Anthropic’s research progress has roughly doubled, with AI accounting for most of the increase, according to an outside assessment discussed by Encode strategy chief Nathan Calvin. The finding raises a useful question: if AI is helping to accelerate AI research, how much evidence should companies share about the process?
Anthropic Watch analysis
What happened
The assessment, published on 5 November after a team of six spent about three months with access to Anthropic’s San Francisco headquarters, internal Slack channels and unreleased models, reportedly found that the company’s technical progress was running at twice its historical average. After accounting for ordinary growth in computing and staff, it put the AI-driven acceleration at 1.7 times.
Calvin discussed the findings on the MTS podcast, in an account published by BigGo Finance on 15 November. He said the outside researchers rejected the explanation that a single unusually successful training run accounted for the change. The account also says Anthropic engineers described their individual coding output rising by as much as twelvefold, while the company trained three unreleased models in six months. Calvin’s suggestion that AI may be helping to improve research processes is his own inference, not a stated finding of the assessment.
Why it matters
A sharp rise in individual coding output did not translate into a twelvefold increase in overall research progress. That gap matters: faster work by individual researchers is not the same thing as an equally fast jump in a lab’s capabilities.
The account says the assessment found no visible slowdown or significant bottlenecks. But it also says much of the underlying evidence was withheld. Calvin argues that competitive concerns help explain the limits on disclosure, while questioning whether those limits leave outsiders with enough evidence to judge the risks or make a case for policy action.
Our read
This is a consequential claim about the pace of AI development, and it deserves more than a passing mention. It is also a claim relayed through a podcast discussion and BigGo’s account, not a licence to treat every interpretation as established fact. In particular, the proposed explanation involving AI-assisted research remains Calvin’s speculation. The reader’s practical takeaway is to separate the reported measurements from the debate about what caused them.
What to watch
- Whether Anthropic or the assessment’s authors release more of the underlying evidence.
- Whether further outside assessments find similar acceleration at other AI labs.
- Whether policymakers pursue proposals for disclosure of how companies allocate computing resources.
Discussion spark: Should frontier AI labs disclose more evidence about how AI is accelerating their own research, even if that could help competitors catch up?
Sources and evidence
Anthropic Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by Anthropic.