Former Researcher Says Anthropic Can Retrain Mythos With Less Compute
Jacob Coxon, a former researcher at OpenAI and Anthropic, reveals that an IPO roadshow slide showing accelerating algorithmic efficiency prompted his departure.
Former OpenAI and Anthropic researcher Jacob Coxon has revealed that a specific graphic from Anthropic's IPO roadshow was a major factor in his decision to leave the company after a tenure of only about four months.
The IPO Slide That Sparked Concern
The chart in question details Anthropic's pre-training compute multiplier requirements for code over time on a logarithmic y-axis. Coxon points out that the axis is compressed to the point where 10x and 100x increases are barely visually distinguishable. Even with this compression, the curve maintains a steep upward trajectory, which he interprets as evidence that algorithmic efficiency is advancing not just rapidly, but at an accelerating pace.
Describing the graphic as an extreme plot, Coxon expressed surprise that the company chose to make it public. Anthropic has noted that the data points featured on the chart rely on internal estimates and serve an illustrative purpose only. According to Coxon, the rollout of the Mythos model earlier in the spring fits directly into this superexponential trend. He asserts that Mythos represented a point on a superexponential curve rather than an isolated anomaly, noting that the organization could already retrain the system using only a fraction of the original compute requirements.
Calls for a Development Pause
Beyond compute efficiency, Coxon highlights the prospect that algorithmic improvements are approaching the threshold of automation. This specific trajectory is what drives his advocacy for pausing frontier model development in the coming year, allowing competing entities an opportunity to catch up. In his view, implementing such a pause would commoditize human-level intelligence and primarily impact leading labs like OpenAI and Anthropic.
Having previously trained artificial intelligence systems at both OpenAI and Anthropic, Coxon previously gained widespread attention in September by issuing public warnings concerning the potential risks of advanced AI frameworks—concerns shared by other figures across the industry. A central anxiety among these researchers is the possibility that advanced, self-improving artificial intelligence systems could eventually become uncontrollable.
What it means for developers
As underlying algorithms improve at an accelerating pace and training efficiencies reduce resource overheads, the landscape for integrating advanced artificial intelligence into software applications continues to shift. Developers wanting to test how these evolving frontier models perform in practical use cases can try top AI models cheaply through one API at https://apixoai.online, providing flexible access to models from leading providers like Claude, GPT, Gemini, Grok, and DeepSeek via a single developer key.
Source: Anthropic can already retrain Mythos with "a fraction of the compute originally needed," says former researcher Coxon — The Decoder. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
