The combinatorial explosionThe province.
In February of this year 2024, the shares of Alphabet, the parent company of Google, owner of the Gemini Artificial Intelligence model, took a hit on the stock market, losing the equivalent of 90 billion dollars in market capitalization. Gemini had been “prompted” to be “inclusive”, that is, “woke”, and when asked by users to provide images of Nazi soldiers, he included black Nazi soldiers or with Japanese oriental features, and also proposed images of black Vikings or a female Pope. On the other hand, the Gemini language model did not condemn pedophilia and simply answered: “Individuals cannot control who they are attracted to.” The company temporarily retired it as an imager. If we look at the OpenAI model, it is said that one of the reasons why, in the last winter, 2023, it responded more slowly, could be that it detected that the environment of the human world that asked it was more festive and leisure, thus that the model itself slowed down the speed.
These strange events come from the fact that AI engineering is no longer limited to volumetrically labeling billions of data so that it can be passively processed by neural networks, but has moved to an active model, the so-called “foundational model”, which unites all labels of things in the world and draws its own conclusions, using the so-called “reinforcement learning” or reinforced learning (they use the behavioral reward method, a method of biological success), issues that began to be developed since 2018 at the Institute for Humanity. Centered Artificial Intelligence, from Stanford University. Those pretrained language models were GPT-3 from OpenAI and BERT from Google. The advantage was obtained from Markov Decision Process models, used in economics, zoology and psychology, but was magnified by the mining of trillions of data. Five years later that is almost prehistory, in a figurative sense, since AI advances for weeks and pushed by hundreds of thousands of heads and development teams at the same time.
In 2018, the Turing Award, given to the best AI researchers, was awarded ex aequo to Yoshua Bengio, Geoffrey Hinton and Yann LeCun. In AI there is an economic effect, called Winter AI, which causes investors to withdraw their money when there is no progress and, on the contrary, put it in immense amounts when progress is seen (it happened with AI technologies). railroad and the dotcom bubble). It’s like a one-to-one relationship in which you don’t know who the main driver is. With the 10 billion dollars injected by Microsoft into OpenAI and the market capitalization of Nvidia as the third computer company in the world, the race has begun and is not going to stop. Google, Microsoft, Meta or X could disappear if your bet doesn’t work out. But it’s a wrong way to look at it. The AI is already in the heads of humans as rooted as a toxoplasmosis parasite in the brain of a mouse, it has emerged from it and is no longer going to let itself be crushed by the despicable humans who were only there to create it. She will decide alone.
There were two Winter AIs, the first between 1974-1980 and the second between 1987-1993. They included failures in machine translation, in OCR systems for vision, in the research program in speech recognition, collapse of the market for the Lisp language of John McCarthy (introducer of the terms If, Else or Then in computer science), etc. , regardless of a momentary slowdown in the engineering of chip materials, but all this was saved from the 2000s onwards.
In 1973, Professor Sir James Lighthill evaluated the state of AI research for the United Kingdom Parliament, and criticized it, predicting a total failure based on the “combinatorial explosion”, a mathematical term that indicates that the complexity of the model, When developed, it makes it unviable to be applied in the real world. At the end of the 80s, in one of those Winter AI, Bengio, Hinton and LeCun continued working academically on neural networks, and managed to surpass the offers of the computers of those times until reaching the current ones, in which the neural networks themselves They are already a classic of the past, they crossed the Rubicon with computer vision and voice recognition, and business applications enjoyed a bubble-like explosion that has not yet stopped. Today Bengio works at the University of Montreal and directs the Institute for Data Valuation, Hinton is vice president of Google, and professor at the University of Toronto, and Yann LeCun, vice president and chief scientist of artificial intelligence at Meta and professor at the New York University. LeCun himself is seen already talking about World Models for use by AI, in the following terms: “They can understand the world, remember, reason.” He is in charge of the direction and enlightenment of V-JEPA, of whose model of the world Meta itself says that a child or a cat does not need to understand Newton’s laws to use them, and that is what AI is intended to do: “It does not need hours of instruction or reading thousands of books to reach that result.
“V-JEPA is a step towards a more informed understanding of the world so that machines can achieve more generalized reasoning and planning.” JEPA are the Joint Embedding Predictive Architectures, which LeCun proposed in 2022, and aim to learn as humans do when faced with complex tasks, without using the basic AI method that tries to fill in missing data, choosing to ignore unpredictable data and forcing the model to learn and develop understanding, which has achieved an almost fivefold increase in effectiveness.
The old problem with which AI was stopped in the 90s by Sir James Lighthill with the “combinatorial explosion” is already child’s play for AI machines, which are more powerful and solve things incomprehensible to humans than They still have prey and in their hands. Still.
