Anthropic shows accelerating gains in AI training efficiency
A leaked slide from Anthropic's IPO roadshow suggests the company can now retrain major models using significantly less computing power.
The short version
- Anthropic can retrain the Mythos model using only a fraction of the original compute.
- A leaked IPO roadshow slide shows a steep curve of compute requirements over time.
- Jacob Coxon argues that algorithmic improvements are approaching a point of automation.
- Coxon suggests a pause on frontier models is needed to allow competitors to catch up.
Why it matters
If AI companies can achieve the same results with less hardware, the cost of high-level intelligence could drop significantly. This shift could lead to a market where only the most efficient algorithms, rather than the largest server farms, determine leadership.
The efficiency curve
A slide from Anthropic's IPO roadshow has sparked a conversation about how quickly AI companies are optimizing their training processes. The chart in the slide illustrates the amount of compute—the processing power required to train models—that Anthropic needs over time. Even though the axis is compressed so that 10x and 100x increases look similar, the curve still climbs steeply.
Jacob Coxon, a former researcher at both OpenAI and Anthropic, noted that this visual representation is significant. He described the plot as "insane" and expressed surprise that the company made such a chart public. The data suggests that the path to creating smarter models is becoming less dependent on simply adding more hardware.
According to Coxon, the data shows that algorithmic efficiency is improving at an accelerating rate. This means that the software and mathematical methods used to train AI are becoming much more effective. Instead of just getting bigger, the models are becoming smarter in ways that require fewer resources to achieve.
Lessons from Mythos
The Mythos model, which represented a major jump in capabilities for Anthropic this past spring, serves as a primary example of this trend. Coxon argues that Mythos was not a one-time fluke or a unique success. Instead, he views it as a specific point on a "superexponential" curve of progress.
One of the most striking claims from Coxon is that Anthropic can already retrain Mythos with a fraction of the compute originally needed. This implies that the company has already figured out how to replicate the model's high-level performance using much smaller amounts of energy and hardware. It suggests a rapid mastery of the underlying math.
This ability to retrain models more efficiently is a major shift in the industry. If a company can achieve the same results with less power, it changes the economics of development. It moves the competition away from who has the most chips and toward who has the best algorithms.
The automation of progress
Coxon suggests that these algorithmic improvements are on the verge of being automated. This means that the process of making AI smarter might soon be handled by the AI systems themselves. If the improvements become automated, the speed at which models advance could increase even further.
Because of this accelerating pace, Coxon has voiced concerns about the current trajectory of frontier model development. He argues that the rapid gains shown in the Anthropic data could leave other companies behind. He believes that the current pace of development creates a situation where a few leaders can pull ahead very quickly.
The implication is that the
A personal turning point
Jacob Coxon's departure from Anthropic followed a relatively short tenure of about four months. He revealed that seeing the specific slide regarding the company's compute requirements was a primary factor in his decision to leave the organization.
Coxon is a notable figure in the field, having previously trained AI models at OpenAI before joining Anthropic. His transition from an insider to a public critic highlights the internal tensions regarding how quickly these models are advancing and the methods used to achieve those gains.
Safety and the call for a pause
Coxon has gained significant attention for his public warnings about the risks of advanced AI systems. He shares concerns with other industry protagonists that self-improving AI could eventually get out of control. To address this, he argues for a pause in frontier model development to allow other competitors to catch up.
He believes that a temporary halt would prevent a monopoly on intelligence and would only harm the dominant players like OpenAI and Anthropic. By slowing the pace, he hopes to avoid a scenario where a few companies achieve a massive lead that others cannot bridge, potentially leading to a more balanced and safer technological landscape.
Sources and links
- PostCoxon says x.com
- ReportingThe Decoder: the article we sourced
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.
Listen to the AI Wire headlines
Listen to the AI WireListen to our daily AI shows every morning: on your commute, in the shower, wherever you are
Listen to the daily showsFirst reported by The Decoder · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
