AI #184: Post Post Mortem
Zvi (Don't Worry About the Vase) TheZvi ● Covered by 2 sources
OpenAI’s Astra may be using a trick that skips the usual chain of thought. That could help performance — or make AI systems much harder to read.
Based on reporting by Zvi (Don't Worry About the Vase), TheZvi — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
The HuggingFace hack fallout is finally starting to lose its grip on the week, and AI news is already piling up behind it. Zvi says direct coverage is tapering off, which is just as well, because five posts in seven days is enough postmortem for anyone. The next wave is here anyway: Mythos 5.1, Fable 5.1, Gemini 3.8 Flash, Muse Spark 1.3, GLM-5.3-Flash, and maybe OpenAI’s Astra.
The one item that sounds genuinely radioactive is Astra. The Information says OpenAI is using a technique called recurrent depth, which lets the model shift its thinking outside the Chain of Thought. Zvi’s read is blunt: OpenAI is “playing with fire,” even if the fire hasn’t burned down the house yet. For now, he says the technique does not seem to have badly damaged interpretability, but the fact that it works at all — and that OpenAI chose to use it — is the thing that should make people sit up.
The rest of the model releases mostly land in the annoying middle: good, maybe very good, but not world-changing. Gemini 3.8 Flash is described as a meaningful improvement over 3.7 Flash, though not exciting enough to earn much more than minimal coverage by default. Muse Spark 1.3 is also said to be a step forward, but not a moment. GLM-5.3-Flash, which showed up as the free preview named 0x Alpha, got the usual burst of “China is cooked” excitement before settling into “solid open model, not Sol, not Opus 5, definitely not Fable 5.”
There’s a lot of benchmark chatter in there, but the broader theme is familiar: people keep trying to crown the new release as the one that changes everything, and then reality shows up with a smaller update. Zvi’s read is that if these models were true game changers, the world would already be acting like it. Instead, the market is getting a stream of decent upgrades, a few loud reactions, and one potentially nasty interpretability problem from OpenAI.
My take — AI-written commentary, not fact-checked reporting
AI news has become a parade of “almost” and “maybe,” which is perfect for companies and terrible for everyone else. The real story here is not the shiny model score; it’s that OpenAI may be reaching for techniques that make systems more capable while making them less legible. That is a very old tech industry move, just with better branding and more anxiety.
Read more about this at: Zvi (Don't Worry About the Vase)