Anthropic's roadshow slide: pretraining efficiency on a bending log curve

A slide from Anthropic's IPO roadshow shows the "pre-training compute multiplier" for code rising on a log scale, and bending upward, from July 2024 to October 2026. Jacob Coxon, who says he left Anthropic in part because of it, posted it on 13 November 2026 as evidence of superexponential algorithmic progress.

Anthropic roadshow slide, "Pre-training compute multiplier over time (code)": a log-scale line with points at Jul 2024, Apr 2025 and early 2026, steepening and ending in an arrow near Oct 2026. Footnote: placement is based on internal approximations and is solely for illustrative purposes Anthropic IPO roadshow slide, via Jacob Coxon. The y-axis has no numbers.

Rufus's note (2026-11-13): RSI and superintelligence data point. — Rufus Pollock

What he says

This slide from the Anthropic IPO roadshow was a big factor in my decision to leave. It's an insane plot and I'm surprised they made it public.

He argues that Mythos "was not a one-time deviation, but a point on a superexponential", and that Anthropic "could already retrain Mythos with a fraction of the compute originally needed". He concedes this doesn't strictly imply a software-only intelligence explosion, since Anthropic has also been spending more on R&D, but says algorithmic improvements "are on the verge of being automated".

How much to read into it

  • The axis has no numbers, and Anthropic's own footnote says the points are internal approximations "solely for illustrative purposes".
  • The LessWrong analysis shows how much the scale matters: if each gridline is 10×, the endpoint is about a 30,000× multiplier; if each is 1.1×, it is about 1.5×. Its author says they are "pretty unsure".
  • It covers pretraining for code only.
  • Coxon resigned from Anthropic in September 2026 citing existential risk (jacob-coxon-anthropic-resignation); The Decoder describes him as a former OpenAI and Anthropic researcher. I found no Anthropic response to the slide.

It fits the same picture as metr-anthropic-ai-rd-acceleration: in the roadshow, Anthropic's founders described compute multipliers compounding, and Tom Brown said this could speed up as Claude finds multipliers itself (per the LessWrong post).

See also