Tootfinder

A look at AI world models, including how they work, what they can do, and what's still unsettled, as startups led by tech leaders like Yann LeCun raise billions (Samuel Axon/Ars Technica)
https://arstechnica.com/ai/2026/07/simulati…

Simulating everything, sort of: The promise and limits of world models
Experts explain how they work, what they can do, and what's still unsettled.

Great to hear some hype deflation from AI researchers:
"[LLMs] basically just accumulate knowledge... They can regurgitate something, you train them to regurgitate, but they're not particularly smart. They don't have an underlying understanding," he says.
"LLMs are largely hopeless for robotics... The claims that somehow by just scaling up LLMs, we're going to reach super human intelligence, that is simply not going to happen."

Artificial intelligence: Yann LeCun works on more flexible AI
Leading AI researcher Yan LeCun has a start-up which is developing a more flexible AI system.

Just realized that the fact that newer large language models keep getting bigger in terms of parameters is kind of a tell about how they work, even as it's also kind of a requirement from the investment standpoint.
Very roughly, models develop complex functional internal state about some sub-domains, and merely memorize many examples in others. In reality it's more complicated than this and even in this simplified metaphor it's a mix between memorization and "real" "understanding" in each domain. But the point is that if companies were really working towards AGI, they'd be feeding more data into models with *fewer* parameters (that's how you force a model not to memorize) instead of building bigger and bigger models (expands the illusion of competence through increased capacity to memorize).
But being the only ones with the hardware to train an even-bigger model is one of their few competitive advantages, and signing new deals for even more hardware is one of the only ways they can signal to investors that they'll retain their advantage and thus not be destroyed by a food of competitors. That's also how they can convince the hardware dealers like NVidia to continue with circular investments. So they have to run in that direction, regardless of the scientific merits.
This is why someone like LeCun would leave that side of things.
#LLMs #AI