This just came out from the Center for AI Safety called Superhuman Automated Forecasting. This is very exciting to me, because I've been saying this was possible for quite some time. It could probably be done even better and by next year it will be much better, because I believe at some point the training cutoff date will be eliminated through ideas like Mixture of a Million Experts, so it won't have to rely on Web searches. And the reasoning capabilities will be much better in the models coming out by end of year or next year.
Combine this idea with autonomous agents who I believe can do simulated market research from traditional questionnaire research to focus groups and the costs of market research plummets in addition to getting real-time insights. Right now scaling autonomous agents up to a large number can be difficult without them losing sight of their goals, but that will soon change and the number of agents will scale from the tens to the millions - all built from sampled detailed data.
Here is another paper on the topic that I need to dig in on called "Approaching Human-Level Forecasting with Language Models."
Oh, and CAIS is calling it "539", which is hilarious.
