Comment by pu_pe
16 hours ago
It was obvious that there would be no space for Yann LeCun after Alexandr Wang came in. He was probably just waiting for the best time to leave.
I cannot judge his research output at Meta but he failed pretty bad at the LLM race. Since so many other organizations succeeded at creating open source models of far higher quality at much lower cost, it would be instructive to understand what exactly went wrong there.
Curious about how much risk Meta leadership was comfortable with when they decided to layer Yann. Perhaps the winds of open research were already blowing a different direction at the company, and he had already indicated that he wanted to leave as a result of that. We can only guess.
Kind of hilarious to me to consider him "failing" with LLMs. Given his remit was a research time horizon of 8-10 years, and the fact that he's gone on record saying that he expects the technology will stall out in the time horizon, it seems he can only take Ws and ties. Indirect influence on open-sourcing the models to propel research forward (which is pretty important for a chief scientist) which added benefit for Meta's other products.
> I cannot judge his research output at Meta but he failed pretty bad at the LLM race. Since so many other organizations succeeded at creating open source models of far higher quality at much lower cost, it would be instructive to understand what exactly went wrong there.
What? Until the Chinese jumped in Llama was the premium open source model. The reason that the Chinese were successful at MOE was just that they were limited with chips and had to think outside the box. US labs are operating on the power law. They also, arguably, distilled from western models (llama).
> he failed pretty bad at the LLM race
Was he even involved in this?
Did they even fail? Llama2 was groundbreaking for open source LLMs, it defined the entire space. Llama3 was a major improvement over Llama2. Just because Llama4 was underwhelming, it's silly to say they failed.
Any exponential growth is failing in a market which demands superexponential growth
No, he said that he was not involved. He had his own research model to develop, his startup will probably continue his work there but I wonder if he thinks its viable in the short term since he's launching a startup. I thought it was a moonshot.