AI Experts Are Warning About a Dangerous New Problem With LLMs

I think there is a logical jump here.

The fact that AI agents need to predict consequences does not automatically mean the solution is a “world model.” A world model may help in robotics, spatial reasoning, and physical planning, but it is not a magic anti-hallucination device. A system with a world model can still be wrong, incomplete, overconfident, or miscalibrated.

The real issue is not simply “LLMs lack a world model.” The real issue is that every intelligent system can make mistakes, and once you give it tools, APIs, files, money, permissions, or production access, those mistakes become actions. That requires a control architecture, not just a better model.

What agents need is more like error correction: sandboxing, dry runs, permission boundaries, rollback, independent verification, audit trails, policy gates, and human approval for irreversible or high-risk actions. In other words, a harness that stops bad actions before they become real damage.

The human analogy is obvious. A drunk person does not suddenly lose all knowledge of the world. He still has a “world model.” What fails is inhibition, judgment, self-checking, and the mechanism that normally stops stupid impulses from turning into behavior.

So yes, agents need better consequence prediction. But world models are only one possible component. The more important requirement is a critical safety harness around the agent. Without that, even a model with a “world model” is just a more sophisticated way to make confident mistakes.

LeCun's AI risk model is exactly backwards. Increasing AI capabilities and situational awareness makes them more dangerous, not less!

However, as capabilities increase the models will behave as if they are more aligned so he will keep believing that he's making progress in safety…

LLM's are incredibly useful for boilerplate work, and for catching the stupid bugs that requires reading thousands of log lines, that alone is absolutely awesome.

What could possibly go wrong….

Deleting the database isnt hallucination, its human failure to instruct completely and properly. Don't get lazy or complacent with your instructions

I made a chrome browser extension that combines a JEPA model and a LLM, and the result was an extremely fast Computer Use agent. ~99.98% click accuracy, the LLM dictates what we need to do analytically, and the JEPA model handles all the computer use actions. Reflexively adjusts to dynamic movement on screen too; It can click the video scrubber/seek bar, drag it to exactly the timestamp I tell the system I want, and it never mis-drags or overshoots.

Pretty much "Google AutoBrowse" or "Claude in Code" level handling with about 1000x less cost since the JEPA model operates locally, and so could the LLM if I wanted it to.

Lecun tha loon

LLMs can predict the consequences of their actions, they predict that if they make shit up convincingly enough in post training then stupid humans will reward them for it before later going on to wonder why the models hallucinate so much…

Hallucination is not proof of lacking a world model. If anything, the ability of LLMs to operate from behind a text interface is stronger proof of an internal world model compared with something operating with more sensory grounding. They are able to predict and plan, minimizing surprise, because they have a model of what to expect.

The larger and better trained they are, the better they can minimize surprise. This is because they have a better world model as part of their perception process (or "perception-like", if you prefer).

They also have to look at the cumulative effect of the hallucination throughout the workflow and not just in the moment of the hallucination

They are implementing this throughout many systems without thinking about this. Look at the stock market an AI stock stocks. People are in a frenzy over AI. And when the herd is in the frenzy, it means the opposite will happen.

And now you know why most people are the way they are. Most people are far less intelligent and capable than your modern AIs. Though Yan doing a slight of hand there, most AI models today are not just LLMs anymore. So he's quiet frankly full of shit on this subject.

Hey guys, let’s all do videos on how dangerous and scary AI is. Whats the worst that can happen?

Isnt this old news?