METR

Formerly ARC Evaluations.

Leaders: Beth Barnes, Chris Painter

Originally I wrote that we hoped to be able to get large funding for METR via non-traditional sources. That happened last year, and METR got major funding. That’s great news. Alas, they once again have to hit the fundraising trail.

METR has proven to be the gold standard for outside evaluations of potentially dangerous frontier model capabilities, and has proven its value even more so in 2025\.

We very much need these outside evaluations, and to give the labs every reason to use them and no excuse not to use them, and their information has been invaluable. In an ideal world the labs would be fully funding METR, but they’re not.

So this becomes a place where we can confidently invest quite a bit of capital, make a legible case for why it is a good idea, and know it will probably be well spent.

If you can direct fully ‘square’ ‘outside’ funds that need somewhere legible to go and are looking to go large? I love METR for that.

To donate, click here. They can be contacted at info@metr.org.

METR evaluates large AI models to look for potentially dangerous capabilities. Its most obvious theory of change—where it finds a scary result and then the AI company pauses development—mainly depends on (1) AI companies giving access to METR (which they often don’t) and (2) AI companies ceasing model development when METR establishes harmful capabilities (which they probably won’t—if there’s any ambiguity, they will likely choose the interpretation that lets them keep making more money).

There’s an indirect but more promising theory of change where METR demonstrates a template for capability evaluation which policy-makers then rely on to impose safety regulations. To that end, METR has engaged with NIST’s AI risk management framework (pdf). This sounds potentially promising but it’s not where I would put money on the margin because:

• I don’t think we should wait to figure out a solid evaluation framework before writing regulations.
• Evaluations are helpful if we want to conditionally pause AI in the future, but not relevant if we want to unconditionally pause AI right now, and I believe we should do the latter.