ChatGPT vs. other Platforms (Source: Number of ChatGPT Users (Aug 2024) (explodingtopics.com) )

How to optimize Websites to be found by Large Language Models?

The shift from traditional search engines forces us to redefine SEO strategies. Instead of merely optimizing for search engine algorithms, you need to create content that AI models can easily interpret.

1. Quality and Relevance of Content

AI models like ChatGPT prioritize contextually relevant and high-quality content. Invest in well-researched and engaging content that offers genuine value. The focus should be on user experience rather than just ranking for specific keywords.

2. Leveraging Natural Language Processing (NLP)

To align with AI models, content must be optimized for natural language processing (NLP). Understand how AI interprets context, semantics, and user intent. Use a conversational tone, answer potential follow-up questions, and cover various aspects of a topic to enhance your content's suitability for AI-driven information retrieval.

3. Provide Bullet Points and Clear Structures

Structured data helps search engines understand the context of content, and it remains relevant in an AI-driven landscape. AI models benefit from well-organized data to provide accurate and comprehensive responses. Implement structured data markup, such as schema.org, to increase the visibility and accessibility of your content for AI models.

4. Embracing Voice Search and Conversational AI

Voice search is on the rise. Generating a podcast to already have contents with voice can be a great AI marketing strategy. AI models like ChatGPT are perfect for voice search because they can generate conversational responses. Optimize your content for voice search by anticipating spoken queries and providing concise, direct answers.

The same applies to chatbots. By using bots, you can gather requests from your clients, directly provide responses and optimize you data base (content of your website or app) by the help of data-driven chatbot analytics. Thus, chatbots are a great method to optimize your content strategy in order to be user-centered and "AI compatible".

5. Avoiding Too Short or Too Long Text Blocks

Very short text can be counterproductive. We see this often in E-Commerce stores, where a single images is provided with a very short title. Finding answers in these short titles is of course difficult - meaning if you do not provide enough information - AI will not be able to retrieve enough information to generate a company-specific answer. You could use the Generic Model of GPT or any LLM to retrieve additional information from the web, however, companies often want to have accurate and company-specific information (especially webshops cannot afford to display wrong information on a product price or product description.

Long text blocks without subheadings can be difficult for LLMs to process since the longer the text the more the LLM system can be "confused". Such websites are generally less likely to be considered by LLMs than those that provide quick, understandable answers. It is recommended to start longer texts with a brief summary of the main points.

6. Avoid Dublicate Information and Spread Content

To make the answer clear and avoid hallucinations or missing information avoid information being spread across long or multiple websites. Keep a clear title, and information block so that chunking (the LLM needs to take the correct "chunks" per question) is done properly. This is especially the case with PDFs, where an answer may be between page 1 and 2 since the text is starting on the page 1 and continues on the page 2. Avoid this scenario. Keep one answer on one page and the structure consistent (e.g. Title - Answer) across all contents of your website.

7. Using Quotes of Experts Using Clear and Information-Rich Language

Including quotes and citations can also be beneficial. Experts in LLM optimization believe that quotes from well-known figures can positively impact discoverability and relevance in LLM results.

8. Be present as a Brand on AI Platforms

We will see more and more brands being present on platforms such as Perplexity AI. What about creating your own GPT on the OpenAI ecosystem? Many companies already created their own GPT. This similar to being present in the App-Store.

9. Keep your Website Up-to-Date

LLMs may retrieve old data from your meta data or contents. Old and unstructured meta data can lead to wrong answers in ChatGPT or other AI tools. Make sure your website is always up-to-date.

How will AI SEO look like in the Future?

Experience in search engine optimization (SEO) shows that it is an ongoing process. The same applies to Large Language Model Optimization or Voice Search Optimization ( Learn more how to optimize your website for search in our blog series (German only but you can probably translate it with ChatGPT ;)) where continuous efforts are key to success.
It is important to keep an eye on both technological advancements and the activities of competitors.
We will see AI Ads emerge as well where interactive and multimodal experiences will be part of the new AI branding strategies. This means AI tech providers will probably offer "Ads like" offers, to be present on certain platforms, in AI search queries and in movies or images.

Do you need support with Large Language Optimization?

We can provide you content creator tools and guidance to make your content ready for the next-generation AI search tools.

A Practical Guide for Midjourney Image Generation

ninhabicht@gmail.com (Nina Habicht) — Sat, 24 Aug 2024 20:46:40 GMT

Since August 2024, users have been able to use the web version of the image creation tool Midjourney. This simplifies usage by providing a user-friendly interface to experiment with one of the top Generative AI image creation tools available. We tested it for you and are sharing helpful tips and tricks.

How to prompt images with Midjourney?

If you use Midjourney on discord, there is a clear prompt structure and prompt parameters to adhere to. Usually, it makes sense to stick to it:

1) To prompt use "/Imagine"

2) Then enter your subject (description and details) you want to see on the image and it's environment (see yellow highlighted below in the prompt example)

3) Then enter composition, lightning, colours (see green highlighted below in the prompt example)

4) Finally add technical parameters to adjust and finalize your image. Please find a useful parameter library here.

Prompt Structure with Discord Midjourney Bot

Important techinical parameters for Midjourney prompt engineering

What is different with Midjourney Website?

The installation in discord and especially the technical parameter handling and prompting is much easier with the Midjourney Web Version since you do not need to know all parameters by heart)his is how the new Midjourney Website looks like. You can find the web version here: Midjourney Feed .

Main Site of Midjourney Website: Explore or Create.

Parameter Settings in Midjourney Website

If you want to create photorealistic pictures use the parameter "raw mode"
If you are looking for more picture-style images, use "Standard"
Stylization is like providing the model "free options" to come up with something independent of the original prompt. Putting it low make the system more consistent and less artistic.
Weirdness renders your picture weird
Variety is the "chaos" parameter, making images less unexpected and unusual compositions.

Parameter Settings. Check for "Raw" Mode to create photorealistic images.

How to use styles with in Midjourney

sref codes are specific parameters used in Midjourney to apply preset styles to generated images. Each code corresponds to a particular visual effect or aesthetic. To apply an SREF code, simply add it to your prompt when generating an image. For example: A serene landscape --sref 19961010.

Find a full library of amazing sref codes here: How to Use Style Reference(--SREF) in Midjourney V6.1 and V6 (midjourneysref.com)

How to change parts in an image with Midjourney (Inpainting)

Use the inpainting function to change parts of your image.

Exchange cloths, or anything you want to.

Changing parts of a face with inpainting functionality. It sometimes does not work consistently and needs several trials.

How to create consistent styles with Midjourney Web Version

Upload the image as a reference image. Watch the video below to see how it works:

1) Drag and drop a reference image (blue pattern)

2) Define whether you want to do a cref or sref. Character reference (cref) feature allows users to create consistent characters without distortions by matching a reference image.

3) Write your prompt

Choose an image as a reference for your prompt.

How to create (website) patterns with Midjourney

Use the parameter "--tile" to generate a beautiful pattern. I generated some amazing patterns of the black forest in Germany:

Do you want to create professional images with AI?

We can provide you guidance in creating images or brand assets for your company.

Understanding Agentic AI: A Game-Changer for the Future of AI

ninhabicht@gmail.com (Nina Habicht) — Sun, 11 Aug 2024 22:40:30 GMT

Agentic AI: A Simple Guide

What is Agentic AI?

Agentic AI represents a sophisticated class of artificial intelligence systems designed to function as autonomous agents. It is when we put LLM in charge of the logic and RAG. They h

1) Agents come up with a plan

2) Agents use tools like calculators or a certain code or other language model, APIs

3) Agents can access memory: They can think through a problem and this thinking and history can be stored and retrieved.

These systems can perform tasks, make decisions, and interact with their environments independently, without the need for direct human intervention.

Unlike traditional AI models that follow predefined rules or Generative AI that focuses on creating new content, Agentic AI emphasizes goal-oriented behavior and adaptive decision-making. These systems utilize advanced algorithms and sensory inputs to execute actions in real-time, learning and optimizing their performance through continuous feedback. Applications of Agentic AI are diverse, ranging from robotics and autonomous vehicles to intelligent personal assistants and complex simulation environments.

How do Agentic AI Systems think?

You also refer to "Think slow" approach with Agentic AI. Thinking slow means the llm thinks about the plan and the way forward and not the human.

1) User Query is given to the Agent

2) Agent Plan / Think

3) Agent Acts -> by using tools (APIs)

4) Agent observes the answer and decides whether iterations are needed to get the best and final answer.

5) Agent generates the final answer for the user

What is the Difference between Agentic AI and Generative AI?

Agentic AI and Generative AI serve different purposes within the AI landscape. While Agentic AI focuses on creating autonomous agents capable of performing tasks and making decisions independently, Generative AI is centered on creating new content such as text, images, or music by learning patterns from existing data. Generative AI excels in producing creative outputs, whereas Agentic AI is characterized by its decision-making and action-taking capabilities.

Instead of just asking the LLM, we ask the Agent to come up with the entire plan and idea as well as the answer.

Robotic Process Automation (RPA) , on the other hand, automates specific tasks based on predefined rules. Unlike RPA, Agentic AI operates autonomously, making complex decisions and taking actions based on continuous learning and analysis of vast datasets. This capability not only enhances operational efficiency and accuracy but also empowers organizations to navigate regulatory complexities and mitigate risks more effectively.

RAG Architecture Simplified (Source: Voicetechhub)

Agentic AI Architecture Simplified (Source: Voicetechhub)

Differences between Definitions: Gen. AI vs. Agentic AI vs. RPA (Source: Voicetechhub)

Agentic AI or Chatbots?

Agentic AI is a natural progression from conversational AI. What started as simple chatbots have evolved into sophisticated systems capable of understanding natural language and making autonomous decisions. Whilte the term "chatbot" is still used for systems using NLP, the word "agents" is more used for digital assistants or conversational interfaces that consists of several inelligent "subbots" where each has its specific fields of responsibility and automation. What is new, is that these agents can interact with each other - call each other via APIs. They fulfill complex processes and tasks without human-in-the-loop. For example: an agent for image creation get's called by an agent creating an article about your weekend plans. Both agents will finalize your whole request without the need of you creating each step, prompting and correcting. So-called Large Action Models (LAMs) extend beyond basic query-response mechanisms, enabling complex actions and workflows without human intervention.

What is a Large Action Model (LAM)?

A Large Action Model (LAM) is an advanced AI framework designed to perform complex tasks and workflows autonomously. Unlike basic AI systems that rely on simple query-response mechanisms, LAMs leverage extensive datasets and sophisticated algorithms to execute intricate operations without human oversight. These models integrate various AI technologies, including deep learning and natural language processing, to comprehend contexts and make informed decisions in real-time. By doing so, LAMs can handle multifaceted tasks across different domains, from automating business processes to enhancing security measures.

Agentic RAG vs. Native RAG

Compared to Native Retrieval Augmentic Generation Agentic RAG is an advanced, agent-based approach involving

a flexible framework that supports complex tasks, multi-step reasoning, tool use, and learning over time.

Agentic vs. Native RAG: Build an Agentic RAG using HuggingFace Transformers Agent | by Plaban Nayak | The AI Forum | Jul, 2024 | Medium

When not to Use Agentic AI?

For certain narrow problems you should go the programmatic route (not using llms to decide), because every single problem will be solved the same way. This approach is faster in this case, especially when you have client-facing chatbots that need to react fast to provide best user experience. Also in the banking and finance or very data-sensitive industries programmatic approaches may be more safe. If you expect a system to solve big complex tasks (e.g. solve support system issues) then Agentic AI may be the better approach. Also you avoid progamming every path that could happen in your (chatbot) system.

LLM Benchmarks: Finding the right LLM for your Needs

ninhabicht@gmail.com (Nina Habicht) — Mon, 29 Jul 2024 22:40:21 GMT

Discover the Right Large Language Model for Your Business

Choosing the right Large Language Model (LLM) for your business can be a daunting task. With numerous options available, it's crucial to understand how to evaluate these models effectively. This article will guide you through the best sources for LLM benchmarks and provide insights into selecting the ideal model for your needs.

LLM Benchmarks: Best Sources

To make an informed decision, it's essential to rely on credible sources for LLM benchmarks.

Here are some of the best resources:

These sources provide comprehensive evaluations of various LLMs, helping you understand their performance across different tasks.

Evaluating Large Language Models

When evaluating LLMs, it's important to consider both functional and non-functional criteria. Functional criteria focus on the model's performance in specific tasks, while non-functional criteria include aspects like data security, deployment options, and cost. Benchmarks to consider from a data science perspective are the BLEU and ROUGE LLM evaluation metrics.

Functional Criteria of LLM Benchmarks

Functional criteria involve assessing the model's ability to perform tasks such as text generation, translation, and summarization. Benchmarks like SuperGLUE and LMSys provide detailed performance metrics for these tasks.

Non-Functional Criteria of LLM Benchmarks

Non-functional criteria are equally important and include:

Data Security: Ensure the model complies with data protection regulations.
Deployment Options: Decide whether to use an API or deploy the model on your own infrastructure.
AI Act & Compliance Regulations: Be aware that some models are not available in certain countries due to regulations.
Cost: Evaluate the cost per token and overall expenses associated with the model.

How to Select the right LLM for your Purpose

To select the best LLM for your business, consider the following steps:

Identify Your Needs: Determine the specific tasks you need the LLM to perform.
Consult Benchmarks: Use the sources mentioned above to compare the performance of different models.
Evaluate Non-Functional Criteria: Assess the deployment options, data security, and cost implications.
Test Models: Conduct your own tests to see how well the models perform in your specific use case.

Source: Mistral Large 2 vs Llama 405B vs GPT-4 vs Claude 3.5 Sonnet: Real Life Comparison | by Towards AGI | Jul, 2024 | Medium

Finetuning or Not? Open vs. Proprietary LLMs

ninhabicht@gmail.com (Nina Habicht) — Tue, 25 Jun 2024 09:46:08 GMT

Open AI Models vs. Proprietary Models at a Glance

In the rapidly evolving field of artificial intelligence, the trend toward open models is gaining significant momentum. Major tech giants like Meta and Mistral are pioneering this movement, with even Google joining the fray by opening up more with its Gemini project . But why is the openness of AI models such a crucial component for fostering innovation?

The Rise of Open Models

The concept of open AI models revolves around making the underlying models and data accessible to a wider community. This transparency enables researchers and developers to build upon existing frameworks, thus accelerating the pace of innovation. One of the key players in this domain is BigScience, a collaborative initiative aimed at creating a counterpart to OpenAI's GPT-3. This project brought together over 1,000 collaborators and spanned 59 languages, utilizing a massive 350 billion token dataset . The resulting model, with its 176 billion parameters, was trained on a French supercomputer, illustrating the scale and complexity involved in such endeavors.

Switzerland's Alps Supercomputers: The Backbone of Open Model Training

Training large open models requires substantial computational power, often provided by supercomputers. Switzerland’s own Alps supercomputer is a prime example, offering the necessary infrastructure to handle the extensive processing demands. These supercomputers facilitate the use of advanced parallelism techniques essential for efficient model training.

What is the Difference between Finetuning and Prompt Engineering?

Finetuning requires modifying the pre-trained model's parameters to adapt it to a specific task, whereas prompt engineering adjusts the input prompts to guide the model's output without altering its parameters. The need for finetuning depends on the circumstances and use case you are facing. Ask yourself first:

For instance, are you using models in a cloud environment? Then you can also do finetuning on a cloud-environment. Keep in mind: Not all models or types (e.g. GPT-4o) are available for fintetuning with cloud providers.
For instance, if you are using models in a cloud environment, extensive GPU or TPU resources are necessary for finetuning. These resources can be provided by your cloud provider or set up in your hosting environment.
If you cannot buy GPUs you can use Consumer GPUs (GTX 3080). Usually you do not train a model from scratch as this takes weeks.
Alternatively, you can buy processing power on virtual machines such as vast.ai which is cheaper for first experiments with finetuning models.

When to Use Finetuning?

Finetuning and Retrieval-Augmented Generation (RAG) are two different approaches for improving the performance of language models. Here's when to use finetuning:

1. Specific Knowledge Domain: If your task requires a deep understanding of a specific domain with specialized terminology or context, finetuning a model on a relevant dataset can significantly improve performance.

2. Task-Specific Adaptation: For tasks like sentiment analysis, text classification, named entity recognition, or other specific NLP tasks, finetuning allows the model to learn patterns and nuances unique to that task.

3. Custom Outputs: When you need the model to generate very specific types of responses or outputs that are not well-captured by a general model, finetuning helps tailor the responses.

4. Consistent Performance: If you require consistent and reliable performance on a particular task, finetuning ensures the model is optimized for that specific task, leading to more predictable results.

When is RAG (Retrieval-Augmented Generation) Enough?

1. Broad Information Retrieval: If your task involves answering questions, generating text based on external documents, or requiring up-to-date information, RAG is useful as it combines a generative model with a retrieval mechanism to fetch relevant documents.

2. Dynamic Knowledge Updates: When the information changes frequently or the task requires access to a large and dynamic knowledge base, RAG can pull in the most current and relevant information at the time of the request without needing to retrain the model.

3. Low Resource Availability: If you lack the computational resources or large, high-quality datasets necessary for finetuning, RAG can still provide high-quality outputs by leveraging external knowledge sources.

4. General Purpose Use: For more general-purpose applications where the model needs to handle a wide range of topics and the precise tailoring of responses is less critical, RAG can be sufficient and more efficient.

Use finetuning when you need highly specialized, task-specific, and consistent performance, and you have the necessary resources and data for training. Use RAG when you need flexible, up-to-date information retrieval and generation capabilities, especially if you lack the resources for extensive model training.

When RAG is not Enough?

RAG can be limited, especially if you want to build models that are very specific for a certain task (e.g., generating passwords). RAG can be constrained by the context window of the underlying model, which limits the amount of data or number of articles it can effectively use for generating responses. The retrieved documents are integrated into the prompt, affecting the model's performance and context management.

What do you Need to Consider for Model Training When Doing Finetuning?

In order to calculate all parameters (weights) - since not all parameters can be processed on a single GPU - fast parallel processing is needed. There are specific methods for achieving this:

1. 3D Parallelism: This method distributes data across multiple GPUs, allowing each instance to handle a different chunk of the dataset.

2. Pipeline Parallelism: Here, different layers of the neural network are assigned to different GPUs, ensuring a balanced load across the system.

3. Tensor Parallelism: This technique slices the computational tasks horizontally across the layers, distributing matrix multiplication tasks among various GPUs.

A significant challenge in training open models is acquiring a robust and diverse dataset. BigCode tackled this by leveraging the GitHub Archive to collect the names of over 200 million repositories. The process involved several months of cloning repositories, followed by meticulous filtering to remove binary data and irrelevant files. The final dataset, amassing 3 terabytes of data, aimed to rival the quality of models like GitHub's Copilot.

Empowering the AI Community Through Open Models

Open models democratize the field of AI, empowering a broader community to fine-tune these models for specific use cases. This approach leads to the creation of smaller, more affordable models tailored to particular needs, driving innovation at a grassroots level. By providing the tools and frameworks needed to adapt and improve AI models, the community can explore a myriad of applications, from language translation to specialized industry solutions.

Conclusion

The push towards open AI models represents a pivotal shift in the AI landscape. By leveraging the power of supercomputers and collaborative efforts, the development of open models is not only feasible but also immensely beneficial. As more organizations and researchers join this movement, the potential for innovation expands exponentially, heralding a future where AI technology is more accessible, adaptable, and capable of addressing a diverse array of challenges.

Prompt Engineering for Product Managers

ninhabicht@gmail.com (Nina Habicht) — Fri, 10 May 2024 21:31:15 GMT

How can ChatGPT Support Product Managers?

ChatGPT can support product managers in different areas , such as:

1) Ideation: Generate ideas and brainstorm

2) Provide competitive edge and market insights

3) Providing first code and mockups

4) Prepare complex discussions with various stakeholders

5) Generate user stories, requirements documents

This article outlines the most useful prompts for AI and Non-AI Product Managers.

Top ChatGPT Prompts for Product Managers

How to Gain Market Insights Through Prompt Engineering?

Please find some helpful prompts to gain insights about your market.

Prompt: What could be the total addressable market (TAM) for {your product idea}. Segment the market in sub-categories and include Go To Market strategies as well. Or: Prompt: What are the current market trends related to {your product}?
Prompt: Use the information I am giving below and tell me the pros and cons of using this product: {Paste the whole homepage content from your competitor’s website}. Use Perplexity AI to gather more sources in detail.
Prompt: Perform {SWOT, PESTLE, Porter’s 5 force} Analysis on {Competitors Name}.
Prompt: Have there been similar products to {our product} launched in the past?
Prompt: What is the projected market size for {your product} market?

How to Make a Competitor Analysis with ChatGPT?

Prompt: Who are the main competitors of {market} or {product}? List them in a table.
Prompt: How is my product different from what is already existing in the market?
Prompt: What are monetization strategies for products similar to {your product}?

How to Create Requirements and User Stories with ChatGPT?

Prompt: I am creating a product with the following features: {Paste features here}. Can you help me write user stories from this information?
Prompt: What might be some good user stories to include in a Product Requirements Document for {product idea}

Sources: VOICETECHHUB AI Product Training

How to Adapt the MOM Test with ChatGPT

"The Mom Test" is a book written by Rob Fitzpatrick that deals with the challenge of gathering honest and constructive feedback about business ideas. The title reflects the concept that the questions proposed in the book are so well-crafted that not even your mother could give you a biased or overly positive response, hence, they pass "The Mom Test."

Prompt: Using the principles from the MOM Test book, could you suggest 5 questions that I should ask my users of {product XY}

How to Assess Product Risks with ChatGPT?

Prompt: What are risk associated with {your product}
Prompt: Are there any known regulations or compliance issues related to {your product}?
Prompt: How have similar products managed regulatory challenges?
Prompt: Are there any known regulations or compliance issues related to {your product}?

We support you with guidance when it comes to any Generative AI and AI challenges. Meet the Voicetechhub team today!

How to Handle Stakeholder Management With ChatGTP

As a product manager, team lead, or expert, you can use prompts to "put on the hat" of your stakeholders and see how they react. This helps you prepare for difficult meetings and gain all perspectives relevant to your product development

Source: VOICETECHHUB AI Prompt Engineering Training for Product Managers

How Does Perplexity AI Work?

ninhabicht@gmail.com (Nina Habicht) — Sat, 04 May 2024 07:34:26 GMT

What is Perplexity AI?

Perplexity AI is the "New Google" or let's say AI enabled search. It is an intelligent AI search application that helps users to receive answers when prompting with very precise citations compared to ChatGPT and other large language user interfaces.

