Foresight, made a machine.
Today we are introducing Foresight Machines. We are building AI systems that model the world around an institution so it can decide which risks to take, which commitments to make, and how much capital they require.
A bank should be able to prepare for a run, an insurer to underwrite a hard-to-price specialty risk, a developer to weigh a power delay before committing capital. Our ambition is to make that foresight a permanent capability of institutions.
Every institution already has a model of the future. A quant expresses it in equations, an analyst in a forecast, an investment committee in the assumptions behind an approval. Demand will hold. Funding will remain available. A supplier will deliver. These beliefs determine what gets financed, insured, and built. An institution can understand each exposure in isolation and still miss what connects them.
A super El Niño can disrupt harvests, lift food prices, and provoke export restrictions. The AI buildout can strain power supply, delay construction, and put financing at risk. A digital dollar can carry a bank’s failure into markets that trade around the clock. Each changes the conditions under which other commitments were made.
SourcesEl Niño and agricultureFood markets and export restrictionsAI and energyCircle’s account
An institution needs a way to follow those connections and test its choices against them. The model behind a consequential decision should be something it can inspect, run, and improve.
The challenge: turn foresight into an instrument
Superforecasting showed that judgment about the future can be measured and improved. In the Summer 2026 Metaculus Cup, AI systems finished first and second overall, ahead of every human entrant. The opportunity is to bring that discipline to the decisions that allocate capital and shape the economy.
SourceMetaculus results
Prediction markets such as Kalshi aggregate beliefs into prices. Robin Hanson’s futarchy goes further, proposing that decisions be guided by markets estimating outcomes under alternative policies. Both make expectations explicit and give people a reason to improve them.
An institution must connect those expectations to its own choices. A forecast of higher interest rates is an input; deciding whether to refinance requires a model of its debt, cash flows, and alternatives. The same forecast can justify different actions for different institutions. What matters is how an event reaches the institution, what it can do in response, and how that response changes the outcome.
Weather forecasting offers a precedent. Observations feed models of the atmosphere; forecasts are checked against the weather that arrives; errors inform the next improvement. This discipline creates value wherever a better view of tomorrow changes what people do today: routing aircraft, dispatching power, planning harvests, and moving goods.
SourcesHow forecasts are madeForecasts in use
Markets and institutions introduce a further difficulty: people respond. A depositor withdraws because others might, a lender tightens terms as risk rises, a government intervenes to contain a shortage. Their responses change the conditions under which other decisions are made.
We want to bring the discipline of weather forecasting to this world of interacting decisions. Two capabilities are opening the way: maintaining models as the world changes, and improving them through a record of what they get right and wrong.
Unlock 1: Models that stay current
Much of the knowledge needed to model an institution sits in contracts, filings, operating reports, and the heads of specialists. Turning it into a working model has required painstaking translation. Keeping it current has required doing that work again.
Language models can take on more of this translation: extracting an obligation from a contract, a dependency from an operating report, a constraint from a regulation. Connected to current observations, these capabilities make it more practical to maintain a model as the world changes.
A contractual payment is an obligation; a power limit is a physical constraint; expected withdrawals are an estimate of behavior. Each needs a traceable basis. The machine must distinguish them while calculating how they interact: whether funds arrive before a payment falls due, or a power delay threatens a project’s financing.
It must preserve what was known when a decision was made, along with what changed afterward. An institution should be able to examine both its current understanding and the evidence on which an earlier commitment rested.
Together, these relationships form a shared model of the institution and the systems it depends on. Different choices can be compared against the same evidence. When an assumption changes, the machine should show which commitments need another look—and which missing information could change the decision.
Unlock 2: Learning from outcomes
A model earns trust by exposing itself to correction. Its expectations must be recorded before events unfold, its errors preserved after they do, its revisions tested before they are relied upon.
History can accelerate that process. Past events provide situations to reconstruct, expectations to test, and outcomes to learn from. We are building the machinery to recover the evidence available at each point and test a model without giving it knowledge of what followed. This would let successive versions learn from a body of experience already available, while reserving separate cases to test whether they have improved. Expectations recorded ahead of new events provide the continuing test.
For a reserve model, that means recording estimates of withdrawals, cash availability, and funding costs, then comparing them with what follows. If withdrawals repeatedly arrive faster than expected, the record should expose the error. A proposed correction must improve performance on cases that were not used to develop it.
This is what we mean by self-correcting: a process that makes mistakes visible and tests whether a change remedies them.
Judging a decision requires another distinction. We observe what followed the action taken; the effects of alternatives must be estimated. A fortunate outcome can reward a poor decision, and an unlucky one can punish a sound decision. The machine therefore needs a record across many cases, with forecast accuracy assessed separately from evidence about what its actions changed.
- Observe: evidence, with its source and time.
- Model: the institution and its dependencies.
- Compare actions: consequences under stated assumptions.
- Record expectations: the action, the model version, and what was expected. This record is never altered.
- Score expectations: against what actually happened.
- Test revisions: on separate cases. A revision that proves better becomes the new model; otherwise the model stays as it was.
Every choice leaves a record that outcomes can score. A revision changes the model only if it proves better on cases it has not seen, and the original record is never altered.
Foresight as an institutional capability
An institution should be able to carry what it learns from one decision into the next. A model built to assess a counterparty can inform a financing choice. Evidence gathered for a supply commitment can inform the next investment. As tested relationships accumulate, the institution gains a more coherent understanding of the systems it depends on.
Better foresight can change which risks an institution takes and how much capital it commits. An insurer could find terms that make a hard-to-price specialty risk worth carrying. A business could make a new market viable by changing its partner, timing, or initial investment. The machine should reveal what makes an opportunity worth pursuing—and when those conditions no longer hold.
As their record earns an institution’s trust, the machines should take responsibility within defined mandates: propose an allocation, carry it out, and monitor the conditions that justified it. When those conditions weaken, the machine should identify what needs to change. Each expansion of its authority should rest on demonstrated competence.
A shared model supports different institutional decisions.
Weather forecasting earned its place in the economy through use, measurement, and improvement. We want foresight about the systems people build to earn the same standing.
The hardest decisions are not unknowable. They are unrun.