This is a special post for quick takes by StefanHex. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.
I am surprised that people are updating so strongly on the METR AI R&D acceleration assessment.
My impression is that the most convincing evidence is the fact that the authors are convinced of their conclusions (because most of the object-level evidence is redacted).
How should I model Anthropic's influence on this? People keep saying we should model Anthropic more as a corporation with normal incentives, and having an IPO this week seems like a huge pressure on Anthropic-the-corporation. I guess Anthropic's valuation substantially depends on AI R&D acceleration and RSI, and that a third party's assessment has some influence here.
I expect that the METR team did the best possible work here, and I expect that individual Anthropic employees probably acted in good faith. But I am worried that Anthropic-the-corporation still had an unknown amount of influence, such as what information to grant, who to assign to work with METR[1], or what to redact.
In the report itself I am worried by statements that express METR's conclusions changing multiple times. To be clear, changing one's mind is good and expected. But I'm worried that a party with unilateral information control (Anthropic-the-corporation) can therefore influence the conclusions (e.g. naively: "stop giving new info as soon as METR is convinced").
However, our understanding of the status of different bottlenecks changed several times as we received more detailed information, so we are not confident in our current assessment.
Over the course of our exercise, we substantially changed our interpretation of key data as more context was shared with us, in some cases leading to multiple reversals on critical questions. We believe fuller access would have dramatically sped up that process of proper interpretation, and we think there likely remain points which we may have misinterpreted due to insufficient context.
I'd feel more comfortable with calling for action and amplifying what people are saying about short timelines & RSI if I were less worried about the above. (For example, if there were more-convincing object-level evidence that could safely be published.)
Again, to be clear, I am not blaming the METR employees for this (or even the Anthropic employees). The cause of concern is the complicated epistemic environment and corporate incentives.
METR's adversarial robustness section clarifies that they chose many of their interview subjects, which I like, but it's not clear whether this applies to their two primary contacts. ↩︎
But I'm worried that a party with unilateral information control (Anthropic-the-corporation) can therefore influence the conclusions (e.g. naively: "stop giving new info as soon as METR is convinced").