Haiyang

[Haiyang @ockddddw]

@agarwl_ Hi I’m very interested in understanding RL

[Haiyang @ockddddw]

@BtreeWw 真有人能做到多线程学习吗👀

[Haiyang @ockddddw]

@Zen_with_AI 你进新组不还是要擦别人的屁股….？

[Haiyang @ockddddw]

@steven_voyager @0xTykoo 应该是某种不靠谱的projection，这才几天啊，单价还那么便宜，怎么能卖到100m

[Haiyang @ockddddw]

@joshua7be @cherylnatsu 宿管大妈存在的意义是啥….？

[Haiyang @ockddddw]

@NekoStranding Facebook一开始也是屎山啊，业务跑起来后有现金了再慢慢重建新架构

[Haiyang @ockddddw]

@ChinmayKak @TianfuF There’s no correct explanation so far I think. Most principled interpretation is some sort of MWU, which makes some sense. But after all this is not something as obviously principled as IS.

[Haiyang @ockddddw]

@maouirr @rdesh26 Something like tailRL just google it

[Haiyang @ockddddw]

@maouirr @rdesh26 Well it works for continuous reward in a follow-up paper

[Haiyang @ockddddw]

@solidSF @yifanzhang_ For a finite game like go, one of the player, black or white, should have guaranteed winning strategy. If you claim to solve go for 7 or 9, can you certify which of the black/white has the guaranteed winning strategy?

[Haiyang @ockddddw]

@adxtyahq they can just ask an agent to do it no?

[Haiyang @ockddddw]

@tastytatab @bonniewinds0 人是肯定会救的，生产���料不一定要救吧

[Haiyang @ockddddw]

@ar0cket1 @maouirr @willcb I remember it to be TailRL

[Haiyang @ockddddw]

@yifanxu_ephai @zengbozb 你回复的这个一看就是ai。。。

[Haiyang @ockddddw]

@deliprao @harshagundal Well, there is no way you can get true confidence anyway. If the objective is to minimize perplexity, and if the model do really minimize it, then it will return the true class probability and therefore the true ‘confidence’… so i wouldn’t say LLM confidence is nonsense.

[Haiyang @ockddddw]

@blex1_ @MomenBuilds @Exocija That’s crazy….

[Haiyang @ockddddw]

@MomenBuilds @Exocija I from the token usage, I think maybe 50$ or even less? Haven’t checked the recent price of ds models.

[Haiyang @ockddddw]

@manicode Peak rage baiting

[Haiyang @ockddddw]

@natolambert Have you tried use this approach to scale your RL to use Riemann conjecture? I bet it would keep being all wrong and no signals will ever be detected..