Botos Csabi

Encode Fellow of the 2025 cohort. PhD from University of Oxford

[Botos Csabi @csaba_botos]

Jeee 🐦‍⬛

I am very proud of our joint effort with @sreejan_kumar Kumar on the project "Reason to Play"

LRMs show human-like rule discovery, and their hidden states predict human brain activity during gameplay.

Interactive demo + paper:
https://t.co/4ouwGoqgDf

[Sreejan Kumar @sreejan_kumar]

I'm excited to share my new work joint with @csaba_botos: how well do frontier LRMs like DeepSeek-V4 learn novel video games and predict brain signatures associated with game learning? We found that they achieve SOTA on both game performance and brain predictivity! Thread 1/7

[Botos Csabi @csaba_botos]

My rockstar team:
@sreejan_kumar, Austin Andrews, @LHuntNeuro, Chris Summerfield, Josh Tenenbaum, @somnirons, @marcelomattar and the brilliant @momchiltomov

The project was funded by: https://t.co/8jk5G3Az1i

[Botos Csabi @csaba_botos]

Jeee 🐦‍⬛

I am very proud of our joint effort with @sreejan_kumar Kumar on the project "Reason to Play"

LRMs show human-like rule discovery, and their hidden states predict human brain activity during gameplay.

Interactive demo + paper:
https://t.co/4ouwGoqgDf

[Botos Csabi @csaba_botos]

@sirbayes I think it's all about detecting the likelihood of environment drift. If you know you're safe then you can amortize away - but then why bake the harness in the weights instead of distilling the weights into pure code and tune the weights to act as a fuzzy sentinel for drift?

[Botos Csabi @csaba_botos]

Isn't harness building, context editing and prompt search analogous to LoRA, but in a semantically meaningful subspace?

It's just cheaper (no need to host the model), safer (catastrophic forgetting less likely), more interpretable (you can read now what lora would do!)

[Botos Csabi @csaba_botos]

@SamuelAlbanie I'd be very interested in an architecture difference breakdown with the competitor in same scale regime Qwen3.5:

https://t.co/5ZHBRu4RZd

[Botos Csabi @csaba_botos]

it's funny to see how mankind got advantage by fencing off a plot of land, taking it into his head to say ‘This is mine’ and now thousands of years later to advance science we need to remove the metaphorical fence, let messy nature back in to access more fruits of our land

[Botos Csabi @csaba_botos]

shower thought: how would you go about teaching a dog to teach other dogs a trick?

[Botos Csabi @csaba_botos]

@francoisfleuret do you mean iFFT being the decoder?

[Botos Csabi @csaba_botos]

all the soap opera around NeurIPS, ICML and ICLR is a direct cause of real world forces acting on an artificial community. Having a paper can get you a job, a grant, a promotion, regardless whether the paper itself is pure garbage or seminal breakthrough.

[Botos Csabi @csaba_botos]

@micahgoldblum @AdityaHumnabad1 they should write LEAN

[Botos Csabi @csaba_botos]

https://t.co/1tFnSoXJQ9

@karpathy Probably, the cerebellum *is* important, just mostly overlooked

@carobellum has an awesome atlas published listing even higher level functions: social, linguistic and spatial domains https://t.co/zdxt2nJibs

[Botos Csabi @csaba_botos]

the bonus bi-product would be some excellent training data for the future AI scientist agents.

Same way as we learned the dark arts of ML by just reimplementing famous papers in the first year of the PhD.

[Botos Csabi @csaba_botos]

Neurips R2 thought: why is the main objective for building AI scientists to accelerate new discovery while the reproducibility crisis is already crippling many fields.

Maybe less ambitious but more robust and likely to be impactful work would be automating reproducibility check.

[Botos Csabi @csaba_botos]

oh my! real art is not running off the edge of the world but stopping where everyone else walks past

[Andy Keller @t_andy_keller]

We made wave dynamics flexible by adding learned damping and natural frequency encoders, allowing hidden state dynamics to adapt based on the input stimulus. On simple polygon images, we found the model learned to use these parameters to produce shape-specific wave dynamics:
6/14

[Botos Csabi @csaba_botos]

@francoisfleuret a professor making a meme about this pain just feels surreal. big time

[Botos Csabi @csaba_botos]

Amazing experience with the prettiest view @NeurIPSConf presenting Label Delay https://t.co/LtQvCqAwUs

the original session was 3 hours - stayed 2h more for follow-up questions, people's interest in the topic was way above my expectations :) feeling awesome!

[Botos Csabi @csaba_botos]

Yay 🦜

Heading to Vancouver on Monday to present my first ever @NeurIPSConf paper.

I am super duper excited to finally join the cool kids’ gang.

Check out the tutorial I procrastinated over for 3 months:
https://t.co/h3KQ2eCCRd

[Ferenc Huszár @fhuszar]

Call For Tasks for IOAI 2025: ML olympiad for high schoolers.

I know you all have strong opinions on the ML topics future computer scientists should engage with to build mastery. This is a high-impact opportunity to help us by proposing tasks.

https://t.co/0kaGockZ9O

[Ferenc Huszár @fhuszar]

Teams of Hungary took 5th and 6th place in the first International Olympiad in AI taking the top 2 spots for silver 🥈.

Congrats to the participants who brought a lot of self-taught expertise, adaptibility to the event continuing to build our tradition of problem solving. 🇭🇺🇭🇺🇭🇺