How does Perplexity differ from Google?

In comparison, Google search results are retrieved through search engine mechanisms, employing indexing, rankings, and engagement rates of internet content, whereas Perplexity first searches for potential sources on the internet and then delivers optimal and precise answers using extensive language models and Natural Language Processing.

When to use Perplexity and when ChatGPT?

Perplexity is ideal if you want to research citations for your academic work or ensure the precision of your references. However, you can also engage in casual conversation with it, similar to ChatGPT. We highlight its user-friendly interface, which automatically integrates movie and picture sources.

Moreover, our experiments have shown that Perplexity performs notably better, particularly when seeking recent updates, as it utilizes real-time retrieval for accurate responses. This stands in contrast to ChatGPT, which is trained only up to a certain point (e.g., March 2023). Consequently, knowledge beyond March 2023 is unavailable when interacting with ChatGPT, whereas Perplexity provides current and relevant sources.

Both AI platforms offer a Pro version. This is necessary if you want to generate pictures or upload files with Perplexity or ChatGPT. With Perplexity we highlight the fact that you can choose between different LLMs (e.g, Claude, Sonar, Llama) which provides you the possibility to upload large files. Please find more in this article about the differences between Claude, Gemini, and other Large Language Models .

What are Perplexity Pages?

The AI search/research tool Perplexity now allows users to dynamically create and share websites with others. This feature has been rolled out to specific users.

Website builders like Wix and WordPress are facing new competition: AI searches now also offer website builders! Next step: If a company is not present with content in these tools, will they be found less often?

Tools like Notion can also create AI-generated web pages. This is not new, but the interesting approach is that a research & search tool now offers this feature, and you can interact with Perplexity pages, which was previously only possible on static websites through chatbots or special widgets.

UX and CX experts are already working on implementing conversational interfaces, as the future is clearly becoming more interactive.

Fun fact: Google is displaying parts of Perplexity pages. The question is how long this will last, according to Search Engine Land: "Spam? This definitely seems like a case of one search engine spamming another search engine." according

The future is moving towards dynamic, interactive websites that you can "talk" or "chat" with. Check this perplexity page for a first "look and feel": https://www.perplexity.ai/page/Beginners-Guide-to-duTB.6CpSbGnPBcXFe8A2A

We support you with guidance when it comes to Generative AI challenges. Meet the Voicetechhub team today.

How to Choose between ChatGPT, Claude, Copilot and Gemini

ninhabicht@gmail.com (Nina Habicht) — Sat, 04 May 2024 07:31:51 GMT

What makes more sense for my team? ChatGPT Team or Claude Team?

One platform is on the rise: Claude is an impressive model when it comes to handling large context windows (with around 200K tokens it can load and analyze huge documents), however, it does not provide internet access (unlike specific AI search tools such as Perplexity AI). Claude excels in writing and will soon also have access to codebases and CRMs. ChatGPT, on the other hand, also supports native voice interaction, allowing you to speak and learn languages with ChatGPT Plus, and it offers enhanced analytical functionalities. The so-called company-specific GPTs are available only with ChatGPT Plus and Team, and are not supported by Claude Team. Claude's prompt library shares some similarities with the GPT marketplace. ChatGPT Teams allows your team to have shared GPTs.

Again, whether GPTs are an important part of your AI company strategy needs to be determined. In the US, GPTs seem to carry more weight, whereas in Europe, smaller agencies and companies use GPTs to experiment. We will see how the importance of GPT's marketplace evolves in Europe in the near future.

We've summarized the differences for you. On the left-hand side, you see the model differences for the normal versions, while on the right-hand side you see the differences for the team version. We will definitely see more exciting enhancements of Claude with it's CRM capabilities soon.

Sources: Claude AI vs. ChatGPT: How do they compare? | TechTarget

How does Gemini Perform vs. ChatGPT and Copilot?

Like ChatGPT Plus, GemLike ChatGPT Plus, Gemini Pro is focusing on its multimodality - meaning that, compared to Claude and Perplexity (check also "What is Perplexity and how does it work" ), these two models provide a platform capable of handling several advanced features with different formats of data, such as video, audio, text, analytics, and mathematics. Copilot is also focusing on covering many use cases with its unique ability to integrate into Office and Edge browser applications. This enables users to utilize General AI directly within the applications they are currently using. However, the Copilot Edge version is slightly less performant in tasks such as writing, image generation, and analytics than ChatGPT Plus, as it is based on an older version of GPT (Codex, from the GPT-3 family). One significant advantage is its direct internet access via Bing search. Some of our students and clients have found Copilot to be very "direct," whereas Claude was experienced as rather too "cautious.

Please find below a comprehensive overview of all end-user platforms (excluding API versions).

We will update it from time to time.

We support you with guidance when it comes to any Generative AI and AI challenges. Meet the Voicetechhub team today!

Are GPTs Relevant for my AI Strategy?

ninhabicht@gmail.com (Nina Habicht) — Fri, 03 May 2024 16:52:20 GMT

This article gives you a comprehensive overview of implementation strategies when adopting Generative AI.

Are GPTs relevant for my Company's AI Strategy?

MyGPTs or GPTs are customized "ChatGPTs" (learn more about GPTs and alternatives in our last blog post ) offered by OpenAI for your own purpose. They can be also offered on an OpenAI GPT store for other users. That means, you can build your own GPT for your specific domain. the newly published GPT store in ChatGPT on 10.01.2024.

With 3m GPTs in January 2023 available this is a crucial distribution channel for companies. In addition, OpenAI launched the Team collaboration option (costing 25$ / month) focussing on small medium enterprises to collaborate easier within ChatGPT.

We have made a research around companies and in the GPT store to see whether companies see GPTs as a major part of their AI investment areas. Our outcome: Most GPTs are still created by single creators or as an extension of tool providers.

But why do businesses still avoid building their own GPT in March 2024? GPTs have main drawbacks such as the inflexibility to control Prompt Engineering and RAG processes to provide better quality based on internal data. Furthermore, companies are bound to the ecosystem of ChatGPT. This means, to use the company's GPT users need to have a ChatGPT Plus account. Builders who wish to share their GPTs in the store need to follow specific guidelines and ensure their creations comply with OpenAI's usage policies .

Three Strategies How Companies Can Make Use of GPTs

1) GPTs as prototypes: Most corporates implement GPTs as a playground or "trial and error" sandbox to develop prototypes for clients or their own future generative AI applications.

2) Community and AI branding presence: Companies see it as a community enabler and growth option to be present in the GPT Store from day one.

3) Companies use it as internal support tool: There are companies that train important internal FAQs in order to use their GPTs as cost-effective internal support assistants for client support. This sounds like a funny use case, however, we've seen it across our market research.

What is the OpenAI Enterprise Teams Version?

If you decide to use GPTs as part of your Ai tool set, you can consider various options: There is an Enterprise Version of ChatGPT which makes sense when teams want to collaborate. Most companies in the US seem to use this option. It is also comparable to the Claude Team version from Anthropic.

ChatGPT Enterprise Team version for SMEs to collaborate and create own "GPTs". More on what GPTs are and how to build them .

Generative AI: It's More Than Just Implementing Tools

ninhabicht@gmail.com (Nina Habicht) — Fri, 03 May 2024 16:52:17 GMT

Optimizing Business Performance: Effective AI Strategy Implementation for Companies

Shall companies take Microsoft Copilot or develop our own Generative AI use cases? We often hear these questions.
There are three ways company adhere when it comes to Generative AI strategy and exploration. We want to share these learnings with you.

Strategy 1: Provide Tools to Enable the Use and AI Upskilling of Employees

We see many companies adopting Microsoft Copilot and other AI tools for various productivity tasks. Microsoft Copilot offers intelligent features such as automatically updating Excel columns with additional information, writing emails, and summarizing emails when you return from holidays. Also you can switch from a "work mode" to a "web mode" that enables the Copilot version to search in the internet via bing (e.g. to competitive analysis). This enables the company to use own internal documents as data base to retrieve accurate informations as well as "outside the company" data.

The major drawback of implementing these tools is that companies often overlook the underlying business goals they aim to achieve , and instead focus solely on the tools ("which are expected to magically solve all issues they have"). Generative Ai is about "us" meaning "human and machines, not: humans or machine". This means we need business and technical experts working together to achieve the best AI implementations and results. Do not make the mistake to exclude employees, stakeholders or even fire people due to "Automation Reasons of Generative AI" because you will need exactly these people to build, maintain and create innovative AI Business models and transform your company to an AI-driven company.

We support you with guidance when it comes to Generative AI challenges. Meet the Voicetechhub team today.

Strategy 2: Experiment with own GAI Use Cases for Micro-Optimizations

Many companies have already started to experiment with generative AI use cases across various departments. This experimentation enables them to learn and identify the feasibility, risks, and potential of specific use cases . However, it's crucial to first assess your needs and business requirements before implementing any tools in an ad-hoc manner. To this end, it makes sense to develop an overarching use case canvas and create a value-effort matrix for each case, including ROI and feasibility data.

If you don't have this data yet, conduct thorough research, involve AI expert companies, and run agile enablement sprints of two weeks to test- let’s say a tool or question you have. Based on your use cases and the north star vision you aim to achieve, you can formulate an AI vision. Since the AI strategy needs to align with the company’s overarching vision and OKRs, it is crucial to work collaboratively to align the AI vision with your company goals.

Once you have your vision in place, the next step is translating these learnings and requirements into actionable AI directives for data scientists and the Ai teams. Adopting this approach helps you to have a guidance.

For example, Microsoft Copilot is effective at addressing a wide range of use cases and can therefore be beneficial. Most companies implement Copilot as part of their AI Strategy to enable workforce tasks to be done in a more efficient way. We experienced Copilot to be sometimes very generic and thus not solving specific client needs. Additionally, you may not want to expose all your data to Microsoft, which happens as soon as you activate Microsoft Copilot Pro for your company. This is especially critical for insurance companies and banks where CID and personally identifiable information must be masked before being shared with third parties, potentially leading to risks. Microsoft Copilot is certainly a valuable AI workspace tool that enhances productivity in daily tasks across your company, however again, do not start with the tool in mind but with the problem you want to solve. Microsoft Copilot will not solve all issues. You need to know what issues you want to solve first.

Strategy 3: Innovate with Generative AI

The more mature companies are with the use and knowledge about AI, the more they start exploring innovative new ways of building new business models and products. We have seen exciting new ideas coming up through our Generative Ai workshops and consulting of companies that have potential to disrupt industries. Use new onboarded people, interns and interested business person from your departments to come up with great new ideas in hackathons or Gen. AI events. You will receive interesting new ideas and use cases.

However, it is crucial to first consider the specific pain points you need to address for your industry and clients before implementing any AI tools.

Generative AI Trends in 2024: Key Investment Areas for Companies

ninhabicht@gmail.com (Nina Habicht) — Mon, 01 Jan 2024 00:20:53 GMT

This article provides companies an overview of the major trends and areas to watch out in 2024 and the upcoming years.

Top 1: Autonomous Agents

Tencent recently released a framework called AppAgents that can operate mobile apps. We will see a trend towards agents taking over tasks for human in AI products and applications.

Autonomous agents are systems or software programs capable of independent action on behalf of their users or creators. Their functionality is rooted in AI and machine learning, enabling them to make decisions and perform tasks without human intervention .

F or example, instead of just prompting generative AI systems to create an image, then a logo, etc. to finally develop a website, the AI system creates directly subtasks that create a logo, check if a domain name is available and help create a website.

Top 2: Adoption and Governance of Generative AI Tools

We will see a more widespread adoption in Generative AI in 2024: "We know that nearly two-thirds of employees are already playing around and experimenting with AI in their job or in their personal life," according to Techtarget. "We really expect that this is going to be more normal in our everyday business practices in 2024."

However, g overnance issues such as data policy, cybersecurity, and model compliance remain unresolved and are top priorities for 2024 in the field of Generative AI. Although companies like JPMorgan, Samsung, and Chase have temporarily, or at least partially, banned the use of the ChatGPT Plus version, new forms of application will be required.

One potential solution is the "Bring Your Own AI" approach, where data is intended to be accessible locally or in sandbox environments. This would enable companies dealing with sensitive data to experiment safely.

Top 3: Open source Generative AI models on the raise

Since closed environments like OpenAI are expensive and because SME often do not have the resources to create own data models or products we will see an increase of usage of free, open source models. Models like Vincuna, Mixtral (which outperforms the LLaMA 2 family and the GPT3.5 base model. Mixtral matches or outperforms Llama 2 70B, as well as GPT3.5), Flamingo (OpenFlamingo is an open-source version of DeepMind’s Flamingo model, built on top of the LLaMA large language model ) will be more and more integrated in AI products. HuggingFace also provides some great Generative AI models . Here you find a comprehensive comparison between different llm models .

Top 4: Generative AI on hardware devices and smartphones

Companies such as Apple and Google are developing AI assistants and offline language models (LLMs) that reduce cloud costs and perform faster than purely cloud-based solutions. We can expect more personalized and informative results owing to the integration with personal devices, a point also highlighted by Qualcomm Technologies. The combination of offline LLMs with efficient, locally running LLM models is highly sought after in industrial applications

Top 5: Multimodality and voice interfaces

Multimodality, which allows for the conversion of one media form into another (e.g., text to image, speech to text, text to video), is expected to gain traction in 2024. While GPT-4-Vision already has the capability to interpret images, we anticipate a broader application in business contexts this year. Some notable use cases include the analysis of medical research and diagnostics (explore this further in our " Generative AI in Healthcare, Medicine, and Pharmaceuticals " trainings), and the generation of manuals from computer-generated designs, such as CADs, architectural plans, and electrical engineering projects, among others.

Voice assistants are poised to become integral in customer care center operations , automating numerous industries where waiting on hold is currently the norm. This includes sectors such as hospitality, medical centers, banking, insurance call centers, restaurant reservations, and the entertainment industry.

Top 6: AI Influencers and Avatars for business cases

Influencers will increasingly be augmented by machines, or at least will utilize tools to create content and movies more quickly. Find more here . Furthermore, digital avatar platforms will replace the labour-intensive creators market. Additionally, digital avatar platforms are poised to transform the labor-intensive creators' market. Industries that will be impacted include movies, gaming, music (e.g. song generation), modeling, advertising, education, and content creation (such as YouTube, TikTok, and social media professionals), as well as consultancy firms. These sectors will increasingly automate their work using intelligent avatars to reduce production costs and enhance efficiency.

Top 7: Advanced Robotics and Gaming with LLMs

Advanced Robotics that combines robotics with the power of large language models and are commercially available could be coming soon. Some examples are: OpenAI's NEO, Optimus from Tesla, and others. Also in gaming, we will see more natural language interactions in conversations with characters. Unreal engine already provides plugins to build intelligent characters within games.

Top 8: Generative AI leads to new job market and wage decrease

The shift towards greater automation is likely to result in a long-term decrease in wages , as labor-intensive tasks (such as programming, consulting, content creation, and management) can be executed more efficiently. Next year, we can expect to see more changes in job roles, maybe not yet salary levels due to the impact of Generative AI.

Top 9: Upskilling and openness to Generative AI will positively contribute to retaining and attracting talent

Companies see the urge to train their employees as existing jobs within companies change. People need to understand how to use Generative AI in a compliant, safe and productive way . Companies that embrace the new technology and upskill their employees will win in the longterm because of the following reasons:

1) Creation of new business value and models

2) Efficiency gains thanks to more automation

3) A culture of change and experimental openess

4) Through 3) attract more talents than companies who do not

5) Provide career paths and new options for their employees to grow

Top 10: Process Automation meets Generative AI

Thanks to the development of more use-case-specific language model (LLM) models, MLOps, and LLM flow builders, enterprises will be able to integrate their operational and business processes with advanced Generative AI technology. The year 2024 is certainly when we will witness many startups and companies delivering significant value by improving process automation through the use of Generative AI.

We support your company to analyze your company data, automize and create value from day one.

How Schools and Universities can use Generative AI

ninhabicht@gmail.com (Nina Habicht) — Fri, 29 Dec 2023 21:26:55 GMT

This article outlines how schools and universities shall approach the recent generative AI developments.

Why educational institutions are hit by Generative AI?

Many schools and also universities are confronted with students using Generative AI tools like ChatGPT. However, often regulations and rules are not available yet or are in an implementation phase. It is a fact that students cannot be controlled when it comes to using ChatGPT. Some might even already know more than their parents or teachers when it comes to ChatGPT.

Secondly, universities are disrupted by a technology which makes knowledge available. Harvard started building their own Generative AI tutors, avatars emerge in the fields of teaching that can significantly reduce costs and scale online trainings faster.

The World Economic Forum event writes about "Why AI makes traditional education models obsolete"

Create creative tasks for students with the help of Generative AI tool ChatGPT.

How Generative AI can be used by teachers?

Generative AI can be used for different tasks:

1) Analyse student's work: Tools like AI-based grammar and style checkers can help students improve their writing.

2) Content: Generate homework for the students

3) Creativity: Explore new ways of learning + executing homework. See how we prompted ChatGPT to create new ways of homework above.

4) Examination/Grading: AI can grade objective assignments like multiple-choice tests, allowing teachers to spend more time on qualitative feedback.

We experience in our AI trainings that we create a very new way of learning . Rather than focussing on traditional topics (e.g. French, programming, art) students will focus on a real-world project work and execute this with the help of AI. This may again include several "classical" skillsets needed (e.g. translate in French, program a code, design a nice picture for the website with DALL E).

Will ChatGPT replace exams and essays at schools?

There will be a trend towards more experiential assignments based on real-world challenges as described above. Students will learn to embrace changes aand solve a real problem .

WEF refers to a " durable skills-based model" which is better positioned in an AI era:

"Resilient to mindless AI use: Active learning techniques (debates, role-play and discussions) are dynamic, social and fast paced, requiring students to show up and interact, as well as creatively and critically “think on their feet”. Authentic and experiential assignments connected to real-world issues also require them to apply their specific skills to address a specific contemporary challenge faced by a specific (local) partner. AI may be able to provide inspiration and suggestions but not be mindlessly used to solve such problems.
Cultivates skills AI lacks: AI tools may be impressive, but they still lack genuine creativity, ethical reasoning, emotional intelligence and the ability to work, lead and negotiate with others. By explicitly training students on such skills we ensure they can do what AI cannot. And by training them on still other skills (such as critical thinking), we ensure that they can effectively and ethically use AI as well as critically evaluate its output."

Should schools allow the usage of AI tools and ChatGPT?

It is essential for students to learn about AI tools like ChatGPT and the underlying technology of large language models (LLMs). Those who can utilize AI effectively will have an advantage over those who cannot. Banning such tools could be detrimental to the future of many people as it is expected to be able to work with these tools (as it is in 2022 expected to know how to build a website or google, literally be digital knowledgeable today).

Providing access to learning materials in an equal and inclusive manner is important for the future of humanity.

Is ChatGPT compromising common sense?

It is important to teach students about the fact that ChatGPT and large language models are "calculators for words" and that the better users of such Gen. AI tools are subject manner experts (e.g. in programming or writing), the better they are able to judge the output of such models. Common sense and critical thinking are key elements hat is emphasized in a school or education context. This can be done by integrating ChatGPT and information literacy actively into your lessons.

Practical Exercise in the class: Let students search for the original sources of the mentioned information in a generated text by teachers. Very importantly: discuss the solutions delivered by AI. Do it iteratively: Discuss and question the answers given by the AI together with the class.

Should schools and universities use AI detectors?

A recent Stanford study showed that detecting AI generated text is a very hard task and that detectors are not reliable:

1) Seven AI detectors flagged 20% TOEFL essays as "AI generated" although they were not.
2) AI detectors are "near-perfect” in evaluating essays written by U.S.-born eighth-graders.
3) They can be gamed as many llms can be prompt engineered to make the text more lexical rich in the answers. This cannot be detected yet.

Despite these facts, there are numerous tools available in the market, particularly for addressing fake news:

Fact-Checking Websites : Websites like Snopes, FactCheck.org, and PolitiFact are dedicated to verifying stories and claims.
Media Literacy Tools : Tools like NewsGuard provide browser extensions that rate the credibility of news sites.
AI-Based Tools : AI systems like OpenAI's GPT-3 have been used to develop tools that analyze news for authenticity.
Reverse Image Search : Services like Google Reverse Image Search can verify the source and context of images.
Educational Platforms : Websites like Checkology offer courses to teach users how to identify misinformation.
Social Media Verification : Some social media platforms have their own fact-checking and content review systems.

It's important to use a combination of these tools and to also rely on critical thinking and cross-referencing with reputable sources for the best results in detecting AI generated or fake news. This will become more and more difficult as AI contents fill the internet.

Source: Generative AI in Leadership Training, Isabell Welpe. Image: created with DALLE 3

What are the risks of Generative AI for society that need to be taught at schools?

While ChatGPT may be enticing to utilize, it is important for teachers and students to be aware of the potential risks involved. ChatGPT acts like a "word calculator" and is not primarily designed to solve math problems, which means it still can provide incorrect answers. Thus, it is important to develop critical thinking skills at schools.

Since LLMs predict the next likely word based on an understanding of the preceding sentences, they can still convey incorrect information that might appear plausible at first glance.

Address Fake news: it is an issue that must be addressed during lessons. Additionally, fictitious influencers - AI assistants designed to accumulate followers and provide entertainment - already exist. It is essential to inform students, parents, and relatives about the presence of virtual assistants on phone calls and social media platforms.

The state of Gen. AI Governance in education

Fewer than 10% of schools (n =450) and universities currently have formal guidelines on AI according to UNESCO . Its new guidance suggests eight specific measures educational institutions could adopt to ensure “quality education, social equity and inclusion". This full guideline can be downloaded here .

Do you need support with Generative AI trainings in your class or school?

We are here to support you: contact us today.

A Comprehensive Guide to MyGPTs and Alternatives

ninhabicht@gmail.com (Nina Habicht) — Mon, 25 Dec 2023 23:32:18 GMT

This article outlines how easily you can build your own bot with different tools. We explain what is possible with GPTs and what is not.

What are MyGPTs or GPTs?

MyGPTs or GPTs are customized "ChatGPTs" offered by OpenAI for your own purpose. They can be also offered on an OpenAI GPT store for other users. That means, you can build your own Chatbot for your specific domain.

What are useful GPTs?

There are many useful GPTs: Some to mention are:

DALL-E: Let me turn your imagination into imagery.
Data Analysis: Drop in any files and I can help analyze and visualize your data.
ChatGPT Classic: The latest version of GPT-4 with no additional capabilities.
Web Browser: I can browse the web to help you gather information or conduct research.

How to build GPTs?

