Making post-von-Neumann useful.
Thermodynamic, neuromorphic and Ising machines have seminars, books, courses, open-source stacks and working silicon. What none of them has is a person who uses one without knowing what it is. This topic asks what that would take, and starts from the one architecture that managed it.
Grant Markhart, Industrial Research Fellow · The Hall Lab, with The Charlot Lab
Binary computing became useful on a date that can be named. In 1979 the computer was sold as a VisiCalc accessory: a buyer walked into a store, watched a spreadsheet recalculate, and bought an Apple II in order to get it. Before that, personal computers were expensive toys for enthusiasts. Not one of those buyers could have described a stored-program architecture, and none of them needed to.
What followed was the same move, repeated. The application deleted the architecture. The interface deleted the task’s mechanics. Play deleted the machine altogether. Adoption is not measured by what a machine can do; it is measured by how much of the machine you managed to remove from the user’s head.
Set every current interface to an alternative architecture against that. You write a QUBO. You write an Ising model. You write a hypergraphical model. The architecture is in the notation, which is the first rung and not the second, and the neuromorphic roadmap literature independently names the identical barrier: commercial adoption waits on how to program these machines at all. The field is paving toward the hardware.
The instrument
So the first question is not what an alternative architecture can compute. It is which of the things people already do has an inner loop that a sampler would serve. We enumerated every human-facing interface use case we could find, across Japanese, Chinese, German and English sources, and scored each on one question: does the thing the user is waiting for require drawing a sample from a constrained distribution?
Does the loop sample? · 102 use cases, scored
The result corrected our own prior. We expected the sampling-shaped subset to be small and concentrated at the entertainment end. It is 39% of the inventory, and the rate barely moves across domains — 30% in industrial process control, 57% in the home. The split is by verb rather than by industry: display, compare, retrieve, classify, recognise, render and sort want a conventional machine; assign, schedule, place, order, partition, complete, reconstruct, generate and resolve want a sampler. So the reason nobody uses one of these machines is not that there is nothing for it to do.
The unlock
Four applications are being actively attempted and are not shipping, and each has its blocker written down in the 2026 literature. Diffusion policies for robot manipulation: sequential denoising, a stated major limitation for real-time control. On-device generative imaging: memory, battery and thermal, with the industry answer being to send the hard half to a datacentre. Live generative game content: latency you can feel, a per-request cost every session, and consistency that drifts when two things must match. Molecular conformer generation: a thousand timesteps for one conformer.
One blocker, four times. None is blocked on accuracy or on whether the method works. Every one is blocked on the number of sequential steps needed to draw a single sample, multiplied by what a step costs in joules and milliseconds. And in all four the architecture is already invisible: nobody using a robot, a camera, a game or a discovery pipeline knows a sample was drawn, and none of them would have to learn. The product hides the machine before it is built. What is visible today is only the cost, which is the one thing an alternative architecture is supposed to change.
Who else is stuck, and where exactly
Four hardware families have now reached the same place from different physics, and each names the same blocker in its own words. Intel’s neuromorphic programme reports that interfacing with conventional sensors, processors and data formats was a bottleneck, that Loihi 2 has no commercialisation plan and remains a research platform, and that software continues to hold back the field for want of a single framework of the kind deep learning has. IBM’s TrueNorth shipped in 2014 and this review did not locate a commercially useful application built on it. Quantum annealing’s advantage is now stated conditionally — a hybrid solver competitive on a limited set of real problems, unable to beat Gurobi on others, with the benefit confined to specific quadratic structures — and a 2026 challenge to the supremacy claim is disputed rather than settled.
The one that reached customers did it by disappearing. Mythic raised $266M and arrived not as a platform but inside somebody else’s product, through partnerships with Honda and with Microchip’s memory arm. That is the accessory strategy, and it is the only one on this list that has worked.
There is a warning here too, and it lands on our own work. Mandrà, Katzgraber and Thomas showed in 2017 that annealer speedup claims on planar gadget problems were measured on instances that minimum-weight perfect matching solves exactly in polynomial time. The Charlot Lab built that solver this month and used planar spin glasses to demonstrate it. The correct reading of such a result is that the instance had exploitable structure, and nothing about hard instances — which is now written into the module rather than left for someone else to point out.
Where it stands
This is a survey and a position, not a result. The inventory is the first deliverable and it is published above, with its counts computed from its own data rather than written down. What it does not yet contain is a measurement: whether a sampler actually closes any of the four blockers, on our own stack, against our own game-engine and codec concept. That is the next piece of work, and until it exists this topic is an argument about where to look rather than evidence that looking there pays. The mechanism is at least published: denoising thermodynamic models replace each neural denoising step with a Gibbs sample from a hardware energy-based model, so the connection between these blockers and this hardware is real rather than an analogy. The reported ten-thousand-fold energy figure is simulated, on a small benchmark, and should be carried as a projection.
The survey behind the interface half of this is What Do You Type? (TR-2026-47), which found that every toolchain in the field returns a sample and nothing that says whether to believe it. This topic asks the question one layer out: not what the interface returns, but whether anyone should have to see it.
Binary computing became useful on a date that can be named. In 1979 the computer was sold as a VisiCalc accessory: a buyer walked into a store, saw a spreadsheet recalculate, and bought an Apple II to get it. Nothing in that purchase contained the architecture, and the arc afterwards repeated the same move, deleting the machine from the user’s head at every step, from architecture to task to metaphor to nothing at all. Adoption is measured by how much of the machine you managed to delete. Every current interface to an alternative architecture moves the other way: you write a QUBO, an Ising model, a hypergraphical model, and the machine is IN the notation. The field is paving toward the hardware. So the binding constraint is not physics, funding or belief; it is that no alternative architecture has ever been sold as an accessory to anything. The engineering change that moves it is a result type rather than another SDK: an answer that carries a bound, a proof status that can say no, a self-distrust field, a cost in joules and a typed refusal, so a caller who knows nothing about sampling can act on what came back. What becomes possible is the first application anyone would buy the machine to get.
One of eight, and only one of them is physics. How we read a frontier →