Relevant: https://x.com/stefanfschubert/status/1695509307444297770?s=46

@agucova maybe we’ll see Yann slowly backpedaling on earlier positions until he effectively agreed with most AI risk arguments?

Similar market I just made: https://manifold.markets/NicholasKross/will-yann-lecun-publicly-endorse-ai

Will Yann LeCun *publicly* endorse AI alignment research in 2023?

50% chance. Similar to https://manifold.markets/PeterWildeford/will-yann-lecun-change-his-mind-abo , but with a shorter timespan. Has to be public (e.g. Twitter), not in like a private convo. I’ll subjectively judge edge cases (e.g. he shows up at an EAG or something). [link preview]

My thoughts on this:

• Yann is clearly very fixated on his current beliefs

• Yann seems to be acting as a motivated reasoner

For someone like Yann to change their beliefs, it would probably take at least one radical change in his environment. Like maybe a major AIS-related accident?

I think my prior for such an event is quite low for 2025 (it's like ~15%) and for Yann to react to that is also not that high (~50%). Change from other causes (loved ones? being fired?) seems also pretty unlikely (< 5%), which leaves me with a pretty low estimate.

@AgustinCovarrubias Yes, I think nothing short of a significant AI-related accident would convince Yann. In a way, he is pretty similar to Eliezer: both are emotionally invested in their beliefs.

@AgustinCovarrubias Another option would be some sort of dangerous capability demonstration by i.e. ARC Evals. I find it likely that they might to some degree be personally targeting Yann LeCun as the audience of their demonstrations.

Hopefully not considering he is the only one of the Deep Learning trio who hasn’t already lost his mind completely.

@BTE Why do you think Yoshua or Hinton have lost their minds? Do you think the AI-risk case is bad?

@hmys I just don’t think Bengio and Hinton have done any real work or lead a real team for years. Hinton especially is full of shit. LeCun is the only one doing real work in both Development and Production using the largest data firehose in history. He is the only one still “close to the metal” to use an imperfect metaphor.

@BTE Bengio is definitely still doing real work (look up gflownets for instance).

He seems very set in his ideas. Unless something drastic happens, I don't expect him to significantly change his stance.

Is it enough for him to cede particular x-risk arguments or does he have to change position towards accepting x-risk arguments in general?

@AgustinCovarrubias He has to broadly do what Hinton did. That is, change position towards accepting x-risk arguments in general and publicly advocate at least once for AI risk. I think it's ok if he's still skeptical of some stuff.