X Debate: Former Anthropic Researcher Jacob Coxon Says a Log-Scale "Pretraining Compute Efficiency" Slide From the IPO Roadshow Was a Big Factor in His Leaving, Drawing 5,200 Likes and About 1.19M Views

Background

Anthropic is on its IPO roadshow, and every chart in the investor materials is being picked apart. Early on November 14 (Singapore time), former Anthropic researcher Jacob Coxon posted one roadshow slide on X, said it was "a big factor" in his decision to leave, and said he was surprised the company made it public. The post quickly reached about 1.19M views and 5,212 likes, and X grouped the controversy into a trending story about the Anthropic IPO chart.

What the slide shows, per Coxon

In Coxon's reading, the slide plots pretraining compute efficiency: how much less raw compute is needed for the same model quality. The y-axis is a log scale and the curve is still bending upward, which he takes as superexponential algorithmic progress. Mythos, he says, was not a one-off breakthrough but a point on that curve, and Anthropic could already retrain Mythos with a fraction of the original compute. He adds that this does not strictly mean a software-only intelligence explosion, since R&D spending has also grown, but argues that algorithmic gains do not need rare genius and are close to being automated.

Numbers that matter

This is not a benchmark, and the slide has no published values; Anthropic has not released them. Some critics point out the chart has no labels or units. Coxon replied that compute multipliers are unitless, so relative distances still matter, and that comparing the gain after Mythos with the gain before it needs only three data points. He says he will not break his NDA to share absolute numbers, out of respect for the trust between the company and its employees. The verifiable numbers are the post metrics: about 1.19M views, 5,212 likes, 444 reposts and 321 replies.

What developers should take from it

For developers, the point is this: if pretraining efficiency really keeps accelerating, a given level of capability gets cheaper faster, and today's flagship abilities could reach smaller, cheaper models sooner. But this is one former employee's interpretation, and Anthropic has not publicly responded. In follow-up posts Coxon also argued for pausing frontier development next year, which is his personal position. For real tool and model choices, rely on official releases, pricing and hands-on results.