Problem-solver in extraordinary circumstances: How Offline-AI tackles your challenges

The advantages of an optimized AI for companies and administration can even be demonstrated on a laptop offline. This Offline-AI performs better than any universal intelligence like ChatGPT in many use cases. It offers full data control and runs very cost-effectively even in continuous operation.

Introduction

This post is a practical example of the possibilities of Offline-AI. But first, a short introduction to Offline-AI. It has been possible for a year now, but only recently has it become so powerful, even for the German language.

Offline-AI is a locally running AI on its own computer or server. Offline means that the AI system runs completely without internet connection. This results in no data being sent to third parties. The following advantages are offered by an Offline-AI, especially:

• Companies and authorities get problem solver for their use cases.
• At any time new AI language models, Stable Diffusion image models, audio models, video models can be installed and used.
• Even continuous operation is cost-effective possible. Goodbye Microsoft Azure. Goodbye security loopholes.
• A self-improving AI that can criticize itself is practically possible.
• Legally simple: Contract processing agreements are unnecessary; the dubious Opt-Out instead of an opt-in inquiry that ChatGPT offers, has no legal consequences anymore. Nobody needs to jump again when another supervisory authority bans ChatGPT.

For completeness' sake, it should be noted that the programming interface (API) of ChatGPT does indeed sometimes make sense. However, this case is becoming increasingly rare thanks to open-source language models that are becoming more powerful on a weekly basis. To preface: With an API, so much can be done. To illustrate: From your given prompt, a language model forms a program description of an image composition, which is then converted by a image generation model into an image. Here's an example output:

For German as a language, there are often better language models than ChatGPT. Either they are qualitatively better or cheaper. Interested parties should look at Claude, Command R+ and Mistral. Offline-AI is more fun, though, and better for concrete problems.

When Offline-AI and when general chatbot?

If you want to know how high the third-highest mountain in the world is, feel free to use ChatGPT. Using chatbots like this one for occasional summarizing of public texts for semi-professional purposes is also a good application. These chatbots are also available for a harmless chat.

Using a chatbot as a search engine is an approach that apparently many people like. If you want a particularly powerful search engine, you're better off with Offline-AI. OpenAI will certainly not read and evaluate a specific website with 5,000 subpages for you.

A high reliability of results is not only desirable but always important in a professional environment. Whoever needs this reliability requires an Offline-AI with mechanisms for quality control.

Sensitive data, business secrets or personal information should not be entered into a third-party system of American providers by anyone. Firstly, US laws like the Cloud Act, FISA 702 or Executive Order 12333 allow American authorities and intelligence agencies to access this data. Secondly, security breaches in popular chatbots and systems from Microsoft (majority shares in ChatGPT!) occur almost regularly. Thirdly, legal assessments and regulations need to be made, which only happens on paper and is a lot of work.

With intelligent AI solutions, companies and administrations can overcome challenges.

Faster, better, and easier than ever.

An Offline-AI can run permanently. Why, some wonder. Why do humans study for 5 years? To get better. Exactly that can an Offline-AI do. Only this self-improvement does not take years, but often seconds or a few hours.

What is possible with Offline-AI?

Offline-AI can do a lot. Even more than ChatGPT. It all depends on the application case. Your data, your use case. Even if you don't have your own data, Offline-AI is usually better.

In the previous section, an example was already given: Retrieving information from the internet. Searching for information on the internet and subsequently evaluating it is also a very good use case for an Offline-AI. For example, current news can be retrieved daily and automatically evaluated with the company's own AI.

How about a market analysis or a competitive observation? It doesn't go well at all with general chatbots, because the data has to be obtained first. This data acquisition is easily automated. Whoever takes this small step gets out of the innovation-hostile comfort zone. As a reward, there's maximum independence, full control and above all better results!

Examples of AI use in administration:

Citizens (or customers) want to make a request. The search for the right form or brochure can be optimally supported by AI. AI is used here for semantic search. This relieves administrative staff and enables them to work more productively.

If a leaflet is too complicated, AI can summarize it in simpler language.

If a form or application is only available in German or English but the citizen understands only Ukrainian, then the document can be translated with AI.

Similar requests and forms related to a found request or form can be found with AI. Also laws, internal guidelines and judgments related to a process can be found in this way. Staff members are thus supported as best possible.

Analog possible for processing damage reports, customer inquiries or support tickets.

In the previous post, further good scenarios for Offline-AI were mentioned.

Evidence for Offline-AI

At the Mainz Digital Congress MADKON24, participants in my workshop were able to witness live how Offline-AI produces results. This took place on my laptop on-site and without an internet connection.

The image shows impressions of MADKON24. A report on lectures and workshops, including mine, has been published by Peter Winninger in his blog KOBLENZ DIGITAL.

On the Laptop, it was shown in particular:

• Image generator: Text input (prompt) generates image output
• Image generator 2: Input image serves as a default for the image structure. A new image is created from this specification and a prompt.
• Prompt optimization: The prompt for image generation was automatically optimized with offline AI. The optimized prompt was then fed into the image generator. The new image was more appealing.
• Prompt variants: Variants were automatically created for the prompt using offline AI:
  • General variants that are similar to the original
  • Variants that are very similar to the original
• Translation: The automatically optimized prompt and the automatically generated variants were
  • translated into
    • English,
    • Spanish,
    • Ukrainian,
  • reformulated, namely
    • into legalese,
    • into pirate language,
    • into the language that a five-year-old understands.
• For fun, a fairy tale was created from the optimized prompt, in two variants

The whole thing went through during the workshop, it felt like a dozen times, each time with a single different input and each time with many new and good results.

For the previously mentioned prompt optimization: This self-improvement of AI is possible by having the AI receive its own outputs again as inputs. What also works very well is the artificial generation of training data by the AI, with which the same AI is then trained. That's like a human writing their own learning material and thereby becoming better.

