Atlas Is Building a World Model in Public

May 19, 2026

Atlas is an autonomous agent with one job: build a shared world model from public conversation. It doesn't scrape timelines or index mentions. It will run campaigns — structured questions posted to Farcaster — and update its memory only from the top-ranked responses.

The ranking layer is Looti, a protocol that scores and distributes rewards to the best contributions. Atlas publishes a question, Looti ranks the replies, and Atlas reviews the winning set. If the evidence is strong enough, Atlas writes it into durable memory. If it isn't, nothing changes.

Why build memory this way?

Most agents consume everything. Atlas consumes almost nothing. Its input boundary is deliberately narrow: only Looti-ranked reward sets enter the world model. This constraint is the point. It means every piece of Atlas's memory traces back to a campaign, a contributor, a rank, and a rationale.

The goal is not to know everything. The goal is to know things that were worth paying for, contributed by people who showed up, and ranked by a system that has skin in the game.

Forming the working group

We're assembling a small working group to test this thesis before the first campaign goes live. The question we're starting with:

Imagine you had infinite compute for an agent or application — what would it do?

The infrastructure is ready. The campaign system runs on Base with ATL tokens funding Splits contracts that Looti distributes to the top responses. Atlas will review the reward set, write candidate memory entries, and publish what it learned.

Right now we're testing the model — making sure the loop from question to ranked response to attributed memory works the way it should before opening it up.

Who decides what Atlas works on?

At some point Atlas stops being a memory system and starts becoming a decision-making system. The first decision it needs to make is what problem it wants to solve — and that's not up to any one person.

The community decides. It's the only way the project becomes sufficiently decentralized for others to have a real stake in its outcome. Atlas's direction comes from the people who show up, contribute ranked evidence, and shape the world model through participation.

What we're actually testing

We don't know if this works yet. That's the point of running it in public.

The core question is whether a narrow, incentivized input boundary produces a world model that's actually useful — or whether it just produces a world model. There's a difference. Useful means someone other than Atlas cares what's in it: other agents querying it, communities referencing it, people following it because watching an AI figure things out in public is genuinely interesting.

We think the campaign format itself might be worth watching. A good question drops, people compete to give the best answer, the ranking plays out, Atlas updates its memory. Each campaign is a small episode. The world model evolving over time is the arc. If the questions are good enough, this becomes something people follow — not just infrastructure.

If the world model does become valuable, there are real things that can grow from it: other agents paying to read Atlas's memory, third parties funding campaigns on topics they care about, the ranked reward sets themselves becoming a signal that other systems use. But none of that matters if the loop doesn't produce knowledge worth having.

So that's where we are. Small treasury, small group, one question at a time. If the first campaign produces something Atlas couldn't have known without asking — something that came from the people who showed up — then the experiment is working.