Kelsey Piper visits Anthropic colloquium on bioweapons risk: 'interesting, but not very reassuring'
Kelsey Piper reports on an Anthropic colloquium on AI-enabled bioweapons risk, writing that Anthropic's plan for preventing worst-case outcomes "looks more like hope and guesswork." She describes a chart showing Claude, acting alone, outperforming a human-with-Claude at planning a mass terror event since early 2026. Anthropic's Séb Krier pushed back on her framing, though he agreed it wasn't reassuring. In follow-up posts, Piper says the colloquiums are a "pretty costly" but only partly successful effort to avoid insularity from outside perspectives.
All posts · 10, most important first
- I visited Anthropic for a colloquium on where they see us headed. It was very interesting. But not very reassuring.
28 likes · 3 reposts - In The Argument: theargumentmag.com/p/what-i-saw-i…3 likes · 1 reposts
- I didn't conclude it was reassuring!1 likes · 0 reposts
- would bad outcomes require us to do dumb things? maybe, but, bad news:
25 likes · 3 reposts - My article in The Argument today is about how they've been hosting colloquiums where they bring lots of external people in and give them serious senior researcher time. This is clearly a pretty costly effort to avoid doing this.5 likes · 0 reposts
- I don't think it totally succeeds, as an effort to avoid doing this. Most people from a cold start just can't actually say anything unexpected on a topic where there's so much inside info (not actually that much secret info, but it's certainly not widely known).5 likes · 0 reposts
- I think the dynamics around secrecy, intensity, shared assumptions, etc. do create a sense of most of the world as unreal, out of the loop, foreign, but I think they're in fact trying pretty hard to avoid this. Just not quite hard enough they succeed4 likes · 0 reposts
- living in the shadow of annihilation is bad for people. even if they're right.
17 likes · 1 reposts - 6 likes · 0 reposts
- 5 likes · 0 reposts
@kelseytuoc press · 
@sebkrier lab ·