Putting auto research to work on how technology and markets evolve

The trajectories of economic, social and technological systems cannot be deduced from first principles alone. Forecasting them requires building and continuously improving models of how they behave, combining quantitative methods with qualitative reasoning and evidence from many sources, including data that are difficult to obtain. Reasoning models increasingly let us automate that research.

We’re building an environment where AI agents construct and refine models of the systems they study: how demand, supply and investment interact, and what those relationships imply for future outcomes. Our first application studies the AI buildout, from chip and memory revenue to spending on data centers and computing infrastructure. With its forecasting method calibrated on outcomes through June, our system scored better than Polymarket on July–September forecasts. In a separate comparison, it was right in 11 of 13 disagreements with published analyst estimates.

---

How it works

Our system combines an ontology layer, which organizes the evidence, with an environment where AI agents build and test forecasting models. These research agents are programs that use reasoning models to inspect previous experiments and propose new ones. The forecasting models combine quantitative evidence with qualitative evidence from multimodal data, including text and video, weighting each source according to what it contributes to the forecast.

We test forecasts by replaying the past one step at a time. At each step, the system makes predictions using only information available at that date. As outcomes arrive, it scores those predictions and returns the experiment records to the research agent, which uses the errors to guide the next revision. Continuing this process as new events unfold would keep its predictive understanding current as relationships change and new trends emerge.

The research agent also identifies evidence worth collecting. In a separate study of GPU rental prices, our agents proposed:

• Reconstructing cloud price histories from software repositories.
• Tracking the availability of individual GPU models.
• Examining reports of sold-out capacity and large compute contracts to help interpret price trends.

In our experiments, agents found our ontology layer easier to work with and produced better forecasts than when using public web indexes. We built it to organize and format evidence specifically for how the agents reason and build forecasting models. It brings together quantitative time series, videos, transcripts, articles and reports, including proprietary sources and primary research, in a form the forecasting models can use together.

---

First results

The AI buildout offers a way to study how a major industry is developing while comparing our forecasts with analysts and markets answering the same questions.

We replayed historical questions using only records available at each forecast date, without access to market prices. The forecasting method was calibrated on outcomes through June 2026 and evaluated on July–September outcomes. The market comparison below excludes forecasts made in each market’s first 24 hours, leaving 189 forecasts across 104 questions. Several versions were evaluated on this test period.

Against published analyst estimates, our system called more outcomes correctly across the same 99 questions:

Our system was right in 11 of the 13 cases where it disagreed with the analyst estimate. This comparison includes both development and test cases and uses our final forecast before each release; analyst estimates come from different dates.

Selected cases where our forecast correctly differed from the published analyst estimate:

Even after allowing each market 24 hours for initial price discovery, our system scored better than Polymarket:

Our system’s advantage over Polymarket grew the further ahead it forecast, with the largest gains two to four weeks before the outcome. We beat the market when it had more uncertainty left to resolve, giving decision-makers more time to act. Both scores measure forecast quality; lower is better.

---

Beyond the AI buildout

The AI buildout is a starting point. This comparison focused on questions with a prediction market to measure against, a narrow slice of what we want to forecast. Many of the questions we want to investigate have no such market. We want to extend this approach to a much wider range of questions about how technology, economies and societies develop, especially those that receive little sustained research today.

Over time, the system could learn to choose the challenges that help it improve: finding gaps in its understanding, gathering the evidence to investigate them and designing experiments around increasingly difficult forecasts. Each result would inform what it studies next. As the ontology layer grows, it would connect knowledge across domains, giving the system more to learn from and more relationships to test against actual outcomes.