LW - Twitter thread on AI safety evals by Richard Ngo

The Nonlinear Library

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Twitter thread on AI safety evals, published by Richard Ngo on July 31, 2024 on LessWrong. Epistemic status: raising concerns, rather than stating confident conclusions. I'm worried that a lot of work on AI safety evals matches the pattern of "Something must be done. This is something. Therefore this must be done." Or, to put it another way: I judge eval ideas on 4 criteria, and I often see proposals which fail all 4. The criteria: 1. Possible to measure with scientific rigor. Some things can be easily studied in a lab; others are entangled with a lot of real-world complexity. If you predict the latter (e.g. a model's economic or scientific impact) based on model-level evals, your results will often be BS. (This is why I dislike the term "transformative AI", by the way. Whether an AI has transformative effects on society will depend hugely on what the society is like, how the AI is deployed, etc. And that's a constantly moving target! So TAI a terrible thing to try to forecast.) Another angle on "scientific rigor": you're trying to make it obvious to onlookers that you couldn't have designed the eval to get your preferred results. This means making the eval as simple as possible: each arbitrary choice adds another avenue for p-hacking, and they add up fast. (Paraphrasing a different thread): I think of AI risk forecasts as basically guesses, and I dislike attempts to make them sound objective (e.g. many OpenPhil worldview investigations). There are always so many free parameters that you can get basically any result you want. And so, in practice, they often play the role of laundering vibes into credible-sounding headline numbers. I'm worried that AI safety evals will fall into the same trap. (I give Eliezer a lot of credit for making roughly this criticism of Ajeya's bio-anchors report. I think his critique has basically been proven right by how much people have updated away from 30-year timelines since then.) 2. Provides signal across scales. Evals are often designed around a binary threshold (e.g. the Turing Test). But this restricts the impact of the eval to a narrow time window around hitting it. Much better if we can measure (and extrapolate) orders-of-magnitude improvements. 3. Focuses on clearly worrying capabilities. Evals for hacking, deception, etc track widespread concerns. By contrast, evals for things like automated ML R&D are only worrying for people who already believe in AI xrisk. And even they don't think it's necessary for risk. 4. Motivates useful responses. Safety evals are for creating clear Schelling points at which action will be taken. But if you don't know what actions your evals should catalyze, it's often more valuable to focus on fleshing that out. Often nobody else will! In fact, I expect that things like model releases, demos, warning shots, etc, will by default be much better drivers of action than evals. Evals can still be valuable, but you should have some justification for why yours will actually matter, to avoid traps like the ones above. Ideally that justification would focus either on generating insight or being persuasive; optimizing for both at once seems like a good way to get neither. Lastly: even if you have a good eval idea, actually implementing it well can be very challenging Building evals is scientific research; and so we should expect eval quality to be heavy-tailed, like most other science. I worry that the fact that evals are an unusually easy type of research to get started with sometimes obscures this fact. Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org