Many write that artificial intelligence is just statistics, and that it's determined by probabilities which word of a text response should be given next. That's nonsense. But what is artificial intelligence in relation to statistics anyway?
A common statement about current AI systems is that they gain their results through statistical calculations. With this, some people, in my impression, want to demystify the capabilities of AI. In my opinion and based on my knowledge so far, this is unacceptable.
Recently yesterday my opinion was confirmed, and that by statements of German AI pioneer Prof. Jürgen Schmidhuber. He is considered the father of numerous mathematical models responsible for the current surge in AI. Notably to be mentioned are the Transformer approach and Long Short-Term Memory (long short-term memory). Without these approaches, especially without Transformer, there would be no ChatGPT (and if so, then as ChatGP, because the “T” stands for “Transformer”).
The German AI pioneer mentioned by Markus Lanz (German TV interviewer) said what I wrote a few weeks ago and which I would like to repeat at any time. He said that the functioning of AI is very similar to how a human brain works.
There are no significant differences between electronic and human brain.
Refers to the functioning.
Just like Prof. Schmidhuber, I believe that AI robots will soon (in 10 to 20 years, I would estimate) have better general abilities than humans. To make this statement is no prophecy. It results from the current possibilities and the simple assumption that computing power will always become cheaper, just as it has been for the last 50 years.
What's it got to do with statistics?
It is correct that AI algorithms rely on statistical calculations. However, this statement is very undifferentiated and not informative. Here are three examples:
A car is obviously a heap of matter. This statement is just as true, as undifferentiated.
Cookies are not text files. I've been trying to clarify this for a long time . Yet, some people say: But cookies are stored in files after all. Then I say: Yes, but those are (today, or generally across all browsers) not text files. Then someone says: Yes, but you can open these (binary-)files with a text editor. Right, you can open any file with a text editor, even if the binary data will then be displayed in an unpredictable and usually incorrect way. Calling everything a file means having exactly zero information content (= entropy = measure for the average information content of a message).
The Quantum Theory is, in my knowledge, the most accurate and best proven theory of human history. It deals with the smallest particles. Any matter on earth (and probably also elsewhere in the universe) is composed of particles that obey quantum theory. Unfortunately, quantum theory is based on chance. Quantum theory is therefore based on statistical processes. In this sense, our entire existence is based on chance. Our daily life is thus determined by statistical processes. Yet I have never heard anyone say that our everyday life obeys statistical processes.
Why do some people say that results generated by an AI would rely on statistical algorithms? Because they don't understand or want to downplay AI, is my thesis. Who can't or won't understand something, acts as if it's not relevant or not so important etc.
Why doesn't AI really rely on statistics?
As shown above, statistics plays a role everywhere. You don't need to mention this separately if statistics is not the dominant factor. In AI other concepts are dominant. To be mentioned are:
- Artificial Neural Networks (ANN) — just like in humans
- Transformers and embeddings -> As in humans
- Recognize structures -> As in humans
Finally, an example for recognizing structures. The German language has many idiomatic expressions that have become established. When a foreigner speaks German, it often happens that he uses an idiom in a way we consider wrong. The same can be said about the German grammar in general. Obviously there are linguistic rules. These have to do with which word is expected to follow the previous word or several previous words. This has to do with statistics.
The core thing is that we as native speakers grasp the context of the German language and a text. Exactly this do modern AI algorithms also.
Impacts on data security
Data security here refers to personal data, i.e. data protection, business secrets and copyright aspects.
Just like with humans, information is not stored exactly in a artificial brain called AI-Model, but rather in the form of weights that calibrate the neural network. Nevertheless, AI systems such as ChatGPT are famously able to recall Facts correctly. How Hallucinations arise, I will describe more precisely in another article.
Therefore, there is a warning to employees at large corporations that are themselves leading in AI, not to feed sensitive data into the AI. For a third party who is unauthorized to receive certain information could obtain this information by cleverly questioning the chatbot.
It's only a matter of time (hardware development) until an artificial intelligence surpasses humans even more than it already does (see refined chatbots that often give much better answers than average intelligent people).
My thesis. I'd gladly bet on that with you.
Sensitive information is relevant both in initial training (“pre-training” named) as well as in further learning (“fine-tuning” named). User inputs can be easily controlled, provided they are not stored and also not used for improving the AI models.
I strongly recommend that companies only develop their own AI systems if company data is fed into them. These local systems are feasible and affordable. Contact me if you need a statement regarding feasibility. What works, in particular:
- Search documents
- Answering questions from own documents
- Audio transcription
- Pattern recognition (researchers have recently discovered new hieroglyphs with AI support)
- Image creation
For image creation a word: It is a huge difference whether you
a) in DALL-E or Midjourney manually have to type an instruction into a prompt, have to pay usage-dependent (or watch out for your free contingent), have to undertake numerous attempts when unsatisfactory results occur, and moreover not know what happens with your input data and the received image Order
b) have your own server on which a custom AI application runs, with which you can generate 20, 50 or even 1000 or 5000 images “free of charge” in response to an instruction. In addition, the generated images can be checked for relevance using an AI. For example, your previous decisions, which images you found suitable for your marketing channels, can be taken into account. So, out of 1000 generated AI images, a list with the 20 best is created. From these, you select as many as you need.
Automatic versus manual, fixed costs versus unpredictable costs, data quality versus chance, individual design options versus blanket solution, data security versus gamble. It's your decision.
Important to know: If you want local AI solutions for your company, you need to know what they're for. There is no one-size-fits-all AI solution. You can't get ChatGPT in a data-friendly format right now! Please see the above-mentioned use cases for possibilities that are solvable with AI systems running on a local server.
Conclusion
The performance that comes out of AI algorithms based on the Transformer approach has as much or little to do with statistics as does the functioning of the human brain.
The main features of modern AI lie rather in the architecture of artificial neural networks, the transformer approach, and the unification of all kinds of data.
Just like humans, AI is unreliable and constantly produces false results (I repeat: just like humans). However, AI systems can also achieve amazing things. One must simply be aware that generated answers in textual form do not represent a reliable result. On the other hand, document searches or image generation are powerful tools that open up many opportunities.
Further contributions to AI:
Key messages
Artificial intelligence is much more than just statistics; it functions similarly to the human brain and will likely surpass human capabilities in the future.
AI relies on more than just statistics; it uses complex structures like artificial neural networks to understand and process information, similar to how the human brain works.
Choosing between using existing AI tools like DALL-E and Midjourney, or setting up your own local AI system depends on your specific needs and priorities. Both options have advantages and disadvantages.
My name is Klaus Meffert. I have a doctorate in computer science and have been working professionally and practically with information technology for over 30 years. I also work as an expert in IT & data protection. I achieve my results by looking at technology and law. This seems absolutely essential to me when it comes to digital data protection. My company, IT Logic GmbH, also offers consulting and development of optimized and secure AI solutions.
