Introducing Foresight Machines
Why we are building AI that models real systems and searches for better ways to arrange them.
A power project can be worth building and still be impossible to finance on the terms proposed. The equipment arrives on one schedule, the revenue on another, and the loan expects both on time. A different sequence of construction, supply contracts or financing might make it work. Finding that sequence means understanding how the pieces move together before anyone commits.
The same problem runs through the economy. An insurer decides which risks it can carry and on what terms. A manufacturer decides how to secure supply. A financial institution decides where to hold funds so it can pay what it owes as conditions change. Each is a commitment made inside a system: resources, obligations, other participants, and events no one controls.
Institutions have strong tools for these questions. But the evidence is spread across contracts, operating systems, forecasts, specialist models and people. Connecting it is slow, and keeping it current is slower. So the alternatives a team considers are usually the ones it had time to build.
Systems also behave in ways no single part intends. A plan can win where it is measured and lose where it isn’t: the buffer that protects one business line can starve another. These effects only appear when the whole system is modeled.
Foresight Machines is building AI to model those systems and search them. A Foresight Machine represents the resources available, the obligations to be met, and the rules and dependencies that govern what happens next. For liquidity, that means payment deadlines, ownership, transfer limits and settlement times. For an infrastructure project, it means equipment capacity, construction dependencies, power availability and financing conditions.
The machine does two things with that model. It projects how the system could unfold under different conditions, using the best available forecasts as inputs. And it searches for different timing, terms, partners or allocations. It tests each candidate, finds what defeats it, and uses that failure to propose the next.
Often the most useful question is not “what will happen?” but “how do we get there?” An institution states the outcome it needs (every obligation met, the project financed, supply secured), and the machine searches for arrangements that reach it, including ones whose first step no one would have chosen. The institution sets the goal, chooses the trade-offs, and keeps authority over every commitment.
AI is making this practical. In the Summer 2026 Metaculus Cup, AI systems finished first and second among 885 forecasters. Language models and coding agents can now read contracts, extract rules and build models that once took specialists months. Systems such as AlphaEvolve show how quickly AI improves when every proposal gets an exacting test. Simulation gives search that test, and every additional unit of computation explores more futures and more candidates.
SourceMetaculus
SourceGoogle DeepMind
We are starting with the liquidity behind stablecoin redemption. An issuer can hold enough reserves overall while the money is not where it needs to be when a redemption arrives. We model how reserve placement, funding routes, service commitments and contingency terms work together, and look for changes that improve reliability, lower cost or make a new service possible. This application has recurring decisions, explicit obligations and mechanics that can be checked. It also makes one distinction concrete: a transfer may be available today, but a new credit facility must be negotiated, onboarded and proven before anyone can count on it.
Trust has to be earned against reality. Every model is built from evidence the institution can inspect. Every recommendation is compared with the institution’s current approach and strong conventional methods under the same constraints, and recorded before events unfold. Operating exercises test whether a plan can actually be carried out. When outcomes arrive, the record shows where the model was wrong, and a correction must improve results on separate cases before it is used. Simulation estimates what alternatives would produce; it does not reveal the true outcome of every path.
The longer question is how much of this machinery can travel. Liquidity, collateral and credit connect to the financing of inventory, equipment, energy and construction. These systems share obligations, resources, timing and dependencies, while each brings its own rules and evidence. We are developing methods that carry what is learned in one system to the next, and measuring what must be rebuilt.
Success means more than a better allocation. A different sequence of purchases could make a project financeable. Different supply terms could make production reliable. Better-designed cover could let an institution carry a risk it would otherwise decline.
We begin in finance because the first decisions must be specific enough to test. The ambition reaches wherever institutions commit resources to an uncertain future. Our mission is to expand what humanity can responsibly make possible.