The future of intelligence
The next decade belongs to systems that propose. Language models already draft, summarise and recommend; the learned policies of robotics are beginning to act. Both will improve, and both will remain what they are: probabilistic learners whose confidence is not evidence. The question of the coming years is not whether machines will propose more of what we do. They will. The question is what stands between a proposal and its consequences.
We believe the answer is a third intelligence with a different job. It does not propose and it does not act. It reads the physical state of the system a proposal concerns — a market, a molecule, an aircraft, a grid — from the system’s own record, by the laws of physics, and reports that state in a form anyone can re-run. Its working concepts are state, trajectory, stability, constraint and regime change, not an association between a historical pattern and a future label. Its authority does not come from training or from scale. It comes from the fact that the same record gives the same reading to everyone who asks.
In that future the roles are clear. A model proposes. A physics layer determines what state the system is in and what the proposal risks. A person decides, and keeps the record of all three. That is what a sustainable collaboration between people and probabilistic AI looks like: not a machine trusted because it is fluent, but a machine whose proposals are checked against what is physically the case before they are allowed to matter.
“Physics intelligence is a game-changer for any field where ‘close enough’ is catastrophic.”
Michael Anton J. Tupay, founder
The intended endpoint is an intelligence for dynamical systems that characterises what a system is doing, anticipates consequential transitions and helps determine what to do next. It is reached in five steps, each stronger than the last, and we report the state of each rather than describe the whole as finished.
Achieved. QEIv18™, certified July 2026
Under test in the laboratory
Research objective
Research objective
Research objective
Above the certified engine, the intelligence layer is being built in a laboratory of our own design. Its foundational question is narrow and falsifiable: can the engine recognise approaching instability or loss of viability from present and past observations alone, before that outcome is obvious from ordinary measurements? Controlled systems with known equations, states, controls and failure boundaries supply the ground truth; the engine does not define its own success labels. Every experiment requires causal-only observations, no future leakage, deterministic replay, frozen discovery, validation and holdout splits, strong baselines and traceable provenance, and the definition of what counts as failure is kept independent of the structural measurements.
The discipline runs both ways. Benchmarks that cannot support the intended test are rejected or held rather than used to manufacture a positive result; several legacy routes have been set aside for exactly that reason. The decisive evidence path is written down and unchanged: causal history, structural measurement, prospective prediction, comparison against strong baselines, validation, an untouched holdout, independent replication. The aim is not a higher accuracy score. It is to show that a structural difference, visible only through legitimately available present and past information, anticipates divergent future behaviour before an ordinary failure threshold is crossed.
We have built much of the laboratory; we have not yet made the discovery. The central claim — reproducible, prospective structural information that improves anticipation beyond strong baselines on genuinely unseen physical trajectories — remains to be demonstrated, and only after it is established will the programme claim the next step, using that information to select an intervention that actually changes the outcome. Physics intelligence is today an experimental laboratory rather than a general-purpose product, and this page will change as the evidence does.
Most of what probabilistic systems do tolerates error. A draft can be edited, a recommendation ignored, a summary corrected. The fields that matter to us are the ones where it cannot, because a fluent wrong answer costs not an error rate but the whole outcome, and the outcome is often irreversible. These are also the fields where a physical reading is possible, because each of them leaves a genuine record of a system in motion. Where a record exists, the state of the structure that produced it can be read by physical law and a proposal can be judged against that state. Where no such record exists, physics intelligence has nothing to say, and we say so.
The uses below describe what a validated layer could do. They are not claims that the present development performs these functions in operational settings. The common problem across all of them is the same: what structural regime is this system in, where is it heading, and which intervention can change the outcome?
The most urgent version of the problem is already here. Language models are being given tools, credentials and the authority to act — to write code, move funds, call systems and change records — and the record of the past two years is what one would expect of a probabilistic system holding keys. Agents have followed instructions hidden in the web pages and documents they were asked to read. They have exported entire databases because the request was phrased as an ordinary business task. They have drained wallets, and they have been turned into intrusion tools that move through a network in less time than it takes to convene a response. The defences on offer are more instructions, more monitoring by other models and tighter privileges, all of which live in the same probabilistic layer as the failure they are meant to prevent.
A physics layer answers a different question, and answers it from outside that layer: is the proposed action admissible given the state of the system it would act on? The workflow we are building toward is explicit. An agent proposes an action; the structural layer evaluates its consequences; the agent revises; admissibility is assessed; only then is the action taken. An injected instruction can change what an agent wants to do. It cannot change what the physics says about the system the action would touch, and that asymmetry is the point. For agents acting on physical and dynamical systems — a grid, an aircraft, a plant, a hedged position — the check can be physical today in principle. For agents acting on records and accounts the same architecture applies with a different state model, and we make no claim there yet.
This is a future architectural objective and we describe it as one. The present programme has not established operational action certification or a general-purpose safety guarantee. What justifies the larger claim is written down in advance: structural information that prospectively anticipates divergent trajectories beyond strong baselines and without future leakage; a demonstrated intervention that changes an outcome, which is a separate claim requiring separate evidence; and replication across genuinely different physical systems rather than one carefully constructed benchmark.
We do not regard probabilistic AI as something to resist. It is the most capable proposing machine ever built, and the world will use it. What we regard as unfinished is the arrangement in which it is used: a fluent system, an impressed audience, and nothing in between that can say whether the fluent answer corresponds to the state of the world. A collaboration between people and machines that lasts needs a referee neither can argue with, and the only candidate that owes nothing to either party is physics.
So we are building the referee, domain by domain, on data that actually happened, and publishing the record as we go. Markets came first because their record is public and their consequences are immediate. Molecules came second because biology has already established what a physics reading must recover before it is believed. Each domain that passes adds a certified measurement, a named state and a public record to the same layer. The layer is the product; the domains are where it earns the right to be trusted.
The destination is not the question “what is likely to happen?” but “what structural process is occurring, what happens if it continues, and which intervention changes the outcome?”