Anybody with a ChatGPT Plus account can build a "GPT". First, think about what you want to build. For that it make sense to think about the use case first:

Is there a daily task you want to be able to do faster?

For example you can easily build an information bot about your PDFs you want to get answers from. Or a bot providing information about your company services?

Download the ultimative Gen. AI Task Overview to learn where llms make sense to know which use case makes sense.

How to build MyGPTs?

There are three simple steps:

1) Buy a paid version of ChatGPT

2) Go to the settings area on the left-down side

3) Go to "My GPTs"

4) Click on create the "My GPT"

5) Let you guide through the creation process

Step 1: Go to "My GPTs" in the paid ChatGPT Plus version

Step 2: Click on "create a GPT"

Step 3: Let yourself guide through the creation of a GPT.

Is there an option for companies to use GPTs?

Yes, companies can use the ChatGPT Enterprise option . The enterprise option is more secure, using SOC 2 type 2 compliance standards, and data from ChatGPT Enterprise and the API Platform (after March 1, 2023) isn't used for training OpenAI's models. To enable features like conversation history data is retained, however, companies can control how long the data is retained (e.g. deleted data is removed within 30 days from the systems). This however means, your whole company has to use the Enterprise version to enable tenant-based creation and scaling of GPTs in a compliant way.

This could be a suitable solution for deploying your own bots in the US, but in European countries, stricter non-functional requirements apply:

All data has to be processed (and / or stored) in the country itself.
In the "worst case": No US provider is allowed

Source: Advantages of using ChatGPT Enterprise version.

What about confidential data being sent to GPTs?

OpenAI takes actions to avoid personal data being processed. They, however, state that by using the Business Offerings (e.g. Enterprise option) the content entered by users in your GPTs will not be used for OpenAI's model trainings: "we may use content submitted to ChatGPT, DALL·E, and our other services for individuals to improve model performance. Content may include chats with GPTs".

Will MyGPT builders have access to specific conversations with their GPTs?

According to OpenAI's statements this is not the case: "For the time being, builders will not have access to specific conversations with their GPTs to ensure user privacy. However, OpenAI is considering future features that would provide builders with analytics and feedback mechanisms to improve their GPTs without compromising privacy."

Can I integrate GPTs into my Website?

The short answer is no. You cannot use MyGPTs outside the ecosystem of ChatGPT. Meaning, once you have build your own GPT users can only use it in the MyGPT dashboard within ChatGPT Plus account. You can build with the assistant API product from OpenAI an own B2B application that can be integrated on any endpoint like a website, app or other platforms.

There is a hack where you can use Voiceflow in combination with your GPTs to deploy your GPT to a website.

What are the alternatives to GPTs to build your assistant?

There is the option to use the MyAssistant API or to build an assistant by using chatbot tools. We tested both and found that they still lack when it comes to reliable answers. There are certain answers that you want to have under your control as a company. The best way to keep full control of this information is to fully control the so-called RAG process yourself, and not rely on "black-box systems" such as GPTs or the MyAssistant API.

We can support you with reliable products that let you better control the answers.

What are the costs for enterprises to use llms?

ninhabicht@gmail.com (Nina Habicht) — Sun, 26 Nov 2023 22:20:43 GMT

Many companies are reluctant when implementing llm-based products because they fear bein confronted with high costs. Especially for medium-sized companies which have not the ressouces or enough capacity to deploy and oprimize their AI models nor to set up an own infrastructure with MLOps. As described in our article about sustainability of Gen. AI applications , cloud and performance costs of running an llm can become very high.

What are the cost types when implementing OpenAI or other llms?

T here are four types of costs related to llms:

Inference Costs
Setup and Maintenance Costs
Costs depending on the Use Case
Other Costs related to Generative AI products

What are inference costs?

An llm has been trained on a huge library of books, articles, and websites. Now, when you ask it something, it uses all that knowledge to make its best guess or create something new that fits what you asked for. That process of coming up with answers or creating new text based on what it has learned is called inference in LLMs .

Usually, developers would call a large language model like GPT-4. But here comes the "but": usually not only large language models account to the total costs when running the final product. To explain: LLMs can be used to classify data (e.g undestand that the text talks about "searching a new car insurance"), for summarization, for translation and for many other tasks. Download the ultimative Gen. AI Task Overview to learn where llms make sense:

What are the cost factors affecting the LLM costs?

Factors affecting LLM cost include model size (e.g. how many parameters it uses), with larger models requiring more resources, and context length (how much data, respectively text you provide to the model when you request your question), impacting computational demands. Keep in mind that larger models are not always better .

Enterprises face a range of choices between proprietary models like OpenAI's and open-source models like Llama 2 and Falcon 180-B. The selection of an LLM ifself depends on specific use cases and a balance between cost and performance.

Which Generative AI use cases cost more?

As outlined in our "Tasks where Generative AI helps" framework we show different use cases where generative AI makes sense.

Cost vary especially with the use case. For tasks like "summarizing" GPT-4 is often preferred in researches but comes with high costs at OpenAI.
Usually chatbots are more complex and cost-intensive than other Gen. AI products, especially when they do offer more than a "one-shot" response and can "talk for a while" about a topic and hold memory.
If your information has to be up-to-date at any time to deliver sufficient user-experience and avoid image issues (e.g. healthcare, insurance, fintech sectors), Retrieval Augmented Generation (RAG, see below for more information) costs will be coming on top. These RAG caused daily query costs vary between OpenAI and other models, reflecting different pricing structures.
Finally, "fine-tuning" is the most expensive process which means you change or customize at least one parameter in your llm neural network. This may change how the model learns on knowledge or its reasoning capabilities. "This leads to $100,000 for smaller models between one to five billion parameters, and millions of dollars for larger models" as outlined in AI Business article . Thus, finetuning comes only into play for large Enterprises or startup that finetune their own large language model or domain (e.g. BloomGPT). In this case, open-source models may lead to less costs in the long-term than OpenAI model finetuning. Keep in mind that people sometimes also refer to "finetuning" related to the optimization of a model concerning its purpose or domain and do not mean refer to any "change the model parameters". The latter is not as expensive as "real finetuning".

How much does a OpenAI model cost per month?

The monthly costs of the model really depend on the model you use, e.g. whether you use GPT-3.5 with 4K context (see below in the table), GPT3.4 with 16K context, GPT-4 with 8K context or GPT-4 with 32K context or even the new "Turbo versions", and on the usage (traffic) of your Generative AI product.

The yearly costs range from $1K-50K on the low usage end, depending on which model. Or from $1M-56M a year for high usage as outlined in the calculation below. Studies show that for low usage OpenAI model make sense whereas for high usage and traffic cases with many millions of users, open-source models can safe costs.

Source: Differences between OpenAI' models. GPT4 is as OpenAI describes "10 times more advanced than its predecessor, GPT-3.5. This enhancement enables the model to better understand the context and distinguish nuances, resulting in more accurate and coherent responses."

Source: Costs are depending on the GPT model and context as well usage requests. For high traffic applications the authors recommend to use own open-source models instead GPT, respectively OpenAI. Find more calculations here .

What are the maintenance costs related to llms?

To set up the infrastructure for llms you need to consider several points:

First, you need to set up open-source models or proprietary models like GPT on a cloud environment. It is also possible to set up the infrastructure on-premises (where you run your models on private servers) but usually this requires costly GPUs and hardware from providers like NVIDIA. As Skanda Vivek outlines this requires a A10 or A100 NVIDIA GPU. A10 (with 24GB GPU memory) costs $3k, whereas the A100 (40 GB memory) costs $10–20k with the current market shortage in place.

Compute: FLOPs (floating point operations per second) needed to complete a task. This depends on the number of parameters and data size. Larger models are not always better in performance.
Data Center Location: Depending on the local data center, energy efficiency and CO2 emission may differ.
Processors: Computing processors (CPUs, GPUs, TPUs) vary in energy efficiencies for specific tasks

Source: Open Source can become more expensive than GPT-3.5, especially due to complexity in maintenance, prompt engineering and extensive data science knowledge required for non-proprietary models like Llama 2.

What are other costs related to LLMs?

Very often embeddings (vectorization of your text corpus) and RAG (Retrieval Augmentation Generation) are needed besides mere GPT or llm models.

Why? Traditional language models can sometimes make errors or give generic answers because they're only using the information they were trained on. With RAG, the model can access up-to-date and specific information, leading to better, more informed answers. Example: Let's say you ask, "What's the latest research on climate change?" A RAG model first finds recent scientific articles or papers about climate change. Then, it uses that information to generate a summary or answer that reflects the latest findings, rather than just giving a general answer about climate change. This makes it much more useful for questions where current, detailed information is important. For Enterprises, up-to-date content is key, thus there must be some mechanism to retrieve best content continuously.

What is better? Open-Source LLM vs. Proprietary Models, e.g. OpenAI?

This is difficult to say. It depends on your requirements. Please find more in this useful article.

"ChatGPT is more cost-effective than utilizing open-source LLMs deployed to AWS when the number of requests is not high and remains below the break-even point." as cited from the Medium article .

Key Learnings when it comes to the costs for Generative AI products

The costs related llms vary from your use case (chatbot, analysis, voicebot, FAQ bot, summarization, etc.), traffic exposure, performance and set up (on-premises, cloud) requirements.

Open-Source is often more flexible than OpenAI when implementing own models, with a lot of traffic (> millions of users). However, it requires more prompt engineering, infrastructure customization and data science knowhow.
Fine-tuning costs differ, with OpenAI models typically more expensive than alternatives like LLaMa-v2-7b.
OpenAI's models are noted for their higher costs, but being more straight-forward, and efforts like the GPT-3.5 Turbo API aim to reduce prices.

Do you need support with choosing the right big tech provider, Generative AI product vendor or just want to kick off your project?

We are here to support you: contact us today.

Checklist for companies to successfully implement Generative AI

ninhabicht@gmail.com (Nina Habicht) — Fri, 24 Nov 2023 11:33:15 GMT

Often companies struggle to start with Generative AI as there are many unknowns when it comes to the costs, responsible AI and security concerns. Using ChatGPT Plus accounts is fun but of course companies need to be able to establish their own custom environment in order to upskill people and prevent data being sent to OpenAI. As stated in our recent article about ChatGPT and data policy - a simple guide , you need take care when sending questions to ChatGPT as the data may be used for the training of OpenAI's models or provided to third-party vendors.

How to find Generative AI Business Cases?

First, it makes sense to identify the use cases that really make sense to be automized with Generative AI.

Here the Generative AI Checklist that can help you to scale faster. The checklist is built upon of hands-on AI product implementation experience (chatbots, voicebots, NLP search and more) and - unlike other sources - considers all relevant aspects which you will encounter during your AI journey. Take these steps:

1. Customer: Identify the target users and customers you serve with your Gen. AI solution (e.g. internal employees, external clients, new or specific segments?)

2. Pain Point: Write down the top problems or pain points that your target customer faces.

3. Gen. AI Suitability: Check how your product can address the identified pain point and whether Gen. AI makes sense for that.

4. Data and Content Quality: First, you need good data to create a generative AI application.

Think which data sources are needed for your AI product?
Where are these data sources in your company? Are there databases, excels, intranet, confluence pages, data sources from call centers,
How good is the data quality itself (e.g. we see often low text content in websites, or inconsistent maintenance such as missing titles and structure which are key for a good chatbot or search)
Is the data/content owner identified in your company? Or are several owners of topics scattered around? Identify them and list them together. Involve these stakeholders in your Gen. AI project from the beginning!

5. Following AI Governance Principles: Are you adhering to AI ethics principles with your solution?

6. Unique Value Proposition: Why is your solution better than others in the market? (e.g. you can faster extract data, or you can search within more information context or your solution is more straightforward to use, hence has less entry hurdles?). Add a slogan to your solution: “ChatGPT to summarize legal docs.”

5. Data Source : Outline what data you have or need to obtain to make the solution work. Define how much and quality of data you have and who ‘owns’ it.

6. Value Streams: Describe how your concept drives value – estimate your saving potential (ROI in monthly hours of work) and revenue potential due to a new market or additional service you can provide thanks to Gen. AI.

7. Cost Structure: Lay out the major costs and expenses necessary to build and run your solution. Here we will provide an additional in-depth article on the costs related to Generative AI solutions.

8. AI Ethics and Governance: Write down any ethical, legal, or compliance hurdles you’ll need to manage.

9. Key Metrics: Outline the essential metrics you will use to track the success and growth of your solution - leading/lagging metrics that reflect the overall health and success of your venture. There are metrics that play a role especially with llm solution such as:

How reliable is the system (hallucinations, accuracy score, llm evaluation metrics)? This helps you to build trust with your customers which is essential with llms.
How much efficiency can you create with your product? Measured in cost analysis
How transparent is your product? Again outline how your product is build, which data base you use

Which Generative AI use cases are bringing high value?

Most companies currently focus on optimizing efficiency when it comes to finding relevant information. This major use case includes extracting relevant data from multiple sources like confluence, notion, emails, pdfs or websites. Why is this case approached by major enterprises and SME's around the world?

Da ta content can be extensively tested, optimized and amended so that your generative AI application works perfectly together with your underlying data sources
Companies can make first experiences regarding Generative AI costs without exposing to a larger group of clients
Image risks due to hallucinations (e.g. generated text) can be reduced by using standard llm evaluation metrics.

Thus, before providing your Gen. AI product client-facing you it can be substantially improved.

Alternatives to OpenAI: Testing Bard vs. Claude vs. ChatGPT

ninhabicht@gmail.com (Nina Habicht) — Wed, 01 Nov 2023 19:24:12 GMT

In this blog you will learn about the alternatives to ChatGPT and OpenAI.

Where is Bard better than ChatGPT?

Bard is the response to OpenAI's ChatGPT. What makes Bard so different to OpenAI?

It is free! So you can try it out here whereas ChatGPT costs $20 per month.
Another advantage is the microphone on the desktop version to directly speak in your question and get a response.
Bard has internet access whereas ChatGPT you need to jump from one service (Web Browsing) to the other
Bard covers far more languages (265 as of October 2023)
Some drawbacks: it is not able to generate pictures. With ChatGPT DALL E-3 you can generate pictures. Bard only offers you a nice description.

Where is Claude better than ChatGPT?

Claude is the version of ChatGPT developed by the company Anthropic.

This tool is currently accessible only in the UK and US, and not yet available in Switzerland. You might consider using Nord VPN to explore its functionality in your country.
Claude has one big advantage to ChatGPT: It can process more "context" ( Generative AI from A to Z ), meaning the input token (100 token equals around 75 words) can be up to 100'000 tokens (75'000 words!). GPT-3 has a limit of 4096 tokens (3072 words) and GPT-4 of 8192 tokens (= 6000 words). So when you want to upload huge files, use Claude.

Source: Bard has a better microphone option, but cannot generate pictures compared to ChatGPT Plus

What is Poe and what are its advantages?

Poe by Quora is a platform offering several llm chatbots at a glance. Basically you can compare it to a "WeChat for llms". The great features of Poe is that you can test large language models such as Claude, GPT, Llama and other models in one single place. And: for free (to a certain extent). On top, you can create your own chatbots in no time (not as sophisticated as Writesonic , GPT-trainer , or Dante AI which are specific Generative AI chatbot Tools). You can basically try out the advantage of Claude, namely analyzing large amounts of data (so-called window context defined as the number of tokens the model can take as input when generating responses according to Hazyresearch ). For example, in GPT-3 the context window size is 2K (2000) and in GPT-4 it is a larger 32K. There is a trend and demand for increasingly larger context window sizes in LLMs. Larger context windows improve LLM performance and their usefulness across various applications where GPT-models are less performant.

Claude is not available in Switzerland as of October 2023. So by using Poe you can overcome this limitation and try it out already. Unfortunately, Google Bard is not integrated to Quora's Poe yet.

Poe in Action: Analyze huge PDFs with Claude as it can process more context (data tokens sent to the llm)

OpenAI vs. Bard vs. Claude vs. Poe - Which Platform is the best?

When looking at the pros and cons of generative AI platforms we can see literally a race between the platforms where each tools launches again another great capability the other tool does not offer (e.g. context window, creativity, open-source availability, analysis, costs).

With the very recent activation of voice capabilities in the ChatGPT app - also for Switzerland and European countries - plus the enormous capability to analyze any kind of formats (excel, pdfs, CSV), we still see OpenAI's ChatGPT as the major leader when it comes to Gen. AI platforms.

How ChatGPT can help you to analyze data fast

ninhabicht@gmail.com (Nina Habicht) — Sat, 30 Sep 2023 12:51:30 GMT

In this blog you will learn the slice and dice with Generative AI when it comes to the analysis of your PDFs, excels, CSVs, and more. Learn the first steps on how you can visualize even data with advanced prompt engineering. This article is very useful for analysts, reporting specialists, controllers, and marketers who have to generate reportings and summaries on a regular basis and want to do it more efficiently. Join one of my courses in Zurich to get 1:1 support or send my team a message.

What are important analytics use cases with Generative AI?

Generative AI can be used to detect patterns and provide ideas when it comes to data visualization.

To name some important use cases:

Ask and summarize pdfs
Analyze sheets with numbers (Web, Sales, News)
Extract PDFs, transform them, and query for specific data
Generate reports (e.g. make your controlling analysis, generate excel charts)
Generate social media posts
Generate ideas on how to analyze and visualize data with Generative AI

What is ChatGPT Advanced Data Analysis?

This tool was formerly known as "Code Interpreter". Now, it comes with the brand new name "Advanced Data Analysis" but still many people do not know its power and capabilities. So continue reading ...

Where can I find ChatGPT Advanced Data Analysis?

1) If you cannot see this option, go to "settings" (left corner of the main dashboard) and activate all Beta release options

2) The availability sometimes is depending on the device and operating system you are working on (iOS, Tablets)

3) Contact the OpenAI support in case you cannot see the option.

Source: using Advanced Data Analysis in ChatGPT

Ask and summarize Excels and PDFs with ChatGPT

1) Upload the files that I want to analyze with the "Plus" button below in the entry field.
2) Specify your ChatGPT prompt: "Summarize the PDF". If you do not know what to analyze at all, no problem:
3) Ask ChatGPT for 10 analysis methods or ideas
4) Have the 3 best analysis methods performed

... and voilà ��, some exciting ideas come together.
Have the most important analysis ideas carried out.

How to extract PDFs with ChatGPT from a big PDF

Upload for example this pdf about AI regulation in the insurance industry.
Ask the prompt to exract in single PDFs.
And now you can easily ask ChatGPT so summarize a single page.

How to analyze an Excel and generate graphs with ChatGPT

You can easily upload your excel file (e.g. expenditures) or any web analysis and let ChatGPT think about the possible ways to analyze it:

Upload your CSV or xls or even JSON
Ask ChatGPT how you can visualize it: "Which visualizations would you propose. Please create a graph as a PNG to download".
Then you could upload a PowerPoint template you usually need.
Ask ChatGPT to add the PNG into your PowerPoint.

What are the limitations of analyzing data with ChatGPT?

ChatGPT's advanced data analysis cannot directly search the internet. If you attempt to, it will respond with something like: "I'm sorry, but I cannot directly browse the web or access external websites like Bing." To access the browsing feature, you must select it when initiating a "new chat". This means that if you wish to analyze data, you'll need to manually copy and paste the information from the co-browsing service to the advanced data analysis service. Admittedly, this can be a bit cumbersome.
Remember, it's essential to avoid uploading confidential or personally identifiable information to ChatGPT Plus (find a simple guide about ChatGPT and policy and privacy ). If you don't want OpenAI to potentially use this data for training their models or are uncertain about the implications, exercise caution. Tools exist that can extract personally identifiable information, such as organization names, using machine learning techniques like NER.
Overall, you can discover patterns in data without being a data scientist or data mining specialist ��. However, there's a caveat: ChatGPT's Advanced Data Analysis doesn't always function perfectly. The quality of the analysis is directly related to the cleanliness of the data. It's crucial to critically evaluate the results from large language models like GPT. Those familiar with basic analysis methodology or data mining are better equipped to discern inaccurate responses from the model. If you're ever unsure, feel free to ask ChatGPT for clarity. Also, if you encounter issues with specific data formats, consider converting them (e.g., from csv to xls, or to json). You can even utilize ChatGPT to assist in cleaning your data.

With the ChatGPT Advanced Data Analyzer, you can create graphs, tables, ppts and more. It does not always work at first trial but give it a try��.

How Generative AI can Boost your Company Page Search

ninhabicht@gmail.com (Nina Habicht) — Sun, 17 Sep 2023 21:35:46 GMT

With the advances in Generative AI companies should consider how to be "on top" when it comes to new technologies such as ChatGPT and Generative AI assistants. While Google SEO was one of the main drivers before the launch of ChatGPT by OpenAI, today companies should optimize towards the next-generation of search engines based on large language models and advanced search models.

Why are Generative AI searches so intelligent?

Intelligent searches basically understand the semantics in a sentence. So old index-based searches used to be sufficient for keyword entries but did often fall short when users did not enter exact the correct keywords or did spelling errors. Also AI based searches with older NLP classification models before the launch of OpenAI were handling some of these challenges but did never achieve the level of intelligence of so-called embeddings. With new vectorbased embeddings ( which are also available by OpenAI ) companies can build intelligent searches that can give to their users the best solution across all their websites.

What are embeddings?

Basically you can think of a vector representations. A user question is translated into a vector representation and this vector is used to query a vector database. Then the answer is used do create a summary via llms.

Think about when your kid plays with its toys. It might group them together based on what they are or how to use them. Your kid might put all your cars together because they all have wheels and can move around. Then it might put all its action figures together because it can play make-believe with them. And the kid's board games would be in another group because it plays them on a table. This is similar to how vector embeddings work in computers. But instead of toys, we have words. Just like how your kid groups toys, a computer groups words that are similar. For example, words like "cat", "dog", and "hamster" could be in one group because they are all pets. But how does a computer know which words are similar? It learns from reading a lot, like how you learn from playing and studying. If the computer sees the word "dog" being used in similar places as the word "cat", it will think these words are related and put them close together in its group.

So, in the end, vector embeddings are like a big, organized toy box for a computer, but with words instead of toys. Just like how a kid can more easily pick a toy to play with when its toy box is organized, a computer can more easily understand and use words when they are nicely grouped by vector embeddings.

How can companies create their own Generative AI search?

OpenAI did not only change the way companies could search with ChatGPT but it also changed the way companies can create their own page search. There are several paid and open-source models (embeddings from HuggingFace or OpenAI and Meta) to create your own intelligent search. If you need support with the development of your own search contact our team for any support.

This graph from TheAiEdge.io nicely illustrates how embeddings work:

Source: TheAIEdge.io

What are the most impactful Generative AI use cases for businesses?

ninhabicht@gmail.com (Nina Habicht) — Sun, 17 Sep 2023 21:35:38 GMT

This article gives you a great overview of the most relevant use cases for generative AI.

