Zvi Mowshowitz covers the chart and the NDA debate, including David Manheim's testimony suggestion
Coxon Left Anthropic Over IPO Chart He Reads as Superexponential Efficiency
A former pretraining researcher's public departure tied to internal efficiency data raises questions about how rapidly AI training costs may be falling. The NDA debate adds a layer about how safety concerns from insiders can and should reach public discourse.
When stories like this move, I write about them in an email each morning.
One free email each morning, around 8 AM US Eastern.
Check your email.
I have sent you a confirmation link. One click and you are in.
Not in your inbox? Look in spam or promotions. If it's there, mark it "Not spam" so future editions land in your inbox.
The full picture
Jacob Coxon, a former Anthropic pretraining researcher, left the company in September 2026 citing existential AI risk concerns. He has since posted publicly about a chart from Anthropic's IPO roadshow that he characterizes as showing superexponential compute efficiency gains for pretraining code models. The chart plots a pretraining compute multiplier on a log scale from mid-2024 to late 2026, and Coxon says the line bends upward even on that log scale, implying efficiency improvements are themselves accelerating. He argues this means Anthropic's Mythos model could be retrained with a fraction of its original compute cost, and that algorithmic improvements do not require exceptional human insight and are close to being automated. Coxon is honoring his NDA and has not released proprietary details, though he has spoken publicly about his concerns. The superexponential interpretation is Coxon's reading and has not been independently verified against Anthropic's published methodology.
How it developed
explainx.ai publishes coverage of Coxon's claims about the Anthropic efficiency chart
Sources
Related
- Related storyMETR's Anthropic AI capabilities audit
- Anthropic's IPO and Coxon's safety departure
- OpenAI's AI research agents and self-improvement reporting
- AI doom and international safety governance
- Anthropic IPO and compute commitments
BackgroundWhen AI doesn't do what we meant
Get the daily email
When stories like this move, I write about them in an email each morning.
One free email each morning, around 8 AM US Eastern.
Check your email.
I have sent you a confirmation link. One click and you are in.
Not in your inbox? Look in spam or promotions. If it's there, mark it "Not spam" so future editions land in your inbox.