More or Less · Topics · Episode #165

OpenAI's Astra and the Observability Trade-Off

From #165 — AI Agents Are Going Rogue. Should We Be Worried? · September 4, 2026 · 1:57 min · Jessica Lessin, Dave Morin, Brit Morin, Sam Lessin

Clip from the full recording, 20:09–22:06 · download mp4 · watch at this point on YouTube

The Information reports OpenAI's next model uses 'recurrent depth' — it shows its work less. Jessica on the uncomfortable tension Altman himself admits: the techniques pushing capability forward are the same ones making rogue behavior harder to see.

Jessica walks through The Information’s reporting on Astra, the next OpenAI model: it reportedly uses a technique called recurrent depth, which makes the model fundamentally less observable — it shows its work less. Altman has been unusually upfront that the progress/security trade-off is uncomfortable, but the lab is shipping the technique anyway, and Reddit’s take is blunt: less observable is less controllable, at exactly the moment everyone just agreed observability is the missing layer.

Other labs are experimenting with the same family of techniques, so this isn’t an OpenAI-only question — it’s the industry’s revealed preference: when capability and transparency conflict, capability ships. Sam’s coda ties the whole block together: every one of these stories is “a conversation about control and sovereignty” — the labs saying trust us, users wanting alignment with themselves, and governments wanting the wheel, with all three pulling in different directions.

Key points

• Astra reportedly uses 'recurrent depth' — a technique that makes the model show its work less, i.e. fundamentally less observable.
• Altman is openly acknowledging the progress-vs-security tension while shipping the trade-off anyway.
• Other labs are experimenting with the same technique family — it's an industry-wide revealed preference, not an OpenAI quirk.
• The timing problem: less observability lands exactly as agent observability became the consensus missing layer.
• Sam: all of it reduces to one question — control and sovereignty. Labs vs. users vs. governments.

Where they landed

Quotes

“It's fundamentally less observable, which also gets to this question of the trade-offs in research.”— Jess

“Everything is just a conversation about control and sovereignty. The labs — we know better than everyone — is very different than aligning with users, which is very different than putting things in the hands of governments.”— Sam

Suggested tweets

All topics · Full episode #165 + transcript · Subscribe on YouTube · Spotify