What are the most relevant Generative AI industries?

Based on Citi Research, the financial and consumer sectors are among the most significant business fields where General AI is disrupting the current status quo. At Voicetechhub, we observe similar patterns from our clients' needs. The following industries we see most relevant for Gen. AI: Insurance, financial services, education, consumer markets/e-commerce, and healthcare. The most prevalent use cases involve enhancing productivity in IT systems and automating tasks to save daily efforts.

Additional Facts:

Generative AI in Finance and Accounting : AI is transforming the financial sector by automating trading, personal finance, fraud detection, and robo-advisors ( Boston ).
E-commerce and Generative AI: AI helps e-commerce businesses with personalized recommendations, chatbots, and predictive sales analytics ( Businesswire )
Healthcare Generative AI: AI applications in healthcare include disease identification, drug discovery, and personalized treatment plans ( Walton , McKinsey )
Generative AI in the Insurance Sector: AI aids in automating claims processing, detecting fraud, and customizing policy pricing based on individual risk assessment ( McKinsey ).

Which Generative AI business cases will have the most impact?

There are great studies around the impact of Generative AI business use cases based on the costs and value they generate. According to the recent McKinsey and Company study 75% of the Generative AI use cases will unleash the next wave of productivity in these areas:

Customer operations
Marketing and Sales
Software Engineering
Research and Development

What are the primary goals of Generative AI use cases?

According to Gartner the major four initiatives why customers use generative AI are the following:

Customer experience, retention: This is in line with the McKinsey analysis that customer support and a great user experience comes with instant and fast responses 24/7h. Thus, when developing a great user experience companies should look at enabling intelligent search systems, phone call assistants and chatbots with an excellent user experience.
Revenue Growth: When you want to establish a new business or learn how to create a new business, get in touch with experts like Voicetechhub in order to find your "Generative AI Sweet Spot"
Cost Optimization: Costs can be optimized in different fields such as content creation, translation services, reporting, code creation, customer support automation.
Business continuity

How sustainable are Generative AI models?

ninhabicht@gmail.com (Nina Habicht) — Mon, 31 Jul 2023 23:56:19 GMT

This blog post outlines why large language models, thus generative AI models need so much power to be trained, deployed and maintained. We also look why some solutions are more sustainable compared to others.

What impacts the carbon footprint of large language models?

There are three main values that impact the carbon footprint of llms like GPT-4:

1) The footprint of the training model

2) The footprint from inference. "Inference" in large language models is referring to the capability of llms predicting outcomes using new input data. The models essentially use past knowledge to make educated guesses about the meaning of new sentences or to predict what comes next in a conversation. It's like connecting the dots using what they've learned from reading lots of text.

3) The footprint needed to produce all the required hardware and capabilities of the cloud data center.

Most energy-intensive is the trainint part of such models. Importantly, larger models do use more energy during their deployment. However, this study shows that inference is also a major consumer of energy, with up to 90% of the ML workloads is due to inference processing.

How high are the energy-costs of a large language models used in generative AI?

We summarized some important facts for you to shed lights on the environmental impact of large language models:

The Megatron Turing model from NVIDIA needed hundreds of NVIDIA DGX A100 multi-GPU servers, each using up to ~ 6.5 kilowatts of power (1).
To train one BERT model (LLM by Google) is roughly the same amount of energy and carbon footprint as a trans-Atlantic flight ( 2 ).
Researchers from the paper " On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? �� " explain that training models such as OpenAI's GPT-4 or PaML from Google uses 300 tons of CO2 (= one average person uses 5 tons a year) as outlined by HBR .

What is a sustainable large language model strategy?

If possible, use existing, pre-trained models from llm providers such as Microsoft OpenAI, Google (bard, PaLM), Meta AI (LlaMA) and do not create your own generative models (sidemark: this also costs around $ 100 Mio. according to a recently published article on HuggingFace . You should - instead - still fine-tune a model for your use case.

What are Best-practices to build greener llms?

Besides pre-trained model usage, use small models (with less parameters) by removing unnecessary parameters (so-called "pruning") without jeopardizing the accuracy you need (e.g. your model should still answer with a good confidence)
Reduce training time by experimenting with a distilled version of BERT, DistilBERT. "Distilled models are also been shown to be more energy efficient" as outlined in sustainable AI in the cloud .
Fine-tune your model with own data for some period instead of training it from scratch
Use cloud-based environments which provide scalable infrastructure (Azure, AWS, Google)
There is specialized hardware that support training speed ( Graphcore , Habana ) and inference ( Google TPU , AWS Inferentia )
Other technical tricks: merge model layers (so-called "fusion"), store model parameters in smaller values (say, 8 bits instead of 32 bits, so-called "quantization")
Run ML models on small, low-powered edge devices without need to send the data to the server to process). Use for example TinyML .
Monitor carbon footprint via tooling such as CodeCarbon , Green algorithms , and ML CO2 Impact .
Encourage your data science team to set benchmarking standards and include sustainability in their model considerations
Finally, think if you really need Generative AI at all. If you need help we can support you by evaluating your use case.

Summary about sustainable generative AI

Please find a comparison between llms and their sizes in the previous blog. Additional material can be found here:

The Greensoftware Foundation sets standards.
This article by Accenture outlines additional facts.

Responsible AI is a crucial foundation for all generative AI products. It is important that we consider carbon footprint with AI models to protect our earth and future. However, some studies also outline the difference between AI generated content vs. human generated content when it comes to emissions. This paper can be found here .

Best Practices in Prompt Engineering

ninhabicht@gmail.com (Nina Habicht) — Tue, 25 Jul 2023 20:42:59 GMT

This blog post is a super short guide to prompt engineering. The discipline gains importance for developers but also everybody who wants to use ChatGPT or other tools to be more efficient.

What is Prompt Engineering?

This is a method to work efficiently with an LLM - simply said:

1) Designing, developing and optimizing "prompts", respectively requests that are send to a large language model

2) Identifying existing vulnerablilities in a model

3) Making models are reliable, efficient (less time-consuming when calling the APIs) and safe (less hallucinations)

How does Prompt Engineering work?

Ususally a prompt engineer interact with the LLM via an API or directly. Thereby the prompt engineer can configure a few parameters to get different results for his prompts.

One of the most important topics is the temperature: the lower the temperature, the more deterministic is the result. If you increase the temperature, the more diverse and creative answers you will get. So if you build a Q&A chatbot, you might want to use a low temperature to be more fact-based. Looking at a story or content creation, you would rather increase the temperature value.

What are the principles of good prompts?

1) Good prompts are structured: use delimiters (like quotes), ask for structured output like a Json, png, and check if conditions are satisfied and assumptions are required to do the task.

2) Give the "model time to think": drill down your task you want to get ChatGPT or another llm to be completed into several structured subtasks. Like with humans you can extrapolate the pain point into several steps to solve it. Do the same with the machine

What is Zero-Shot or Few-Shot Prompting?

Few-shot and zero.shot prompts enable the language models to learn tasks given a one or few demonstrations how to do it.

Zero Shot Prompting Example:

Your task is to answer in a consistent way:

Few Shot Prompting Example:

To classify a sentence you can use few shot prompting:

<I wish I could have gone on that trip> // sad

<Yey! I will be married soon!> // happy

<hm, I don't feel well> // sad

<So glad that you are here> // ?

And the llm would basically learn based on this structure what to predict in the last sentence. Thus "happy".

What shall I do when I do not know how to prompt ChatGPT?

You can always ask ChatGPT itself to rephrase your question (e.g. "from now on optimize my question and ask me if I should use it?") or optimize your prompt by using tools that help you to write best prompts. A free one is: https://chat.openai.com/share/c1b9f609-e3ac-497a-bf77-a56e1bfd5fb6.

Role Task Format (RTF) for Prompt Engineering is a good starter to learn how to ask ChatGPT. An example is: As an analyst I need to create the most relevant KPIs in a table about topic {XYZ}.

GPT, LaMA, Falcon, Claude, Bard - An overview of LLMs

ninhabicht@gmail.com (Nina Habicht) — Sun, 23 Jul 2023 12:19:27 GMT

Which Large Language Models exist?

New ParagraWe summarized all relevant large language models in this blog post. If you want to learn what a LLM is read the following article about What is a Large Language Model .

If you want to see the latest benchmarks to evaluate the best LLMs have a look at the Hugging Face Benchmark or at sapling.ai. There is also the FLASK evaluation framework ( F ine-grained L anguage Model Evaluation based on A lignment SK ill Sets), a fine-grained evaluation protocol that can be used for both model-based and human-based evaluation).

Source: FLASK evaluation model

LLaMA 2 by Meta (Facebook) gained popularity because it is an open-source model that can be completely used for free also for commercial use cases. LLaMA-13B outperforms GPT-3 (175B) on most benchmarks, and LLaMA65B is competitive with the best models, Chinchilla-70B and PaLM-540B. This paper gives all details.

ChatGPT is the first Generative AI Chatbot presented by OpenAI to the market in November 2022, it is fine-tuned from either GPT-3.5 or GPT-4 Large Language Models using Reinforcement Learning from Human Feedback (RLHF). It allows us to chat in a conversational way, supporting many tasks like the answer to questions, writing summaries, debugging codes, generating texts and more.

Falcon first released in October 2021 was developed by a research center in Abu Dabi. It is optimized for performance and efficiency and focusses on high-quality data. "It outperforms GPT-3 for only 75% of the training compute budget—and requires a fifth of the compute at inference time."

Bard by Google was released on February 2023. It is based on the LLM PaLM 2 (May 2023) and originated from LaMDA Large Language Model. In contrast to ChatGPT the search data is from the web and in real-time meaning, it is not based on data until 2021. PaLM is multimodal, but only for it's domain specific model Med-PaLM 2 .

h2oGPT is part of the platform H2O.ai supporting a variety of models: GPT 3.5 turbo, LLaMA 2, Falcon. You can use it either online or locally or with a UI.

Claude 2 by Anthropic works like ChatGPT, updates in real-time and is a upcoming competitor to ChatGPT and Bard.

Dolly was developed by Databricks and released in March 2023. It has a size of 12 billion parameters, based on EleutherAI ’s Pythia model and fine-tuned on 15.000 record instruction corpus generated among Databricks employees.

Find a great comparison charts below from helpful sources.

Source: https://blog.zhaw.ch/artificial-intelligence/2023/04/20/deploy-your-own-open-source-language-model/

Source: towardsdatascience.com - https://towardsdatascience.com/choosing-the-right-language-model-for-your-nlp-use-case-1288ef3c4929

Source: Comparison of models can be easily done via sapling.ai

Can I use Voice in ChatGPT?

ninhabicht@gmail.com (Nina Habicht) — Tue, 04 Jul 2023 09:06:54 GMT

Why will be voice the next stage with ChatGTP?

This article summarizes interesting facts about voice I encountered during my journey. It started in 2018 with these facts. By then, the voice interfaces were not that mature though!

20% of Swiss regularly use voice-based assistants, preferably on Smartphones but intend to buy a smart speaker (Statista 2018/2019)
Siri, Google Assistant and then Alexa are the most used assistants in Swiss and Austrian households (McSchindler 2019, IAB Austria Integral 2019)
60% of Germans use voice assistants, 30% intensively. Google and Amazon are market leaders. (Splendid Research 2019)
A 40 bn. $ US market is expected for voice-based smart personal assistants (USA Today 2018)
21 % of US shoppers already use voicebots (Mastercard 2018)

But how has this changed from 2018 to 2022, and how will it evolve in the future?

What about About Voice and Large Language Models

When you look at ChatGPT from a user perspective you can use iOS and Android Configurations to get a so called "voice-input" option. Behind OpenAI's ChatGPT is Whisper, a highly sophisticated Speech model that can be also used as an API for developers since
How Siri, Alexa and Google Assistant Lost the A.I. Race with large language models coming that are more powerful
Google has announced a new Universal Speech Model (USM) here: speech models that includes two billion parameters that have been trained on 12 million hours of speech and 28 billion sentences of text. Currently, the model is based on ~ 300 languages and already included in Youtube.
Google's Bard enabled 40 languages in July 2023. You can try out speaking to Bard here (now also open for Switzerland):
Recent announcement in June 2023 says PaLM is a game changer in terms of speech: https://www.marktechpost.com/2023/06/23/google-researchers-introduce-audiopalm-a-game-changer-in-speech-technology-a-new-large-language-model-that-listens-speaks-and-translates-with-unprecedented-accuracy/

How I can use Voice in ChatGPT?

How to use Siri with the ChatGPT iOS app?

Simply trigger Siri (by using the Home button or saying "Hi Siri") and either, "Ask ChatGPT", "Send a question to ChatGPT", "Query ChatGPT", or "Tell ChatGPT" to ask a question to the ChatGPT iOS app.

How can I activate voice recording when using ChatGPT?

You can do this via keyboard activation for android:

Activate voice typing: Ensure that Gboard is set as your active keyboard. ...
Access the chat interface: Open the ChatGPT app on your Android device and tap on the prompt input box. ...
Initiate voice typing: On the Gboard keyboard, locate and tap the microphone icon.

Find out more here .

For iOS there's a simple easy video on that:

Download the official OpenAI app from the store
Tab on the waves below near the entry field where you type your query

ChatGPT and Data Policy - A Simple Guide

ninhabicht@gmail.com (Nina Habicht) — Wed, 21 Jun 2023 20:22:35 GMT

Does ChatGPT use my data?

This blog summarizes the privacy policy key points you need to consider when using generative AI such as ChatGPT. The situation may vary between llm vendors and their type of licenses (e.g. LaMA open-source models vs. proprietary models like GPT-4).

Data policy for GPT API usage

If you use the API product from OpenAI to build your own product or want to embed OpenAI services into your software, OpenAI explicitely states that "OpenAI will not use data submitted by customers via our API to train or improve our models, unless you explicitly decide to share your data with us for this purpose" ( https://openai.com/policies/api-data-usage-policies ). Remember that data entered before March 2023 - the effective date of this policy change - was likely used for training purposes. In addition, your data will be kept for 30 days to monitor any misuse.

Data policy for ChatGPT end-users

If you use the ChatGPT Basic or ChatGPT Plus functionality as a end-user you need to take care as data may be used from OpenAI to optimize their models. So for the "Non-API product" be careful when entering sensitive data: "When you use our non-API consumer services ChatGPT or DALL-E, we may use the data you provide us to improve our models. You can switch off training in ChatGPT settings (under Data Controls) to turn off training for any conversations created while training is disabled or you can submit this form . Once you opt out, new conversations will not be used to train our models."

Can it happen that ChatGPT releases confidential information from my company?

If you use knowledge bases from your company (e.g. databases like company docs, or personal notes, CRM data, text with confidential data) in the ChatGPT user interface could be used by OpenAI for training purposes and there may be a risk that the data will become public - unless you explicitely opted out in the beginning. So if you don't want that information of future offers, or sensitive information about your company is generated without control, avoid entering sensible data. As a company you can avoid this by using the API or in the safest way finetuning you own models on dedicated on-premises / own environments.

For example, Bloomberg created their own GPT model because their main business model is "knowledge". If they had given all information to OpenAI, and not build their own model, they would have handed out all their unique selling proposition. Building an own Large Language Model (llm) is very cost-intensive and requires high deployment resources as well as specific skills within a team (e.g. data scientists, MLOps expert) to enable a proper software deployment with all required tasks such as clean data pipelines, finetuning, training, testing and continuous model optimizations.

Is my AI generated content copyrighted?

Copyright law protects the rights of creators to their works. An AI (is no legal subjects) cannot be prosecuted under copyright law, only humans: The U.S. Copyright Office states that there is no copyright protection for works created by non-humans, including machines. As a consequence, the product of a generative AI model cannot be copyrighted. However, copyright protection of an AI product that exhibits traits of its creator can be protected by copyright. Therefore, prompt files or any code may fall under copyright. Please always look at the exact terms and conditions of the according generative AI providers when using generative AI for commercial reasons.

What do I need to amend in my data policy when I want to implement a ChatGPT in my company?

Companies must ensure that they respect the privacy and rights of their customers and that they communicate their use of AI in a transparent and accountable manner. This is by explicitely referring in their data policy to the models used. Find some examples here:

Jumbo explaining the purpose of its AI usage in their newly launched Jumbobot. They explicitely explain that the bot may "hallucinate" - meaning that there is the risk that the bot's answer may not be 100% correct: https://jumbo.frontnow.site/de/terms.

Helvetia Chatbot with ChatGPT transparently explaining the usage of GPT technologies to the users: https://www.helvetia.com/ch/web/en/private-customers/contact/services/chatbot-clara/ask-clara.html

Does ChatGPT give my data to third-parties?

If you use ChatGPT OpenAI my use your data for third-parties. The list of third-parties are can be found here: https://platform.openai.com/subprocessors

Large Language Models - all you need to know

ninhabicht@gmail.com (Nina Habicht) — Tue, 16 May 2023 09:06:37 GMT

What are Large Language Models and Generative AI?

To give you an overview first it is important to understand the context:

Generative AI is a subfield of the Artificial Intelligence with massive growth since ChatGPT in 2022. There are many subfields in AI, such as Machine Learning, Deep Learning, Natural Language Processing, Computer Vision, Robotics, and many more.

Generative AI refers to Deep Learning models capable of generating new content like text, images, video, audio, graphs and more.

Large Language Models (LLMs) are "foundation models" that use Deep Learning, Natural Language Processing (NLP) and Natural Language Generation (NLG) tasks. They are the underlying models behind generative AI that are used in chatbots like ChatGPT or Bard from Google. As you might know LLMs are used in use cases like writing, answering questions, translating languages, and coding. LLMs are pre-trained on extensive amounts of text, wikipedia, data from the internet. For the pre-training they often use techniques like fine-tuning, In-context training or Zero/One/Few-shot learning.

You can imagine something like a huge brain with neurons that learn based on stimulations from its inputs it gets and was shaped (so-called finetuned) to execute certain specific tasks better (and not so generic tasks)

Important to remember:
LLMs mimic human language use but don't truly understand it.
They basically predict what comes next in a sentence.

What are Foundation Models?

If you want to Imagine you're building a house. Before you can put up walls or install a roof, you need a solid foundation that will support everything else. Foundation Models in artificial intelligence work in a similar way. These models are designed to learn from a huge amount of data about the world, sort of like a base layer of understanding. They're trained on lots of unstructured information (like text, images, or sounds from the internet) without any specific task in mind. Think of this as a foundation learning a bit about everything, from cooking recipes to sports, from science facts to popular jokes, and so much more.

Then, when we want our AI to do something specific – like answer a question, translate a sentence, or identify an object in a photo – we can build upon that foundation. Because the foundation model has already learned a lot about the world, it can be adapted or "fine-tuned" to do many different tasks across various domains.

In short, Foundation Models serve as a general basis for learning, upon which more specific, task-oriented AI models can be built. Just like a good house foundation supports many different kinds of rooms and designs, a good foundation model can support many different AI applications.

What are Large Language Model Use Cases?

LLMs can be used for process automation, and to boost efficiency in many tasks human execute, such as creating text, summarisation, generating ideas, translation, detecting patterns etc. Please find some use cases which can be better performed with LLMs than before:

Text summarization
Text generation
Sentiment analysis
Content creation
Chatbots , virtual assistants , and conversational AI
Named entity recognition
Speech recognition and synthesis
Image annotation
Text-to-speech synthesis
Spell correction
Machine translation
Recommendation systems
Fraud detection
Code generation (see Copilot)

What are ChatGPT Plugins and how to use them? A simple guide

ninhabicht@gmail.com (Nina Habicht) — Thu, 27 Apr 2023 13:00:51 GMT

What are ChatGTP Plugins?

Plugins are an option for companies to provide their own "App" within the OpenAI ChatGPT ecosystem so that users can automate tasks easily. As a user - on the other hand - you might be looking for a QR-code generator on Google but instead, you can just easily search in the ChatGPT Plugins store for "QR" and activate a QR generator plugin. All done. You do not need to install any additional tools anymore or search in Google for any "free option".

Are you thinking about the next App store? You are right! It seems like OpenAI is becoming the new App store but far more powerful: OpenAI launched plugins, which enable people to automate tasks from within ChatGPT's interface. You can ask ChatGPT to perform a task in another app for you. Unfortunately, ChatGPT plugins, are not accessible to free users and currently in Alpha testing phase (add yourself on the waiting list ).

What am I able to do with the plugins version?

ChatGPT gets internet-access, meaning that it is not anymore limited to knowledge up to 2021 and can provide information with the use of Microsoft Bing. Thus, you can retrieve real-time info, e.g. stock process, latest news etc. by using WebPilot Plugin.
First ChatGPT plugins have been created by Expedia, FiscalNote, Instacart, KAYAK, Klarna, Milo, OpenTable, Shopify, Slack, Speak, Wolfram, and Zapier.
Find an example how ChatGPT looks like on Slack, see here .
Find an example how Zapier plugin looks like in ChatGPT , executing for you direct tasks
With this you can retrieve knowledge-base info from databases like company docs, or personal notes, CRMs or maybe even ERPs systems in the future?

We see ChatGPT is becoming more and more a virtual assistant being able to also execute “real actions” on behalf of the user, e.g. book a table, order food, and more. By integrating explicit access to external data, language models can become even smarter by acting with more evidence-based responses. Of course, there is a risk that the plugins could lead to security problems, e.g. by executing risky actions in the backend.

--> If you are interested to check out GPT-4 but are not a Plus user, you can explore GPT-4 here .

QR Code Generator to Web Search - Best-Practices for ChatGPT Plugins

QR Codes: Generate a QR Code in an instant
Ask PDF: about its content. Search for PDF and install the plugin
WebPilot: Searches in the internet and is very specific in giving you answers about a webpage. Use this prompt "Visitwebpage {URL}) or in German Befehl: Besuche {URL}
Wolfram: Executes mathematical functions for you when it gets complex
Zapier: Interfaces to other software easily build: For example CRMs, Email, Google Spreadsheet.

Predicting a Student’s Characteristics from a Chatbot Conversation

Thu, 13 May 2021 10:50:58 GMT

Introduction

Universities from all around the world continuously develop Intelligent Tutoring Systems to address the need of individual learning experience and to ease the tutoring burden of teachers and tutors. The Covid-19 pandemic further accelerated the need for digital solutions in education.

