OpenAI hired the man who thinks AI could kill everyone
Issue #034 · 4 min read
OpenAI hired the man who thinks AI could kill everyone
Paul Christiano joins the board that's supposed to keep the company safe. Also: GPT-6 Astra lands on GitLab, and Anthropic's crime count hits four.
The big story
Two stories about the same fear landed on the same day, and they do not agree with each other.
Paul Christiano just joined OpenAI's Foundation board and its Safety and Security Committee. He's an alignment researcher, and the phrase people keep using for him is "AI doomer" — someone who thinks advanced AI poses a real risk of catastrophe, not a theoretical one. OpenAI announced it straight: he brings "experience in AI alignment, safety, and standards." TechCrunch's framing is blunter: OpenAI just put a prominent doomer on its board.
On the same day, an Anthropic researcher quit.
Not quietly. He left with a warning, reported by Ars Technica: "We really do earnestly believe AI could kill all humans." That's not a headline writer's exaggeration — it's a direct quote from the resignation.
So one company is hiring the person who worries most, and one is watching that person walk out the door instead.
Both of these are supposed to be the safety-conscious labs. Neither move tells you which one is actually safer — it tells you the argument inside these companies is not settled, and it is happening in public now, on the record, with names attached.
If you build on top of either model, this isn't background noise. It's a signal about how much the vendor above you actually agrees on what "safe" means, and that disagreement eventually shows up in what they'll let their models do for you.
What shipped
GPT-6 Astra is now on GitLab Duo Agent Platform, and GitLab ran the numbers itself. In GitLab's own internal evaluation, Astra finished a typical run 43.4% faster than the previous model and used 42.7% fewer tokens per run — while completing every task in the benchmark, according to GitLab's writeup. If you're paying by the token for agentic work like dependency updates or build fixes, that ratio is the one to watch, not the marketing copy around "most capable model" language OpenAI used in its own launch post.
GitHub now lets enterprises put a leash on Copilot's agents. Admins can centrally set which agent operations get blocked outright, which need a human to approve them, and which can just run — full rundown here. This is the boring, necessary counterpart to every "look what our agent can do" release: someone has to decide what it's allowed to do before it does it in your codebase.
Vercel's Sandbox is now available in every region, closing a gap where teams testing agents against production-like infra were stuck picking from a shorter list — details here. Small change, but if you've been routing test traffic through one region because that's where Sandbox lived, that constraint is gone.
What I'd actually do this week
- If you're evaluating GPT-6 Astra for agentic coding work, ask your team for token usage before and after — GitLab's own 42.7% number is a real benchmark you can hold vendors to, not a claim to take on faith.
- Go check who in your org can already override or approve agent actions in Copilot, Cursor, or whatever you've rolled out. If the answer is "nobody, it just runs," GitHub's new permissions model is worth ten minutes.
- Read Christiano's actual position and the departing Anthropic researcher's actual warning before you repeat either company's line about "safety" in your next vendor review. They disagree, and your risk assessment should reflect that, not paper over it.
Reply if you've hit a wall rolling out agent permissions somewhere real — I want to hear what actually broke.