@CFGeek Even better, if you will

@johnhewtt In other words, if we assume, prior to seeing the machine's behavior, that the NL term captures the machine's internal representation, does the delta represent how much that assumption needs to be "corrected" to align with the machine's actual behavior (or a desired behavior)?

@geoffreyirving @krishnanrohit It’s the massive hiring during Covid (and subsequent layoffs) that make the chart, not GPT

@lateinteraction @DimitrisPapail Much much appreciated!

@YouJiacheng @stalkermustang @__nmca__ @ahelkky Yep but seems unclear how (“learned scoring fn”); also selection is top 50

@__nmca__ @stalkermustang @ahelkky So what you are saying is that o3 has learned to become its own critic...?

@PreethiLahoti Pinged in DMs :)

@willccbb The exact amount we will never know :) time will tell

@willccbb I reckon this is pretty much what’s behind Pro version and ARC score, which provides them yet another dimension for scaling

RT @natolambert: The fact that we seem to be marching straight towards another cold war, where AI is the defining technology, is hard to em…

@xuanalogue To be fair there seems to be discourse going on around this backtracking being part of the pretraining corpora (in subreddit-esque corpus), which the models learn to exploit during the outcome based RL training they go thru. Someone had found some evidence of this, ill follow up.

@willccbb @aryaman2020 I think a foundational piece of research most forget to mention is Belkin et al: Which came before all the OAI work on DD/Grokking

From

RT @mpshanahan: I agree with @AmandaAskell. It is a fallacy. It can be appropriate to say both a) that LLMs are next-token predictors, at a…

@doomslide @dhtikna @Grad62304977 @stochasticchasm @zephyr_z9 Different base as in?