Article

Former Anthropic researcher issues another warning: algorithmic progress is accelerating, and training more powerful AI may no longer rely on piling up computing power.

Beating AI Flash News: This September, Jacob Coxon, who previously worked on pre-training research at OpenAI and Anthropic, resigned, angrily denouncing the two companies for rushing toward superintelligence capable of self-improvement, calling it "gambling with our lives." These remarks prompted U.S. lawmakers to call for stricter regulation, while Musk even suspected he was cooperating with regulators to build public support.

Now, Coxon has posted another image from Anthropic's IPO roadshow, saying it was one of the key reasons that led him to resign. "This chart is insane—I didn't expect them to make it public."

Anthropic originally used this chart to showcase its technological advantage: as algorithms continuously improve, less and less computing power is needed to train models with equivalent coding capabilities.

However, Coxon noticed that although the vertical axis uses a logarithmic scale, the curve remains increasingly steep. This means that, if the data is accurate, the rate of improvement in training efficiency itself is also accelerating. Based on this, he believes Anthropic's algorithmic progress has entered a phase of "hyper-exponential growth."

Mythos may just be an ordinary stop along this steep curve. Anthropic can now train models at a level comparable to Mythos using far less computing power than initially required. If the same amount of computing power were invested, it might even be possible to train AI that is stronger still.

Moreover, improving training algorithms does not necessarily require brilliant researchers to constantly come up with breakthrough ideas; this kind of work is approaching a stage where it can be automated. During the roadshow, Anthropic co-founder Tom Brown also mentioned that in the future, Claude may be able to independently discover ways to improve training efficiency. These new methods could be used to train more powerful models, which in turn could continue searching for further optimization techniques.

However, the chart does not disclose specific multiples of computing power; it only shows three data points, and the positions marked below the chart are for illustrative purposes only.

Coxon argues that cutting-edge AI research should be paused in 2027 to allow competitors to catch up. This would gradually turn human-level AI capabilities into commonplace commodities available to everyone, breaking the advantage that OpenAI and Anthropic have built by relying on their leading models.