The Forecasting Inflection
Every strategic decision is implicitly a prediction about the future. Backing a technology is a prediction about where the market is going. Entering a new geography is a prediction about regulatory and competitive dynamics. Pricing a round, timing a launch, choosing which competitor to worry about… Decision makers are always forecasting, whether they call it that or not.
For the longest time, the only way to predict the unfolding of future events was gut feeling or, to use a more elegant term, “having judgment.” Judgment was not a scientifically reliable tool, and naturally generations of economists, quants and intelligence analysts produced gradually more sophisticated mathematical models to compute what’s next. But nobody, so far, has been able to reduce the inherent complexity of any sector, much less human society as a whole, to a set of variables you could plug into a model. Whether a technology gets adopted depends on founder psychology, regulatory mood, competitive responses that haven’t happened yet, a blog post that shifts an entire community’s opinion overnight. The evidence that actually drives outcomes is overwhelmingly qualitative, unstructured, contextual, ambiguous, and you cannot coerce it into a mathematical framework.
Prediction, it turns out, is fundamentally an act of judgment: the ability to read a messy situation, weigh conflicting signals, and develop a feel for where things are heading. Some people are extraordinarily good at this. Philip Tetlock spent twenty years tracking tens of thousands of real-world predictions against outcomes and found that a small group of forecasters (working with nothing but public information) consistently outperformed intelligence analysts with classified access, pundits with credentials, and prediction markets with real money on the line. His work proved that superforecasting is a real, learnable skill, and that it depends on judgment as much as information.
But judgment doesn’t scale. There are only so many superforecasters in the world, forecasting any given question takes significant time, and there is only so much domain expertise one superforecaster can aim to possess.
Until now.
Two years ago, no AI system came close to matching human superforecasters on calibrated prediction benchmarks. One year ago, the best systems were in the middle of the pack. Today, raw frontier models have only a marginal gap with human superforecasters in ForecastBench, and it is closing fast. In November 2025, Bridgewater’s AI lab claimed parity with superforecasters. A team including Tetlock showed that twelve LLMs matched 925 human forecasters in a head-to-head tournament. On the Metaculus benchmark, a single LLM with a generic news API already rivals the collective intelligence of hundreds of human forecasters combined.
Whatever you think about LLMs, these are not the results of a stochastic parrot. To make calibrated predictions about geopolitics, technology adoption, and economic shifts, a model must have encoded something real about how the world works: cause and effect, incentive structures, competitive dynamics. And as new architectures emerge that reason more deeply about complex systems, this capability will only improve.
But today’s models forecast with nothing but a web search and a prompt. A human superforecaster can call a government agency for data that isn’t publicly indexed, read paywalled research, aggregate information across dozens of specialized sources, and build context over weeks. The gap between what models can do and what they could do with the right evidence infrastructure is enormous. Give them access to structured, continuously updated, domain-specific information, and the advantage over any individual human becomes overwhelming.
As far as we know, no lab trained these models to forecast. It is an emergent capability, which means it can be deliberately sharpened. Build the right training environments, evaluation frameworks, and feedback loops, and the judgment that appeared on its own becomes something you can systematically optimize for with every iteration.
We are at the tipping point for machine prediction, but the infrastructure to harness it doesn’t exist yet. It’s time to start building. Prediction is upstream of every strategic decision, and better prediction changes everything downstream.
The Arms Race
If prediction is upstream of every strategic decision, then better prediction is the ultimate competitive advantage. And once it becomes clear that machines can do it, there is no going back. Every executive, every strategy team, every investor and analyst will have to adopt these systems, because the alternative is to constantly make worse decisions than everyone who did. Gut feeling stops being a viable strategy the moment your competitors have something quantifiably better.
Even when frontier models inevitably match the best human superforecasters out of the box, the race for machine prediction will have only just begun. That level of performance, once widely available, becomes the new consensus. The edge is everything above it: the margin built through better evidence and sharper judgment than anyone else can access. The pressure to push further only grows, because the baseline keeps rising and the decisions riding on each incremental gain keep getting larger.
In the race for superhuman forecasting, there will be no ceiling.
This race is already underway. Bridgewater, the world’s largest hedge fund, has built an agentic forecasting system around evidence gathering and trained judgment. Most of the world has not caught on yet, but that is the nature of an inflection: by the time it is obvious, the early movers have already compounded their lead.
Having access to frontier foresight will soon stop being a competitive advantage and start being a prerequisite for making consequential decisions. The only question is who builds the infrastructure first, and in which domain they go deep.