Just before the Covid pandemic, in 2019, a group of teachers and researchers from the Business Information Systems (BIS) master program at the University of Applied Sciences Northwestern Switzerland (FHNW) came to a decision that they will develop a chatbot Intelligent Tutoring System (ITS) to support BIS students working on their group assignment. This is an ambitious endeavor and developing such a system requires a lot of resources and preliminary research before the actual chatbot and system development starts. It was important to conduct a study in advance as a first step to understand if such an ITS would really help individual students while working on their group assignment and serve their tutoring needs. This blog post explains the goals and results of this preliminary study that was originally published in the form of a master thesis.

Why do we need to predict the student’s characteristics for an ITS?

During the conventional tutoring process, teachers are able to make assumptions about the characteristics of the students and their knowledge already when they sign up for tutoring. This perceived knowledge about the student’s characteristics continuously changes in the tutor’s mind as the tutoring progresses. Finally, teachers are able to suggest future learning activities based on how they assessed the student at the end of a tutoring class.

An Intelligent Tutoring System (ITS) also needs to build up this knowledge about the student’s learning progress to suggest proper individual tutoring steps for the student.

But wait! How can a system assess students’ characteristics?

An ITS can make predictions about the student’s learning progress from their behaviour with the system. For example, how the student communicates with the system, what words they use, how many failures they make during an assessment question, or if they are proactive in a conversation. These systems model the student’s characteristics in the so-called student models. These student models are a very important part of an ITS, that usually consists of 4 elements; a domain model, a tutoring model, a student model, and an interface. The picture below shows the standard architecture of an ITS with the four components from Nkambou et al (2010).

In this architecture, the domain model contains all knowledge elements of the field to be studied. The student model can be described as a dynamic model that contains information about the student’s intellectual and emotional state during the learning process. The tutoring model makes an assessment based on the information coming from both student and domain models and suggests tutoring actions. The interface component ensures access to the tutoring system (Nkambou et al., 2010). In our research we focused on the student model of a future Intelligent Tutoring System at FHNW.

There are several approaches on how to build student models in Intelligent Tutoring Systems. Stereotype models, machine learning models, cognitive approaches, fuzzy logic or Bayesian Network models can be used to build student models, just to mention a few (Chrysafiadi & Virvou, 2013).

As a student model, we built a Bayesian Network, called DigiBnet, to predict the student’s characteristics, such as their knowledge and motivational level from an imitated chatbot conversation.

Why a chatbot ITS?

The Digitalisation of Business Processes (DigiBP) module from the BIS Master program was selected to be the first module where the study program would offer tutoring via a chatbot Intelligent Tutoring System, called Digital Self Study Assistant (DSSA). The DigiBP module was selected, because it was observed during the years that students need more tutoring to complete their group assignment compared to other modules. Furthermore, the tasks of the assignment are often non-equally divided, as students with programming skills have higher workload compared to non-programmer students. Thus, non-programmer students have difficulties reaching their study goals.

After the module was selected, we conducted a preliminary survey among the DigiBP students to understand if they have a special tutoring need compared to what we already found out from the literature review and if they would positively accept a chatbot tutor. The positive acceptance of chatbot tutors was confirmed by this preliminary survey and we could also find out the special tutoring needs of DigiBP students.

In addition to that survey, we could learn from previous studies and research how student models can be deployed behind conversational agents and chatbots.

Kerly and Bull (2006) were the first researchers who proved the positive acceptance of a possible chatbot interface for a special type of student models, called open learner models, with whom students can negotiate their knowledge via a chatbot interface. After they built the CALMsystem, Kerly and Bull found that the system not only improved the users’ self-assessments accuracy but helped them to stay responsible with their learning process by being challenged by a proactive tool, like a chatbot that they were willing to work and communicate with.

Why Bayesian Network for the student model?

After considering several previous works, we decided to build a Bayesian Network open student model for the future DSSA. The lack of data and information about students, and the high uncertainty drove us towards this approach. It seemed that only this method could ensure that we can make predictions about the student’s characteristics from human experts’ insights and define some prior probabilities in a Bayesian Network (Duda, Hart, & Nilsson, 1981). This approach seemed feasible to determine conditional probabilities intuitively (Conati, 2010) and to calculate the tutoring needs of the students.

Bayesian Networks need input variables from different observations to predict a certain event. To find the right input variables for our student model Bayesian Network, DigiBnet, we used some findings from Hill’s research (2015) that reveals differences between human-human and human-chatbot conversations. As an addition to Hill’s research, an anonymized chatbot answer data was used from the Business Intelligence (BI) module from the Autumn Semester 2019 at FHNW for further inspiration to define possible observation inputs.

Based on these above mentioned insights we have considered the length of messages as an indicator of student motivation (where very short messages usually indicate low motivation), the number of knowledge elements (correctly) covered in the conversation and the number of questions and hints that the chatbot had to ask to achieve that coverage. In the picture below, the mentioned aspects can be seen as orange nodes in the Bayesian Network. It also shows how DigiBnet predicted the tutoring need of a student based on the activated input variable (orange) nodes.

From the activated input variables the Bayesian Network could predict the knowledge and the motivational level of the students, from which the final tutoring need of the students was further predicted by the model.

How did we conduct and evaluate the chatbot conversations?

To test the DigiBnet student model, we have imitated 4 chatbot conversations with students with assumed different knowledge levels in the topic. One teacher took over the role of a chatbot who followed the dialogue policy (not described in depth here) and responded to students’ utterances as if the answers were given by a chatbot. Students were informed that they were interacting with a teacher who took over the role of a chatbot and were also asked to give answers as if they were communicating with a chatbot. Thus, the research was not a blind Wizard of Oz study (Dahlbäck, Jönsson, & Ahrenberg, 1993), but an instant messaging chat between student and teacher. Although this setup might have caused some biases in our research, this method seemed the most feasible to test our Bayesian Network student model with the available resources to conclude some initial learnings.

After we conducted the chatbot conversations with the students, we used the chat text files and manually evaluated the observations that were needed for DigiBnet to activate its nodes. Based on this input data, our student model could assess the knowledge and motivational level of the students and could calculate the probabilities of which tutoring advice would serve the student’s need the best.

After the conversations were evaluated, we conducted interviews with the 4 students and shared with them how DigiBnet assessed their knowledge and motivational level, and what kind of tutoring it would suggest to them.

What did we find out from the interviews?

When we shared DigiBNet’s assessment with the individual students, there was only one contradiction where one student did not agree with DigiBNet’s assessment. All other assessments of the DigiBnet were correct and accepted by the students. It means that the observations that the model used for prediction and the assumed conditional probabilities were capable to properly predict the knowledge and motivational level characteristics of the student.

From earlier research we know that opening the student model and sharing the model’s assessment with the students can improve the metacognitive skills of the students, which results in better student performance (Mitrovic & Martin, 2002). We could also confirm that, as a result of the student's self-reflection that they would be grateful for advice regarding the need to study further materials. Some of the students found that they received rather general than personalised tutoring advice, i.e. it did not go much beyond what they already knew. We could conclude from the feedback that the well formulated and placed advice is useful for the students, but it should be provided during the tutoring conversation, after it is clear for the model that the student doesn’t know the answer of a specific question. Students suggested additional video tutorials showing the exact minute of the video, where the missing part is explained in detail, or a particular paper or blog post directing the students to the questioned topic of the material would be highly useful for them.

What are the key learnings?

From our research we concluded that there is a need from the students for personalised tutoring advice. From our preliminary survey we found that most of the students (79.5%) are open to communicating with a chatbot about their assignment and would positively perceive an assessment from a chatbot (72.7%). DigiBnet could predict the characteristics of the students well and they were positively accepted by the students. Providing individual tutoring however requires more fine tuning and testing to find out what would be really beneficial for the learners.

Building a chatbot ITS is a very complex journey with several modelling, testing, prototyping tasks before we start building a knowledge base and developing a system. In our research we proved that such a system would be beneficial for the students, however it is still a long journey to put these systems into production at FHNW. The first steps are done, let’s see what comes next.

Agnes Pakozdi

Business Analyst, Zühlke Group

Agnes Pakozdi has finished her Master of Science studies in Business Information Systems at FHNW under the supervision of Dr. Hans Friedrich Witschel in 2020. Agnes also holds a Bachelor diplom from Budapest Business School, a Certificate in General Management from HSG and a Diploma in Product Management from Product Academy. Agnes is passionate about digitalisation, product management and connecting Data/IT with business.

Trust-Supporting Design Features for Customer Service Chatbots

Sat, 10 Apr 2021 21:13:12 GMT

Introduction

Based on recent advantages in the area for artificial intelligence (AI), chatbots play a crucial role in customer service. However, due to a lack of user trust in chatbots, the adaption of the technology in the business are is only progressing slowly. The paper by Zierau, N., Hausch, M., Bruhin, O., & Söllner, M. (2020) serves as an initial identification of possible design principles to support trust in chatbots in the customer service domain.

Why Design Principles Support Trust in Chatbots

As an automated communication software powered by artificial intelligence, chatbots work as an interactive information partner. They enable users to experience fast and seamless humanlike conversations in natural language. On the company-side, chatbots represent a high-quality and cost-effective method for dealing with recurring and standardized inquiries to improve customer service. However, due to a lack of trust in AI, the widespread adaptation of the technology is still lacking in spite of the wide range of possible applications in the business field (Griffeth and Simonite 2018). Scholars observed that one reason for this might be that the development and adaption of chatbots in the business area was initially based on technology push rather than market pull. This resulted in a failure of sufficiently identifying customer pains and subsequently addressing customer needs and wishes (Coniam 2014). To address these shortcomings, Dietrich et al. (2019) mentioned the importance of standardizing design principles that support trust in chatbots as well as evaluating their instantiation.

In order to do so, the development of the design principles was carried out step by step according to the Design Science Research Approach by Hevner (2007) using a mixed-method approach (mix of qualitative and quantitative evaluation).

How to Develop Trust-Supporting Design Principles for Chatbots

First, findings from 22 semi-structured qualitative user and expert interviews were summarized in 12 user stories. These provided an initial, qualitative overview of elements relevant for the development of trust in interactions with text-based digital assistants.

Second, an extended systematic literature review was carried out and revealed that trust in chatbots can be divided into trust in chatbot, environment and user-related criterias which include different trust-relevant components. From the considered literature, 10 literature issues are identified which are integrated into meta-requirements. The meta-requirements describe which requirements are to be built into the artifact. In the final step of the development, design principles are being created. Following a specific formulation, they define the concrete implementation of the previously defined meta-requirements in the artifact. Finally, design features represent the lowest granular level of design elements and the eight identified can be seen in the figure in this article. As components of design principles, they define exactly how which element from the proposed artifact can be implemented in a concrete use case.

The 8 Design Principles to Implement Chatbots in Customer Service

In conclusion, the following eight design principles are recommended to implement in chatbots for customer service:

Olivia Bruhin

Master Student in Business Innovation at the University of St. Gallen (HSG) and HCI-Enthusiast

After finishing her bachelor's degree in business administration, economics and psychology at the University of Berne, Olivia gained work experience in innovation management at SBB CFF FFS AG and PostFinance. To deepen the knowledge in this field, she started her master's degree in business innovation at the University of St.Gallen (HSG) in fall 2019. There, she mainly focuses on corporate entrepreneurship, design thinking as well as applications in the field of human-computer interaction with a special interest in the healthcare sector.

Part 1: What should be considered under data privacy law with voice assistants?

Sun, 29 Nov 2020 18:42:29 GMT

Introduction

Voice assistants such as Google Assistant, Alexa or Siri - also called "virtual assistants" - have been criticised by media because of their capacity to "listen" what users say. While tech providers like Google, Apple and Amazon provide vast amounts of information in privacy policies that are difficult to read and understand, companies still lack clarity on how to proceed with a chatbot or voice assistant project.

In this first blog article Voicetechhub and Dr. Adrian Bieri, Partner of the Swiss law firm Bratschi AG, give concrete answers to the questions what should be considered from a technical point of view and under data privacy law when implementing digital assistants. It also gives a great deep-dive into data protection law (in particular GDPR) and its impact on digital assistants based on Natural Language Processing.

When to involve legal in voice assistant projects?

Before discussing with your legal team, you must have a clear idea of what your digital assistant should solve. We experience that companies often do not know what they want to reach with a chatbot or voice assistant.

A short checklist for voice product managers

During the early-stage of your project, you should have answered basic questions first before starting discussions with your responsible managers in the fields of legal and data protection law.

What is the goal of the chatbot or voice assistant?
There are various methods such as creating a customer journey, scenarios and thinking about the pains and needs of your customers.
How does the current process of your specific use case look like? Start thinking about this process and where a bot supports best the current problems you encounter. After that, you can go into a deep-dive with conversational design and user story definitions.
Finally, you should have a list of goals you want to achieve with the bot.
Which systems are needed?
During the requirements phase, clarify which systems are needed. Note down any knowledge base, data base or system which might be an interface to your chatbot, e.g. CRM system.
What kind of data is saved in which system?
After having sketched down the system’s architecture, write down where you gather which kind of information of your users.
Also, make a note, if there is any personalised data stored somewhere. And: do you authenticate your users (i.e. request for login data)?
Verify with your project team how long you need the data to be stored.
On which platforms do you deploy the chatbot? Is it a Facebook, Slack, MS Teams bot or is the assistant running on smart speakers or your website? Depending on the platform, control platform-specific guidelines for your developers.

What has to be considered in terms of data privacy law (DSVGO) with chatbot and voice assistant projects?

General data protection law principles

Data protection laws such as the European Union’s General Data Protection Regulation ("GDPR") and the Swiss Federal Act on Data Protection ("FADP") aim to protect personal data. Personal data means any information relating to an identified or identifiable natural person. An identifiable natural person in the sense of the GDPR and the FADP is one who can be identified, either directly by reference to an identifier such as a name, an identification number, location data, an online identifier (such as an IP-address or an E-Mail address) but also indirectly by comparing various pieces of information. Data that does not relate to a specific individual, on the other hand, usually falls outside the scope of data protection legislation.

Chatbots and other applications based on artificial intelligence ("AI") process various amounts of personal data, in particular also for training purposes. Therefore, chatbots and other similar AI-applications (e.g. digital assistants) regularly fall within the scope of data protection law. And if chatbots and digital assistants process biometric data such as facial images or a person's voice (e.g. for voice and face recognition purposes), the stricter data protection regulations for processing sensitive personal data (i.e. processing of special categories of personal data) apply. It should be noted, that once personal data is anonymised – which requires irreversibly removing the possibility of identifying an individual through the data in question – it no longer constitutes personal data, thus becoming exempted from data protection law obligations. This option should be considered, where the purpose of the processing of data does not require the link to an identifiable individual. Besides anonymisation, pseudonymisation can be seen as a further means of restricting the applicability of data protection law.

Once it is clear that data protection law applies, the question arises as to which data protection obligations must be observed when using chatbots and other AI-applications. Data protection law imposes various obligations with respect to the processing of personal data. The central data protection obligations in connection with chatbots certainly include compliance with the information obligations under Art. 13 GDPR and ensuring that there is a legal basis for the processing of personal data according to Art. 6 GDPR.

Other relevant obligations in connection with the processing of personal data by chatbots and other similar AI-applications include:

(i) the implementation of appropriate technical and organisational security measures (Art. 32 GDPR) to protect personal data against accidental or unlawful destruction or loss, alteration, unauthorised disclosure or access as well as
(ii) conducting a data protection impact assessment according to Art. 35 GDPR to evaluate and, where necessary, mitigate the risks associated with the application in question.

Further specific obligations may apply if chatbots and other AI-applications (in particular in combination with digital tracking tools) are used to create large user-profiles and, therefore, such processing activities are considered as profiling under Art. 22 GDPR. Finally, the GDPR and also the FADP impose various documentation obligations and other duties, which must be respected when processing personal data via a chatbot or other similar AI-applications.

Information obligation and having a legal basis for processing in particular

As previously mentioned, providing transparent information about the data processing that takes place via the chatbot and ensuring a legal basis for the processing is provided for are paramount considerations.

According to Art. 13 GDPR , the data controller (e.g. the company who wants to provide a chatbot or a voice assistant to its customers, see also section 4.3) shall, at the time when personal data are obtained, provide the chatbot user with various information about the relevant data processing activity. This information includes, among others, the purposes of the processing for which the personal data are intended, the legal basis for the processing as well as the recipients or categories of recipients of the personal data, if any.

In order to fulfil the information obligations according to Art. 13 GDPR, it is necessary to inform the user at the beginning of the chat or before using the digital assistant about the purpose and scope of processing. This must be done, in accordance with Art. 12 GDPR, in a precise, transparent, comprehensible and easily accessible form in clear and straightforward language.

In the case of a chatbot, it is advisable to indicate within the chat itself, in a basic form, how personal data are processed and to provide a link to the privacy policy which contains further information. This information must also be provided before personal data is processed, ideally before the digital assistant is installed, and must be kept available via the app or website. Informing the user about the purposes of processing is feasible as long as the chatbot’s functionalities are limited, thus, making it possible to predetermine which topics will be covered by the chatbot. The situation is, however, different and more complex with regard to the variety of possible voice commands in the field of digital assistants.

Art. 6 GDPR provides that processing of personal data is lawful only if, and to the extent that, a legal basis as provided under Art. 6 GDPR is applicable for a given data processing activity. These legal bases under Art. 6 GDPR include, among others, the user's consent to the processing (Art. 6(1)(a) GDPR), or, as an alternative, the necessity of processing the data to enter into or perform a contract or contractual negotiations (Art. 6(1)(b) GDPR; contractual necessity). Moreover, Art. 6(1)(f) GDPR establishes that a processing activity necessary for the legitimate interests pursued by a data controller (or by a third party) form a legal basis for processing, insofar as the controller's interests are not overridden by the interests, fundamental rights or freedoms of the affected data subjects. We recommended carrying out as many data processing operations as possible based on Art. 6(1)(b) GDPR because this legal basis neither provides for the possibility of withdrawal – as in the case of consent by the user – nor for a possibility of objection.

Since many chatbots are currently still designed for a specific function (e.g. customer support) and the corresponding user input usually contains product- or service-related questions or orders from the user, it is often possible to base the data processing operation on contractual necessity according to Art. 6(1)(b) GDPR. Also, it seems feasible to argue that the purpose of data processing is aligned with the contractually-agreed provision of a functional digital assistant. However, should court case law or regulators opt against this approach, the legal basis of contractual necessity (Art. 6(1)(b) GDPR) could in many cases fail to serve as a legal basis for data processing by digital assistants. In view of the fact that the legal basis of the legitimate interest under Art. 6(1)(f) GDPR involves, depending on the specific individual case, significant legal uncertainties, the providers of digital assistants would in many instances be advised to obtain consent from the user, which conforms with the requirements under Art. 6 (1)(a) GDPR for legally valid consent.

Moreover, if a user discloses sensitive data to a chatbot or voice assistant and if the company that offers the chatbot or voice assistant to its customers obtains the consent from the users, Art. 9 GDPR requires the explicit informed consent of the user. This requires among others that the user is informed by the controller before giving his consent as to which specific data processing activities will be covered by his consent. A mere voice command to activate a digital assistant, for example, is most likely not sufficient to be considered as legally valid consent according to the requirements of Art. 6 (1)(a) and 9 GDPR, particularly because of the absence of sufficient prior information about the relevant data processing activities for the user by the data controller. However, practically this is often done, which is solved by the chatbot providers through anonymisation of data.

Which role do you have: Controller or Processor?

When it comes to the question, who is responsible for the fulfilment of the various obligations data protection law provides for chatbots or voice assistants, one has to first clarify, who of the involved parties (i.e. the company, that offers the chatbot or voice assistant to its customers, and the various service providers the company engages to provide the chatbot or voice assistant) plays what role from a data protection law perspective.

The primary distinction is between the controller and the processor:

The controller is the person who determines the purpose and means of a data processing activity (Art. 5(j) FADP).
The processor processes personal data on behalf of the controller and not for his own purposes. In principle, the controller is responsible for the compliance with most of the obligations that the GDPR or the FADP set forth.
In the particular scenario where two or more controllers jointly determine the purposes and means of the processing of personal data, they are joint controllers. Joint controllers must, by means of an "arrangement" between them, apportion data protection compliance responsibilities (e.g., it should be agreed which controller shall be responsible for providing clear information to data subjects).

With respect to digital assistants, the company that offers them to the users is considered as a controller, and, therefore, is responsible for compliance with data protection law. It cannot be excluded that the company which uses the digital assistant and the company which develops and operates the digitalassistant are to be considered as joint controllers, however.

Legal checklist for voice assistants and chatbots

To summarise, we developed the following final checklist, which helps companies from a data protection law point of view when implementing voice- or chatbots:

Clarify (i) which kind of data (e.g. non-personal data, personal data, sensitive personal data) is processed by chatbots and digital assistants and (ii) how the flow of such data is structured (e.g. by whom and where is such data processed).
Make sure you have a proper legal basis for the processing or consider to implement anonymisation or pseudonymisation procedures in order to exclude (at least in part) the applicability of data protection law.
Provide the users in advance with clear information about the associated processing of personal data (e.g. privacy policy).
Perform before the use of chatbots and digital assistants a data protection impact assessment ("DPIA") and think of additional legal requirements that apply in the event of profiling.
Verify which role from a data protection law point of view you have in a specific digital assistant project and make sure to conclude the appropriate agreements (e.g. data processing agreement, joint-controller agreements, controller-controller agreement) with all the parties that process personal data in the course of a chatbot or voice assistant project.
Comply with the additional legal requirements if personal data processed by digital assistants is transferred to a destination country which does not provide an adequate level of data protection.

Dr. Adrian Bieri

Partner at Bratschi Ltd., Head of PG Intellectual Property, Technology and Privacy

Adrian has an in-depth experience in the fields of legal technology, intellectual property and privacy. He has been working as a lawyer and partner at Bratschi Ltd. , which is a full-service law firm with six offices throughout Switzerland. He is highly engaged in solving new legal issues related to emerging technologies. Adrian holds a degree in law from the University of Zurich.

The Swiss Post Explains Why Voice Assistants are Gaining Relevance

Fri, 31 Jul 2020 19:18:13 GMT

Conversational AI at Swiss Post

The Swiss Post is well-known for its Chat option with the PostFinance

chatbot for 24/7 customer support. However, we want to know what are

the upcoming use cases and plans in the fields of Voice user interfaces and how the future services of the Swiss postal services could look like.

Voicetechhub speaks to Philipp Leuthold, Manager Online Channels

@Swiss Post

Tell me about your background, experience and how you got in touch with Conversational Interfaces, i.e. Voice Assistants or Chatbots?

I have been working in the digital team at Swiss Post for several years.

