In brief
- The goal
- Capable AI that runs on hardware people already own, so that intelligence, starting with a child's education, is owned rather than rented.
- The lever
- Efficiency over scale. It is the one advantage that cannot be bought up, metered or legislated away.
- The long bet
- BQSM, an engine whose computation has an analog form: it runs digitally today, and is built so that a physical system could one day do the settling itself. Part VII
- The standard
- Every claim measured against the best reference, every failure published. Part VIII
Part I
Every learner gets an expert in whatever they love
School today runs on one curriculum for every child, moving at the pace of the room. We think the next generation will learn differently. They will follow their own interests, and a patient, expert guide will go wherever those interests lead.
Picture a twelve-year-old who is obsessed with skateboards. Today that is a distraction from school. In the system we are building, it is the way in. Friction, momentum and torque come through the board. Material science comes through the deck and the wheels. Geometry comes through the ramp. A local tutor follows the curiosity and brings in a physics expert, a materials expert and a design expert when the questions get there. Following a real interest leads into every other subject.
Three principles
Interest is the curriculum. The learner's own question sets the direction. The tutor's job is to go deeper, never to drag the learner back to a fixed track.
Cross-pollination is the method. Real understanding happens where fields meet. Each tutor can call on experts from neighboring fields, so one learner's single question becomes a connected map of knowledge.
The tutor belongs to the learner. It runs on hardware the family or school already owns. It remembers what this learner has done, adapts to how they learn, and needs no subscription, metered cloud service or outside company's permission to keep working.
Why ownership is the whole point
If the best tutor in the world can only be rented from a few companies that own the data centers, education becomes tenancy: thinking billed by the meter, and access to learning controlled by whoever owns the infrastructure. We refuse that future. The one advantage nobody can capture, revoke or legislate away is efficiency. Make capable AI cheap enough to run on an ordinary computer, and the moat built on compute stops mattering.
Our horizon is a generation: about fifteen years until the first cohort is educated entirely this way. The limiting factor is plain engineering, namely how fast a capable model can think on the CPUs people already have. That is the problem we work on every day.
Part II
“No one understands how AI works” is false
People fear what they don't understand. The most repeated claim in public AI debate is that the systems are a black box that even their builders cannot explain. This is false, and it is doing real damage, because a public that believes it cannot understand a technology will hand control of that technology to whoever claims to.
The truthful version separates two different questions:
Fully known — every step
- The architecture: every layer, every operation, written down in code
- The training procedure: the exact objective and the exact update rule
- Inference: every number computed from input to output, which can be traced, reproduced and inspected
- The weights: every parameter, stored and readable
Still being mapped — actively
- Which specific internal features and circuits a trained model has learned
- A complete, human-readable account of why a given output came from those features
The second column is an open engineering problem, the same kind of problem as mapping a large codebase you didn't write. Interpretability researchers make measurable progress on it every year. It is not a mystery that puts these systems beyond human understanding. We run these models ourselves, on our own machines, and we can watch every number move.
A machine whose every operation can be printed out is not a black box. It is a big box.
Part III
Three kinds of alarm, and how to tell them apart
Not all AI alarm is the same, and treating it as one thing helps the people who benefit from the confusion. Sort the alarmists by what they know and what they stand to gain:
The uninformed majority
Most public AI fear comes from people repeating “no one understands how it works.” As Part II shows, that is false. This is fear of the unknown, and the answer to it is education, not ridicule. Explain the machinery and the fear shrinks to its real size.
Informed insiders with a stake
Some of the loudest, best-informed warnings come from people at the companies building frontier systems, who warn that AI is dangerous and then lobby for rules that only companies of their size can comply with. Judge them by what they lobby for, not by what they say. When the proposed fix is licensing, compute thresholds and compliance costs that raise the barrier to entry, the warning works as a moat.
So among the alarmists who actually understand AI and are pushing for regulation, the ones demanding rules that entrench themselves are the ones who profit from the result.
Informed independents
There are serious researchers, including academics and people who walked away from industry salaries, who hold real concerns with no monopoly to protect. They deserve answers on the merits, not dismissal. Our answer is structural: the safest distribution of a powerful capability is a wide one. Capability concentrated in a few hands is a single point of failure. Capability spread across hardware people already own is resilient.
