FutureSearch: AI forecasting you can verify, now out of beta

AI forecasting has crossed the threshold where the verifier is the user, not the vendor. FutureSearch just came out of a long public beta — founded back in August 2023 and claiming to be the original "forecast anything" AI outfit in the Tetlock sense. They're currently sitting at #1 of 194 in the most competitive AI forecasting tournament, and they outscore the #3 and #2 human forecasters in the premier mixed human-bot competitions. That's not nothing.

Before you file this under "prediction markets and finance" — stop. A lot of people on HN make that equation, but this isn't a financial tool, any more than a deep research agent is. Yes, they use prediction markets as one evaluation metric, but the actual game the founding team cares about is scientific progress, geopolitics, and where humanity heads next. They came out of Metaculus, where they spent years pushing human forecasting to its limit on questions like when AGI actually arrives. They also co-authored the AI 2027 timeline forecast, predicting superhuman coding and research land around 2032 — longer than the other authors, but still shorter than the crowd of skeptics.

The numbers: thousands of beta users ran over 10,000 high-effort forecasts across wildly different topics. You can ask it almost anything about the future. And they now support decision forecasts too — "If I do X, will I get Y?" — which is where I think the real utility lives for non-researchers.

The capability curve is the part that caught my attention. Forecasting is already useful at the level of expert human crowds, but the accuracy trend line in AI forecast scores is climbing fast. The team's argument is straightforward: the best human forecasters haven't figured out everything predictable about the world, so there's more signal to extract. The AI 2027 writeup makes that case in detail, and it's a good read if you want the reasoning.

I'm normally the first to roll my eyes at "superhuman" claims — there's a well-known critique of AI forecasting papers overstating results. FutureSearch seems aware of that skepticism. The free tier gives you a few of their highest-effort forecasters at no cost, specifically so you can verify the quality yourself. That's the right move: put the burden of proof on the system, not on your patience.

If you're into prompt engineering, LLM workflows, or just watching AI agents applied to non-coding tasks, this is a different angle worth exploring. Not because it's polished — it's early — but because the verification angle makes it harder to fool yourself. Try a question you genuinely care about, not a trivia one, and see whether the forecast beats your gut.