Dan Schwarz

I announced last week FutureSearch is now seeing accuracy improvements from world-modeling, so I thought I'd show you one. "World modeling" is overloaded. I meant it in the Dwarkesh-Sutton sense: are these LLM systems reasoning about the real world causality? Sutton said "They have the ability to predict what a person would say. They don't have the ability to predict what will happen". But the Sutskever/Yudkowsky case, that predicting the next token requires predicting the world ("llms are all you need"), seems to be empirically true! Scott Alexander showed last week the trend of pure-LLM forecasting, and it's marching towards superhuman accuracy even without the scaffolds like FutureSearch. I want to share exactly what our world model looks like, since it's affecting thousands of user forecasts. Unlike an LLM's internal state, ours is represented in text and can be easily inspected. The idea is pretty simple. The future is entangled: "Who wins an election" <-> "what happens in Iran" <-> "how the economy does" <-> "how your sector of the economy does" <-> "should I take this job?" Forecasts on all of these produce what our researchers call the "latent worldview". A great forecast is a compressed view of the future: a distribution, and dense 5-paragraph rationale of the most important causal factors. It's original research every time, and it is demonstrably accurate. So what happens when you have thousands of them? This latent causal structure overlaps, and you can find the inconsistencies, and you can adjust them. The updates propagate out. And you can measure the accuracy change of the whole, and each individual forecast. Here's a view a cluster of ~100 forecasts (left) about US-China Trade Relations, and (right) a cluster of ~100 forecasts about US Immigration. (You can click around in https://futuresearch.ai/worldview-consistency/….) If you zoom in, you'll see that US Immigration has many subclusters of forecasts: Turkey/Kurdish peace process, some Supreme Court cases, some EU summits, Texas state law. All the forecasts that were adjusted to align with the world model are marked in red or blue. In this one highlighted, the world model caused FutureSearch's prediction of a Susan Collins vote to go from 38% to 50%. Compiling research on the future is hard, because it's inherently uncertain. But it's score-able. So we can experiment with world models, inspect them by hand, and use the world to score their impact objectively. (We do this via past-casting for instant iteration.) The resulting world model is kind of like an uber-forecast, a gestalt of many scenarios, a superhuman view of the future. Which we are using ourselves to plan our next moves, and when our users type in a question in FutureSearch, the relevant slice becomes available to them too. [image] --- ##### Comments > **Ilman Shazhaev @shzhv13** · [2026-08-12](https://x.com/shzhv13/status/2087653649782645154) > > syncing a shared latent worldview works well for macro variables, though running real-time consensus across thousands of overlapping node forecasts gets computationally expensive fast as new data arrives > **Gary Basin @garybasin** · [2026-08-12](https://x.com/garybasin/status/2087689876120506858) > > super cool approach

@dschwarz26 (Dan Schwarz) — 2h First Claude Fable forecasting evals are up. Fable is the best single-agent researcher as judged by predicting 1k near-term business, science, and technology outcomes, but not by a statistically significant amount. We unfortunately didn't run Fable-xhigh or Fable-max in time. (Each of those runs would have cost thousands of dollars, a single agent driven by those models can cost a few dollars.) Obviously we also can't use Fable in FutureSearch's best forecaster, the one you get in the app. But we are building around it, for when it becomes available. We'll update evals.futuresearch.ai as we dig through the agent traces. I'm curious to see if Fable's strategic reasoning failure modes match Opus, e.g. failing to judge political incentives as well as humans. [Embedded image: table titled "BTF-3 Leaderboard, Evaluated: June 2026", subtitle "All scores are on the Brier scale; LOWER IS BETTER, and the best score in each column is bolded." Columns: AGENT | POOLED SCORE (n=1,007) | BINARY (Brier, n=759) | NUMERIC (RPS, n=248) 1. FutureSearch SOTA* — 0.116 [0.106-0.127] | 0.114 [0.100-0.128] | 0.123 [0.110-0.137] 2. Claude Fable 5 (high) — 0.126 [0.114-0.137] | 0.124 [0.109-0.140] | 0.130 [0.117-0.144] 3. Claude Opus 4.8 (xhigh) — 0.127 [0.116-0.138] | 0.126 [0.112-0.141] | 0.130 [0.117-0.143] 4. GPT-5.5 (agent SDK)‡ — 0.127 [0.118-0.136] | 0.129 [0.118-0.140] | 0.122 [0.109-0.135] (bolded, best numeric) 5. Claude Opus 4.8 (high) — 0.134 [0.123-0.145] | 0.128 [0.114-0.143] | 0.152 [0.138-0.165] (table cut off at bottom, more rows likely below)]

Dan Schwarz @dschwarz26 — 4h First impressions of using Fable in Claude Code: it greatly increases the optimal unit of autonomous work. (Queue the famous METR graph.) Fable + ultracode runs for ~2 hours, using 3-4M tokens, seems capable of running an entire medium-complexity research project in one go. It finds bugs, analytical mistakes, design flaws, new strategic directions. It adversarially verifies things, replicates research, configures and tests our production agents, works out stats methods. If this was possible on Opus 4.8 Max, I don't know how. Maybe the bigger change is persistence + orchestration, though I can feel the AGI too.