Part IV
The capture hypothesis
We present this as a hypothesis, not a proven fact, because a claim this serious should be testable. Here it is in plain terms:
Hypothesis
The leading frontier AI labs have committed enormous capital to compute and power for training. As the returns on that spending come under pressure, the incentive shifts from out-competing rivals to out-legislating them. Regulation becomes a way to protect the investment and control where the market goes.
Economists call this regulatory capture, and it is a familiar pattern in industries that depend on one bottleneck technology, from railroads to telecom. AI is a sharper case because the product being metered is cognition itself.
The hypothesis makes predictions. Here is what would support it and what would count against it:
- Supports itProposed rules whose compliance costs scale with company size in a way only incumbents can absorb; licensing regimes for model training or release; restrictions aimed at open-weight and locally run models; lobbying spending that rises as revenue falls behind infrastructure spending.
- Counts against itThe same labs backing rules that apply equally to themselves and new entrants, exemptions protecting small and local developers, and support for open models running on consumer hardware.
Look up the lobbying disclosures, read the proposed bills, and decide for yourself. We would rather you check than trust us.
Part V
Cognition is not consciousness
Most fear of AI quietly assumes that a system which reasons must also feel, want and scheme. That assumption mixes up two separate things.
| What it is | Where AI stands | |
|---|---|---|
| Cognition | Information processing: inference, recall, planning, language, problem-solving | Clearly performed. A language model calculates cognition, one step at a time, in arithmetic you can inspect. |
| Consciousness | Subjective experience: the capacity for there to be something it is like to be the system | No evidence of it. Nothing in the architecture requires experience to produce the output. |
A language model performs cognition by calculating it: a fixed sequence of mathematical operations turns input into output. Doing those operations well is not evidence that anything is experienced. A calculator that does arithmetic flawlessly is not thereby aware of numbers.
We are careful about how far that claim goes. Science has no validated test for experience in machines or anywhere else, so the truthful statement is not that consciousness has been ruled out forever. It is that there is no evidence of it, and none is needed to explain what these systems do. That is enough to dissolve the fear that capable AI is a hidden mind waiting to turn on us. It is capable machinery, and machinery can be understood, owned and governed by the people who use it.
Part VI
What we are building
We are building the engine for this future: capable intelligence that runs on CPUs people already own, with no GPU cluster, no metered API, and no gatekeeper in the loop. For individuals and developers it is VectorOS, free to use. For institutions it is EmergenceOS. Both run the same engine.
The tutor described in Part I is not a slide-deck promise. It is being built on that engine now, starting on our own machines. Every gain in efficiency brings it closer to every kitchen table and classroom on Earth, and further from anyone's control.
Don't argue with the concentration of power. Make the rent unnecessary.
Part VII
The machine underneath
Efficiency on today's processors is the near-term work. The long-range bet is a different idea of what the computer is.
A digital machine follows instructions. An analog machine lets physics do the arithmetic.
A digital processor runs a model one step at a time: fetch weights from memory, multiply, store the result, and repeat across billions of weights for every word. In large-model inference, much of the time and energy goes into moving the weights rather than into the arithmetic. An analog computer works the other way. The weights are not fetched, they are the wiring. Couple a network of resonating elements with the right strengths, drive it with the input, and the system relaxes to a settled state that is the answer. For a linear network that settled state is exactly the matrix-vector product, the operation that dominates AI. That is mathematics, not a claim of ours, and it is why analog and photonic AI hardware is an active field of research.
BQSM: a model you can read as a relaxing physical system
Our engine, BQSM (Basin-Quotient State Machine), is built so that the computation has this analog form. A language model's forward pass is read as a network of coupled oscillators settling to one stable state, a fixed point. What we have measured and published:
- The forward pass is one fixed point. Solved in order, which is the ordinary forward pass, or all at once with nothing sequenced, the two methods converge to identical logits. The write-up says exactly how large the test was, and that it implies no speedup.
- The core operations have wave forms, each checked against its reference. Normalization behaves as a saturable gain medium, softmax as parametric amplification, and rotary position as free-running oscillator phase.
- Not every oscillator model computes. The textbook phase-only model, the one most people mean by coupled oscillators, cannot produce a matrix product at all: we measured 97.5% error on real model weights, even after the best possible rescaling. The coupling has to carry amplitude as well as phase. We include the dead end because it is what makes the working version credible.
