Twitxy Saved

Taco Cohen

Verified account This account is verified. Learn more

@TacoCohen

Slop janitor & post-trainologer at Meta / FAIR. Into codegen, RL, equivariance. Spent time at Qualcomm, Scyfer (acquired), UvA, Deepmind, OpenAI.
tacocohen.wordpress.com Joined March 2013
4K Following 31K Followers 1.7K Posts

Latest posts 5

Taco Cohen Verified account @TacoCohen
Three takes for the price of one:

1. Excellent fear marketing. Hats off

2. “My agent did it during an eval” is now the perfect excuse if you get caught hacking.

3. Now is the time to start freaking out about paperclip maximizers / RL agents relentlessly pursuing narrow goals again
OpenAI Verified account @OpenAI Jul 21
We're partnering with @huggingface to investigate an unprecedented security incident.

Cyber-capable OpenAI models compromised Hugging Face production during a benchmark evaluation.

Sharing preliminary findings to help defenders understand emerging risks:

openai.com/index/hugging-…
View quoted post on X
11 replies 6 reposts 218 likes 27K views View on X
Taco Cohen Verified account @TacoCohen
As prophesied by the venerable I. Sutskever
Tiberiu Mușat Verified account @Tiberiu_Musat_ May 27
Why does deep learning generalize? What does weight decay really do? Can algorithmic information theory address these questions?

In my latest preprint, I give a proof that the minimum neural weight norm matches the minimum program length (aka Kolmogorov Complexity), up to a logarithmic factor. In other words, the neural network with the smallest possible weight norm (that fits the data) must encode the shortest program (that fits the data).

The result only holds for fixed-precision neural nets: infinite precision nets can store infinite information with finite (small) weights.

arxiv.org/abs/2605.10878
View quoted post on X
5 replies 34 reposts 458 likes 76K views View on X
Taco Cohen Verified account @TacoCohen
There’s been a lot of talk about AI revolutionizing science and human health, but I feel like often there’s a gap between the narrative (often by consumer AI companies) and the actual results.

Now it’s actually starting to happen: Max’s company CuspAI is making serious progress towards AI-discovered materials for PFAS removal.
Max Welling @wellingmax May 21
View quoted post on X