Impossible yesterday. Probable today. Possible tomorrow.
Every technology you use was impossible once. Not metaphorically: there was a year when the people who understood it best could tell you exactly why it could not be done. Then somebody named the thing that was in the way, moved it, and the impossible became ordinary. Automation and machine intelligence everywhere, on this planet and past it, is that same story still running. It is not here yet. This page is about how to get there faster.
Why was anything ever impossible?
Heavier-than-air flight was impossible until 1903, and the reason was not mystery. It was power-to-weight: engines were too heavy for the lift a wing could make. Name that number and the problem stops being philosophy and starts being a shopping list. Every frontier has one of these. The work is finding out which.
The Wright brothers did not wait for a better engine. They built a wind tunnel, measured lift and drag on two hundred wing shapes, and found that the published coefficient everyone had been designing against was wrong. The constraint moved because somebody measured it. That is the pattern, and it repeats.
So when we study something that does not work yet, the interesting output is never whether it works today. It is which single thing is in the way, how far that thing has moved before, and what becomes possible on the other side of it. Answer those three and you are holding a plan.
Assuming today is the endpoint.
Every constraint below has moved before and will move again. The thing that reliably does not move is the belief that the present arrangement is final, and that belief is not an engineering position. It is an economic one: capital against benefit, asked to justify why an alternative to the status quo deserves to exist at all.
That question is a reasonable one to ask of a purchase. It is the wrong question to ask of a frontier, and asking it early is how a field talks itself out of the work. The steam engine was a curiosity with a bad efficiency figure. The transistor was a laboratory result with no market. In both cases the honest accounting at the time said wait, and in both cases the accounting was measuring the wrong year.
So this Institute does not spend its time arguing that the future is coming. The economics are already doing that. We spend it on the physical question underneath: energy, compute, and the loop between them, which is what sets how fast any of it arrives. The constraints below are not reasons for caution. They are the map.
Eight places, and only one of them is physics.
The other seven all move, and the record of how far they have moved is public. Five sit inside the machine and three sit in the world it has to arrive in. Knowing which one you are pushing on is the difference between working hard and going fast.
A conservation law, a thermodynamic floor, a diffraction limit. The only constraint that does not move.
What moves it: Nothing moves it. The useful surprise is how far away the floor usually is: erasing a bit at room temperature costs about 3 zeptojoules, and a 2026 processor spends roughly a billion times that. The distance between the floor and today is the whole opportunity.
How you spot it: It can be written as an equation with no free parameters. Landauer's kT·ln2 is one; the speed of light is one. Almost nothing is.
No substance yet has the property the design needs, at the temperature or the price it needs it.
What moves it: A new substance, and they arrive in steps rather than on a schedule. When neodymium replaced ferrite, everything sized around the old field strength was redesigned at once. Ask what number the new material would have to hit, and you have a research programme.
How you spot it: Progress arrives in steps, from discovery, and is hard to schedule, but the target is a number, and you can say what number.
The joules per task are wrong by a factor you can state, on a budget you can measure.
What moves it: Efficiency, which has moved by factors of a thousand more than once. A robot that needed a tether runs on a battery. A model that needed a hall of machines runs on the one in front of you. This is the constraint that has moved the most, and it is still moving.
How you spot it: It is the constraint most often mistaken for physics. A factor of a thousand in efficiency has happened repeatedly; a violated conservation law has not.
Every part is understood and the integration is not done: tolerances, interfaces, manufacturing, reliability.
What moves it: Person-years and revisions. Nobody disputes it can work, so the question is how many turns of the crank. When it moves, a demonstration becomes something you can buy.
How you spot it: Nobody disputes it can work. The question is only how many person-years and how many revisions.
The representation or the method does not exist yet, though nothing forbids it.
What moves it: One idea, sometimes overnight, by orders of magnitude. It is the most volatile of them all, which is why a forecast that assumes today's methods ages fastest.
How you spot it: The most volatile of the five. It is where a single idea can move a field by orders of magnitude overnight, which is why forecasts that assume today's methods age worst.
Counting our own topics against this list turns up something we did not set out to find. Four separate lines of work — pricing a decision in joules, reading a machine's power trace, a reactor with nobody on site, a spacecraft that re-plans in flight — arrive at the same sentence: the machine works, and what is missing is evidence a second party can check rather than trust. Measurement is the most common thing in the way after engineering itself, and it is the cheapest of all of them to move.
Three more, on the way to the field
The five above decide whether a thing can be built. These three decide whether it ever reaches anybody, and for an institute working on airspace, farms and factory floors they bind at least as often. They move on their own clock: you cannot engineer your way past a regulation, and no amount of money compresses a training pipeline the way it compresses a production run. Knowing which clock you are on tells you what to work on this year.
The quantity that would settle the question cannot currently be measured, or has no agreed definition.
