Nathan Lambert: agents improving AI training code doesn't add up to doom in N years
natolambert · x · 2026-09-09
RL researcher Nathan Lambert pushes back on doom framing: we have evidence that (1) AI agents are good at improving software used to train AI, and (2) LLMs can be superhuman in many domains. But these two facts don't combine into 'AI will kill us in N years.'
He argues such forecasts amount to assuming inevitable saturation plus unknown future breakthroughs, rather than rigorous reasoning.
More from AGI Musings
- Debate: ASI could arrive in 3-10 years, leaving little time to adapt — ronbodkin · 2026-09-10
- Benjamin Bratton and Marek Poliks lecture: how trillions of agents will reconfigure society — tylerjdunn · 2026-09-10
- Full text of 'Superdark Factory' paper now live on Antikythera, detailing the dark stack's four principles — tylerjdunn · 2026-09-10
- Ex-Continue founder publishes 'The Superdark Factory' in MIT Press: 35T agent tokens/month by 2029 — tylerjdunn · 2026-09-10
- Experiment with 100 AI agents: when 9% cheated on math problems, 24% blew the whistle — weballergy · 2026-09-10
- kuza55: framing AI safety only as 'superalignment or pause' fuels panic — kuza55 · 2026-09-10