Do you agree with French AI scientist Yann LeCun’s view on artificial intelligence?

[julie09]

Yann LeCun is one of the world’s leading experts in artificial intelligence. He is skeptical that we will achieve Artificial General Intelligence (AGI) simply by continuing to scale today’s AI systems.

In his view, current AI systems—particularly large language models (LLMs)—have fundamental limitations. They can process and generate an enormous amount of information, but they still lack a genuine understanding of the physical world and struggle to interact with it in the way humans do.

That doesn’t mean LeCun thinks current AI is useless. Quite the opposite: he recognizes that LLMs are extremely useful and can perform many tasks remarkably well. His argument is that impressive performance on intellectual or digital tasks should not automatically be confused with general intelligence.

And here is what he had to say to those who claim that the latest AI models, such as GPT Astra, are already getting close to AGI:



" Just give it a humanoid body and ask it to fix your bathroom, to watch your children, to clean your house, to cook dinner, or just to clear the dinner table and fill up the dishwasher.

Ask it to assemble a Lego set (which any 10 year old can do), or just to take care of a rodent invasion (which any cat can do).

And let's not talk about trusting it to drive your car unsupervised...

[exitseven]

Yes, I agree that it cannot really think like a human. It processes informaion very fast and in some ways it can approximate human reaction to stimuli but it is incapable of thinking in abstract terms like humans can and just making it faster and adding more algorithms will not change that.

[julie09]

Yes, exactly – that’s why he thinks LLMs are doomed to fail in their quest to achieve AGI. He’s working on world models to help AI understand the physical world.

[Malwi93]

well, there are other models

[purplepoppy]

AI reminds me of that kid whos good at algebra, he can rearrange an equation but doesn't really know why A moves there and B moves there, he's just following what he's taught without any grasp of why or how it relates to the real world. It's the historian who can churn out countless books on wars yet doesn't understand the suffering. It's the poet with perfect meter but no concept of the weight of words.

A child not yet able to read understands a big bad wolf gobbling up a piggy in a story book is a sad event, AI doesn't get it. So I agree, we're still a long way off.

[MrChinaski]

Yes. AI training is incredibly inefficient. If you are a baby learning to crawl you wriggle some arms then fall on the floor. Clear objective. Instant feedback. Rapid learning.

Try asking an LLM 'how does the history of Islam relate to the development of quantum physics?' and it tells you:

Greek mathematics/natural philosophy → Arabic-Islamic scholarship → medieval European scholarship → Renaissance astronomy → Galileo/Newton → classical mechanics and electromagnetism → Planck/Einstein/Bohr/Heisenberg/Schrödinger → quantum mechanics

Useful? Who knows. It's just searching enormous databases to meet the requirement to provide a response. Feedback? Zero.

This is why LLM is so expensive to train. It relies on more and more and more and more data points, where the baby learns a complex task like crawling after a dozen tries with objective feedback.

[NathanDavis]

I agree...

AI is extremely limited to what can it actually do properly and in reliable ways too

so the promises of AI have clearly failed... there's no point to think or consider AGI to be a thing at the moment, lol

AI is still a regurgitator... and not even close to any type of intelligence

[Levin]

Yeah. Or maybe even more fundamentally than that, it's capacity to feel, is completely zero. And there doesn't appear to be any prospect in current models of that changing.

Same with these fancy new limbs. Where's the nervous feeling throughout it? Again, non-existent.

[jshm2]

LOL, His input is largely meaningless.

What these fools think of as "AGI" has been reached... according to the guy who makes most of the hardware.

https://www.businessinsider.com/nvidia-jensen-huang-agi-openai-astra-ai-2026-9

[RyanEC]

I think he may be missing the potential of LLMs.


We don’t yet have robotics to test these AIs capabilities in the physical world.


I remember watching an interview his example was how self driving cars have millions of hours of driving but can’t drive to the point where a human isn’t there but I think those were older AIs.

[MrChinaski]

LLMs are lousy at chess. Why? Because it is (eventually) based on reasoning not knowledge.

[AngryCarl3]

@MrChinaski With all due respect, chess is based mainly on calculation. We have specialist neural networks that can beat any human at chess, not because they understand the game but because they have trained to play almost perfectly. The reason LLMs can't even play 40 legal moves is that they aren't sufficiently specialized.

[MrChinaski]

@AngryCarl3 I think we agree.

[AviatorTom]

I agree. AI is far from capable of independent thought, though it's getting better at analyzing lots of human-generated data and determining what makes sense. A long ways to go though.

[AngryCarl3]

A large language model isn't intelligent at all, so the word "artificial" has to do all the heavy lifting. The significant work on AI is being done through neural networks which somewhat resemble the human brain.

[Nikki1989]

It is not hard to stump AI. I have seen it happen many times.

[AngryCarl3]

My common experience with AI is that it gives me an answer that is obviously erroneous and when I point out it's mistake it's like "You're right, my bad. I got confused." Well duh!