Former Anthropic researcher issues another warning: algorithmic progress is accelerating, and training stronger AI may no longer require stacking computing power.

Beating AI News Flash: In September of this year, Jacob Coxon, who previously worked on pre-training research at OpenAI and Anthropic, resigned, angrily accusing both companies of racing toward self-improving superintelligence and essentially "gambling with our lives." His remarks prompted U.S. lawmakers to call for stronger regulation, and Musk even suspected he was coordinating with regulatory propaganda efforts.

Now, Coxon has posted a chart from Anthropic's IPO roadshow, saying it was one of the key reasons that prompted his resignation. "This chart is insane. I can't believe they made it public."

Anthropic originally used this chart to showcase its technological advantage: as algorithms continue to improve, the compute required to train a model with equivalent coding capabilities keeps decreasing.

But Coxon noticed that the vertical axis in the chart already uses a logarithmic scale, yet the curve still keeps getting steeper. This means that, if the data is accurate, the rate of improvement in training efficiency is itself accelerating. Based on this, he argues that Anthropic's algorithmic progress has already shown "super-exponential growth."

Mythos may just be an ordinary stop on this steep curve. Anthropic can already train a model at the same level as Mythos using far less compute than initially required. If the same amount of compute were invested, it would be possible to train a stronger AI.

And improving training algorithms does not necessarily require genius researchers to keep producing breakthrough ideas; this kind of work is already approaching the stage where it can be automated. Anthropic co-founder Tom Brown also said during the roadshow that in the future Claude is expected to discover methods to improve training efficiency on its own. New methods can be used to train stronger models, and stronger models can in turn continue to find new optimization methods.

However, the chart does not disclose the specific compute multiples; there are only three data points, and a note below the chart states that the positions are for illustrative purposes only.

Coxon advocates pausing frontier AI research and development in 2027 to let competitors catch up. This would gradually turn human-level AI capabilities into an ordinary commodity available to everyone, breaking the advantages that OpenAI and Anthropic have built through leading models.