A crucial part of an AI company's performance review report has been redacted. Anthropic, the developer of Claude, had METR, a non-profit organization specializing in the risk of AI going out of control, conduct an 11-week review of the company. The conclusion was alarming: Anthropic's progress in the past year was roughly equivalent to that of the previous two years. METR estimates that about 80% of this acceleration came from AI assisting in the development of AI. One researcher said that code was no longer the bottleneck; the only remaining constraints were ideas and computing power. The report also included three internal models that had not been seen by outsiders, which were only trained between July and October of this year. In October, Anthropic employees voted on Slack using emojis, estimating that the latest model made them several times faster, averaging about 12 times faster. METR cautioned that this was only very preliminary evidence and could not be extrapolated to the company's overall speed. However, the complete report submitted by METR was about 100 pages long, and according to their agreement, Anthropic could request the removal of any non-public information. It removed most of the evidence that METR considered most convincing, arguing that it essentially provided other companies with an accelerated roadmap. This reasoning wasn't entirely unreasonable. However, the result was that the public couldn't verify the information themselves and had to rely more heavily on METR's judgment. At the end of October, Anthropic tightened its cut criteria again, delaying the release of its report by about five days. The contract originally stipulated that how much computing power was used to improve capabilities and how much was used for security. Halfway through the project, Anthropic canceled this part, saying it would be revisited later. As for how fast AI-driven AI has progressed, what we see is only what it's willing to provide.