crossposted from https://bengoldhaber.substack.com/p/are-there-examples-of-communities

Many of our biggest problems are collective reasoning problems. What policy are we going to adopt, what should our company build next, where are we all going to go eat lunch today? If we can use AI to reason better collectively, perhaps we can solve these (I fear that planning a lunch for over ten people is AGI complete)¹.

Given that I’m interested in using AI to improve our epistemics, in particular improving our collective epistemics, I’m looking for places, particularly communities, that have started to adopt AI to improve how they know the world.

A great example would be a community where AI is embedded deeply into the way the group “knows things” or “makes decisions”. For instance:

• A discord server where people are using an LLM to adjudicate disputes.
• A group of people that automatically add an LLM generated example into the mix when considering options.
• A forum where bots steelman arguments and add evidence.

Two almost examples:

• One example that is not quite a community - X.com. I wrote about the way users on X are turning to Grok as an oracle to resolve factual disagreements and to dunk on people who are wrong. Community notes, which is very popular and seemingly trusted, used a bridging algorithm to bring in perspectives from both sides, and now includes bots that write AI notes. It’s probably the best example of a trusted AI system improving an epistemic platform. But this is a platform as opposed to a specific community.
• Another, kind of almost example is the way that Tyler Cowen, SSC, and I’m sure others are linking to deep research reports to provide fact posts/basis for some of their claims (example from the TC<>SSC USAID Debate). This feels like an emergent community behavior insofar as you think there is still a blogosphere.

But I don’t have many examples of communities that have adopted them. Communities are an interesting abstraction layer, in my opinion, when looking for adoption, because they are less prone to being “fake”.  Communities, with most of us not being paid to participate in them, are more sensitive to whether the thing actually helps.

My best bet on where examples might be are in the forecasting community. I.e., active groups of forecasters on Manifold, Metaculus, Polymarket. In part because the groups are already organized around mechanism design, and in part they’re nearby to AI builders which makes them likely early adopters.

I’m most interested in examples that are not AI related. I’d be more impressed by an example of AI improving epistemics in a research community of geologists than one of AI researchers, because I worry the fan boy effect of being excited about AI doesn’t select well for actual usefulness.

1

I hope to build a “Where are We Going to Go to Lunch” benchmark, which will be the gold standard for whether your tool for thought actually solves a real coordination challenge AND people want to use this. If anyone actually wants to build this I’d probably pay you some number of dollars.