More artificial than intelligent
..AI is more A[rtificial] than I[ntelligent].
First time I heard this phrase in a critical review of the game Fallout 76 by a YouTuber Internet Historian. I would not have thought that it will come to so accurately describe a whole field of information technology and an impressive financial bubble coming down at the markets a couple years later.
“It is alive!”
Contrary to the popular belief, Artificial Intelligence (AI) does not think – it has no facilities for that. Contemporary AI, which in popular culture has become synonymous with the technology behind modern chatbots – generative pre-trained transformers (GPTs) –, can be boiled down to statistical models.i
GPTs are overtly elaborate estimators for the next value. Such estimators in mathematics are known as statistical models. One could think of the most primitive statistical model like a trend line – a linear function. Given the last value is known, the following can be fairly certainly predicted. With some noise, but following the trend.
In that, GPTs are similar. When they are presented with an input, there is no “thinking” that occurs, they simply predict the likeliest to follow a word.ii Prediction is the core purpose of machine learning. That is why AI is used in simulations, climate modelling, particle physics and such.iii AI is no better thinker than a human, because it is no thinker at all, but it definitely handles patterns well.
Search query suggestions, despite using a completely different technology, are similar to GPTs in a way. Input gets consumed, and it returns a list of the likeliest to follow queries, except a large language model would return a single result – the likeliest – and repeat. The same case is text prediction on phone's keyboards. Tapping on phone's keyboard suggestions for long enough would produce an essay. Essay of gibberish, of course, but parts of sentences would make sense, as, if the likeliest word is followed, a coherent sentence would inevitably get chained.
This also explains why AI model training requires GPUs rather than CPUs – they trade in excess precision for parallel processing power.
Why (continuously) mislead the public?
Not that the entrepreneurs would have much choice.
They have made inflated promises to their shareholders about the so called artificial general intelligence, backing down from which wouldn't come lightly. Those promises are the only reason why they can raise such astronomical amounts of investment. At this point, backing down from them is exactly what might cause the bubble to pop.
Moreover, this brings money and a lot of it. Those who will sell their shares before the bubble pops will have turned multimillionaires. Some companies have had their share prices rise a dozen times. The same is the case with the executives, the administrators. Those who find refuge elsewhere, before the tsunami comes, will have made their career. And it doesn't matter that AI is financially unfeasible, the financial speculation is what matters.
Especially considering that it is not the AI that is the problem as much as the financial bubble around it. Thank God for machine learning – it is a marvellous technology, it is just not being used the best.iv The man has a tendency to rush so far beyond his time that he forgets to use the maximum of what he has already.
What comes next?
Bubble will blow, and then the real AI will come.
Same as it was with the dot-com bubble.
Update (15.11.2025.): Congratulations to those looking forward to Artificial General Intelligence. No more em dashes (—).
i Different from classical models, GPTs are non-linear, high-dimensional and hierarchical. The method by which they estimate, are vastly more complex compared to classical models, but as a didactic analogy, it is valid.
ii To be more precise, a token. If in human language the building block of a text is a word, then for AI that is a token. Sometimes a token is as long as a word, sometimes shorter, sometimes longer.
iii It should be noted that for such kind of predictive modelling, different architectures are employed, for instance graph neural networks (not GPT).
iv One potential use case that I have fallen in love with is accessibility. How much could the lives of people with various visual, hearing or mental impairments be improved with this technology. I am eager to witness its application for such use-case.