Machine
Foresight*
Run the world forward before you decide.
State T₀ Futures T₁ → Tₙ
Foresight Machines maintains a live state of the world, runs consequential decisions through possible futures, and learns from what actually happens.
* The future, as an input.
Every decision contains a model of the future.
An investment memo, a credit file, an underwriting book, a board strategy — each is a model of a world that does not exist yet. Today that model is static: it disappears into prose, spreadsheets and committee judgment. Software learned to answer. It never learned to anticipate.
The decision
Acquire a 1.8 GW data-center portfolio at $1.8B?
What it is quietly betting on
- Power availability
- Financing conditions
- Hyperscaler capex
- AI utilization
- Construction completion
- Local permitting
- Natural gas prices
- Grid additions
- Lease renewal rates
- Terminal cap rates
A handful of these carry most of the downside. The machine’s job is to find which — put a number on each, and commit those numbers before you sign.
So we build a different kind of computational object — a foresight machine — with five properties:
- 01 State A persistent, time-indexed belief about what is true now.
- 02 Transition A model of how that state can change.
- 03 Intervention An action, an assumption, or a shock — introduced by you.
- 04 Distribution Many futures, weighted. Never one answer.
- 05 Resolution Claims committed before the outcome. The machine learns from what reality does.
Foresight Machines builds foresight machines.
Made before
the answer was known.
Every claim the Machine commits is cryptographically sealed before the outcome is known — hashed, timestamped, anchored beyond quiet edits. When reality resolves it, the seal opens, the score posts, and the Machine learns.
The bar is calibration: when the Machine says 70%, about 70% of those events should happen. The Record is where anyone can check.
A decision should not disappear after it is made.
The Record is the
training loop.
The next token is not the next state.
A language model continues a sequence. A foresight machine advances a state — under the action you are contemplating, against the evidence it can see.
A language model
P( tokent+1 | tokens )
A foresight machine
P( St+1…t+n | St, A, E )
S — world state · A — contemplated action · E — evidence
Decisions you can run.
Fifteen questions from five domains — each one a decision, an assumption, and a horizon. Bring your own.
We start where uncertainty becomes capital allocation. The same machine extends to insurance, real assets, energy and national security.
Systems that learn from time.
In 1654, Pascal and Fermat traded letters about dividing the stakes of an unfinished game — and began the mathematics of probability. The unfinished game is now the world.
Project 1654 is our residency for researchers and builders whose machines make claims that time can test. Fellows get a salary, compute, and the live state and the Record as experimental infrastructure — and every project commits at least one claim.
Nothing at Foresight Machines goes ungraded.
Apply to Project 1654