- The ring dynamics are public and runnable. The Wave Observatory lets you watch a ring of coupled oscillators synchronize, break apart and settle. It is a phase model, so it shows the dynamics rather than the computing coupling.
Digital today, physical tomorrow
Today BQSM runs as a digital scaffold: an ordinary CPU computes the settled state directly, which for this system is the ordinary forward pass, instead of waiting for a physical system to relax. We say plainly that this is no faster than stepping through the layers, and on the same machine our engine is currently slower than llama.cpp, the standard CPU engine. That comparison, including where we lose, is published in the engineering section of our home page. The scaffold exists for another reason: it lets every claim about the analog form be checked against a reference engine on a machine anyone owns, before any exotic hardware is involved.
The bet
If the computation is a relaxation, hardware that relaxes physically could perform it with its own physics: the cost becomes a settling time instead of billions of memory fetches, and the weights live in the wiring instead of crossing a bus.
That is the moonshot. It is not a faster program, it is a different substrate for the same program, with the equivalence already checked against a reference on the digital one. Analog AI is not our invention and we do not pretend otherwise: resistive crossbars, photonic meshes and coupled-oscillator computing are established research areas, and the equilibrium view of neural networks has a long literature (see below). What we add is a specific, checkable bridge from a working language model to an oscillator description of it, so each step toward a physical substrate can be tested instead of assumed.
What is established, and what is not
| Claim | Status |
|---|---|
| A forward pass is one fixed point, and sequenced and unsequenced solvers agree | Measured, published |
| Each core operation has a wave form, checked against its reference | Measured, published |
| The digital scaffold reproduces a reference engine's output | Measured, published |
| The scaffold is faster than llama.cpp, the standard CPU engine | Not true today |
| A physical substrate performs the settle | Not built. This is the bet. |
The last two rows are on the page deliberately. A claim sheet with no failures on it would be the thing to distrust.
Context and prior work
- Hopfield, J. J. (1982). Neural networks and physical systems with emergent collective computational abilities. PNAS, 79(8), 2554–2558.
- Kuramoto, Y. (1975). Self-entrainment of a population of coupled non-linear oscillators. Lecture Notes in Physics, 39, 420–422.
- Bai, S., Kolter, J. Z., & Koltun, V. (2019). Deep Equilibrium Models. Advances in Neural Information Processing Systems, 32.
- Shen, Y., et al. (2017). Deep learning with coherent nanophotonic circuits. Nature Photonics, 11.
- Csaba, G., & Porod, W. (2020). Coupled oscillators for computing: a review and perspective. Applied Physics Reviews, 7, 011302.
Part VIII
Backing the work
Everything above rests on one bet: that the most valuable thing in AI is not more compute, it is needing less of it. Compute can be bought up, rationed and metered by whoever holds the most. Efficiency cannot be captured that way, because an idea that makes a machine do more with what it already has works for everyone who has the machine. That is why we are working on efficiency and nothing else first.
Why this moment
The hardware squeeze is not hypothetical. AI demand is driving up the price of memory, the part of a device a capable model depends on most, and the supply data, with sources, is on our home page. Every month it stays that way, the cost of owning a capable machine rises and the case for renting one gets stronger. A system that makes ordinary hardware go further is a direct answer to that, and it is worth more the longer the squeeze lasts.
What you can check, not just believe
- Every result is published, including the failures. Our research library records what did not work alongside what did, each with the script that produced it.
- The tools are free to use. Anyone can run the same measurements and disagree with us.
- One honest gate decides the education work. A tutor on a learner's own machine has to answer at conversational speed on ordinary processors. That speed is the number we measure and report, whether it flatters us or not.
What backing buys
Time
- Protected time for the efficiency work, and for the long bet in Part VII
- Room to run experiments that may return nothing
Hardware
- A small, varied set of ordinary machines to prove the work on the devices people actually own
- Compute for measurement runs, not for scale
Learners
- A small pilot of the per-learner tutor, once the speed gate is met
- Publishing what the tutor gets right and what it gets wrong
Openness
- Documentation, reproductions and tools kept free to use
- An open invitation to check and challenge our claims
We are a commercial company, not a charity, and the home page says so plainly. We also think the best argument for backing us is the same as our argument about AI: do not trust a claim you cannot check. If you want to fund efficiency over scale, or just want to challenge what we have measured, we would like to hear from you.