The reason why the 'father' of AI left Meta

Yann Lecun, 65, is the chief AI scientist at Meta. Photo: Bloomberg .

Yann LeCun, chief scientist in charge of foundational AI research at Meta, will soon be leaving the company, according to multiple reliable sources. At 65, he is a pioneer in the AI scientific community, possessing a formidable knowledge base.

Meanwhile, CEO Mark Zuckerberg said Meta is making internal progress that is very close to superintelligence, after continuously poaching the best people from competitors. Therefore, the departure of a top scientist from a company that is spending heavily on AI has caused public opinion.

In fact, Lecun has previously expressed very different beliefs about AI. He recently became famous for arguing that large language models (LLMs) are no longer worth pursuing, will lead to distraction, and ultimately will be a dead end, no matter how much investment is made.

A WSJ analysis of LeCun’s career suggests some of the other possibilities stem from differences in direction. Midway through this year, 28-year-old Alexandr Wang became head of AI at Meta, and LeCun’s boss. The company later brought on another relatively young chief scientist, Shengjia Zhao.

In Meta's announcement, Zhao's new role was highlighted based on a breakthrough related to his ability to increase model size. LeCun, on the other hand, no longer believes this.

According to FT, Lecun will launch a startup focused on “world models”, which he believes will answer questions about AI. Having participated in the development of Meta’s AI smart glasses, he is very confident in a future where people will have to interact with wearables, not LLMs, because they do not understand the world .

Yann Lecun once presented an experiment that showed the lack of sensitivity of LLM. Photo: Shutterstock.

“We can’t even replicate the intelligence of a cat or a mouse, let alone the intelligence of a dog,” he said. World models are AI models designed to understand and simulate the world, based on how objects and events behave, and the consequences of actions.

Scientists believe that world models will allow the construction of systems that can plan actions in a hierarchical manner to achieve a goal. These systems will have stronger safety features, because the ways to control them will be built right into them.

LeCun offers a thought experiment to illustrate what world models, or any human, can do, but LLM cannot. Imagine a cube floating in front of you, and rotating it 90 degrees around its vertical axis. He says that LLM can write out textual explanations, but it doesn't help the user actually interact with the cube.

LeCun has been working on world models at Meta. In his speech at the AI Action Summit, he described a dream model that could estimate the current state of the world. Instead of predicting each successive token like LLM, this model predicts the resulting state of the world after a user performs a certain sequence of actions.

Source: https://znews.vn/ly-do-cha-de-ai-roi-meta-post1603403.html