video sounds like a sales pitch https://www.youtube.com/watch?v=ngBraLDqzdI
“LLM are good programmers, they are not software architects, they are not computer scientists”
- comment: it really depends on the quality and amount of training data available X-D for example most models in 2026-07 fail to program tetris in bash, while succeeding in html+js.
- that seems exactly the dilemma AI companies are facing right now: all the text based training data has been used (well there is new data been added every day like this blog post)
- for Yann LeCun thus the next level of AI, is AI that does not need training data to solve a problem
- give it a new problem it has no training data for and see how well it solves it
Q: what does Yann LeCun mean with “world model” ?
ChatGPT 5.0 mini: He means an AI’s internal simulation of how the world works: a model that can predict what will happen next if it takes an action, not just predict the next word in a sentence. In LeCun’s framing, the point is to learn abstract representations of the environment, then use them for planning, reasoning, and control. <citation
src="1,2,3"></citation>
So the short version is: LLMs predict text; a world model predicts consequences in the world. `
Q: how far is progress in this area?
Quite a bit of progress, but it’s still early.
The field has moved from small latent dynamics models toward much more capable systems for video simulation, embodied control, and planning.
In the last couple of years, surveys and workshops describe world models as spanning latent-state prediction, video generation, 3D/embodied simulation, and decision-making; notable 2025–2026 work includes V-JEPA 2, Genie 3, and related robotics/world-model papers. <citation
src="6,8,9,10,15"></citation>
The main jump is that world models now do more than make plausible videos:
some can support action-conditioned rollouts, planning, and robot policy learning in constrained settings.
But they still struggle with long-horizon consistency, precise intervention handling, and open-world robustness;
even the more advanced systems are usually research previews or domain-limited benchmarks, not general-purpose “mental models” of the real world. <citation
src="6,11,15"></citation>
So the honest summary is: promising and fast-moving, but not solved. We have convincing prototypes and useful narrow systems; we do not yet have a world model that is broadly reliable enough to replace today’s best agent stacks. <citation
src="6,11,15"></citation>
Yann LeCun quotes Linus Torvald’s joke (?) for GNU Linux “world domination”.
In the 1990s Linus Torvalds used to give a talk called World Domination 101 on the early steps he believed Linux would need to take to achieve “world domination — fast” [1].
- http://www.linuxworld.com/linuxworld/lw-2000-03/lw-03-vcontrol_4.html (not available)
- https://liw.iki.fi/liw/talks/linux-world-domination-paper.pdf
-
In 1992 when Linux was at release .95A, Greg tried the OS again. This time he finally got it to boot. When he saw the $ prompt appear for the first time, he knew immediately that "this changes everything." (src)
liked this article?
- only together we can create a truly free world
- plz support dwaves to keep it up & running!
- (yes the info on the internet is (mostly) free but beer is still not free (still have to work on that))
- really really hate advertisement
- contribute: whenever a solution was found, blog about it for others to find!
- talk about, recommend & link to this blog and articles
- thanks to all who contribute!