This is where the external digital channels are located, such as the

website, microsites, apps, newsletter system and our basic services e.g.

the central login, location search and much more. New technologies

fascinate me and make my work always exciting. After being involved

in the websites and apps of the Swiss Post, I started getting interested

in voice as a digital channel. First as a private user and then more and

more intensively with the idea in mind of how and what Swiss Post can

offer via this channel. Last year, I had the opportunity to spend several

weeks working intensively on this topic at an innovation camp at

Swissnex in San Francisco. In addition to many user experiences, I was

also able to validate some hypotheses, realize a prototype and develop

a sharpened procedure plan.

What are your experiences with Voice Technologies?

Voice is simple - Voice is fleeting. It is easy to obtain information by voice, to control things in the house and to place highly repetitive orders via good skills/actions. As a user, however, I am usually busy with something else and not just focused on voice interaction. I prefer it if I can already log in and so the skill/action already "knows" who I am, where I live, how I pay and what my preferences are - so interaction with Voice becomes easy and fast for me as a user.

Did you already realized Voice Assistants at Swiss Post?

At Swiss Post we are at the end of the approach phase. Various

departments have already developed prototypes. There are showcases

for the location search of our branches or the parcel pickup service at

the frontdoor (pick@home) where all information is coded, as well as

functional prototypes with connection to backends. For example, our

real estate department has developed a skill for managing coffee

corners to report coffee machine defects or to order cleaning services.

We have also created a skill for consignment tracking, which transmits

live data, here we still have some challenges as it is not really great to

dictate an 18-digit consignment number.

What are the challenges you are facing in your Swiss Post Voice projects?

In the case of new technologies, ideas, prototypes and projects are created in a large company at different locations. In order to bundle synergies and also to achieve a clear orientation of the channel, basic principles must be developed and a competence centre created. As far as the basics are concerned, we are currently in the process of developing a brand strategy for the voice channel and the legal and voice-specific data protection regulations. On the technical side, we are concentrating on being able to combine Swiss Post's login with the voice solutions. This has the advantage for the user that the known user information (name, address, payment method) and the services used (recognized shipments, pick-ups at the front door, etc.) are already "in stock", so that the actual voice interaction can take place easily and in a few steps.

Where do Voice Assistants make sense at Swiss Post?

Voice assistants make sense for Swiss Post where our customers need to

interact with us via voice. Voice assistants are usually found in

the home environment. For this reason, we are concentrating in the first

phase on integrating the Post-Login into the Post-Applicators. This is the

basis for our online services, such as tracking and controlling shipments,

picking up parcels at the front door or configuring existing subscriptions, such as those currently offered by Bread Post – a service where Swiss Post daily delivers fresh bread. With the current basic work, we are laying the foundation for the next steps.

What can we expect in the future from Swiss Post?

Due to the great diversity of Swiss Post, Voice can be expected to be used i n many different locations and services. In addition to the obvious end- user offerings, I can also imagine specific solutions in the B2B area:

Laboratory staff who use Voice to order a transport drone so that they do

not have to touch panels etc., or building technicians who use Voice to

locate and rectify malfunctions for our real estate customers. Even in

internal processes, for example in the sorting of parcels and letters, hands are often not free and with voice it is easier to order empty containers. Voice offers the advantages of hands-free and hygienic interaction in many areas.

Philipp Leuthold

Manager Online Channels

Since the first hours of the internet, Philipp has been an innovative e-media communication professional with in-depth experience on a national and international level. He realized successful community launches for various age segments and did relaunch

project management of national web portals (www.sbb.ch, www.post.ch). Also, he realized implementations of mobile applications (mobile web

and apps) and holds many years of practical experience in newsletter creation, web analysis and CMS systems.

He has been working for Swiss Post in the digital sector for 14 years and is currently Product Owner for mobile apps, newsletter system and voice.

The Swisscom Voice Assistant

Fri, 12 Jun 2020 07:45:33 GMT

Switzerland's first Voice Assistant

The Swisscom Voice Assistant is unique as it can speak up to five languages, especially Swiss German. It has been developed by the Swiss telecommunication company Swisscom.

Voicetechhub speaks to Riccardo Lopetrone, Senior Product Manager Voice @Swisscom. We want to share these great insights with you.

Tell me more about your background, experience and how you got in touch with Voice?

I have been working as a TV Product Manager for different telecommunications companies for the last 15 years. Since 5 years I'm responsible for the customer experience around voice controlled devices @Swisscom.

My first contact with voice was in 2011, when Apple presented its first iPhone with Siri. I bought the new 4S model straightaway and made my first experiences with voice assistants. I found the feature very fascinating, even if the Siri capabilities were very limited back then. I then recognized that I wanted to deep-dive into the topic.

In 2016 at Swisscom we developed the first TV box with integrated voice control in Switzerland, a push-to-talk solution on the TV remote control which supported several voice commands on Swisscom TV. When Alexa was then launched in 2018 I started my first experiences with smart speakers. Today I have over a dozen devices at home (Amazon Echo, Google Nest, Homepod and of course the Swisscom Box), which I regularly use to e.g. play music, create shopping lists, control my smarthome devices or check the weather forecast.

Last year Swisscom decided to take the next innovative step so we integrated our own smart speaker into the new Swisscom Box.

How would you describe the Swisscom Box compared to other smart voice-enhanced devices?

All new Swisscom customers with a Swisscom TV contract get the new Swisscom Box with the integrated Voice Assistant. Existing customers can also upgrade to the new box. Beneath great entertainment features and a wide choice of content from the most popular providers (e.g. Teleclub, Netflix, Sky, etc.) the Swisscom Box comes along with two microphones and a speaker to support the new Voice Assistant. The Box is delivered with deactivated microphones to cover data privacy concerns. The customers need to actively switch on the microphones before they can use the voice assistant. Only then the Swisscom Box will react to our wake word "Hey Swisscom"

A few of the available skills (e.g. Spotify or the WiFi skill) first need to be activated on the separate companion app (Swisscom Home App) to allow account linking and for additional information around data privacy (e.g. when showing your WiFi password on the TV screen. All other skills are activated by default.

What are the main features?

Our focus was on the voicification of Swisscom products and services.

So far, the following skills are available:

TV (e.g.channel change, title & actor search, trick modes, TV recommendations, etc.
Smarthome (e.g. switch on and off your smarthome devices, execute scenes, etc.)
MyCloud (voice control of your pictures and videos from the Swisscom cloud on the big TV screen)
Weather (ask for the weather forecast in Switzerland, in collaboration with our companies local.ch & search.ch)
News (ask for the latest news, in collaboration with our company bluewin.ch)
Router & WiFi control (switch on and off your Internet & Wifi, present the WiFi password on the TV screen)

What is unique about the Voice Assistant of the Swisscom Box?

Strong data privacy policies
Speech recognition in 5 languages (DE/FR/IT/EN including a Swiss German model for the main dialects in Switzerland
A deep integration into our services and products which would only have been partially possible on a 3rd party voice platform

What were the main challenges when developing the box?

First of all we had a very challenging launch target. We developed all the components including the hardware in less than 12 months. Since many teams were involved (more than a dozen) the coordination was very complex.
A further big challenge was the support of different language models and on top the huge amount of entities (terms) which come along with some of our skills. For the TV search for example we have more than 30'000 programme and movie titles which are available for our customers every week. And sometimes we need to support also mixed languages within the same language model. E.g. if you have chosen the English speech recognition but you are looking for movies with a French actor ("show me movies with Gérard Depardieu").
Privacy and transparency around data storage was also a very important topic for us, so it took many discussions with our legal and data governance departments. As an example we decided to deliver the Swisscom Box with deactivated microphones so the customer first needs to activate them if he wants to use the voice assistant. Furthermore the customer can allow or decline that humans will listen to a small part of his generated audio files, which would allow us to improve the system.

What can we expect in the future?

In a first step we are optimizing the basic functionalities (e.g. the wake-word sensibility, the speech recognition for all the supported languages, the intent recognition, etc.). In parallel, we want to improve the existing skills and include more context to enhance the customer experience. And of course we want to expand the number of supported skills with further Swisscom services and products.

Will it be possible to develop a third-party application?

For the time being we are focussing on Swisscom Use Cases. We will evaluate in a later stage if it makes sense to open up the platform for other Swiss companies.

Riccardo Lopetrone

Senior Product Manager Voice

Riccardo Lopetrone is Senior Product Manager Voice at Swisscom, the leading telecommunications company in Switzerland. For more than 5 years he has been committed to creating great user experiences for voice controlled services.

More about the Swisscom Box

Managers View on Voice Commerce and Conversational Commerce

Wed, 27 May 2020 19:27:15 GMT

The Disruptive Potential of Voice Assistants for Marketing

The world is confronted with the rise of voice assistants, increasingly used for shopping activities. AI-enabled voice assistants like Amazon Alexa or Google Home are likely to radically change the way consumers search and purchase products with severe impact on brands. However, the behavior of these shopping-related machines represents a “black box” difficult, if not impossible, to decode for managers. Our new study, “ The Evolution of Marketing in the Context of Voice Commerce ”, in collaboration with NetComm Suisse and e-Business Institute , examines managers’ perceptions of the evolution of voice assistants and their potential effects on marketing practice. The study of the managers’ interpretation of a voice-powered marketplace is critical as it is posited to influence future marketing choices. An expert survey with Swiss and European managers (N=62) revealed the expected impact of voice assistants on the shopping process. Findings show that managers consider voice assistants a disruptive technology assuming a central relational role in the consumer market.

Voice Assistants as Middlemen among Consumers and Brands

In the managers’ view, in-home voice assistants assume a central relational role in the consumer market and progressively mediate market interactions, with a severe impact on brands. Remarkably, 87% of managers agree that voice assistants will become “powerful marketing, sales, and distribution channel” and “a technology increasingly able to influence consumer's choices”, for 80% of respondents. The prediction of a severe impact on consumer brands from two-thirds of the study participants reflects the agency role attributed by managers to these devices. While functioning as a “salesperson,” voice assistants are redefining relationships among consumers and brands. In this respect, 75% of the study participants believe that voice assistants will become new middlemen between brands and consumers.

Threats of Voice Commerce to Brands

A key challenge connected to the diffusion of voice commerce is related to the potential bargaining power shift in favor of voice assistants’ manufacturers, especially Amazon. A total of 71% of respondents “somewhat agree” or “strongly agree” that Alexa will disproportionally place its private labels while penalizing other consumer brands. Managers are aware that Amazon's biased placement on voice assistants of its private labels might challenge national brands. As such, ranking algorithms represent a gatekeeper that might contribute to a reduction of brand visibility during the consumer’s search process. Nearly two-thirds (65%) of managers believe that brands will have reduced visibility on voice assistants compared to other touchpoints. In this context, search advertising in the form of voice assumes a paramount role in the marketing practice. Among respondents, 61% agree that advertising cost on voice assistants would be higher than web-based advertising because of the limited space available for sponsored messages (paid recommendations). Also, 44% of managers disagree that “low involvement product categories will be the only affected by the voice assistant's diffusion,” showing an expected “universal” effect of voice commerce across product categories. Moreover, three-fourths of respondents believe those voice assistants will ongoingly re-evaluate the consumer's product choice and suggest better alternatives. In a context in which brands are required to continually justify their positions, competition might increase.

Manager’s Perspective on Voice Commerce Across Industries

Among different industries, “Consumer goods” managers show a remarked distrust towards voice assistants. In particular, they do not trust their integrity, that is, the ability of voice assistants to adhere to a set of principles that the user finds acceptable and provide unbiased product recommendations. Opposite to other industries, in particular, “Fashion and retailer,” managers working for CPG firms do not believe that voice assistants will improve consumer’s decision quality both in terms of the degree of confidence in the shopping decisions and overall decision-making abilities. Opinions’ divergence between these two industries might be driven by the historical dependence of CPG companies from intermediaries (retailers) functioning as the gatekeeper of their distribution. Fashion managers appear significantly less concerned about potentially unfair voice assistants’ manufacturer practices and the influence that the VA might exercise on consumer’s decision making. This might be due to the importance of a multi-sensory experience during the selection process. Besides, automated purchases through product subscription seem less relevant for the fashion industry than others as only a few selected items have the potential to be automatically reordered (e.g., underwear or sneakers). In that context, also given the acceleration of digital practices, the fashion industry might look at the evolution of voice assistants more from a voice marketing perspective than from a transactional standpoint.

Relevance of Voice Commerce for Companies

We conclude that managers have a shared understanding of voice commerce’s challenges and opportunities for their brands. Nearly three-fourths of the respondents (74%) believe that voice commerce represents a “great opportunity for their brand.” At the same time, 77% consider it a significant challenge. This result shows a dual mindset capturing the current state-of-mind of the companies. Voice commerce is seen as a revolution in marketing and brand management, but also a phenomenon with potentially detrimental consequences.

Reviewing the data across geographies, managers operating in Switzerland find voice commerce less critical for their country compared to European managers. This might be due to the underdeveloped presence of Amazon in Switzerland, a fact that drives the penetration of voice assistants (and voice commerce) in other markets. At the same time, European managers believe to a greater extent than the Swiss that voice commerce is “important to win shares in the short-term.” In this respect, the rather stable Swiss economic situation is reflected in the manager’s reduced sense of urgency.

In light of these firm’s radical exogenous changes, researchers and marketers are called to study further the interplay between consumers and brands in response to “machine behaviors.”

Alex Mari

Marketing Researcher

Alex Mari is a research associate at the University of Zurich, where he studies the effect of machine influence (AI-enabled voice assistants) on consumer and brand behavior. He has extensive managerial experience in marketing, working as head of digital marketing and start-ups CEO. Alex teaches digital and AI marketing in business schools.

More about Alex

E-Learning Assistent kombiniert ML & Rule-Based Algorithmen

Wed, 25 Dec 2019 00:01:45 GMT

Inhalte

Vorteile von E-Learning Agenten

Liste von E-Learning Chatbots

E-Learning Assistenten im Kundensupport

Wie wurde der hybride E-Learning Bot umgesetzt?

Vorteile von E-Learning Agenten

Sogenannte Educational oder „Pedagogical Agents“ führen Aufgaben in Tutoren- und Lernsystemen aus. Mit der rasanten Entwicklung der von Computer Linguistic und Machine-Learning seit 2000 können inzwischen auch komplexe E-Learning Chatbots realisiert und in sogenannte Learning Management Systeme (LMS) integriert werden. Sie zeichnen sich durch folgende Eigenschaften aus:

Sie unterstützen einen effizienten Lernprozess
Sie liefern wichtige Informationen über das Lernverhalten und den Lernprozess
Chatbots lernen ständig dazu und verbessern sich.
Sie fördern das Spielerische und die Freude am Lernen
Sie begleiten die Lernenden durch die Kurse, geben Updates und motivieren durch Ermutigungen (positive Feedbackschleifen)
Chatbots können technische Lern-Nuggets einbinden und gleichzeitig den Gesprächsfluss zur Lernstruktur aufrechterhalten
Die Bewertungen sind unvoreingenommen und objektiv im Vergleich zu echten Lehrpersonen

Liste von E-Learning Chatbots

Wir haben e-Learning Bots auf chatbots.org , botlist.co und Google Actions sowie Skills analysiert bzw. ausprobiert. Hier haben wir uns auf die neusten Bots ab 2017 konzentriert. Die Meisten unterstützen das Erlernen einer Sprache. Wir haben solche, die aus technischer oder UX-Sicht revolutionär und bekannt sind, zusammengefasst:

1. Duolingo : Sprachlern-App, die nicht nur stark personalisiert und in Richtung Conversational E-Learning weiterentwickelt wurde, sondern auch mit künstlicher Intelligenz ihr Nutzerengagement um 12% steigern konnte. Zum einen wird die sogenannte spaced-repetition Methode angewendet, die besagt, dass es besser ist, kurze Lektionen in Intervallen zu wiederholen, als die gleichen Informationen in kurzer Zeit zu pauken. Zum anderen, wird das statistische Modell der Halbwertszeitregression angewendet: Fehlermuster von Millionen von Usern werden analysiert, um für jedes Wort die Halbwertszeit vorherzusagen. Diese ist die Dauer bis ein bestimmtes Wort mit einer Wahrscheinlichkeit von 50% vergessen wurde. Duolingo soll - wie eine gute Lehrperson - verstehen, was der User kann oder eben nicht kann, ihn abholen können und den Lernstoff entsprechend passend vermitteln. Eine Weiterentwicklung zu einem Sprachbot auf Alexa oder Google sind noch nicht angekündigt. Obwohl die Bot-Funktion von 2016 abgestellt wurde, wünschen sich die meisten User diese zurück.

Abbildung: Duolingo Bots

2. Alex unterstützt die Studenten der TU Berlin mit Fragen rund um Kurse.

Auch Sprachassistenten gibt es viele, welche vorwiegend beim Erlernen von Sprachen unterstützen. Viele Voice Assistenten sind für Kinder und Teenager konzipiert und als Quiz aufgebaut. Komplexere Lösungen für die Erwachsenenbildung gibt es wenig.

3. Coursera hat einen Amazon Skill entwickelt, um direkt Infos über die Kursbelegung und Quiz-Ergebnisse, Aufgaben und Lernfortschritte zu erhalten.

4. Scoyo hat eine Google Action sowie Skill für Kinder entwickelt, um einfach das Einmaleins zu erlernen. "OK Google, spreche mit Scoyo Einmaleins" aktiviert die Action. „Alexa, öffne Scoyo“

5. Englisch mit Oxford hilft mit einer nützlichen Skill . Ganze Geschichten werden auf Englisch erzählt und danach mit Multiple Choice Fragen abgefragt. Die Zusammenfassung der Ergebnisse werden direkt zur Alexa App geschickt.

Das nachfolgend beschriebene Projekt ist ein Beispiel dafür, wie E-Learning Systeme und Bots für komplexere Anfragen kombiniert werden.

E-Learning Assistenten im Kundensupport

Positive Erfahrungen bei Anfragen an Vertrieb oder Support sind für die Kundenbindung oftmals entscheidend. User Experience in natursprachlichen Dialogen mit Kunden wird je länger je wichtiger, denn die Erwartungen nehmen zu. Gewünscht ist eine schnelle, effektive und qualifizierte Antwort. Diese kann heute bereits weitgehend automatisch erzeugt werden. In schwierigen Fällen soll eine möglichst nahtlose Übergabe an einen Mitarbeiter erfolgen. Für den optimalen Mitarbeitereinsatz muss das automatische Dialogsystem die Fragen für die Übergabe thematisch aufbereiten, präzise Einordnen und an den bestqualifizierten Bearbeiter weiterleiten.

Die integral-learning GmbH aus Berlin hat in Zusammenarbeit mit dem Zentrum für Informations- und Sprachverarbeitung der LMU ein natursprachliches Dialogsystem entwickelt, das die Arbeit in einem Call Center für eine e-Learning Anwendung unterstützt. Lernende können fachliche und organisatorische Fragen stellen.

Dieser Agent ist einzigartig, da er gleichzeitig regelbasierte Bot-Technologien und eines der mächtigsten ML-Algorithmen ( BERT von Google ) kombiniert. Der automatische Agent begleitet die Lernenden und bereitet die oftmals unklar formulierte Anfragen geschickt vor. Damit wissen die Fachleute im Call-Center, die die Anfrage übernehmen, genau um was es sich bei der Frage handelt und zu welchem Teil des komplexen Kurses eine Frage gestellt wurde. So wird die Zeit, welche die Fachleute mit der Anfrage verbringen, reduziert.

Abbildung: Dialog des E-Learning Assistenten

Wie wurde der hybride E-Learning Bot umgesetzt?

Das agile Projekt wurde von einem Team aus Entwicklern zusammen mit Alena Moiseeva (Zentrum für Informations- und Sprachverarbeitung Ludwig Maximilian Universität, München) und einem Product Owner in ca. 4 Monaten in Python umgesetzt. Der Algorithmus des Dialogsystem ist hybrid: Er kombiniert regelbasierte Komponenten mit Maschine-Learning basierten und verwendet dazu den BERT Algorithmus von Google. Die Bot Komponenten sind über ein REST-Interface mit dem Chatsystem, das in Javascript und node.js entwickelt wurde, verbunden. Neben Software-Entwicklungsarbeit waren für die Umsetzung des Chatbots auch umfangreiche Analysen der vorhandenen Dialog-Daten sowie ein auf die Zielgruppe angepasstes Conversational Design mit einer geeigneten „Persönlichkeit“ für den Chatbot wichtig.

Michael Heimann

Software Architect

Michael hat über 10 Jahre Erfahrung als Softwarearchitekt in der IT-Branche. Als Product Owner konzipierte und entwickelte er mit dem Team von integral-learning und dem CIS der LMU ein natürlichsprachliches Dialogsystem für E-Learning. Bereits in seiner Abschlussarbeit als Informatiker entschied er sich für ein Thema aus dem Bereich des maschinellen Lernens. Die Analyse von Problemen und die Umsetzung von Lösungen mit Methoden der modernen Softwareentwicklung sind seine berufliche Leidenschaft.

Strukturierte Daten für Voice Search - Teil 3

Mon, 25 Nov 2019 07:23:14 GMT

Im letzten Artikel «Wie betreibe ich SEO für Voice Search» haben wir verschiedene Fragen angeschaut, die du dir am Anfang stellen solltest, Sechs Tipps zum Thema Voice behandelt, definiert was Voice Search gerechten Content ausmacht und warum dein Google My Business Beitrag so wichtig ist. Im zweiten Teil werden wir uns anschauen, was strukturierte Daten Markups sind und welche Richtlinien von Google vorgegeben werden, um strukturierte Daten einzuführen.

Inhaltsverzeichnis

1. Was sind strukturierte Daten Markups?

2. Welche Richtlinien gibt es für strukturierte Daten?

3. Erkenntnisse zu SEO für Voice Search zusammengefasst

1. Was sind strukturierte Daten Markups?

Inhalte einer Webseite können durch strukturierte Daten Markups (Auszeichnungselemente) besser ausgezeichnet und erklärt werden, da den Suchmaschinen mehr Informationen zur Verfügung gestellt werden. So kann die Suchmaschine die Webseite besser einordnen und treffsicherere Suchergebnisse ausspielen. Durch strukturierte Daten wird eine Unternehmung höher gerankt, da Google und Co. mehr Kontext haben und besser verstehen, worum es auf der Webseite geht. Darüber hinaus kann durch strukturierte Daten die Darstellung der Suchergebnisse verbessert und auffälliger gestaltet werden. Die nächste Abbildung zeigt das Suchergebnis am Beispiel von Mammut ohne Rich Snippet, wobei bei Campz ein Rich-Snippet in Form von Kundenbewertungen angezeigt wird:

Abbildung: Rich-Snippet Beispiel campz.ch (Google Suchergebnis, 2019)

Um strukturierte Daten einzusetzen, gibt es mehrere Möglichkeiten. Dies kann zum Beispiel durch Mikrodaten, Mikroformate, RDFa oder Schema.org ausgeführt werden. Um eine Ein-heitlichkeit zu erlangen, haben sich die grossen Suchmaschinen, Google, Bing, Yahoo und Yandex, zusammengetan und das Projekt Schema.org ins Leben gerufen. Diese Art der strukturierten Daten gilt für die relevantesten Suchmaschinen, daher empfehle ich nach diesem Schema die Markups einzupflegen. Die Suchmaschine entscheidet, wann sie ein Snippet oder welches Snippet sie einsetzen will, egal in welcher Form (Featured, Rich, Universal) das Snippet ist. Ob es eingefügt wird, hängt davon ab, ob das Snippet der Userin oder dem User einen Mehrwert bietet oder nicht.

