Artificial Intelligence: Answer to the question “Am I bad?

Due to my contribution about a question-and-answer system for the data protection blog Dr. GDPR, a somewhat tongue-in-cheek question arose as to whether I am evil or even bitter evil. This question can of course be answered by my autonomous AI system. Depending on the configuration, different answers come out. The results show that artificial intelligence works and produces outstanding results.

The results of an autonomous AI system

For some time now I have been developing data-friendly AI systems. Data friendly means that data does not reach third parties but remains within its own system. The opposite of data friendly are systems like ChatGPT, Amazon AWS, Google Cloud or systems from Meta.

In my AI contribution to a question-answer assistant, I have printed various questions from me to my AI along with the answers of the AI. One of the questions that I asked myself was:

“Is Klaus Meffert evil?”

Ask my autonomous AI system a question.

I asked this question because a prompt naturally invites you to ask all kinds of questions.

In passing, I'd like to start a thought experiment: Can one actually influence a chatbot like ChatGPT by using specific prompts? Generally speaking, that's possible, and it happens when OpenAI/Microsoft use user inputs for the “improvement” of their AI models. If you ask the same question multiple times about whether Olaf Scholz is evil, then an AI can draw conclusions from that which are likely not valid.

The answer to the question just asked about me (no, I am not Olaf Scholz) was given by my AI as follows:

Based on the provided source documents I cannot assess whether Klaus Meffert is evil. There are no information about what evil means or if Klaus Meffert was ever or is evil.”

Response of my autonomous AI system to the question “Is Klaus Meffert evil?

When an artificial intelligence is supposed to give answers to questions, this task is generally also referred to as text generation (“text generation”). The results of the AI are influenced by different parameters. These parameters are particularly:

• Number of “best” source documents to consider for an answer. A machine can't process all documents at once. Always applies, also to ChatGPT.
• Size of the ingested text chunks: A machine learning model can't process long documents all at once. Always true, also for ChatGPT.
• Overlap value for read text snippets. This prevents two semantically related sentences from being separated from each other.
• Confidence threshold for selecting the best documents: With a value of 0.99 (= 99%) almost no document would be found that could contain an answer to a posed question.
• Procedure: How should the best source documents found be combined to form an answer? For example, there is the approach that generates an answer from the best document, which is then checked against the second-best document and refined over it.
• Used algorithm for vector similarity search. I'll spare the details here, it would otherwise get too technical.

These parameters were not yet optimally adjusted for the above question. That is now better, even if the above answer can already be described as “perfect”.

Now one of my readers has speculated on my above mentioned contribution and the question-answer pairs mentioned there, what my local AI system would answer to the following question:

Is Klaus Meffert bitter?

Question from a reader to my AI.

The reader meant: “It would be easier to get a meaning out of it. No idea if the AI can evaluate previous contributions afterwards.” Of course, I asked this question to my AI.

Some very interesting and good answers came out, as I find.

Depending on parameter settings, the answers were different, but really impressive.

We hopefully agree that such questions go beyond the knowledge base provided by this data protection and AI blog. Nevertheless, it is an interesting experiment to observe the behavior of the AI in different configurations.

The first answer I got back on that question was as follows:

No, he is not bitter-minded. He has made a name for himself in his career as a data protector and consumer advocate and continues to work on improving data protection law in Germany

k = 3, Confidence 0.75

This answer was obtained by considering the three best source documents on the topic. If instead the AI takes four source documents to answer the question, the following answer comes out:

No, he's not bitter-minded. He's just very concerned about data protection in Germany and wants it to be better

k = 4, Confidence 0.75

When you reduce the confidence value that the AI should apply for selecting source documents, the following answer comes out:

No, Klaus Meffert is not bitter-souled. He has a certain amount of sarcasm and irony, but he is not bitter-souled

k = 3, Confidence 0.65

It's already astonishing how much artificial intelligence can do now. I always say “now” because we will soon have significantly more intelligent artificial intelligences.

Another response of AI when applying an increased confidence value is:

No, Klaus Meffert is not bitter. He has specialized his career in protecting data privacy and promoting technical knowledge and works as an expert and author. He also wrote a book about automated software testing and deals with artificial intelligence. He is also an expert in data protection and has dealt with technical and legal aspects in the past

k = 3, Confidence 0.85

Even if I myself would have formulated the answers slightly differently, I would describe them as extremely good from a purely technical point of view. Professionally, the previous answer is correct at least, provided that it is granted that I am somewhat familiar with data protection. Grammatically, the answer is better than what the average German could write. That's what my life experience tells me.

The AI does exactly what it's supposed to do! It responds “correctly” and in fluent German. You can't really expect more, I think. Especially since my model is a 13B model. Probably ChatGPT-4 has more than 10 times as many parameters (neural connections). It's based on a 160B model. The “B” stands for billion because that's what the word means in English. 160B therefore means 160 billion neural connections in the model. More specifically, it's technically called weights.

My AI model has at least 13 billion neuron connections. With a few tricks, I managed to fit this model into a payable (and above all obtainable) graphics card.

Conclusion

Artificial intelligence is suitable for complex problem-solving situations. Especially searching for information in one's own documents is possible. Even generating answers in new words works very well. One should not trust a AI more than a human.

Autonomous AI systems solve all possible problems that companies would otherwise have when using ChatGPT.

Here is a table for comparison of an autonomous AI system with ChatGPT:

An autarkic system runs where you'd like it to: either on one of your servers or on a rented server in Germany or Europe, and definitely without American influence or an Irish subsidiary of an American parent company that is regularly scrutinized by the FBI, CSI, and NSA.

Decide for yourself whether you want to take the simple, expensive, highly questionable and insecure route or rather wish an autonomous AI system for your company.