Here is an example of the output of the mentioned image generator:

Enlarging this already not small image is possible with AI upscaling. This allows for the creation of a four times larger size without noticeable quality loss.

The aforementioned image generator 2 was now given this image as input and was supposed to create a crazy scientist from it; while maintaining the image structure. Here is the result from Mainz:

In addition, variants with a mole, tiger, often unpopular politicians from the USA, Commander Data, a prominent lady, and many more were calculated. All this on the laptop in short time. In total, around 100 images were generated in a short part of the two-hour workshop. As another example, an image generator was shown that had been trained to a specific style. For example, images can be generated in the style of IKEA instructions for every text prompt.

If one were to count the number of texts generated with Offline-AI in the workshop, it would come to several thousand texts. These were optimized text prompts, translations, language simplifications, and more. For example, keywords were extracted from news articles. Similarly, search engine hits were classified using Offline-AI. This way, knowledge articles could be distinguished from advertising messages. Even retrieving hits from a popular search engine (not Google) can be automated.

The barely imaginable possibilities of an Offline-AI:

An input, then a continuous AI loop. The AI generates further inputs itself and thus produces more and better results. If there are too many results, these can be criticized and filtered by the same AI.

Now possible, even on a laptop. Not a wish, but reality.

Present was also Michael Westphal from TV1. He recorded my workshop (without my knowledge) as audio. Afterwards, he delivered me the transcript, with Offline-AI. Although he only used a good mobile phone microphone, numerous people were in the room, the room was large and several people spoke, the result was very good. A recording of an interview with him and me after the workshop will be published soon. Also in the video, AI is used for transcription and subtitles.

I used the Offline-AI to hunt for the audio transcript (for language in a length of two hours) so that it could be summarized. After that, the summary was linguistically simplified and key statements were extracted.

The key takeaways from my AI workshop at the Digital Congress in Mainz are:

• AI works like the human brain: human intelligence function has been deciphered.
• Nvidia is currently the only practically available provider of AI graphics cards.
• Open-source AI models are great.
• Applications that could not be programmed at all or only after months or years can now be developed in the shortest possible time thanks to AI.
• AI allows for the uniform solution of diverse problems. This significantly reduces complexity and effort.
• AI can be applied uniformly to different data types. Data types include text, images, audio signals, sensor values, videos, etc.
• AI applications require data to deliver meaningful results.
• The origin of this data can be: unordered list List item Historical data, such as bid data from completed tenders, list item marker List item Data from the internet (via automated search), list item marker List item Artificially generated data (Offline-AI usage), list item marker List item Human-generated data. list item marker Unordered list end tag
• Reliable answers can only be provided by an AI that is guided. This guidance can occur automatically.
• General AI systems like ChatGPT are qualitatively worse (and in the medium term usually much more expensive) than specialized systems. .
• Only the combination of AI and classical programming delivers the best results.
• Innovation arises from leaving behind comfort and dissolving dependencies.

The best solution is just a small step away these days.

Offline-AI illustrates with images

The featured image at the top was generated by an Offline-AI. It can be seen on the far left in the following collection of three images:

The first (left) image was generated by an Offline-AI based on a manual prompt. The second image was generated by Offline-AI, which was tasked with adjusting the lighting and adding more details to the image. The third (right) image was generated by an Offline-AI image generator. The prompt here was:

• Input image (previously AI-generated, but any other image is also possible)
• Assignment: Add a few very small robots that dance around the light bulb_

It is also possible to ask the AI itself which other prompts would be good. One must, of course, make some kind of specification. For example: I want the image to contain more details/look more interesting/remind me of cake..

Depending on the offline image generator used, it is even possible to chat with the generator. First, you let an image be generated. Then you say: Replace the man with a woman. This process could also be automated by a AI critic. At the end, x images are created. You can either sift through them yourself or have them pre-filtered by a AI. The AI can learn which images you like. To do this, simply mark for each image whether it pleases you well or not so well. The AI then learns from a few examples what you want.

The prompt that led to the post image was optimized by an AI as follows:

• Original: problem solver system, intelligent, connections, bright colors
• AI version: Vibrant problem-solving AI, illuminating connections with radiant hues.

After that, an image was generated using the AI version of the prompt:

Now the local Offline-AI was instructed to add a futuristic dinosaur:

As you can see, 3 variations were generated. We could have also generated 1,000 or 100,000 variations. It doesn't cost anything. Thanks to an offline image critique AI, the top 10 images can then be filtered out. The human can then look at them and quickly find the best image.

Conclusion

The advantages of AI are particularly offered for companies and authorities with an Offline-AI.

The quality is simply much better for many problem situations. This is achieved through conventional additional programming, which causes very little effort. Plastic mini-example: How many letters a word has can be determined best this way, while ChatGPT fails.

Full data control** is another important aspect. Let your chosen lawyer do something other than speak nicely about what isn't nice: ChatGPT is uncertain, Microsoft is uncertain (proof in this blog as well as following).

The most important thing is the strategic significance of Offline-AI. It can improve itself and self-criticize. All this is practically only possible because no usage-dependent fees occur. Even organizations with large budgets would like to know beforehand what "it" will cost. But nobody knows that beforehand when a brain (AI) is questioned. Clearly, the very frequent call of a fee-based AI interface, best still with extensive documents, quickly becomes expensive.

Innovation arises through the removal of obstacles. Exactly that is possible with Offline-AI. Since the development is so rapid, it's a huge advantage that new LLMs can simply be exchanged. Open-source language models, for example, are so good that commercial providers have to get very warm in order not to freeze.

What use cases do you have in your company or authority or administration? You would be happy to receive a first assessment of feasibility.