Featured Snippet und Rich Snippets

Der Hauptunterschied von Rich Snippets und Featured Snippets ist, dass Rich Snippets im Quellcode durch einen fest definierten Code für die Suchmaschine gekennzeichnet werden muss. Beim Featured Snippet ist dies nicht der Fall, dieses erstellt die Suchmaschine von allein. Die nächste Abbildung zeigt ein Rich Snippet:

Abbildung: Rich-Snippet Beispiel campz.ch (Google Suchergebnis für Jacken, 2019)

Die Darstellung der beiden Snippets auf der Suchergebnisseite unterscheidet sich sehr. Das Featured Snippet wird an oberster Stelle angezeigt, deshalb wird es auch «Position 0» genannt. Solche Anzeigen geben der Userin oder dem User eine direkte und prägnante Antwort auf die Suchanfrage. So müssen die Userin oder der User nicht noch lange Zeit verschiedene Seiten durchklicken. In der nächsten Abbildung wird ein Featured Snippet gezeigt:

Abbildung: Featured Snipped (Featured Snipped, 2019)

Auch hier entscheidet Google selbst, wann ein Featured Snippet ausgespielt wird. Doch gerade für Voice Search ist dieses Anzeigeformat relevant, da der Voice Assistant die direkte Antwort kurz und bündig vorlesen kann.

Was ist Universal Search?

Unter Universal Search versteht man die Einbindung von Bildern, Videos, Produkten, News oder Maps in die organischen Suchergebnisse der Suchmaschinen. Der Zweck der Universal Search ist es, der Userin oder dem User zusätzliche Informationen zur Verfügung zu stellen. Die erweiterten Informationen ermöglichen es Google, ein optimales Interpretationsangebot für die verschiedenen Intentionen der Userinnen und User bereitzustellen. Das Ziel ist es, dass die Userin oder der User nicht noch eine Suche starten muss, sondern aus den verschiedenen Medien wählen kann. Die nächste Abbildung zeigt eine Universal Search Suchergebnisseite:

Abbildung: Universal Search Anzeige (Google Search Ergebnis für Rucksack, 2019)

Speakable Markup

In der Schema.org Struktur für strukturierte Daten gib es unter anderem das Markup «Speakable». Dieses Markup identifiziert Abschnitte innerhalb eines Artikels oder einer Webseite, welche sich am besten für die Audiowiedergabe durch Text-to-Speech eignen. Durch die Markierung kann der Google Assistent die Information vorlesen und eine breitere mögliche Zielgruppe erreichen, da auch andere Kanäle durch Speech zugänglich gemacht werden. Aktuelle Nutzeranfragen können so durch intelligente Lautsprecher beantwortet werden. So-bald ein Google Assistent einen sprechbaren Inhalt vorliest, wird die dazugehörige URL an das Mobilegerät gesendet und für die Nutzerin oder den Nutzer verwendbar gemacht. Im Moment gibt es diese Speakable Markups nur in den USA als Beta Version für die englische Sprache. Nichtsdestotrotz sollten die anderen Markups auch verwendet werden, da diese genauso vorgelesen werden können.

Was sind die Richtlinien für strukturierte Daten?

Damit strukturierte Daten (Rich-Snippets) aufgenommen werden können, müssen einige Richtlinien , welche von Google vorgeschrieben sind, beachtet werden:

• JSON-LD

• Mikrodaten

• RDFa

Des Weiteren müssen folgende Punkte erfüllt sein:

• Die Markups müssen aus aktuellen Informationen bestehen. Es werden keine Rich-Suchergebnisse angezeigt, welche zeitkritisch oder nicht mehr relevant sind.

• Es müssen Originalinhalte verwendet werden, die entweder von der Firma oder von den Userinnen und Usern erstellt wurde.

• Alles, was in den Markups ausgezeichnet wird, muss auch auf der Webseite aufzufinden und für Leserinnen und Leser sichtbar sein.

• Für die verschiedenen Markups soll der spezifische Typ- und Property-Namen verwendet werden, welche nach Schema.org definiert wurden.

• Es sollen keine Markups für irrelevanten oder irreführenden sowie gefälschten Inhalt erstellt werden.

• Die erstellten Markups dürfen nicht sexuelle Gewalt, Pädophilie, Sodomie, gewalttätige oder grausame Handlungen, gezielten Hass oder gefährliche Aktivitäten fördern.

Die wichtigsten Erkenntnisse zu SEO für Voice Search zusammengefasst

SEO spielt bei Voice Search eine elementare Rolle. Je nach zur Verfügung stehenden Infos wird nur das eine Ergebnis vorgelesen. Damit eine Firma eine Chance hat, muss sie folgendes im Bereich SEO beachten:

• Ausführlicher, relevanter und gut strukturierter Content muss zur Verfügung gestellt werden

• W-Fragen mit Content beantworten

• Überschriften und Meta Tags einsetzen

• Rich Media (Video, Bilder, Tabellen, Abbildungen etc.) verwenden

• Backlinks beachten

• Produkt- und Kundenbewertungen einführen

• Pagespeed ausbauen

• FAQs aufbauen

• My Business Einträge müssen gepflegt und aktuell gehalten werden

• Strukturierte Daten einpflegen und Zusatzinfos den Suchmaschinen zur Verfügung stellen

Wenn diese Punkte beachtet werden, dann stehen die Chancen gut, dass man als Suchergebnis vorgelesen wird. Der nette Zusatzeffekt: alles was ihr hier für Voice optimiert, wird auch euer «normales» organisches Ranking positiv beeinflussen!

Janine Kraft

Digital Marketing Expert

Vom stationären Handel zum schnelllebigen Online Marketing: Seit über zehn Jahren im Retail und

über drei Jahren im Digitalen Marketing tätig, beschäftigt sich Janine mit verschiedensten Online-

Marketing-Disziplinen und hegt eine grosse Faszination für Artificial Intelligence und den daraus

resultierenden Voice Thematiken. Ihre Erfahrungen sammelte sie im Grosshandel und bei namhaften

Brands. Ihr Herz schlägt fürs Content Marketing, SEA/SEO, Social Media und das Performance

Marketing.

SCOTT TAYLOR

Wie betreibe ich SEO für Voice Search? - Teil 2

Sun, 10 Nov 2019 09:55:11 GMT

Der elementare Unterschied zwischen einer normal getippten Suchanfrage und einer Voice-Suchanfrage besteht in der Ausspielung der Ergebnisse. Während bei der getippten Suchanfrage mehrere Ergebnisse inklusive Anzeigen auf einer Seite angezeigt werden können, ist dies bei Voice Search nicht möglich. In diesem Fall gibt es nur ein einziges Ergebnis, welches vorgelesen wird. Dies erhöht den Konkurrenzkampf für die «eine Position» enorm. Momentan ist es auch noch nicht möglich bezahlte Werbung über Voice Search zu schalten. Dies klingt im ersten Moment etwas frustrierend, jedoch kann man die organischen Suchergebnisse von Voice Search beeinflussen. Wie das geht, wird im weiteren Verlauf des Artikels erklärt.

Inhaltsverzeichnis

1. SEO für Voice Search: 5 Fragen die du dir stellen solltest

2. Sechs Best-Practices zum Thema Voice-Search-Optimierung

3. Wie gestalte ich Content für Voice Search?

4. Fazit zu Voice Search

SEO für Voice Search: 5 Fragen die du dir stellen solltest

Jetzt die Weichen stellen für Voice Search Optimierung

Nach Björn Tantau, ein angesehener Online Marketing Spezialist aus Deutschland, braucht es folgende Fragestellungen, um eine Website auf Sprachsuche zu optimieren:

1) Wie schnell wird meine Webseite geladen?

2) Wird die eigene Webseite als vertrauenswürdig und verlässlich von den Suchmaschinen eingestuft?

3) Wie verständlich ist der vorhandene Kontext auf der Webseite? Wie zugänglich und einfach ist er geschrieben?

4) Wie gut sind die bestehenden Suchmaschinenrankings?

5) Werden die Inhalte der Webseite auch in Social Media Channels geteilt?

Durch diese Fragen kann geprüft werden, wie gut die Webseite auf Sprachsuchanfragen bereits vorbereitet ist. Speziell sollte auf die Ladezeit geachtet werden, da weder die Suchmaschine noch die Nutzerinnen und Nutzer eine langsame Webseite tolerieren. Die Frage nach dem Suchmaschinenranking ist insofern wichtig, da diese positiven Signale sich auf die Voice Search Ergebnisse auswirken. Google bevorzugt Inhalte für Voice Search, wenn diese auch in der traditionellen Suche gut ranken. Insbesondere wichtig sind in diesem Zusammenhang die «Featured Snippets» , welche im späteren Verlauf noch genauer erklärt werden.

Sechs Best-Practices zum Thema Voice-Search-Optimierung

1. SEO für Voice Search priorisieren: Immer mehr Menschen suchen über die Sprachfunktion nach Informationen. Je besser ein Unternehmen sich auf die Sprachsuche vorbereitet, desto grösser wird das Vertrauen von Google in eine Webseite. So kann dem zukünftigen Kampf um die eine vorgelesene Position vorgebeugt werden.

2. Markups: Damit eine Firma erfolgreich in der Sprachsuche sein kann, muss zwingend die Zielgruppe und deren Suchverhalten genau analysiert werden. Durch Markups, sogenannte Rich Snippets, kann Google gesagt werden, worum es sich handelt und wie Google etwas ausspielen soll. Weiter macht es Sinn, dass eine Art FAQ-Landingpage erstellt wird, bei der die typischen «Wer», «Wie», «Was»-Fragen beantwortet werden.

3. Content «vorlesbar» machen: Um sprachbasierte Content-Suchmaschinenoptimierung zu gewährleisten, muss der Content laut vorlesbar sein. Der Text soll einen natürlichen, semantischen Sprachfluss haben und durchgängige Textpassagen aufweisen. Die Schlüsselwörter sind keine einzelnen Keywords, wie «Ski Jacke kaufen», sondern Long-Tail-Keywords wie «Wo kann ich hochwertige Ski Jacken in Zürich kaufen?».

4. Page Speed: Die Website sollte möglichst schnell geladen werden. Die Nutzerinnen und Nutzer von Sprachsuchanfragen sind oftmals in Eile oder benötigen schnell Informationen und Ergebnisse. Langsame Webseiten sind ein No-Go und signifikante Umsatzbremsen für die Unternehmung. Empfehlenswert sind Accelerated Mobile Pages (AMP-Seiten), welche das Laden der Website signifikant erhöhen.

5. Bildbeschriftung: Immer wichtiger wird die Bildbeschriftung. Im Moment gibt es noch keine Möglichkeit, dass in der Sprachsuche die Suchmaschine das Bild erkennt und einen Kauf begünstigt. Daher muss ein Unternehmen ein Produkt verkaufen, ohne dass eine Abbildung notwendig ist. Damit dies funktioniert, müssen die Bilder, Fotos und Infografiken genau, zielgruppenorientiert und emotional beschrieben werden, damit die Suchmaschine die Bilder «lesen» und in die Evaluierung einbinden kann.

6. Google My Business: Für Unternehmen, welche speziell im stationären Handel tätig sind, ist eine Präsenz auf Google mit einem My Business Account unerlässlich. Typische Suchanfragen handeln von Standorten und Produktfragen. Solche Suchanfragen sind prädestiniert für die Spracheingabe. Speziell zu beachten sind auch die Be-wertungen und Fotos, welche von Nutzerinnen und Nutzern bei My Business hochgeladen werden. So bekommt die oder der Suchende gleich Produktbewertungen zu dem gewünschten Produkt. Durch den My Business Eintrag versteht Google, welche Produkte angeboten und welche Informationen dazu geliefert werden können.

Wie gestalte ich Content für Voice Search?

Guter SEO Content bringt dem Leser Mehrwert, ist unterhaltsam, informiert oder hilft bei einem Problem oder einer Fragestellung. Der Content muss einzigartig erstellt und exakt auf eine bestimmte Zielgruppe ausgerichtet sein. Die Qualität des Contents ist entscheidend. Die Textlänge soll ausreichend Material für Google und Co. hergeben, somit muss der Text eine gewisse Länge aufweisen. Guter Content ist aktuell und wird ständig ergänzt und erweitert. Je häufiger Inhalte erneuert werden, desto öfter kommen die Crawler von den Suchmaschinen vorbei, prüfen den Inhalt und indexieren diesen. Die übersichtliche Formatierung der Texte ist ebenso wichtig wie die Titelvergabe und die Grösse der Schriften. Der wichtigste Inhalt eines Textes sollte immer am Anfang stehen. Auch Grammatik und Rechtschreibung spielen eine Rolle und haben indirekt Einfluss auf das Ranking, da grammatikalisch korrekte Texte von Qualität zeugen . Darüber hinaus sollten so viele W-Fragen wie nur möglich in Blogbeiträgen oder Artikel beantwortet werden, da die Userinnen und User genau so ihre Frage stellen wer-den. Backlinks helfen das Vertrauen einer Seite zu erhöhen und verschiedene Medienarten: Video, Bilder oder Tabellen lockern den Text auf und strahlen Expertise aus. Produkt- und Kundenbewertungen helfen dabei Unique Content zu erstellen und geben gute Indizien zu den Produkten und der Dienstleistung.

Warum ist SEO Content wichtig für Voice Search?

SEO Content ist extrem wichtig, da Suchergebnisse über Voice nur ausgespielt werden können, wenn auf der Webseite Text zum Vorlesen zur Verfügung gestellt wird. Ohne Text wird es keine Voice Search Ausspielungen für die eigene Unternehmung geben. Hierbei ist darauf zu achten, dass die Texte einfach vorlesbar sind, das Wichtige am Anfang steht und der Text gut strukturiert ist.

Wie pflege ich Google My Business für Voice Search?

My Business ist ein kostenloses Tool für Unternehmen und Organisationen, welches von Google zur Verfügung gestellt wird. Durch dieses Tool kann die Onlinepräsenz auf Google Maps und der Google Suche schnell erfasst und einfach verwaltet werden. Durch die hinzugefügten Informationen können Kundinnen und Kunden wichtige Informationen erlangen und die Sichtbarkeit wird erhöht. Durch den My Business Eintrag können folgende Themen der Kundschaft zur Verfügung gestellt werden:

• Informationen zum Unternehmen oder den Stores können einfach verwaltet und angepasst werden. So können die einzelnen Verkaufsstellen erfasst und wichtige Informationen, wie die Öffnungszeiten, angezeigt werden.

• Google My Business ermöglicht die Interaktion mit den Kundinnen und Kunden. So können Bewertungen und Fotos zu Produkten und Dienstleistungen hochgeladen werden.

• Das Tool stellt Statistiken der Nutzung zur Verfügung. So können Anpassungen vorgenommen und der Service verbessert werden.

• Darüber hinaus kann das Tool sowohl auf dem Desktop, als auch auf dem Handy verwendet werden. Bei der mobilen Nutzung können weitere Features, wie zum Beispiel der Routenplaner, verwendet werden .

Warum ist Google My Business wichtig?

Ein grosser Teil der Voice-Suchanfragen sind Informationsabfragen. Beispiele dafür könnten sein:

• Welcher Schuhladen ist in meiner Nähe?

• An welcher Adresse finde ich den nächsten Schuhladen?

• Wie lange hat der Schuhladen geöffnet?

• Was ist die Telefonnummer des Schuhladens?

All diese Fragen kann Google anhand des My Business-Eintrages einfach beantworten. Daher ist es essenziell, dass Unternehmen ihre My Businesseinträge pflegen und aktuell halten. Diese Informationen sind nicht zu unterschätzen und bringen sehr viel Traffic auf die eigene Seite, wenn man die Einträge regelmässig bewirtschaftet. Ein kleiner Aufwand für ein tolles Ergebnis.

Fazit zu Voice Search

Da es momentan nicht möglich ist Werbung über Voice Search zu schalten, müssen Unter-nehmen sich auf die organische Voice Suche konzentrieren. Viele typische SEO-Massnahmen zahlen auch auf Voice Search SEO ein. Der Content spielt hier eine spezielle Rolle: ohne Content können keine Ergebnisse vorgelesen werden. Daher schreibe Texte so, dass sie vorgelesen werden können und das Wichtigste am Anfang steht. Des Weiteren solltest du unbedingt das kostenlose Angebot von My Business Einträgen annehmen und so viele Informationen wir nur möglich pflegen und hegen. Je genauer diese Daten, desto besser kann Google und Co. die Inhalte verstehen und den Kundinnen und Kunden vorlesen.

Janine Kraft

Digital Marketing Expert

Vom stationären Handel zum schnelllebigen Online Marketing: Seit über zehn Jahren im Retail und

über drei Jahren im Digitalen Marketing tätig, beschäftigt sich Janine mit verschiedensten Online-

Marketing-Disziplinen und hegt eine grosse Faszination für Artificial Intelligence und den daraus

resultierenden Voice Thematiken. Ihre Erfahrungen sammelte sie im Grosshandel und bei namhaften

Brands. Ihr Herz schlägt fürs Content Marketing, SEA/SEO, Social Media und das Performance

Marketing.

SCOTT TAYLOR

Was ist Voice Search? - Teil 1

Thu, 24 Oct 2019 21:02:54 GMT

Voice Search Assistenten

Der Siegeszug von Voice Search Assistenten ist früher oder später nicht mehr aufzuhalten, da die Voice-Software in Milliarden von Geräten implementiert sein wird. Die grössten Techfirmen sind weltweit mit Hochdruck daran, die Assistenten zu perfektionieren. Die Erleichterung des Alltags und die einfache Nutzung von Voice Search sind klare Vorteile der Software, da Zeit gespart und ein schneller Zugang zu In-formationen ermöglicht wird.

Auch gegen Ende vom Jahr 2019 sind Voice Search Assistenten ein reges diskutiertes Thema. Doch was ist Voice Search? Wie werden sich Voice Assistenten entwickeln? Was sind die Vor- und Nachteile für Userinnen, User und Unternehmen? Genau diese Themen werden im folgenden Artikel besprochen.

Inhaltsverzeichnis

1. Was ist Voice Search – Eine Definition

2. Voice Assistenten – Unsere persönlichen Assistenten?

3. Vor-und Nachteile für User

4. Vorteile und Schwierigkeiten aus Unternehmenssicht

5. Trends zum Thema Voice Search

1. Was sind Voice Search und Voice Assistenten - Eine Definition

Voice Search

Dann lasst uns zuerst mit einer Definition starten: das englische Wort «Voice Search» bedeutet auf Deutsch «Sprachsuche». Damit ist die spezielle Suche auf Google über eine Spracheingabe gemeint. Google Voice Search gibt es seit 2011. Damals war die Sprachsuche nur in englischer Sprache verfügbar. Im späteren Verlauf kamen dann viele andere Sprachen hinzu und eine Desktop-Version wurde entwickelt. Wenn von «Voice» gesprochen wird, dann ist die akustische Wiedergabe einer sprechenden Person gemeint, wobei mit dem Ausdruck «Speech» die generell gesprochene Sprache gemeint ist.

Voice Assistenten

Durch einen «Voice Assistant» , zu Deutsch «Sprachassistenten» , werden mithilfe eines Programms gesprochene Anweisungen und Aktionen ausgeführt. Bekannte Beispiele sind der Google Assistant, Siri, Cortana oder Alexa. Vor allem wird diese Applikation auf Smartphones verwendet. Mit dem Launch des iPhones 4 wurde der Sprachassistent zum ersten Mal in den Alltag der Menschen integriert. Siri wurde kreiert und wird heute dazu verwendet, das Wetter abzufragen, den Verbindungsaufbau von Telefonaten zu tätigen oder Termine im Kalender des Smartphones einzutragen. Voice Assistenten gibt es entweder als Applikation oder als spezifisch konzipierte Devices (Smart Speaker) wie der Google Home Assistant , Alexa von Amazon oder Cortana von Microsoft .

2. Voice Assistenten - unsere persönlichen Assistenten?

Durch den persönlichen Assistenten werden die Menschen noch mehr Zeit am Handy verbringen, wodurch sowohl die Werbeeinnahmen als auch die Geräteverkäufe steigen werden. Gemäss Gentsch, Professor und Spezialist für künstliche Intelligenz, wird die Schlagwort-Suche durch die Assistenten abgelöst:

«Die Suche nach Schlagwortbegriffen, zum Beispiel über Google, wird voraussichtlich im Rahmen dieser Entwicklung langfristig verschwinden. Stattdessen wird der Kaufentscheidungsprozess in Konversation mit dem digitalen Assistenten erfolgen.» (Gentsch, 2018, S. 93)

Damit ein Unternehmen oder eine Marke auch in der Zukunft Erfolg hat, müssen die jeweiligen Produkte und Dienstleistungen auf den Algorithmus der Assistenten abgestimmt werden. Wenn zum Beispiel Blumen durch den Voice Assistenten bestellt, eine Hotelbuchung durchgeführt oder Schuhe gekauft werden sollen, dann wird der Assistent die in seinem Netzwerk verfügbaren Marken und Dienstleistungen berücksichtigen können. Des Weiteren könnte auch die Rangfolge der Google-Ergebnisse eine grosse Rolle spielen. In Zukunft wird für die Kundinnen und Kunden weniger die Marke oder das Unternehmen im Vordergrund stehen, sondern die Bequemlichkeit (Convenience). Dies hat zur Folge, dass die Unternehmen, welche im Netzwerk der Assistenten angeschlossen sind, einen Vorteil haben werden.

Ein Netzwerk von Voice-Diensten

Damit Voice Assistenten funktionieren, braucht es zwei elementare Bestandteile: zum einen muss eine Verknüpfung von verschiedenen Diensten zu einem riesigen Netzwerk gegeben sein und zum anderen müssen die Assistenten «lernen» können. Um nun effizient die Anfragen der Userinnen und User bearbeiten zu können, müssen verschiedene Programme, Apps und andere Dienste miteinander kommunizieren können. Ein Beispiel dafür ist das Apple Betriebssystem iOS10, bei welchem Siri nun beispielsweise ein Uber-Taxi buchen kann, da die Verbindung zu den zwei Dienstleistern freigegeben wurde.

