Fable 5.1 was released this week. Astra might be coming as soon as tomorrow. Sam Altman, in a long post on X that should really be a blog post:
We have been living with the tension between being excited and anxious about progress for some time, and it is still discordant for us. We know it is much more discordant for other people. And yet, we believe strongly that the world needs to understand where AI is going and how models perform in the real world. More importantly, we believe the world will need aligned AI to manage the future phases of this transition.
In other words, shit is getting real, and the only way out is through. It’s not practical to pretend it’s not happening or to expect that an elite few people will control the release of capable models, especially as open weight models from China will keep advancing no matter what the frontier labs do.
Sam Altman expanded on this approach of iterative deployment in an interview with Alex Heath. I think Alex did a great job with his questions, covering a range of issues including data centers.
The Information broke a story last night that Astra uses recurrent depth or a so-called looped transformer. The technique effectively runs information through the same block multiple times, with more reasoning happening internally instead of in a chain-of-thought in English that can be inspected by humans.
I wonder if this was a purposeful leak, to get some potentially controversial news out of the way early, and it unexpectedly blew up. I don’t know enough about the technical details to judge whether it is worrisome. Jakub Pachocki of OpenAI pushed back on X, though he also wrote that chain-of-thought monitoring is generally becoming more fragile:
OpenAI has worked to preserve and utilize chain-of-thought monitoring since our very first reasoning models. We deeply care about this technique, as it can give us a view into how model alignment generalizes from its training distribution. I do think it is fragile and unfortunately trending in a negative direction, for reasons not contingent on architecture changes that I will write about soon. But there are things we can do to strengthen it, and it’s a core goal of our current research program.
Astra does appear to be a major step forward. GPT-5.6 Sol is very good in my experience. I’m pretty excited to see what Astra can do.
