Yann LeCun predicts "new paradigm of AI architectures" within 5 years(techcrunch.com)
techcrunch.com
Yann LeCun predicts "new paradigm of AI architectures" within 5 years
https://techcrunch.com/2025/01/23/metas-yann-lecun-predicts-a-new-ai-architectures-paradigm-within-5-years-and-decade-of-robotics/
10 comments
LeCun has been saying basically this same thing since 2021-22.
It’s a lot easier to say what something is not than what it is.
His comments now are really sounding a lot more like sour grapes from someone who is ultimately behind in this race.
When was the last time he launched something truly new and innovative that isn’t just a copycat? Genuinely asking.
I used to pay a lot of attention to him but he’s like the Jim Cramer of AI. Always wrong.
It’s a lot easier to say what something is not than what it is.
His comments now are really sounding a lot more like sour grapes from someone who is ultimately behind in this race.
When was the last time he launched something truly new and innovative that isn’t just a copycat? Genuinely asking.
I used to pay a lot of attention to him but he’s like the Jim Cramer of AI. Always wrong.
> These “limitations” inhibit truly intelligent behavior in machines, LeCun says. This is down to four key reasons: a lack of understanding of the physical world; a lack of persistent memory; a lack of reasoning; and a lack of complex planning capabilities.
aligns with (or is based on) Demis Hassabis' assessment from yesterday on missing cognitive capabilities for AGI: long-term memory, reasoning, hierarchical planning. He then goes on to suggest scientific creativity may be essential.
https://www.youtube.com/watch?v=yr0GiSgUvPU https://news.ycombinator.com/item?id=42817089
https://www.youtube.com/watch?v=yr0GiSgUvPU https://news.ycombinator.com/item?id=42817089
Yann's predictions so far are worse than stochastic monkeys
Still better than than Hinton.
Dude panicked so bad after the DeepSeek R1 release he came back to X to pat himself on the back.
Why would anyone panic over DeepSeek R1? It’s cheaper and open, but still not remotely general intelligent.
Their breakthroughs can be copied, and probably surpassed by someone else. Could be Meta, could be another company in China.
But companies like Meta, OpenAI, Google and MS is still in a far better position to monetise their AIs. They have the brand, the customers, the data centres. You could have the best model in the world and it’d mean nothing if you can’t run it and sell its results on a large scale.
From what I’ve seen the people within these companies know that they’re not always gonna be ahead. They know that the models they’re developing now will be outdated very quickly. So the game isn’t really about always having the best model. It’s about how their model integrates with other services and apps that people are using. It’s about having the data and data centers to train a new world class model when a new AI architecture comes along. The model can always be switched out for something better when it comes along.
Their breakthroughs can be copied, and probably surpassed by someone else. Could be Meta, could be another company in China.
But companies like Meta, OpenAI, Google and MS is still in a far better position to monetise their AIs. They have the brand, the customers, the data centres. You could have the best model in the world and it’d mean nothing if you can’t run it and sell its results on a large scale.
From what I’ve seen the people within these companies know that they’re not always gonna be ahead. They know that the models they’re developing now will be outdated very quickly. So the game isn’t really about always having the best model. It’s about how their model integrates with other services and apps that people are using. It’s about having the data and data centers to train a new world class model when a new AI architecture comes along. The model can always be switched out for something better when it comes along.
Because it's roughly on par with OpenAI's o1 which is not something the gigantic and extremely lavishly resourced GenAI org at Meta has managed to even come close to so far. And DeepSeek's non-CoT model, DeepSeek V3 makes LLaMA4 irrelevant. And they've achieved both of these feats at a small fraction of the overall cost. You can bet there's a lot of panic at Meta and elsewhere. There's no quick fix. Truth in engineering always prevails in the end.
https://www.youtube.com/watch?v=MohMBV3cTbg