Leaked Anthropic IPO slide shows superexponential pretraining efficiency gains, sparking debate
sebkrier · x · 2026-11-14
- A slide allegedly from Anthropic's IPO roadshow shows pretraining compute efficiency improving superexponentially over time (log-scale y-axis).
- The poster says it factored into their decision to leave: Mythos was one point on a superexponential curve — the same model could now be retrained with a fraction of the original compute.
- The implication: algorithmic progress is nearing automation, hinting at a software-only intelligence explosion.
- Economist bengolub pushed back: "this is not a serious argument" — AI people are getting used to discussing high-stakes matters without rigor.
A rare glimpse at Anthropic's internal scaling narrative plus a substantive fight over how the industry reasons about it.
More from AGI Musings
- Scott Alexander revives '1960: The Year the Singularity Was Cancelled' — slatestarcodex · 2026-11-14
- Dev pushes back on week's anti-GPU-market narrative, backs open source AI — sloppenheimer · 2026-11-14
- Musk: Agentic AI traffic will be 1,000x human internet traffic within 5 years — rohanpaul_ai · 2026-11-14
- On Purpose: AI expands creativity by unlocking ideas that never had a chance to exist — sachasayan · 2026-11-14
- Heaven-design debate: boredom complaints are failures of the lower rungs — jd_pressman · 2026-11-14
- 90% of job applications are obvious bots, says recruiter; human emails win — minhnguyen · 2026-11-14