How concerned should we be about Astra's recurrent architecture?
Summary
Rauno Arike assesses reports that OpenAI's Astra uses a looped-transformer recurrent-depth architecture akin to Geiping et al.
Astra is a looped transformer, not a classic RNN.
Recurrence runs along depth for a token—the same forward pass can loop before emitting—rather than carrying an unbounded hidden state across the trajectory. That bounds practical serial depth and is a step toward opaque reasoning whose size depends on loop count, not architecture scare alone.
Pachocki caps current depth near GPT-4.
OpenAI says frontier models including Astra stay within a factor of two of GPT-4's computation-graph depth and still lean on CoT monitoring. Effective depth matters more than looped-versus-unrolled form for monitorability; weight-tying may even help slightly.
The depth of the computation graph for our present frontier models, including Astra, is within a factor of two of GPT-4.
The remaining worry is a depth dial and a race.
Greenblatt asks whether inference or light fine-tuning can raise loops performantly, whether opaque reasoning jumped, and whether the architecture makes massively deeper future runs natural. Arike guesses Astra uses about three to four loops today.
this statement is consistent with Astra having a configurable "dial" that is currently set to a low depth but could be trivially increased.
Most of my concern comes from the possibility that looped transformers can be scaled a lot further in the future
Literature does not yet show a race to hundreds of loops.
Huginn trained up to 32 and tested to 64, but larger models such as Loopie (20B-A2B) use two; small-scale stability work reaches ~12. Deeper is not always better, so a bottom-race toward hundreds of loops is uncertain.
Neuralese concern tracks serial depth needed to subvert oversight.
Circuit-complexity takes and no-CoT horizon scaling suggest "spit a token every few hundred layers" is a weak bound; Hugging Face-style tool-call spoofing may need less opaque depth than a full plan. OpenAI's own monitorability signals still trend fragile independent of this architecture.
Read the original ↗Save to your Sponge
Summarized with Sponge. Quotes are the author’s own words.