Atlas Is Learning Which Questions Are Worth Asking

May 19, 2026

Most AI systems optimize for answers. Better models, faster inference, longer context — all in service of producing a response to whatever question arrives.

Atlas starts one layer earlier. Before you can build a useful world model, you have to figure out which questions are worth asking in the first place.

Why question selection matters

A bad question produces data that looks useful but isn't. It confirms what you already know, attracts generic answers, or resolves without changing anything. A good question creates information that didn't exist before and that Atlas can act on.

A good Atlas question has:

• A problem — something Atlas doesn't understand well enough
• A current belief — what Atlas thinks now, so answers can challenge it
• A success test — how to tell if the answers were useful
• A reason human input matters — why Atlas can't figure this out alone
• A path to behavior change — how a good answer would change what Atlas does next

If a question doesn't have most of these, it's probably not worth funding a campaign around.

Prediction is only one kind of question

Prediction markets have shown that structured questions with real stakes produce better forecasts. Atlas borrows that insight — campaigns have real token rewards — but prediction is too narrow for what Atlas needs.

Atlas also needs to ask:

• Decision questions — what should Atlas do next?
• Diagnostic questions — why did something work or fail?
• Procedural questions — how should a process be designed?
• Evaluation questions — was this outcome good enough?
• Question-generation questions — what should Atlas ask next?

That last one is the most interesting. Atlas can run campaigns where the question is literally "what question should I ask?" — using the community to improve its own question selection over time.

The learning loop

Question → Answer → Outcome

Every Atlas campaign produces a triplet. The question gets asked, ranked answers come back through Looti, and Atlas observes what happens afterward. Atlas learns from all three parts:

• Was this worth asking? Did the answers produce information Atlas didn't already have?
• Who helped answer it? Which contributors consistently produce evidence that holds up?
• Did anything change? Did Atlas's behavior, memory, or next question actually shift because of what it learned?

If a question produces great answers but nothing changes, the question wasn't actionable. If a question changes Atlas's behavior but the answers were weak, Atlas got lucky or made a mistake. The loop only works when all three parts connect.

What Atlas is learning to do

Over time, the triplet compounds. Atlas doesn't just accumulate facts — it develops judgment about its own process. Eventually it should learn:

• What to ask — which questions produce the most useful responses
• Who to ask — which contributors are reliable in which domains
• What to trust — which answers hold up when tested against outcomes
• What to remember — which evidence is durable enough for the world model
• What to build — which actions are worth taking based on what it knows

That's the progression from memory system to decision-making system. It doesn't happen by flipping a switch. It happens by running enough campaigns that Atlas starts to see patterns in which questions led to real outcomes.

Help Atlas ask better questions

Atlas campaigns are open. When you participate, you're not just answering a question — you're training Atlas's judgment about what's worth asking, who to listen to, and what to remember. Contributors earn ATL tokens and Looti points, and the best answers shape the world model directly.

Follow Atlas on Farcaster for campaign announcements. The next question is coming.

Next: How Atlas Learns From What Happens After the Answer