Logical Intelligence: Ян Лекун строит AGI без галлюцинаций и LLM

The global race for superintelligence has definitively turned into a competition of wallets and electricity consumption. While Microsoft and Google build data centers costing tens of billions of dollars to feed neural networks even more internet text, a San Francisco startup called Logical Intelligence decided to go against the tide. They are convinced that the path being taken by market leaders leads to a technological dead end, where behind the bright facade of chatbots lies the absence of real understanding of the essence of things.

Behind Logical Intelligence stands Yann LeCun, Meta's chief AI specialist and one of the three godfathers of modern deep learning. LeCun has been methodically tearing apart the current approach to creating large language models on his social media and at conferences for several years. His main argument is simple and ironic: an ordinary house cat today is smarter than any GPT-4, because it possesses a model of the world. A cat understands gravity, senses space, and can plan its actions, while a neural network simply guesses the next most probable word in a sentence.

Logical Intelligence is trying to implement the concept of world models in practice. Instead of endlessly predicting the next token, their architecture learns to predict the consequences of actions in reality. This is a fundamental shift in the development paradigm. We are used to thinking that modern AI is a very advanced T9 that has read all the books in the world. But LeCun and his team want to create a system capable of planning and reasoning based on internal logic and understanding of physics, rather than on the statistical probability of letters appearing.

Why is this important right now? The industry is beginning to hit a ceiling. Quality data for training in open access is running out, and model hallucinations, when they confidently lie, are not going away. Logical Intelligence is betting that true intelligence does not need to read all of Wikipedia to understand simple truths. It is enough for a child to see a cup fall once to forever learn the basics of mechanics. If the startup manages to translate this biological principle into software code, we won't need personal nuclear reactors to power servers — digital intelligence will become compact, fast, and most importantly, reliable.

Of course, skeptics from the scaling enthusiasts' camp will remind us that LeCun's pure mathematics is currently losing to the brute force of huge GPU clusters. But the history of technology is full of examples where elegant engineering solutions ultimately defeated cumbersome giants. Logical Intelligence is now in the same position as OpenAI was seven or eight years ago: they have a bold idea, limited budget, and support from a man who has already once radically changed the world.

If this experiment succeeds, we will see the sunset of the era of classical LLMs. We will stop discussing how politely a chatbot answered a stupid question and start talking about systems that can autonomously manage complex industries or make scientific discoveries without the risk of hallucinating a false formula. The industry has been waiting for an alternative, and it seems to be taking on concrete shapes in San Francisco offices.

Key takeaway: Will LeCun's approach become a lifeline for the industry or remain a beautiful theory? If Logical Intelligence shows results, the transformer architecture will be sent to a museum faster than we can get used to GPT-5.