You were working on AI long before LLMs became a mainstream approach. But since ChatGPT broke out, LLMs have change into almost synonymous with AI.
Yes, and we’re going to change that. The general public face of AI, perhaps, is generally LLMs and chatbots of assorted types. But the newest ones of those aren’t pure LLMs. They’re LLM plus a whole lot of things, like perception systems and code that solves particular problems. So we’re going to see LLMs as sort of the orchestrator in systems, a bit bit.
Beyond LLMs, there may be a whole lot of AI that’s behind the scenes that runs an enormous chunk of our society. There are assistance driving programs in a automotive, quick-turn MRI images, algorithms that drive social media—that’s all AI.
You’ve got been vocal in arguing that LLMs can only get us to date. Do you’re thinking that LLMs are overhyped lately? Are you able to summarize to our readers why you think that LLMs aren’t enough?
There may be a way wherein they’ve not been overhyped, which is that they’re extremely useful to a whole lot of people, particularly in the event you write text, do research, or write code. LLMs manipulate language rather well. But people have had this illusion, or delusion, that it’s a matter of time until we are able to scale them as much as having human-level intelligence, and that is solely false.
The truly difficult part is knowing the true world. That is the Moravec Paradox (a phenomenon observed by the pc scientist Hans Moravec in 1988): What’s easy for us, like perception and navigation, is tough for computers, and vice versa. LLMs are limited to the discrete world of text. They’ll’t truly reason or plan, because they lack a model of the world. They’ll’t predict the results of their actions. This is the reason we don’t have a domestic robot that’s as agile as a house cat, or a really autonomous automotive.
We’re going to have AI systems which have humanlike and human-level intelligence, but they’re not going to be built on LLMs, and it’s not going to occur next yr or two years from now. It’s going to take some time. There are major conceptual breakthroughs that should occur before we now have AI systems which have human-level intelligence. And that’s what I’ve been working on. And this company, AMI Labs, is specializing in the subsequent generation.
And your solution is world models and JEPA architecture (JEPA, or “joint embedding predictive architecture,” is a learning framework that trains AI models to know the world, created by LeCun while he was at Meta). What’s the elevator pitch?
