Pragma Research

Minds that learn by living

Every AI system in use today stops learning before it ships. It is trained once, its settings are locked, and after that it can only repeat what it already knows.

Cadence brains keep learning the entire time they run, the way an animal does. They are small enough to sit inside a robot and learn that robot’s job while doing it.

Bernhard Mueller · floatingpragma.io · Seed round, $4,000,000 · September 2026

Cadence · Pragma Research1 / 20
The thesis

Intelligence is what a body does in a world

No animal is trained and then switched on. It is born into a world, acts in it, is hurt by it, and changes because of it, every day, until it dies. There is no training phase and no moment when the learning stops.

Everything we want from a general intelligence follows from that arrangement. Coping with a situation nobody prepared for. Picking up a new job without a team first collecting thousands of examples of it. Keeping yesterday’s skill while learning today’s. Working again after being damaged.

A machine that cannot change after release does none of this. Making it bigger does not help, because the limit is in how it is built.

Cadence · Pragma Research2 / 20
The gap

What "frozen" costs

Today’s AI is trained in a datacenter. Its settings are then locked and copied out to wherever it is used. Nothing that happens afterwards changes it.

What a living mind needs What today’s AI does
Learn while it acts Learns once, then the settings lock
Learn from its own life, in order Learns from shuffled examples gathered in advance
Keep yesterday’s skill Teaching it a new task can erase the old one
Run inside the body Runs on rented servers, reached over a network
Cadence · Pragma Research3 / 20
The gap

What that costs in the field

A robot arm goes into a factory with a locked-in controller. The grippers wear, the parts shift slightly, and the success rate slides. Fixing it means collecting fresh examples, shipping them to a datacenter, retraining, testing, and pushing an update to the fleet.

retrainretrain80%85%90%95%100%month 0month 6month 12month 18Illustrative. Shape of the problem, not a measurement.

That repair cycle takes weeks, and the factory’s data has to leave the building. A Cadence brain adjusts during the shift, and nothing leaves the machine.

Cadence · Pragma Research4 / 20
The rule

Where the learning rule came from

Cadence started from a physics question. If the universe is made of small regions that each keep their own records, compare them with their neighbours, and fix the disagreements they find, what does that produce? We derived that it produces one shared, consistent world, with nothing in charge of it.

Turned into a learning rule, that says: each connection in the brain changes using only the activity of the two cells it joins, plus one broadcast signal saying whether things went well. Nothing ever looks at the whole network at once, so nothing has to store a trace of the computation and walk back through it. That is what removes the training phase.

broadcastBackpropagationone pass back through every layer, and the graph must be keptCadencethe two cells it joins, plus one broadcast signal

About the physics. It is formal proofs a computer checks step by step, like a compiler for mathematics, and more than 10,900 of them pass. The physics community has not accepted it, and this deck does not ask you to. The rule is judged on what it does.

Cadence · Pragma Research5 / 20
The object

How a Cadence brain works

A conventional network passes a signal through a stack of layers once, and an answer comes out the far end. A Cadence brain behaves more like a physical system. Its cells push on each other until they stop moving, and the state they come to rest in is the answer. We call that settling.

signals arrivethe cells push on each otherit comes to rest, and that state is the answer
Cadence · Pragma Research6 / 20
What exists

Built, running, and open to inspection

The library is public and installs in one line. Anyone can download it and check the claims.

Demonstrations that run in a browser, each showing the whole brain beside the task it is doing: a robot arm, a Connect Four player, a composer, an artist.

A Connect Four brain beats AlphaZero, the method that taught itself to beat the best chess and Go programs, trained here with its published recipe. Cadence wins 92 and 99 of 100 paired games against two configurations of it.

Brains built from real biological wiring. A map of a fruit fly’s brain containing 138,639 nerve cells, and the complete 302-cell nervous system of the roundworm C. elegans, both loaded and run as Cadence brains. The worm learns to associate a smell with food.

Every number in our paper is read out of a result file whose fingerprint is recorded, so a reader can trace it back. The paper also names ten things the design does badly.

One person built all of this, self-funded, at $5,000 to $10,000 a month.

