Jacob Coxon says an Anthropic IPO roadshow slide on pretraining efficiency was a big factor in his resignation

X · Personal post

Read the source (opens in a new tab)
What happened
Jacob Coxon, a former Anthropic researcher, posted a slide he said came from Anthropic's IPO roadshow and wrote that it "was a big factor in my decision to leave."
Key points
  • The slide plots pretraining compute efficiency on a log scale, which he says shows "continued superexponential algorithmic progress."
  • He writes 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 original compute.
  • He says this does not strictly imply a software-only intelligence explosion, but algorithmic improvements "are on the verge of being automated."
  • In a follow-up he argues that pausing frontier development next year "would commoditize human-level intelligence" and "would only harm OpenAI and Anthropic."
Evidence
Personal post by a former Anthropic researcher, with an image of the slide. The slide's footnote says its data points are "based on internal approximations" and "solely for illustrative purposes."
Uncertain
That the slide comes from Anthropic's IPO roadshow is Coxon's account. The chart gives no values on its vertical axis, and Anthropic's response is not part of this source.

See it in the latest record