Unverified: Jacob Coxon Shares a Slide He Says Is From Anthropic's IPO Roadshow
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. It shows the compute efficiency of pretraining over time. The y-axis is a **log scale**: they are seeing continued superexponential algorithmic progress. Mythos was not a one-time deviation, but a point on a superexponential. They could already retrain Mythos with a fraction of the compute originally needed. It doesn’t strictly imply a software-only intelligence explosion- they have been spending more on R+D. But algorithmic improvements do not require ridiculous creative genius and are on the verge of being automated. You do the math.
Context
If the slide is genuine, it would be a rare look at how Anthropic describes pretraining efficiency to investors, but the document is unverified.
Jacob Coxon posted on November 13 a chart he said is from Anthropic's IPO roadshow, and wrote that the slide was a big factor in his decision to leave. He said the chart shows pretraining compute efficiency on a log scale and that Anthropic is seeing continued superexponential algorithmic progress. Anthropic has not confirmed the slide in the posts scanned, so the document and his reading of it are unverified. A footnote on the image itself says the data points are based on internal approximations and are solely for illustrative purposes.
Deep dive
Jacob Coxon posted on November 13 an image of a chart on the pretraining compute multiplier over time, with code noted in the title. He said the slide is from the Anthropic IPO roadshow, that it was a big factor in his decision to leave, and that he was surprised they made it public. Anthropic has not confirmed the slide, its origin or Coxon's account in the posts scanned. The document should be treated as unverified.
Coxon said the chart shows the compute efficiency of pretraining over time, that the vertical axis is a log scale, and that this means Anthropic is seeing continued superexponential algorithmic progress. He also wrote that Mythos was not a one time deviation but a point on a superexponential, and that they could already retrain Mythos with a fraction of the compute originally needed. Those interpretations are his.
The image itself includes a footnote that says the placement of the data points on the graph is based on internal approximations and is solely for illustrative purposes. The visible dates on the horizontal axis run from Jul 2024 to Oct 2026. The post does not include the rest of a deck or a source file.
What to expect next
Watch for Anthropic to confirm, deny or ignore the slide, and for any official roadshow or prospectus materials that show the same chart. If other people who attended the roadshow corroborate the image, that would strengthen the claim that it is genuine. Until then, treat both the document and the claim of superexponential pretraining progress as unverified. A later official chart with labeled axes and dates would also help readers check Coxon's reading.