Cadence · Pragma Research7 / 20
Evidence

What it does better, and why that matters

Arm reaching targets it was never shown
99.4%
against 98.0% for the standard method, which needs a stored buffer of past attempts
Choosing what to remember, four memory tasks
94.6%
against 49.0% for a conventional model given more memory than Cadence had
Updating a 138,639-cell brain after a small change
6.3×
faster than redoing the whole thing, because effort follows what changed
Function recovered after part of the brain is destroyed
85%
against 50%. The rule that learned the skill also repairs it
Practice needed to learn to balance a pole
40k
attempts against 137,000, for the same skill

Every number here is read out of a result file in our paper whose fingerprint is recorded, so a reader can trace it back to the run that produced it.

Cadence · Pragma Research8 / 20
Evidence

What it does badly, and why we are showing you

What was measured Result What it means
Time to make one decision 13 milliseconds, against 1 for a conventional network Settling costs more than passing through layers. Too slow for a fast control loop today.
Predicting the next character of text Behind a small transformer, the design behind ChatGPT, at every size tried Language is not where this design wins. We are not building a chatbot.
Learning from a reward that arrives much later Well behind the standard method for learning by trial and error Long chains of cause and effect are unsolved here.
Holding many separate skills at once Loses the older ones unless given a stored buffer The weakness most likely to sink the approach.
Size of the largest brain built 70 million connections A frontier language model is roughly 10,000 times larger. Nothing here shows this design scales that far.

Each of these is named in our own paper with its measurement. Every work package in this plan attacks one of them. A company that showed you only the previous slide would not survive a technical review.

Cadence · Pragma Research9 / 20
Embodiment

The ladder of bodies, and what each rung tests

Robot arma second job learned while working, the first one keptDexterous handmany fingers at once, learned from touchFour legsbalance and recovery under a changing loadFull bodymany limbs, and goals that take a roomMeasured in simulation · each rung above is a published result, including the rungs that fail

Ten cheap arms run in parallel, because what is being measured is what happens over a working lifetime, and one arm gives one lifetime at a time.

Cadence · Pragma Research10 / 20
The hardware idea

Why this design points at unusual hardware

A Cadence brain thinks by letting a system relax until it stops changing. On a normal computer we have to imitate that relaxing, by grinding through arithmetic over and over until the numbers settle. That is the main reason one decision costs 13 milliseconds instead of 1.

Some physical things relax by themselves. Light bouncing inside a shaped mirrored cavity falls into a stable pattern in the time it takes light to cross it, and no arithmetic happens at all.

An animalvoltages across nerve cellsabout 20 wattsA computernumbers in memory, in a looparithmetic, and 13 msAn optical cavitya light fieldthe time light takes to cross

Our mathematics says which shapes settle into which computation. That is what connects the brain design to the hardware.

Cadence · Pragma Research11 / 20
The hardware idea

The Echosahedron

LED shining inforward biasedthe same LED reading outreverse biased, measuring the lightTwelve windowsone at each vertex, each one a portA clear bodylight bounces and settles into a pattern

A thing with a boundary, a set of ports, an inner state and a way to read it. Which is the same description as one region of a Cadence brain. Connect several bodies through their windows and the assembly is a brain running in light.

Status, stated plainly. The mathematics exists and the design exists. No device has been built. This plan prints the bodies, builds one chamber, and measures whether it computes anything useful, against control experiments designed to catch us fooling ourselves. If it does not, we publish that and continue with the software version, which is what the rest of the plan rests on.

Cadence · Pragma Research12 / 20
The plan

Eighteen months, and when each answer arrives

0369121518months1 Fast decisionsunder 1 ms2 A learner that climbsrises with practice3 Real bodiesarmthree bodies4 Optical carrieroperatorone cortex in light5 Wiring evolvesbeats hand-drawn6 Scale and proofsthe curvea gold diamond is a gate, and every gate can fail

Nothing waits on everything else. The work that makes decisions fast runs first because every body depends on it, and the optical bench starts once its design is frozen.

Cadence · Pragma Research13 / 20
The plan

Six work packages, every one able to fail

Work package What “done” means If it fails
1 Make decisions fast enough for real bodies One decision in under 1 millisecond, on a chip small enough to sit inside a robot Brains supervise a fast controller instead of being one
2 Make the brain keep improving as experience accumulates Performance rises with practice instead of flattening, measured over a million decisions Ship the one-shot memory alongside a small conventional learner, and say so
3 Put brains on real bodies An arm learns a second job during operation and keeps the first. Then a hand, then a legged robot Narrow the scope to arm and hand
4 Test the optical hardware A built chamber computes something repeatable and useful, against controls Publish the negative result. The software path does not depend on it
5 Let the wiring design itself A layout found by evolution beats our best hand-drawn one on a body it never saw Keep hand-drawn wiring and report the idea as unproven
6 Find out whether it scales Publish the curve from 70 million connections upward, whatever shape it has Publish the flat curve
Cadence · Pragma Research14 / 20
The plan

