Algorithmic improvements are reaching a point where they can be automated, according to Jacob Coxon, a former researcher who previously trained AI models at OpenAI and Anthropic. Coxon, who brought AI risks into the mainstream and drew wide attention with public warnings about advanced systems, cited a slide Anthropic released for its IPO roadshow as the primary reason he left the company after roughly four months.
The disclosed visual exposes a rapid compounding of algorithmic efficiency that undermines massive capital expenditure budgets. While frontier labs secure infrastructure to train flagship systems, efficiency gains are accelerating so quickly that the cost to replicate capabilities drops sharply shortly after deployment.
The Compressed Roadshow Metrics
The chart released for Anthropic's IPO roadshow shows the compute volume the company requires over time to train its models. On this chart, the axis is compressed so that a 10x and a 100x increase barely look different, yet the curve still climbs steeply. In Coxon's assessment, this slope demonstrates that algorithmic efficiency is improving at an accelerating rate rather than merely advancing quickly.
"It's an insane plot and I'm surprised they made it public,"
Coxon wrote regarding the disclosure. The mechanical reality behind that chart implies that the compute required to deliver landmark milestones drops rapidly once the architecture is established. What does this steep curve mean for the longevity of frontier systems?
It means that the capital barrier protecting frontier releases decays at speed. Anthropic's Mythos model represented a major capability jump in the spring, yet Coxon argues that Mythos was merely a single point on a superexponential trend. According to Coxon, Anthropic could already retrain Mythos with a fraction of the compute originally needed. This collapsing compute requirement transforms what was once considered an unassailable infrastructure moat into an asset that can be reproduced for significantly less capital.
The Push for a Development Pause
Some might argue that accelerated algorithmic efficiency simply lowers operational costs for everyone rather than undermining market leaders. Under that view, established labs like OpenAI and Anthropic merely benefit from cheaper compute while maintaining their scale advantages.
A pause would "commoditize human-level intelligence. It would only harm OpenAI and Anthropic."
Yet Coxon's assessment directly challenges that logic by showing that lower compute thresholds erode enterprise moats. Because algorithmic improvements are nearing complete automation, Coxon wants frontier model development paused next year so competitors can catch up. In his written assessment, such an intervention would commoditize human-level intelligence and would exclusively harm OpenAI and Anthropic.
Can OpenAI and Anthropic maintain multi-billion-dollar infrastructure valuations when competitors can replicate flagship systems like Mythos next year using only a fraction of the original compute?
What it means
Anthropic flaunted its compounding algorithmic efficiency during its pre-IPO roadshow, yet that exact curve directly destroys the economic moat around its training clusters. The moment an architecture stabilizes, fast-moving optimizations drastically reduce compute demands—letting Anthropic reproduce its spring Mythos model for a mere fraction of the original budget. While Anthropic targeted a trillion-dollar valuation, this collapse in computing requirements strips front-runners of defensibility and hands rivals an extraordinarily cheap path to frontier capabilities. Will private capital continue underwriting massive infrastructure bets for OpenAI and Anthropic once flagship systems like Mythos can be duplicated next year for a fraction of the cost?
Our stories: Anthropic Eyes $1T IPO on $65B Revenue Run-Rate (August 18).