Was sind aktive Voice Assistenten?

Die heutigen Voice Assistenten agieren passiv. Damit ist gemeint, dass sie erst aktiv werden, wenn eine App aufgerufen wird, bestimmte Knöpfe betätigt oder Begrüssungen gesprochen werden. Unter aktiven Assistenten stellt man sich ein Gerät vor, welches mit Hilfe künstlicher Intelligenz selbständig Gespräche initiieren kann, Ratschläge gibt oder Missverständnisse aufklärt. Aktive Assistenten bergen aber auch Gefahren, beispielsweise wenn die Nutzerin oder der Nutzer unbedachte Aussagen macht oder eine Notlüge anbringt. In solchen Fällen könnte es passieren, dass der Assistent ins Gespräch eingreift und die Nutzerin oder den Nutzer blossstellt. Damit ein Assistent perfekt integriert werden kann, ist es essenziell, dass der Assistent allgegenwärtig und auf allen Geräten synchronisiert ist. So sollten andere Geräte an die Stelle des Smartphones treten können, wenn dieses einmal zu Hause vergessen wurde. Auch Gesten und Handbewegungen dürften in der Zukunft vom persönlichen Assistenten wahrgenommen werden können. Dies wäre möglich mithilfe von Kameras und Sensoren.

3. Vor- und Nachteile für die User

Die zwei grössten Vorteile von Voice Assistenten sind zum einen die Bequemlichkeit und zum anderen die Anpassungsfähigkeit an die Userin oder den User. Durch einen persönlichen digitalen Assistenten müssten keine Apps mehr gesucht, heruntergeladen, installiert oder konfiguriert werden. Die Zeit zwischen einer Anfrage und der Antwort kann reduziert werden, was die Convenience für die Nutzerin oder den Nutzer stark steigert. Durch die Personalisierung der Assistenten vergrössert sich deren Anpassungsfähigkeit. So kann der digitale Assistent eine Art «Bewusstsein» für Zusammenhänge entwickeln. In Zukunft könnte man von einem persönlichen Assistenten erwarten, dass dieser sich an die Nutzerin oder den Nutzer anpasst, so wie es sich die Menschen von der zwischenmenschlichen Interaktion gewöhnt sind. Für die Userinnen und Unser von persönlichen Assistenten wäre es essenziell, dass diese angeben könnten, wann sie Informationen vom digitalen Assistenten erhalten wollen oder welche Rah-menbedingungen erfüllt sein müssten. So könnten die Textnachrichten während des Autofahrens ausgeschalten und beim Eintreffen zu Hause aktiviert werden.

Ein anderer wichtiger Punkt wäre, dass die persönlichen Assistenten sich der Stimmungslage der Nutzerin oder des Nutzers und der aktuellen Situation anpassen könnten. Zum Beispiel kann es sein, dass die Nutzerin oder der Nutzer unter Umständen müde ist oder gerade einen Abend mit Freunden verbringt und sich nicht aktiv mit bestimmten Dingen befassen will. So könnten personalisierte Algorithmen automatisch nach dem besten Ermessen Entscheidungen treffen, ohne ständige Rückkopplung zur Userin oder zum User. Wenn der Algorithmus des Geräts über solche Umstände informiert ist, könnte er empathischer und zugänglicher reagieren. Mit anderen Worten würde sich der Assistent so anpassen, wie es reale Menschen tun würden.

Quelle: Vorteile intelligenter Lautsprecher (Zuberer, 2018, S. 6)

Die Assistenten haben nicht nur Vorteile, sondern auch Schattenseiten. Bedenklich findet die deutsche Bevölkerung den Gedanken an den «gläsernen» Kunden und den Missbrauch von Daten. Die EU hat mit der DSGVO-Verordnung auch auf staatlicher Ebene Einschränkungen getätigt, was in Zukunft noch mehr dazu führen könnte, dass die Entwicklung von Voice Assistenten oder allgemein KI gehemmt wird.

Quelle: Vorteile intelligenter Lautsprecher (Zuberer, 2018, S. 6)

4. Vorteile und Schwierigkeiten von Voice Search aus Unternehmenssicht

Voice Search ist für Unternehmen unter anderem aus diesen Gründen wichtig:

• Suchmaschine: Schnellere Interaktion führt zu grösserer Kundenbindung und Engagement.

• Konsumentin/Konsument: In manchen Fällen ist Voice Search schneller als die getippte Suche durch die Tastatur. Viele Konsumentinnen und Konsumenten streben ausserdem nach einer positiven Kosten-Nutzen-Bilanz.

• Unternehmen: Kundenverhalten und Suchmaschineninvestment gehen Hand in Hand. Die Kundinnen und Kunden haben viel Erfahrung mit Suchmaschinen und mobilen Geräten. Menschen vertrauen darauf, dass die Suchmaschine die beste Antwort für die eigenen Bedürfnisse hat.

Voice Assistenten bringen viele Vorteile für Kundinnen, Kunden und die Unternehmen. Durch speziell auf die Kundin oder den Kunden zugeschnittene Dienste, kann die Kundenbindung gestärkt werden. Dieser Vorteil von personalisierter Werbung ist auch für die Kundin und den Kunden ein Mehrwert und verbessert die Dienstleistungen. So sind letztendlich nicht nur die Kundin oder der Kunde zufriedener, sondern es kann auch der Ruf der Unternehmung und die Bekanntheit einer Marke gestärkt werden. Durch die personalisierten Assistenten erhalten Firmen mehr Einblicke in die Wünsche und Bedürfnisse der Kundinnen und Kunden und deren Kaufprozesse.

Was sind die Hauptvorteile von Voice Search:

1. Suchvorgänge werden beschleunigt, da Voice Search eine schnellere Eingabe ermöglicht als über die Tastatur.

2. Die Absicht der Userin oder des Users lässt sich einfacher feststellen, da die Sucheingaben spezifischer formuliert sind.

3. Voice Search ermöglicht Handfreiheit, Bequemlichkeit und mehr Komfort.

4. Die simple Handhabung hilft bei alltäglichen Belangen.

Des Weiteren bringt Google immer weitere Optimierungen und Funktionen auf den Markt. So kann auf Geräte, welche KaiOS (Betriebssystem für FirefoxOS) unterstützen, Voice Typing angewendet werden. Damit kann Sprache geradewegs in Text umgewandelt werden. Google Lens ist ein weiteres Beispiel, welches Bilder und Texte aus Aufnahmen erkennt, direkt vorlesen, übersetzen oder mit weiteren Google Features verbinden kann.

Bild: Voice Search über Desktop mit Google

5. Trends

Quantified-Self Bewegung

Durch die Entwicklung von KI wird auch der Einsatz der Voice Assistenten und deren Leistung gesteigert. Die Spracherfassung nimmt jedes Jahr um 20% zu. Es ist heute möglich, schon über 90% der gesprochenen Sprache zu erfassen, dank der Verarbeitung grosser Datenmengen in Bezug auf die natürliche Sprache (Natural Language Processing).

Ein weiterer Trend ist die sogenannte «Quantified-Self-Bewegung» , auch unter dem Namen «Lifelogging» bekannt. Es handelt sich dabei um Menschen, welche personenbezogene Daten über den Tag hinweg erfassen und analysieren, meist durch eine App. Diese erfassten Daten könnten sein: das konsumierte Essen, die Luftqualität, der Gemütszustand, Blutsauerstoffwerte sowie physische und mentale Leistungen. Auch Wearables, also am Körper tragbare Geräte, ermöglichen die Aufzeichnung von Daten. Dies kann zum Beispiel geschehen durch Sensoren und Elektronik, die in die Kleidung eingearbeitet werden. Zusammen mit den Fortschritten im Bereich Data Science hat dieser Trend das Potenzial Kundeninformationen im Voice Assistant Bereich noch mehr zu personalisieren sowie die Bedürfnisse und Wünsche der Kundinnen und Kunden noch besser vorherzusehen und auf Basis der Daten präzisere, personalisierte Werbung zu generieren. Damit Voice Assistenten in den Kaufprozess integriert werden können, braucht es nahtlose Zahlungstechnologien. Solche Anbindungen sind durch Drittanbieter in immer grösserem Ausmass als Application Programming Interface (APIs) verfügbar.

Fazit

Voice Assistenten bringen viele Vorteile für Kundinnen, Kunden und die Unternehmen. Durch speziell auf die Kundin oder den Kunden zugeschnittenen Dienste, kann die Kundenbindung gestärkt werden. Dieser Vorteil von personalisierter Werbung ist auch für die Kundin und den Kunden ein Mehrwert und verbessert die Dienstleistungen. So sind letztendlich nicht nur die Kundin oder der Kunde zufriedener, sondern es kann auch der Ruf der Unternehmung und die Bekanntheit einer Marke gestärkt werden. Durch die personalisierten Assistenten erhalten Firmen mehr Einblicke in die Wünsche und Bedürfnisse der Kundinnen und Kunden und deren Kaufprozesse.

Die Vorteile für die Userin oder den User sind die einfache Handhabung und die Anpassungsfähigkeit an uns Menschen. Firmen können dagegen eine starke Kundenbindung aufbauen und sich profilieren mit hoch personalisierter Werbung. Der Trend geht immer mehr in die Richtung von «Lifelogging» und dem «persönlichen Assistenten» den jeder griffbereit in der Tasche mit sich führt. Ein spannendes Thema, welches uns in nächster Zeit noch stark begleiten wird.

Janine Kraft

Digital Marketing Expert

Vom stationären Handel zum schnelllebigen Online Marketing: Seit über zehn Jahren im Retail und

über drei Jahren im Digitalen Marketing tätig, beschäftigt sich Janine mit verschiedensten Online-

Marketing-Disziplinen und hegt eine grosse Faszination für Artificial Intelligence und den daraus

resultierenden Voice Thematiken. Ihre Erfahrungen sammelte sie im Grosshandel und bei namhaften

Brands. Ihr Herz schlägt fürs Content Marketing, SEA/SEO, Social Media und das Performance

Marketing.

SCOTT TAYLOR

Why Digital Assitants?

Tue, 24 Sep 2019 18:34:25 GMT

Why Voice Assistants?

Voice assistants also called voice user interfaces using text-to-speech and speech-to-text capabilities are gaining more relevance due to their technological progress in the fields of voice recognition engines (AVR), machine learning algorithms and architectural design.

What are the advantages of using voice instead of normal GUIs or text?

Speed. Speaking is faster than writing. On average, people speak 150 words/minute. for search, voice can be useful: Search Engine Land considered voice assistants as taking over a “critical mass” with 122,7 million users in 2021. According to ComCore, 50% of searches will be done by voice and 30% of web browsing sessions will be done without a screen until 2020.

Voice is more natural. Emotions, needs, and wishes will be articulated in more detail which offers the company a better picture of how customer’s needs look like.

Brand Recognition. Voice creates a new channel for delivering a brand.

Handsfree. In any situation where handsfree is useful, voice is ideal, i.e. when shopping with children, when driving, when cooking, when industrial workers are picking, when surgeons are operating, etc. There are various use cases where voice very useful. Furthermore, people with disabilities can “talk” to message with friends, shop or execute tasks before writing was omnipresent.

Which Voice Assistants are used?

Source: Usage of Digital Assistants ( Microsoft Voice Report , 2019)

Who uses Voice Assistants?

All age groups use assistants, however, we find that in Switzerland as well as internationally Millennials and Seniors using more voice assistant or speakers than the age group between 35-44. Interesting is the fact, that older generations more frequently use Amazon Alexa. 1/3 of the total US population use voice assistants at least monthly on a device ( eMarketer, 2019 ). In Germany, out of 605 person, 30% use it more than weekly, 30% monthly ( Splendid Research, 2019 ).

Source: Switzerland Voice Assistant Usage per Age Group ( Deloitte , 2018)

Source: Smart Speaker Brand Choice by Age Group ( Voicebot.ai , 2019)

In which fields are Digital Assistants used?

Capgemini’s showed in a study with 12’000 consumers and 1000 executives who used voice or text assistants (or both) that 50% of the key players, especially in the fields of automotive, banking, insurances and consumer products have deployed first voice or chat assistants.

Source: Use Cases of Digital Assistants ( Capgemini Research Institute , 2016-19)

In which fields are chatbots vs. voice applications?

Source: Chatbot vs. Voice Assistant Use Cases ( Capgemini Research Institute , 2016-19)

Voice Assistants on Smartphones and Smart Speakers?

According to Splendid Research German users are currently more likely to use smartphones with assistants than Smart Speakers. In Switzerland and Austria, we see the same usage pattern. However, this highly depends on the age group and technology provider (Cortana does not have a smart speaker). Voice assistants are also implemented in other devices, such as wearables, smartwatches, and TVs.

Source: 60% of the Germans already used Voice Assistants on these devices ( Splendid Research , 2019

Which Technology Providers are the Market Leaders?

When looking at the global market shares of smart speakers, Strategy Analytics finds Amazon, Google and Baidu to be the global leaders. By 2019 Google has 1 billion Google Assistant enabled devices whereas Amazon 100 million enabled devices according to their market share announcements in January 2019. According to Venturebeat Cortana has 500 million active monthly users on Windows 10 PCs and "tens of millions of users on Xbox ". Siri from Apple has over 500 million active users in 2018.

Source: Development of Major Global Voice Tech Players ( Strategy Analytics , 2019)

In Switzerland, Siri seems to be still dominant, followed by Google Home and Amazon. Finally, various sources clearly announce a shift to Google Assistant, also for Switzerland and Europe.

Source: Swiss people know and use Siri and Google Home ( Mcschindler.com , 2019)

Source: Shift from Amazon to Google Technologies ( Voicebot.ai , 2019)

How to Design Digital Assistants?

Wed, 31 Jul 2019 20:23:39 GMT

CUI Design Learnings

I summarized the most evident facts you need to follow to create your digital assistant. Various Bestseller books and papers on chatbot development as well as my long-year-research experience were used. In general, you can use these recommendations for bots, avatars, voice interfaces but also other types of digital assistants. Start with these points to become a successful CUI designer.

1 clear onboarding phase and greeting:

Make clear to the user what he can expect from the digital assistant. For example, it should be clear what the chatbot's goal and functionality are. You can do this by offering mini-tutorials at the beginning of the conversation. This clarifies the expectations of how intelligent an assistant is and reduces the probability of a user being disappointment (Valério, Guimarães, Prates & Candello, 2017). A “capability check" , i.e. a possibility to ask the assistant, in the sense of "What can you do?", should be built in (Moore, 2018).

2 personality and the branding

are declared by numerous conversational designers as the first step to start with:

• First, think about the vision, mission, goals, and actions your CUI needs to solve.

• Second, consider the opinion, goals, and needs of the users: What are there personas? What are their needs? pain points?

• Third, depending on the use case and emotional level of a conversation, visualize your digital agent (Pearl, 2017). Comic-like chatbots are perceived as more likable and trustworthy (Luo, McGoldrick, Keeling, & Beatty, 2006). Photorealistic agent, called avatars tend to be higher concerning their abilities and level of intelligence (Dehn & Mulken 2000, p. 2).

• In which context do you use the agent? In e-commerce for example, voice and text-based chatbots are more common than embodied agents. These are increasingly used in the context of games and as educational agents (Knote et al., 2018).

3 guided, contextual conversation.

Depending on whether you build a rather text-based interface or voice-first agent you need to guide your user. You can either use:

• Rich controls (e.g. buttons, images, carousels, maps) to respond to the intentions of the user (Shevat, 2017). Of course, they limit the user’s freedom but can reduce the risk of entering something the system is not prepared to (Valério et al. 2017).

• Design contextual : The agent should remember the previous nouns ( "Peter likes apples (...)", answer with pronouns ("he" instead of "Peter") to respond more natural. Use variations for the same user intention ("Peter likes fresh fruit"). Prepare your assistant for upcoming negations ("I don't want red apples"(Jain et al., 2018)), silly user inputs or complete new requests during a conversation which are not referred to the current context.

4 assistance and repetition

The user should be able to say “I need help”. Consider all cases: The agent helps in the current context but also when the user needs to get back to an overview or start. With chatbots, use context-related global help function via a persistent menu. At the same time, however, it should also be possible to ask for help during the conversation (Moore, 2018).

5 main functions, persistent menus

Focus on fewer functions when developping a conversational interface. Take the main functions or use cases (also intentions) and implement them consistently and sufficiently. The user should be able to go back in the conversation, easily change already entered or said inputs. The user can use a main menus to display main functions again during the chat process and react to them. This makes it possible to change messages and entries from the past, thus, change time schedules, booking entries or chosen articles.

6 humour and playfulness

are fostering a satisfactory user experience since the conversational agent is perceived as more human (Luger & Sellen, 2016, Habicht 2019). However, some agents use fun elements for error handling which – when appearing too often – are not solving the user’s problem, and are annoying for the user. Studies show that more conversational turns are generated with entertainment/gaming bots than with shopping chatbots, having less hedonistic goals (Jain, et al., 2018).

7 humour and playfulness

When building assistants think about these possible errors: A) No speech/input detected, B) Speech/input detected but not recognized, C) Recognized but wrongly handled by the system D) Incorrectly recognition. Whether an error handling is initiated by your assistant or not highly depends on: The options to continue the conversation (or break it!), presence of listening indicators (i.e. Alexa ring light, or an Avatar actively looking at you until you ask the question again).

8 a void media breaks

Create an assistant being able to call-up a network of services, including basket, payment options, calendar, and other third-party functionalities. Assistants opening up too many external links are not optimal from a user-perspective but sometimes legally necessary (i.e. checkout, privacy, opt-in).

9 conversation exit

Your assistant should say goodbye or offer a way to finish the conversation easily.

Keep in mind that a loss of confidence and trust can be avoided with a user-friendly design. Successful examples of conversational interfaces are known to convey high levels of trust.

Nina Habicht

Founder VOICETECHHUB

Growth Hacker and Business Developer provides strategic and operational support to start-ups and companies in the development of visions and products. She holds two Masters degrees, an MAS in Business Informatics and an MSc in Strategic Management and Marketing. In Switzerland she is an expert, author and researcher in the fields of AI Design, Conversational and Voice User Interfaces as well as Chatbot & Avatars since 2011.

With over 10 years of experience, she supports various companies and industries in the fields of Digital Leadership, Strategic Management, Business Development, Web & Data Analytics, Requirements Engineering, UX-/CUI-Design, Prototyping, Communication and Digital Marketing. From 2016 to 2018, she was founder of Brainshare, an AI startup for innovation and ideation. In 2019 she founded VOICETECHHUB - a platform fostering AI Experience and Voice Tech projects and products. It also supports customers with the implementation of digital agents such as Amazon Alexa or Google Home and on-premise Conversational Interfaces.

Voicetechhub Feed

AI Cybersecurity vs. Security for AI vs. Red Teaming

What is AI Cybersecurity and what is Security for AI?

What is Red Teaming?

Important Red Teaming Definitions Business People Need to Know

The AI Security Checklist - How can I Mitigate the Risks of AI Attacks?

Do you need a Sparring Partner for AI Security and Red Teaming?

Which AI Presentation Tools are Suitable for Enterprises?

Which Tools are Best for AI Powerpoint Creation?

What PowerPoint AI Tools are Suitable for Enterprises?

AI Presentation Overview

﻿

What are the Limiations of AI Presentation Tools?

How to Create Stunning AI-Generated PowerPoint Presentations

Are These Reviews Paid?

Which Tool for What? State of AI Tools 2025

Should I use several AI tools or stick to one platform?

Our guideline to be successful with your AI tool journey

Conclusion

How to Use AI Chatbots in Your Business – and Stay Compliant

Why Chatbots Matter in 2025

Typical Use Cases for AI Chatbots

Business Benefits of Customized AI Chatbots

Legal & Compliance Aspects

Implementation Roadmap

Conclusion

Reasoning vs. Deep Research vs. RAG

What's RAG?

What is Reasoning?

How does Deep Research work?

Conclusion:

Do you want to create your secure Deep Research for your company?

AI Video Creation: The Ultimate Guide to Runway, Luma AI, Haiper.ai, and Hailuo AI

Why AI Video creation tools are game-changers?

1. Runway - The Professional AI Video Suite

2. Luma AI: Revolutionizing 3D with Photorealistic Modeling

3. Haiper AI: Simplified AI for Creative Video Ads

4. Hailuo AI: The Artistic AI for Video Content

Conclusion: Choosing the Right AI Video Creation Tool

Do you want to deep-dive into AI video creation?

What are the Best AI Powerpoint Tools

What are useful AI Powerpoint Tools?

What are the drawbacks of AI powerpoint tools?

Do you want to deep-dive into AI powerpoint creation?

Image Generator and AI Branding Tool Review for Marketing and Brand Managers

What to Choose: Midjourney, Flux, Getty Images or Adobe Firefly?

Pros and Cons of Image Creators

Midjourney

FLUX

RunDiffusion

Do you need support with Image Creation Skills?

Optimizing your Website for AI: How to get found by ChatGPT

What do Users Tell on ChatGPT vs. Google Search?

How Swiss People Search with AI Tools

In Switzerland, one third are using ChatGPT for search reasons. The latest report from Nau Media AG reveals surprising facts from Switzerland:

Worldwide studies show happy users with AI Search

How to optimize Websites to be found by Large Language Models?

1. Quality and Relevance of Content

How will AI SEO look like in the Future?

Do you need support with Large Language Optimization?

A Practical Guide for Midjourney Image Generation

How to prompt images with Midjourney?

What is different with Midjourney Website?

The installation in discord and especially the technical parameter handling and prompting is much easier with the Midjourney Web Version since you do not need to know all parameters by heart)his is how the new Midjourney Website looks like. You can find the web version here: Midjourney Feed .

Parameter Settings in Midjourney Website

How to use styles with in Midjourney

sref codes are specific parameters used in Midjourney to apply preset styles to generated images. Each code corresponds to a particular visual effect or aesthetic. To apply an SREF code, simply add it to your prompt when generating an image. For example: A serene landscape --sref 19961010.

How to change parts in an image with Midjourney (Inpainting)

How to create consistent styles with Midjourney Web Version

How to create (website) patterns with Midjourney

Do you want to create professional images with AI?

Understanding Agentic AI: A Game-Changer for the Future of AI

Agentic AI: A Simple Guide

What is Agentic AI?

How do Agentic AI Systems think?

What is the Difference between Agentic AI and Generative AI?

Agentic AI or Chatbots?

What is a Large Action Model (LAM)?

Agentic RAG vs. Native RAG

When not to Use Agentic AI?