What exists after eighteen months

  1. A brain that learns on three real bodies. A robot arm, a dexterous hand and a legged robot, each picking up a new job while working and keeping the one it already had, with no retraining run and no data leaving the machine.
  2. Software that robot makers can install, fast enough for a control loop, with a guaranteed worst-case decision time and a command that returns the robot to its original behaviour.
  3. An answer on optical hardware. A chamber built and measured, and one brain region run in light and compared against the same region in software. A negative answer is published as readily as a positive one.
  4. A wiring diagram found by evolution that beats the best one we drew by hand.
  5. Version 1.0 of the open library, stable enough to build products on.
  6. The papers and the proofs, with every number traceable to the file it came from.
Cadence · Pragma Research15 / 20
Defensibility

Why this is hard to copy

The learning rule was derived, not tuned. It fell out of an argument about consistency and was then checked by machine. It cannot be read off a results table and reproduced, because the reasoning is what produces it.

The optical path needs both halves. A photonics lab can build a cavity and has no theory saying what the cavity computes. An AI lab has the computers and will not build cavities. The theory that joins them is ours.

The open library is the distribution. It is already public, with the measurements and the proofs attached. Adoption compounds while the paid work sits on top of it.

Everything is checkable without trusting us. The proofs run on anyone’s computer. Predictions were published and time-stamped before the measurements arrived, so they cannot be quietly adjusted afterwards. Ten weaknesses are named in our own paper, and there is a public prize for disproving the framework.

Cadence · Pragma Research16 / 20
Business

Where the money comes from

The library stays free and open. It is how developers find us and how outsiders check the work. It is not the product.

The product is a finished brain for a particular job. A wiring diagram, a training curriculum and a tested runtime for, say, a warehouse picking arm. A robot maker licenses that per robot. Building one takes us months and would take them years, and that is what they are paying to skip.

The first paying use is fleets already installed. A robot whose success rate is drifting down as its hardware wears. Our brain rides on top of the maker’s existing controller and closes the gap during operation, without altering the controller that was already approved for safety.

Optical hardware is a second business and exists only if work package 4 returns a number.

Roughly 4.66 million industrial robots are working worldwide, and about 542,000 new ones were installed in 2024. No revenue is assumed inside the 18 months.

Cadence · Pragma Research17 / 20
Team

Six hires, no managers, and a back office run by AI

Role Starts Why then
Principal investigator (B. Mueller) month 1 Leads the research and writes code with the team
GPU and systems engineer month 1 Work package 1 gates every body that follows
Machine-learning researcher month 1 Work package 2 is the long pole
Robotics researcher month 4 Once there is a simulator to work against
Photonics and hardware engineer month 7 Once the chamber design is frozen
Operations lead month 1 Company setup, payroll and reporting, run with AI tools
Marketing and developer relations month 7 Once there is a robot to show
Formal proofs, half time month 7 Proofs follow the mechanism

Bernhard Mueller. Pwnie award 2009, the security field’s research prize. Author of the OWASP MASVS, the mobile security standard Google Play’s independent app review certifies against. Built Mythril, the most widely used security analyser for Ethereum contracts. Speaker at DEF CON 27 and HITB. Thirty years of changing field every five years, from hardware to reverse engineering to mobile to neural networks to blockchains to physics.

Cadence · Pragma Research18 / 20
Budget

$3,739,232 over 18 months

Personnel$1,595,000Compute$484,300AI inference$390,000Robots and lab$353,000Operations$202,800Dissemination$184,000Fabrication and optics$129,500Contingency, 12 per cent$400,632Total$3,739,232

Every price was checked in September 2026 and carries its source. The 3D printers are bought rather than outsourced because the print quality and surface finish of an optical chamber are part of what is being measured.

Cadence · Pragma Research19 / 20
The ask

$4,000,000 for 18 months

Six hires, four robot bodies, the optical bench, and the six work packages on the previous pages. Smaller versions of the programme exist, and which work packages a smaller round buys is a conversation rather than a table.

The code, the proofs, the data and the papers are released openly. The finished brains and the runtime are the company.

bernhard@floatingpragma.ai · floatingpragma.io · github.com/muellerberndt/cadence

Cadence · Pragma Research20 / 20