Evidence & Judgment
Building infrastructure that can actually win this race means solving two problems at once. The best evidence in the world is useless without judgment sharp enough to extract signal from it, and the sharpest judgment is wasted if it only sees what a web search returns. Every robust result in machine forecasting so far points to the same conclusion: evidence quality and judgment quality are the two variables that determine accuracy, and they compound each other.
The evidence infrastructure is the biggest bottleneck so far. When researchers tested an agentic forecasting system with and without search, the results were stark: on live prediction markets, removing search access caused a 3.6x degradation in accuracy. But generic, non-agentic search barely moved the needle. What mattered was structured, adaptive evidence gathering: a system that knows what to look for, where to find it, and how to follow one source to the next. Researchers building competing systems have independently reached the same conclusion: the quality of the research matters more than the quality of the judgment or calibration sitting on top of it.
And yet the evidence that actually drives outcomes in most domains remains almost entirely untapped. Consider what a model with a simple web search cannot see: hiring velocity on LinkedIn that reveals which teams are scaling and which are quietly shrinking. Sentiment shifts across developer communities like GitHub discussions, Discord servers, and niche forums that signal whether a framework is gaining real traction or fading. Paywalled analyst reports and proprietary market data. Earnings call transcripts parsed not for the headline numbers but for the subtle hedging that precedes a strategy pivot. Regulatory filings across jurisdictions, patent applications, procurement records, expert interviews. These signals exist, they are continuously updating, and they are scattered across formats and sources that no general-purpose search engine knows to index. AI-native search APIs index public information faster and more intelligently than traditional search, but the evidence that drives forecasting is a fundamentally different problem. Much of it is not public at all: it must be aggregated from social platforms, generated through structured elicitation, or acquired from proprietary sources that no search engine will ever reach. And forecasting requires the ability to reconstruct what was knowable at any given point in time so that judgment can be trained on temporally-coherent signals rather than hindsight-contaminated ones. This is diametrically opposed to what modern search optimizes for, which is always surfacing the most recent information.
Assembling this kind of evidence base, one that is structured, queryable, temporally aware, and continuously expanding across both public and private sources, is an infrastructure problem that the current search landscape is not built to solve.
But evidence alone is not enough. A system that gathers ten thousand signals and cannot tell which ones matter is no better than one with ten. The system needs to learn judgment from raw information: how to map diverse, often contradictory evidence into a calibrated probability that a specific outcome will occur. This goes beyond a retrieval problem or a summarization problem. It requires a model that understands causal structure, that can weigh a strong signal from a weak source against a weak signal from a strong source, and that updates its beliefs proportionally to the actual information content of new evidence. Recent work has demonstrated that a model trained through reinforcement learning on resolved prediction outcomes can match frontier systems seven times its size, because training on real-world resolution data teaches the system which reasoning patterns actually lead to accurate predictions and which are noise. Another line of research achieved the highest score on ForecastBench by maintaining a structured belief state that tracks how each new piece of evidence should update the running probability. The lesson is consistent: judgment can be systematically sharpened, and the sharper it gets, the more it reveals about which evidence to gather next.
These two variables create a flywheel. Better evidence lets the system make more precise predictions. Resolved predictions generate training signal that sharpens judgment. Sharper judgment identifies which evidence sources actually carry information, focusing the next round of evidence gathering. The system improves on both axes simultaneously, and the gap between a system running this flywheel and one that is not widens with every cycle.
In practice, a system built on these principles takes a question and returns a calibrated probability with full provenance: what evidence was gathered, how it was weighed, and which factors pushed the estimate in each direction. Not a number, but an explained belief that can be interrogated, challenged, and trusted, one that updates continuously as new evidence arrives.
Why Technology First
Both evidence infrastructure and trained judgment are inherently domain-specific. The signals that matter for geopolitics are entirely different from those that matter for biotech or technology, and judgment trained on one does not transfer cleanly to another. Technology is the ideal first domain because it is volatile, consequential, and exceptionally rich in untapped signal. The evidence base is enormous, continuously updating, and scattered across repositories, developer communities, job postings, product launches, earnings calls, and regulatory filings that no one is systematically aggregating. Every VC, every corporate strategy team, every public market analyst pricing a technology company is implicitly forecasting adoption curves, competitive outcomes, and platform risk, and making those calls on pattern matching and intuition. And unlike most problems in AI, the returns on building this infrastructure do not diminish as the system improves.