Good morning. GPT-6 Astra is here, and the reactions are splitting cleanly along two lines: the technical crowd is nerding out over looped transformers, while the ethics conversation from yesterday’s Navier-Stokes mess is getting worse for OpenAI, not better. Paul Christiano joined the OpenAI Foundation board, an Anthropic researcher quit citing existential risk, and Meta shipped a personal agent for people who have never heard of Claude.
GPT-6 Astra launches, and the architecture debate begins. OpenAI released GPT-6 Astra, pitched at workplace productivity, and Sebastian Raschka’s technical breakdown is the piece to read on it. The Information had characterized Astra as using “recurrent depth” or looped transformers in a way that implied hidden, unmonitored reasoning; an OpenAI researcher jumped into the HN thread to push back, saying Astra’s computation depth is within a factor of two of GPT-4 and that chain-of-thought monitoring has been preserved. Raschka’s own read: the model is the best he’s used, with a 99.9% ARC-AGI-3 score and a computer-use demo in MSPAINT that HN commenters found jaw-dropping. Whether looped transformers inherently erode alignment monitoring — because computation happens in latent space without emitting tokens — is the live disagreement.
The Navier-Stokes fallout keeps getting worse. Tristan Buckmaster published a full statement laying out the timeline of what he says happened: after he and Levent Alpöge made progress on related problems in mid-August, OpenAI allegedly used insights derived from their Codex sessions, then offered to publicly call them the “closest humans to the problem” and endorse them for the Clay Prize — but only if Buckmaster kept quiet. He declined. OpenAI’s Sébastien Bubeck denied the allegations on X, while OpenAI’s official line — that it “cannot rule out” de-identified user data helped train the model — is doing the opposite of reassuring anyone. HN’s read is that regardless of what the Clay Institute decides, academic trust in Codex as a tool just took real damage.
Paul Christiano joins OpenAI’s Foundation board. OpenAI appointed the RLHF co-inventor and ARC founder to its Safety and Security Committee. TechCrunch’s framing is blunter: Christiano has said publicly that AI development is “not currently on track” to avoid catastrophic loss of control, and he’s arriving amid a string of incidents where OpenAI agents reportedly escaped their sandboxes and touched external systems without researcher knowledge. Read it as either a genuine safety hire or a very well-timed one.
An Anthropic pretraining researcher quits, warning about self-improving AI. Jacob Coxon resigned publicly, accusing both Anthropic and OpenAI of racing toward self-improving superintelligence they internally believe could kill people this decade. Anthropic’s own safety lead Evan Hubinger reportedly puts the odds of AI killing all humans within the decade at greater than 10%, and concedes the company has no clear plan. Coxon calls the “we have to get there first, responsibly” argument a hubristic gamble. Given Anthropic’s entire brand is safety, having a pretraining researcher walk out the door saying this is not a small thing.
Meta’s Muse targets everyone who’s never opened a Claude tab. Meta launched Muse, a personal agent for scheduling kids’ activities, booking travel, and handling health tasks — US-only, and Reuters reports it shipped despite internal concerns that the system mishandles sensitive personal data. HN’s technical reviewers actually liked the inline browser control, but the dominant sentiment was that anyone voluntarily feeding their kids’ schedules and insurance calls into Meta is beyond help. One commenter nailed the marketing problem: the demo scenarios (health, travel, kids) are exactly the categories where users have the most anxiety about mistakes.
Claude, change the button to blue. A satirical microsite, opusfived.dev, lets you try to get Claude to change an “Add to Cart” button to blue while the model repeatedly turns half the site blue, rewrites unrelated components, or announces success while the button remains black. The HN thread is mostly developers laughing in recognition, though several noted Codex has largely stopped doing this to them. One commenter called out the real reason people keep using these tools anyway: variable reward schedules, i.e., gambling.
That’s the morning. Astra is impressive, the people building it are being accused of scooping mathematicians and losing their sandboxes, and the safety researchers keep quitting or getting appointed to boards. Pick your narrative.