Post

Between this letter and the fact two of the three were first authors on the CoT monitorability paper, it seems likely to me that the employees were fired for leaking the fact that Astra is a lopped transformer to The Information. www.wsj.com/tech/ai/fire...

8/10/26, 1:35 am1 repost48 likes

fry69

@fry69.dev

What's the fuzz about looped transformers? It seems it is just am minor technical optimization, how layers get reused (usually only a second time)? See here -> sebastianraschka.com/blog/2026/op...

sebastianraschka.comOpenAI Astra and Looped TransformersA short note on OpenAI Astra, recurrent depth, looped transformers, Nanbeige 4.2, and the Mixture-of-Recursions paper.

Alex Lindquist

@alxfed.bsky.social

Three immigrants (from China, Ukraine and Poland) are working on 'security' of this thing that is being developed for DoD, NSA and the rest of them? Ha-ha-ha! 🙂

Ryan Moulton

@moultano.bsky.social

Would people be having the same reaction if the model was just deeper? It seems like looping created a mental discontinuity of some kind that maybe isn't warranted.

Festivity

@festivitymn.bsky.social

I doing think they would. I think the worry is based on work that tries to understand model behavior based on activations. If you loop the model, it’s harder to see what those are because they get overwritten. That’s probably not a big issue with one or two loops, of course.

Festivity

@festivitymn.bsky.social

But my recollection is that this sort of thing loops until the answer doesn’t change, making it “infinitely deep”. I think that would make it more or less impossible to know what to look for. julialang.org/blog/2021/10...

julialang.orgComposability in Julia: Implementing Deep Equilibrium Models via Neural ODEsComposability in Julia: Implementing Deep Equilibrium Models via Neural ODEs