Mariano Kamp

ML and long walks on the beach, no beach though.
Working for @awscloud. All words, singing and dancing are mine.

[Mariano Kamp @mkamp]

@sunilmallya Re: the most informative data available: failures.
Maybe combine a GEPA like mechanism that reflects in the language space, but then suggests new rollouts that with SFT can create some signal in the model, that RLVR could then bring to the surface?

[Mariano Kamp @mkamp]

@sunilmallya This is such an insightful and precise article. Congrats Sunil!

[Mariano Kamp @mkamp]

@purbeshmitra Yes, using a larger model, so that the smaller (maybe latency constrained student) can benefit from the extra capacity of its teacher. Still on-policy, still the student does the sampling of the trajectory. But the teacher then creates logprobs on the student's logits as its own.

[Mariano Kamp @mkamp]

@purbeshmitra I am wondering about the teacher being allowed to use thinking: high, while the student is only allowed to use thinking: none. This could be helpful for the tutoring, but it could also create hallucinations as the student then learns based on thoughts it did not come up with.

[Mariano Kamp @mkamp]

@purbeshmitra This is awesome! Thanks for the explicit writeup of the steps on gh. I am curious, did you try to give the teacher (more) reasoning tokens or tried a higher capacity teacher?

[Mariano Kamp @mkamp]

@siyan_zhao @anzhit What do you think about letting the teacher get additional thinking tokens, but only use the thinking tokens for the teacher conditioning, the actual trajectory is sampled from the student and the logprobs of the teacher are used to tutor the student?

[Mariano Kamp @mkamp]

@sunilmallya @SahikaGenc You fixed friction as in you added the concept in your simulator?

[Mariano Kamp @mkamp]

@Alex_Ambroz @anon_opin @duolingo As a native German speaker I can tell you: this is not helpful at all.

Unless you shout it. Then it will create friendships left and right. It’s how we Germans are. You are welcome.

[Mariano Kamp @mkamp]

@Zakrea_312 @scaling01 Say more.

[Mariano Kamp @mkamp]

@swyx @cognition Great news. Congrats.

[Mariano Kamp @mkamp]

@rasmus1610 Just double checking. Read your missing manual to Simba and wanted to know when you post something about GEPA :)
It seems to be a relatively recent trend to have a blog without a feed. I experienced this with @UnslothAI, and a few others, as well.

[Mariano Kamp @mkamp]

@rasmus1610 QQ: Your blog does not have an RSS/Atom feed, or does it?

[Mariano Kamp @mkamp]

@karpathy How to add the blog to a newsreader? Here NNW on iOS.

[Mariano Kamp @mkamp]

@AmandaAskell Could pin the receipt of a large trump coin purchase on the jacket.

[Mariano Kamp @mkamp]

@girlbossintech @corbtt Stupid question, but why RL over SFT (DPO like)? Easier to use the score from one sample/one answer?

[Mariano Kamp @mkamp]

@heitor_lessa How do you like mirrors now? :)

[Mariano Kamp @mkamp]

@rohanpaul_ai Also then, bigger Ns would be the result of the formula?

[Mariano Kamp @mkamp]

@rohanpaul_ai I am wondering about your thoughts here. This does not consider the effects of KV-Cache, or does it? The very low MFU for the output tokens should go up a bit when using KV caching and the general cost should go down from 2DN to 2D and a smaller N? Does that make sense?

[Mariano Kamp @mkamp]

@giffmana In terms of accessorizing your new purchase I got you covered. Say good bye to the can opener and hello to:

[Mariano Kamp @mkamp]

@giffmana Proves that you don’t have a constructing worker’s mind as you didn’t open the bottle on the edge of your TP. (Not on this photo at least).