Kevin Gimpel

[Kevin Gimpel @kevingimpel]

@inductionheads @InverseMarcus hm? the SP paper was written in late 2020 and published in 2021 while the original chatgpt was released nov 2022. agreed that RL methods are ancient and that parrot claims are now obsolete, but the parrot claims in 2020/2021 were not made about chatgpt.

[Kevin Gimpel @kevingimpel]

@inductionheads @InverseMarcus but the training did not always use that harness, e.g., original seq2seq paper (as well as many others obviously). it's not required. using a decoding harness during training is a distinction one can make when describing methods.

[Kevin Gimpel @kevingimpel]

@InverseMarcus hard to answer w/o a definition of an LLM, as some think about them as a model family & others as systems with specific types of training criteria (string prediction). the latter may say current things are LLMs in terms of model family but not training.

[Kevin Gimpel @kevingimpel]

@yoavgo @petergostev agreed. when we discuss the SP paper, we should note that it comments only on "systems which are trained on string prediction tasks", which would exclude most RL as used today and many other things.

[Kevin Gimpel @kevingimpel]

@mathelirium @DavidPe51482177 but aren't you making some assumptions about the variance in those many dimensions? what if variance is near-zero in nearly all dimensions?

[Kevin Gimpel @kevingimpel]

@jsuarez @andreiC_dev what are you maximizing when automatically tuning the coefficients? is it still some sort of automated metric/reward? if so, then why not just use that metric/reward directly?

[Kevin Gimpel @kevingimpel]

@brian_jabarian @HonglinB @alexolegimas some detectors go beyond binary classification though, e.g., QuillBot does 4-way classification

[Kevin Gimpel @kevingimpel]

@tinkerteller @yoavgo how about universal transformers (https://t.co/bsZyWaOAwI)? they can "can choose the number of processing steps based on the input data"

[Kevin Gimpel @kevingimpel]

@ayaannmalik @Kangwook_Lee this output is a little misleading though.. the problem is that the o3 bar's height was erroneously set to equal the height of the 4o bar, not that the "without thinking" bar was drawn taller than the o3 bar.

[Kevin Gimpel @kevingimpel]

@orionweller gotcha thanks! yea makes sense if you're considering baselines that are good for both types of tasks. i haven't used encs much for retrieval, but for classification i am still waiting for something to outperform debertav3-large. modernbert also did not outperform it.

[Kevin Gimpel @kevingimpel]

@skirano correct answers implies supervised data created with manual effort (or synthetic data generated by templates.. still requires manual effort). i just don't see how this feels any different from the requirements of supervised finetuning.

[Kevin Gimpel @kevingimpel]

@skirano huh? SFT still helps though. DeepSeek-R1-Zero is not as good as DeepSeek-R1. and DeepSeek-R1-Zero still requires labeled training data in order to compute rewards. the feedback you mention is correct answers.

[Kevin Gimpel @kevingimpel]

@zjasper @karpathy @deepseek_ai ~100 people is small?!

[Kevin Gimpel @kevingimpel]

@jeremyphoward is something off with the x axis? 10ms/token is extremely slow.

[Kevin Gimpel @kevingimpel]

@sharifshameem @abhi_venigalla "sometimes ~2 sec"? what does that mean exactly? for the 10x to be accurate, the mean or median would have to be 2sec, right?

[Kevin Gimpel @kevingimpel]

@complingy @aryaman2020 ya did some stuff for MT in 2013, but i'd recommend looking at this paper:
https://t.co/BfdlxU9rgs

[Kevin Gimpel @kevingimpel]

@fredahshi somehow @kalpeshk2011 stole your "i" ;)

[Kevin Gimpel @kevingimpel]

@fredahshi @GoogleAI congratulations Freda!! :))

[Kevin Gimpel @kevingimpel]

@kalpeshk2011 @PreethiJyothi1 congrats, Kalpesh!!