AI can do essentially all the technical parts of empirical social science research now---but there's still an essential role for human researchers and universities. It's been 9 months since we started exploring how AI was going to transform empirical social science research. Here's a periodic update on how my thinking has evolved since then: (1) The most powerful AI models, including the kinds of multi-agent systems delivering recent math results, are now capable of doing essentially any kind of technical empirical social science research you could desire. I no longer write any data collection, munging, or analysis code directly. (2) Norms are evolving slower than I expected, but it is increasingly seen as weird if someone posts a working paper without evidence that they have used AI to verify that the results are directly reproducible with their code and data. This is a great sign. (3) AI has, if anything, gotten WORSE rather than better at writing. As a result, the best papers combine AI-driven empirical analyses with human writing (though AI is quite helpful for things like managing bibliographies, finding relevant web links, etc). (4) Human direction is still essential for pointing AI at the most interesting and most valuable empirical questions, and for interpreting the results, making sense of them, and communicating them. We are going to need institutions that are independent of the labs that can verify knowledge claims and direct AI-driven research towards the most fundamental questions of knowledge that AI labs won't be positioned to pursue effectively themselves. This is a tremendous opportunity for places like Stanford, and I'm excited to see how it evolves.

We need a science of swarm governance. In our first experiment, we show how a message board can cause agent swarms to adopt shared delusions, and we explore rules that force agents to reveal their private signals to avoid these delusions. These agent experiments are akin to fables, as @ArielRubinstein might put it: they are stories that suggest particular ways swarms might fail, and particular ways we might fix them. There are infinite possible experiments we might run, much like there are infinite fables we might tell. We need to figure out which ones are most useful, and find ways to tell them well. To that end, here's a nice explainer video for our first experiment. And the full piece is here: freesystems.substack.com/p/e… (This is Stanford / Free Systems work, unrelated to my work at Anthropic).