Ex-Anthropic staffer leaks IPO roadshow slide showing superexponential pretraining efficiency gains
anshulkundaje · x · 2026-11-14
- A former Anthropic employee says this IPO roadshow slide was a factor in his decision to leave: it plots pretraining compute efficiency over time on a log scale, indicating continued superexponential algorithmic progress.
- The author argues Mythos was not a one-time deviation but a point on that superexponential curve — the model could already be retrained with a fraction of the original compute.
- A tiny footnote notes "the placement of the data points on the graph is solely for illustrative purposes," which the reblogger mocks.
- It doesn't strictly imply a software-only intelligence explosion since R&D spending also grew, but algorithmic improvements don't require genius and are on the verge of being automated, the author concludes.
More from Companies & People
- Daniel Cremers, MCML director and TUM professor, joins X — CSProfKGD · 2026-11-15
- Humanoid startup Foundation misses payroll as founder borrows $4.6M for a new home — PTrubey · 2026-11-15
- FT: Nvidia lacks insurance to shield itself from shareholder suits after Huang's exit — Polymarket · 2026-11-15
- Tim Sweeney Slams Steam's AI Disclosure Labels as a 'Scarlet Letter' for Developers — mark_k · 2026-11-15
- Would you pay $2,999 for unlimited Claude coding? Community debates a dedicated device — jasonkneen · 2026-11-15
- 2009 clip resurfaces of Jensen Huang explaining GPUs to a baffled Charlie Rose — CatAstro_Piyush · 2026-11-15