←
Botos Csabi370 posts
Botos Csabi avatar

Botos Csabi

@csaba_botos

curriculum design @GoogleDeepMind botcs.github.io/academic

◷ Joined May 2017▧ 75 Photos and videos
Botos Csabi avatar
Botos Csabi@csaba_botos·Oct 8

I know this is a wild bet, but please, take the data and see if this check out. There's just so much more to learn from it: botcs.github.io/reason-to-pl… huggingface.co/datasets/csbo…

Botos Csabi avatar
Botos Csabi@csaba_botos·Oct 8

... than the most sophisticated cognitive theory based models. There's no incentive for this stuff to happen. I think there may be some abstract universal circuitry that all intelligent systems have to independently rediscover.

Botos Csabi avatar
Botos Csabi@csaba_botos·Oct 8

they were trained on coding, math and all the domains, where the high validation score often meets extreme scrutiny due to benchmaxxers, and I just can't imagine that the labs went out of their way to train their models so that they end up predicting brain signals better ...

Botos Csabi avatar
Botos Csabi@csaba_botos·Oct 8

I have a really hard time wrapping my head around how and why frontier models would ever show human-brain alignment on games (unlike perception tasks, see Brain Score leaderboard)

Botos Csabi avatar
Botos Csabi@csaba_botos·Sep 17

After 10 incredible years of not being at @GoogleDeepMind, I’ve decided to quit. The decision to stop not working there was not taken lightly. I'll be joining @sirbayes' team to continue the work on curriculum design, towards a general computational model of teaching.

Post attachment
Botos Csabi avatar
Botos Csabi@csaba_botos·Aug 27

wow. actually true. the same way GH is crumbling now, under yottabites of slop code, HF will soon see a surge of everyone uploading agent trained slopcheckpoints for RSI flagplanting.

Idk how to explain this but Microsoft buying GitHub feels identical to NVIDIA buying HuggingFace

Botos Csabi avatar
Botos Csabi@csaba_botos·Aug 25

Listening to @SemiAnalysis_ and @dwarkesh_sp Dylan rants made me realise: at one point Dylan was complaining about "releasing too much alpha" - however if he didn't share anything no one would care listening. So maybe the meta is to share just enough to influence the market?

Oliver Habryka avatar
⟳ Botos Csabi retweeted
Oliver Habryka@ohabryka·Aug 20

Applications for the first round of lightconecommons.com are closing this Sunday! We'll likely facilitate $35M of grants by end of October. The application process is minimal and extremely flexible. If you have an ambitious project of almost any scale, send in what you have.

Botos Csabi avatar
Botos Csabi@csaba_botos·Aug 17

worry not, there are some interesting news on the horizon... I've been looking into the edits done by RLVR on different domains.

0:28

IMO OLMo 1-3 still underrated in their contributions to science. I wish we could’ve spent more time amplifying this while building them.

Fanfei Li avatar
⟳ Botos Csabi retweeted
Fanfei Li@fanfei__li·Aug 14

We hope LittleLearner + LittleCurriculum will serve as a controlled sandbox for studying learning beyond pretraining: from continual learning and interpretability to curriculum design and RL-driven discovery. 📄 Paper: arxiv.org/abs/2608.13545

Botos Csabi avatar
Botos Csabi@csaba_botos·Aug 14

Until that point, I believe working on a heavy weight data driven computational model of teaching is going to benefit both flesh and metal.

Botos Csabi avatar
Botos Csabi@csaba_botos·Aug 14

I predict that soon we will have a branching point in frontier AI, where one camp will target super human performance in every aspect and the other will aim to close the gap in every success and failure modes to humans.

Botos Csabi avatar
Botos Csabi@csaba_botos·Aug 14

Why 3-5 years? - We are _FAR_ from having fully reverse engineered our learning algorithm. - Models are _FAR_ from being able to infer our relationship to discovering new material, executing on known concepts.

Botos Csabi avatar
Botos Csabi@csaba_botos·Aug 14

I am searching for a computational method that can virtually A/B test 10k+ curricula, model my learning process in each hypothetical environment, and select the one that structures my understanding of the world in a way that I will feel motivated to fill in the gaps myself.