Yann LeCun has been warning for years that Generative AI is stupid. The current models, he claimed a year ago, are no more intelligent than a domestic cat. This speech has become especially uncomfortable especially because LeCun, considered one of the godfathers of AI, was until now one of the most responsible for this segment in Meta. Now everything seems to point to an imminent departure that will see LeCun found his own startup.
Why is it important. While AI companies strive to train AI models by collecting more data and spending billions of dollars on computing power, LeCun is clear that this strategy is a dead end. It is something we have been talking about for a long time and that other experts like Andrej Karpathy have also have warned: This scaling of resources previously allowed notable leaps in performance. Not now.
He knows what he’s talking about. In 2003 LeCun joined New York University and later founded the institution’s Data Science Center. In 2013, Mark Zuckerberg recruited him to lead his new AI division at Facebook called FAIR (Fundamental AI Research). In 2018, LeCun, along with Geoffrey Hinton and Yoshua Bengio, won the Turing Awardthe highest honor in computer science, for his contribution to the study of neural networks.
LLMs must give way to “world models”. LeCun has prophesied that within three to five years no one in their right mind will be using today’s LLMs. Instead of them, the architecture that will triumph will be that of the so-called “world models”which learn from the environment through visual information, similar to how a baby does, in contrast to LLMs, which are predictive models based on vast text databases.
Internal tension. That vision of LeCun has ended up being a problem in Meta. Mark Zuckerberg does not seem to have the same opinion, and in recent months he has made it clear. with a bet multimillionaire in which ended up signing talent and creating its new superintelligence division that precisely reinforced the role of the LLM that LeCun sees as useless.
An uncomfortable situation. These signings have caused the FAIR Group that LeCun led to lose prestige, resources and weight in the organization compared to that new AI research organization led by the new rising star, Alexandr Wang.
Exit in sight. Last week the first rumors appeared that LeCun is planning his departure from Meta to create his own startup. Precisely this new company would explore the creation of those models of the world that this scientist and researcher wants to develop in depth. If he executed that step, it is very likely that the investment world would support that vision and offer him sufficient funds to work on it. It has happened with startups Ilya Sutskever and Mira Muratithat without having visible product They have achieved multi-million dollar financing rounds.
LeCun seems to be right. The evolution of LLMs seems to confirm LeCun’s theory that they are not the valid way to achieve truly notable advances in the field of AI. What current models do is not so much solve problems as locate past instances of solved problems to use probability and apply answers.
Don’t even think about pursuing LLMs.. In recent months, LeCun’s work in Meta has become more blurred and he has been seen participating in several conferences. In one of them gave a message clear to those aspiring to get into this field: “if you are a PhD student in AI, you should never work with LLMs.”
Image | Goal
GIPHY App Key not set. Please check settings