Yann LeCun argues LLMs are a dead-end.
At the end of last year, he left Meta to start up his very own French world model company, AMI Labs. It’s worth billions.
via NY Times
Oh good. Another concept to understand.
What are world models? From Paul:
Ok, so imagine you’re playing with building blocks.
Right now, most AI is like a friend who can describe blocks really well or even draw pictures of them. But if you ask them to remember where the blue block went when it rolled behind the red one, they might forget or get confused.
A world model is like the AI has a permanent mental map.
The AI can keep track of things even when they move or go out of sight.
Just like you know your bedroom is still there when you close your eyes, world models help AI remember that things exist and how they work.
But, and this is the jump, it can also model thousands of scenarios within that mental map. To work out and test different parameters. Before building something in the physical world.
Think about a dog.
- Current AI: Shows dog running, but when that dog, Rex, goes behind couch, the collar might vanish or the couch might turn into something else.
- With world model: AI remembers “Rex has collar, couch stays a couch” even when the camera moves.
Remember our O’Reilly scenario. This is LeCun’s argument.
From Nvidia this week:
“Industry world models” – super-smart computer simulations that understand how real factories, drugs, materials, and machines actually work in the real world.
Imagine you’re a PM at a car company. Right now, you might:
- Build a prototype car
- Test it
- Find problems
- Rebuild it
- Repeat (costs millions, takes months)
With Nvidia’s world model:
- Design the car in the computer
- The AI predicts exactly how it’ll behave based on physics
- Test thousands of scenarios instantly (crashes, weather, different roads)
- Fix problems before building anything real
- Only build what actually works
The key difference: Instead of generic AI that guesses, this AI actually understands gravity, materials, engineering rules – it knows how the real world works.
Who knows what this means. But it feels like we’re seeing incremental rather than large leap improvements with the new LLM releases. Not a sign of AI being a generative technology. At least not across all industries.
Gordon MacRae Blog 
Leave a comment