The new dream AI is said to be driving Yann Lecun away from Meta

One of the most important AI scientists in Big Tech who wants to erase the current way of building AI-Level Ai for humans. What we need, Yann Lecun has shown, are not big models of language, but “models of the world.”

Lecun, the head of AI science of ‘AI basic research’ at Meta, is expected to leave Meta soon according to many reports from reliable outlets. Lecun is a 65-year-old veteran in the world of AI Science, and he has always been seen as having no limits working without power working as the main brain of AI in the largest companies in the world.

Why is he leaving the company that has been destroying the miracle, putting in talented professionals with the best skills from other firms, and, according to the July blog post about CEO Zuckerburg, making the quick elimination of CEO Zuckerburg, making the CEO’s outbreak called “Superintelligence now visible” now?

He had actually been answering the question for a long time. When it comes to Human-Level intelligence, Lecun has recently become famous for saying that LLMs as we currently understand them – are no longer worth repeating, no matter how inspiring they are. He said in April of last year that “LLM is basically an off-ramp off, a distraction, a dead end.” (Arch ai critic Gary Marcus ripped into Lecun for “fighting back” by protecting LLMS from Marcus’ own hacks and captures.)

A Wall Street Journal analysis of Lecun’s careen published on Friday points to other possibilities about the reasons for his departure because of this belief. This past summer, a 28-year-old kid named Alexandr Wang – an LLM-based developer – became the head of AI at Meta, making Upstart LLM Lecun’s Boss. And the meta brought another young scientist to work above the lecun this year, Shengjia Zhao. Meta’s announcement of Zhao’s new role brings out the “success” that is clearly not presented. Lecun says he has lost faith in moderation.

If you’re wondering how Lecun can be a great scientist when Zhao is also a scientist, it’s because Meta’s AI Operation feels like it has an eccentric org chart, split into many, many different groups. Hundreds of people have been booked in the past month, apparently in an effort to streamline all of this.

A financial report on Lecun from earlier this week suggests that Lecun will now receive a start-up focused on “Global Models.”

And, Lecun wasn’t shy about why he thinks the world’s models have AI answers. He gave a detailed speech about this at the AI ​​action conference in Paris back in February, but it got kind of announced by the US, said bellicose President Jelico

Why does Yann Lecun impress the models of the world?

As written in speech – Lecun, who worked on meta ai smart glasses, but -I To a significant extent in the llama llm of the meta

It’s amazing how ray-ban meta glasses can help the disabled. https://t.co/w3zxcfttle

– Yann Lecun (@YYecun) September 30, 2024

We will need to communicate with future problems As if they were human, You think, and llms just don’t understand the world like people. With LLMS, he says, “We’re not going to recreate Cat intelligence or rat intelligence, let him do amazing intelligence.

Lecun offers a thought experiment to illustrate what he thinks can drive – if you know how – a model of the world, and it’s something he thinks anyone can do:

“If I say to you ‘Imagine a cube floating in the air in front of you. Okay now turn this cube 90 degrees around the vertical axis. What does it look like?’ It is very easy for you with the grace of having this rotating cube mental model. ”

With very little effort, an LLM can write a dirty limerick about walking around a cube, sure, but it doesn’t really help to connect with someone. Lecun Avers that this is due to the difference between the textual details and the details taken from working in many parts of the world which are not compatible. When llms are trained on the amount of text that will take 450,000 years to read, said Lecun, a 16-year-old child is already awake, with their eyes or 14

These, by the way, are just measurements of Lecun giving his speech, and it should be noted that he has given others. The numerical output indicates, however, that LLMs are limited in ways that Lecun thinks World Models cannot be.

What model does Lecun want to build, and how will he build it?

Lecun has already started working on meta-world models, including making an introductory video that imagines a rotating cube.

Lecun’s dream model as described in his AI Action Summit speech consists of “state modeling,” in the form of some kind of abstract representation, well, everythingor at least everything that is relevant in the current context, and instead of predicting a sequence of actions, tokelen, “predicts the resulting global situation that will occur after taking that sequence of actions.”

World models will allow computer scientists of the future to build, he says, “Programs that can organize actions – they can be purposeful, and programs that can communicate.” Lecun also realizes that those programs will have more powerful security features, because the ways we control them will be built into them, rather than being black boxes that emit beautiful text.

In what Lecun says is Classical AI – similar to the software used in a search engine – all problems are reduced to efficiency. His model of the world, he suggests, will look at the current state of the world, and seek to adapt to a different situation by finding appropriate solutions. “You want a power function that measures the inequality, and given X, find AY with the lowest power for that x,” Lecun said in his speech.

Again, these are just reliable reports from mature information about Lecun’s plans, and he is not even convinced that he is starting something new. If all we can gather from Lecun’s public statements sounds real and bleak at this point, it should. Lecun screams as if he has a moonhot in mind, and suppresses another chatgpt-like burst of mysterious abilities. It can take years – or literally forever – not to mention the distribution of billions of dollars, for anything truly amazing to create.

Gizmodo reached out to Meta to comment on how Lecun’s work fits into the AI ​​company’s work, and will update when we hear back.