What moves it: An instrument, a shared denominator, or a published protocol. It is usually far cheaper than the research it is holding up, and when it moves, an argument that ran for years becomes a number anyone can check.
How you spot it: Everyone argues from different numbers and nobody is lying. Two credible sources disagree by an order of magnitude and neither can reproduce the other, which is a statement about instruments rather than about the world. The fix is an instrument, a denominator or a published protocol, and it is usually cheaper than the research it is blocking.
The thing works and is not permitted yet, or the path to permission is undefined.
What moves it: A rule that reads a quantity it does not read today. Name the rule and the number, and the work becomes concrete: 23 CFR 490.409 grades a bridge by the minimum of three integers, so a fourth measurement changes nothing until the rule reads it.
How you spot it: The demonstration exists and the deployment does not, and it binds hardest where a rule was written for a different technology. Say which rule, and say what number it would have to carry. "Regulation is slow" is a complaint; "23 CFR 490.409 grades a bridge by the minimum of three integers, so a fourth measurement changes nothing until the rule reads it" is a constraint, and it tells you what to build.
The people who would operate, maintain or certify the thing are not there, or are not trained, and the pipeline that would produce them has a rate.
What moves it: A training pipeline, whose rate is knowable in advance: throughput, instructor supply, hours to competence. When it moves, the technology reaches the people who would use it, which is the only place it becomes real.
How you spot it: Roles stay open at rising salaries while the technology is unambiguously ready. It is countable, which is what separates it from a mood: training throughput, instructor supply, certification latency, hours to competence. The slowest of these to move and the most predictable, because a pipeline's length is known in advance. Teach the cohort and the constraint lifts on a date you can name.
Each of these still comes back to a number. Which quantity would the rule have to read? What rate does the pipeline run at? What would the missing instrument measure? Answer that and a soft-sounding obstacle turns into something you can build, schedule and check.
One habit is worth picking up early: notice when an energy or engineering constraint is being described in the language reserved for physics. "This cannot be done" is a strong claim, and it is almost always false: what is meant is that it costs a thousand times too much energy, or that the integration has not been done. Those are the two constraints that have historically moved the most.
What does it look like when this works?
Three times our own measurements came back looking like bad news. In each case the number was right and the first reading of it was wrong, and finding the constraint underneath turned a dead end into the next thing to build. Here they are, including the one where we were wrong in public.
An earlier stage of a generated world left 0.000% trace in the finished one, in two independent receivers, at every settling budget.
What it looked like at first: "Staged generation does not work."
What was actually going on: The constraint was algorithmic and one line deep: a mark stopped being held when its stage ended, so the cell relaxed back to whatever the physics preferred. Carrying the constraint forward made the earlier stage survive at full strength, in both receivers, at every budget. The measurement that produced the null is the same one that verified the fix. the measurement →
Zero of eleven headline efficiency claims in the 2026 AI-energy literature survived audit; the real gains are 1.4–3×, not 7–96×.
What it looked like at first: "Edge efficiency is hype."
What was actually going on: The audit also found which term binds: on-device inference loses on joules and wins on interconnection, and mixture-of-experts costs energy at the edge rather than saving it. Those are engineering and architecture constraints with numbers attached, and they say where the next factor comes from. the Charlot Lab →
A covert channel of 5.0 bits per second survives our own detector, about 18 kbit an hour, ample for a key.
What it looked like at first: "The certificate architecture is broken."
What was actually going on: The floor is a property of that detector, not of the approach, and the report says so and says what would move it. Publishing the number about our own architecture is what makes the next detector measurable against something. the report →
What if it is all solved by 2035?
Then it will have been solved by people who are working on it now, and the useful question is what they will have needed to know. Nothing arrives finished. It arrives because somebody found the binding constraint and moved it, and that is a skill you build by practising it on the constraints in front of you today.
A student who can look at a system, say which of the eight is holding it back, and design the measurement that would settle the question is already doing the work. Not preparing for it. The gap between now and solved is made entirely of that kind of afternoon.
This is why the courses here are built on instruments rather than on slides. You do not learn that a sampler has a mixing time by being told; you run two chains from opposite extremes and watch where they meet, and afterwards you know what the number depends on and what would change it. The same instrument that teaches the concept is the one that measures the frontier, which is the point: there is no separate simplified version for students, because a simplified version would not be able to surprise anybody.
It is also why the negative results are published. A field where only the successes appear is a field where every newcomer re-runs the same failed experiment, and the binding constraint stays hidden behind a decade of unpublished disappointment. The fastest way to make something possible tomorrow is to say precisely, today, what makes it merely probable.
A standard, so it can be failed.
Where a piece of our own work does not yet meet this, the gap is a task rather than a secret. The corpus is audited against this standard rather than assumed to meet it, because a standard nobody checks is a preference.