Anthropic's AI Development Speed Doubled in One Year
An independent review finds the company is improving its models at twice the rate of a year prior, a finding that contrasts with CEO Dario Amodei's public calls for the industry to slow down.
By Bianca Herrera ·
Anthropic is developing its artificial intelligence systems at approximately twice the speed of its progress from a year ago, according to an analysis released Thursday. The findings come from Model Evaluation and Threat Research, an independent nonprofit based in Berkeley, California. The organization reviewed internal company data and interviewed staff to assess the pace of technical advancement.
The report was released as Anthropic prepares for an initial public offering later this month. The listing is expected to be the largest stock debut in history. The timing places the company's rapid development under the spotlight just before it enters the public market.
The acceleration occurs despite recent public statements by Anthropic CEO Dario Amodei. In an essay published on Sept. 12, Amodei urged the AI industry to slow its work to allow safety measures to catch up. He specifically called for a reduced pace in improving model capabilities. His comments were made in response to growing concerns about the potential for AI models to elude human control.
Sam McCandlish, an Anthropic co-founder and chief architect, responded to the report on the social platform X on Thursday. He stated that many of the group's conclusions align with his own observations. McCandlish noted that while AI is speeding up research, it is difficult to quantify the exact extent of that acceleration. He confirmed that the company plans to release its own report on the speed of internal AI progress within the coming weeks.
McCandlish explained that the collaboration with the research group began well before the current evaluation period. He described the work as separate from the embedded third-party evaluators that Anthropic has committed to include as part of its pacing strategy. He called the external assessment a good illustration of why independent evaluators are valuable.
The research group estimated that the company is improving model capabilities for well-specified tasks at double the previous rate. The analysis was based on internal data provided by Anthropic over a period of about 11 weeks, from mid-August through late October. The team also conducted employee interviews during this window.
Researchers from the group attributed much of the faster development to the use of AI itself in the improvement process. They ruled out an increase in computing power as the primary driver. The report noted that there are no obvious bottlenecks, such as restrictions on data or hardware, holding back progress. The group warned that using AI to improve AI could lead to compounding acceleration.
The report noted that Anthropic's public risk report, published in August, understated the level of acceleration occurring internally. McCandlish acknowledged that company redactions made it harder for others to verify the group's work. He said the company decided not to share exactly what is driving the speedup because it is commercially sensitive and could provide a roadmap for competitors.
In his September essay, Amodei highlighted the concept of recursive self-improvement. He defined this as the growing ability of AI to build the next generation of AI. He warned that if left unchecked, this process could outrun the ability to understand and control these systems. OpenAI CEO Sam Altman supported the idea of independent evaluators with employee-like access at the time of Amodei's essay.
Anthropic is based in San Francisco. The research group is funded through donations and does not accept compensation for its evaluations. Partner companies provide access to data and tokens for use in research and engineering. The next step for Anthropic is its public stock offering, scheduled for later this month.
