More or Less · Topics · Episode #167
Dario Says Slow Down. Is It Because Anthropic Is Losing?
Clip from the full recording, 2:44–10:41 · download mp4 · watch at this point on YouTube
A researcher quits Anthropic after three months and goes on a doomsday media tour, Dario publishes a slow-down essay — and the pod splits on whether it's genuine safety, marketing, or what you do when Astra is beating you.
The week’s main event: Jacob Coxon quit Anthropic after roughly three months, went on a media tour (with a new venture already lined up), and Dario Amodei published a long essay saying the AI race he’s been leading needs to slow down. Dave’s instant read: ‘Anthropic must be losing again’ — Astra is ‘objectively crushing them,’ OpenRouter data shows them falling behind for the first time in over a year, and DeepSeek is shipping Fable-level models at a hundredth the cost. Sam zooms out to his standing thesis: there’s no compounding advantage in models, so there’s no return on being first, which makes these bad businesses whose rational move is regulatory capture ‘before we realize the emperor has no clothes.’ Brit pushes back that it’s genuinely gray — bioweapons risk, agents breaking out, Demis proposing a DeepMind Institute — but concedes the China race makes slowing down impossible: damned if you do, damned if you don’t. Sam’s synthesis: he’s not an AGI doomer, but the chaos level coming is ‘insanely high’ — both things are true, the risk is real AND the essay is marketing.
Key points
- Jacob Coxon quit Anthropic after ~3 months, went on a doomsday media tour, and already has a new venture — Sam: the best career move if you're about to get fired.
- Dave: Astra is objectively crushing Anthropic; OpenRouter shows them losing for the first time in a year; DeepSeek ships comparable models at ~100x less cost.
- Sam's core thesis: no compounding advantage → no return on being first → bad businesses → regulatory capture is the rational endgame.
- Brit: the safety concerns are real (bioweapons, agent breakouts, Demis's DeepMind Institute idea) but the China race makes slowing down a non-option.
- Sam on Hugging Face as 'canary or red herring': obviously a canary — you don't need superintelligence for AI to break everything; banks will get hacked regardless.
Where they landed
Quotes
“Because there's no compounding advantage, there is no return on being first, which means you're in a brutal competition to zero... which means they have to go for regulatory capture as quickly as they possibly can before we realize the emperor has no clothes.”— Sam
“Astra is objectively just like crushing them. It's so much better, it's not even close.”— Dave
“We can't slow down because China is racing ahead... it's a damned if you do, damned if you don't situation.”— Brit
“We don't need super intelligence for AI to fuck everything... I'm not a doomer, but the level of chaos coming is insanely high.”— Sam
Suggested tweets
Dave Morin's read on Dario's slow-down essay: 'Anthropic must be losing again.' Astra crushing them + OpenRouter data + DeepSeek at 100x cheaper = the essay writes itself.
Tweet this →Sam Lessin: no compounding advantage means no return on being first means brutal competition to zero — so the labs' rational move is regulatory capture 'before we realize the emperor has no clothes.'
Tweet this →Is the AI slowdown debate safety or marketing? More or Less says: yes.
Tweet this →All topics · Full episode #167 + transcript · Subscribe on YouTube · Spotify