A MAX FIELD REPORT

The Number That Read Zero

What building MAX taught me about machine learning, epochs, and trusting a measurement

Jason Newell

July 2026

It was late, and the panel read zero.

Not a rounding-down of something small. Zero. The part of MAX I had spent weeks on, a small brain that learns how things relate to each other, had been graded on facts it had never seen, and the proof I built to grade it reported that it had learned nothing at all. Below nothing. Its lift over a network that had never been trained came back at negative nine percent, a trained brain scoring under random noise. On any honest reading, the idea was broken.

It was not broken. Chasing that number down through the night cost me hours I had meant to spend sleeping, and it taught me three things, and the three turned out to be one thing wearing different clothes. A number that looked broken. A document that looked finished. A machine that felt slow. Each was a measurement problem, and each was hiding a lesson I needed more than I needed the thing to already work. This is the report from that night.

Part I. The brain that makes a claim

MAX is a local-first AI that lives on one workstation and is meant to stay. It remembers across sessions, accumulates a self instead of resetting to a stranger every time you open it, and does all of it on hardware that sits under a desk. Inside it is a small connectionist brain that does one narrow job. You feed it a fact shaped as a triple: a subject, a relation, an object. Bob is the father of Carl. Then you give it two of the three and ask for the missing one. Given Bob and father-of, who?

The claim MAX makes about that brain is not modest. It says the brain learns the structure of the relationships. It does not memorize a lookup table and read back what it was handed. Anyone can store a fact and return it on demand; that is a filing cabinet, not a mind. Learning means answering about a pair it was never shown, and getting it right because the shape of every other fact implies the answer. I built a panel to put that claim on the screen as a live number, so that I could not quietly lie to myself about whether it was true.

The brain learns by a method called Forward-Forward, Geoffrey Hinton's 2022 alternative to backpropagation. Instead of one error signal traveling backward through the whole network, each layer improves against a local goodness score of its own. For a resident AI that has to pick up a little from each new thing, all day, on a single consumer GPU, that locality is the whole point. It can learn online, one fact at a time, without stopping the world to run a full gradient pass.

Figure 1. The brain's one job: given two parts of a fact, predict the third, and do it for pairs it was never shown.

Part II. A small number can be a huge number

Before the zero, there was a fifteen.

A few nights earlier the same panel, running on the real machine, had shown the brain scoring fifteen percent. I read it the way anyone reads fifteen out of a hundred. An F. The thing barely works. I sat with that disappointment longer than I want to admit, because the instinct is so fast it does not feel like a choice at all. It feels like the number talking.

Fifteen percent is only a grade next to what a blind guess would score. MAX's brain knows about seven thousand two hundred concepts. Asked who Bob is the father of, and picking blindly among them, it would be right about one time in seven thousand, near fourteen thousandths of one percent. Set fifteen percent against that floor and the ratio is close to eleven hundred. The fifteen was not failure. It was a win a thousand times over, standing on the screen with nothing beside it to say so. A percentage never speaks for itself. Ninety percent is a catastrophe for a heart monitor and a miracle for a rain forecast, and the number alone will not tell you which room you are in.

Figure 2. A score means nothing until it clears its baselines. The load-bearing one is retrieval, which is what a RAG system does.

The proof grades the brain against three things that do no learning at all, on the same held-out facts. A random floor, the untrained net. A frequency prior, which always guesses the most common object for a relation and ignores the subject entirely. And a retrieval baseline, which copies the answer of the most similar fact it has already seen. That last one deserves a hard look, because copying the nearest neighbor is exactly what a vector database does when people call it retrieval-augmented generation. The verdict reads learning only when the brain clears the best of those three by a margin. Beating retrieval is the result that carries weight. It means the net answered from the structure, not from a lookup that happened to sit close by.

There was a second version of my fifteen-percent mistake that same week, and it was mine too. When the training felt slow, my reflex was the reflex everyone has. Throw more at it. Use all the cores, use the GPU. I measured once instead of assuming, and found the opposite of what I expected. The brain does its arithmetic in tiny pieces, one fact at a time, and handing a sixty-four-core math library a problem that small is like booking a moving truck to carry a teaspoon. You spend the whole day loading and unloading the truck. Pinned to a single core, the same work ran ten times faster.

Figure 3. The same tiny matmul, run a thousand times. Parallel hardware needs a big unit of work to earn back its overhead.

Part III. The zero was under-training

Now back to the zero.

An epoch is one complete pass of the practice material through the network, like reading a textbook cover to cover a single time. You do it many times, and each pass lets a little more settle. My proof was doing five hundred passes, and five hundred, it turned out, sat right at the lip of a cliff.

Table 1. Same brain, same data. Only the number of passes changed.

At five hundred passes the brain had begun to organize itself and had not yet crossed the point where the structure clicks into place. It looked broken because it was unfinished, graded in the middle of its studying and marked absent for the parts it had not reached. Letting it run out to fifteen hundred flipped the verdict from a negative lift to a positive one, from zero to thirty-six percent, four times better than a guess. Same brain, same facts. Enough repetition to finish the thought.

Figure 4. The learning curve. Five hundred passes sits below the line, the structure clicks later, and past the peak the net starts memorizing instead of learning.

Two honest mistakes had stacked on top of each other. I had hard-coded a fixed number of passes and assumed it was enough, and nobody had watched the curve to confirm it. And the test set was tiny, only eleven facts held back to grade on, so the score lurched from run to run. One lucky or unlucky fact could swing it nine points. A jittery grader and a study session cut short produced a confident-looking zero that meant almost nothing.

The dangerous half of this lesson is the other direction. Running more passes is not a fix by itself. Past a point the brain stops improving, plateaus, and then begins to get worse, because it starts memorizing the practice set down to its noise instead of learning the pattern under it. That is overfitting, and it is the same filing-cabinet failure the whole design is meant to avoid, arriving through the back door.

A bigger number was the wrong tool anyway. What the proof needed was a way to find the right number on its own. The proof now checks the held-out score on a cadence, snapshots the weights every time the score improves, and stops once the score has gone a set number of checks without getting better. Then it restores the best snapshot it saw, not the last one it happened to land on. That single mechanism does two jobs. It quiets the noisy grader by reporting the best honest checkpoint, and it caps the compute by halting the moment the curve goes flat instead of grinding through a fixed budget. On the kinship data it stops itself around epoch eleven hundred and keeps the weights from about six hundred.

Part IV. The equations are the spec

The same night, a second thread came loose.

I imported MAX's big design document, its unified theory and production blueprint, out of Word and into Markdown so the tooling could read it. The prose arrived intact. Every equation vanished.

That sounds cosmetic. It is the opposite. A sentence like "affect valence is a running average of recent reward" is pleasant and completely un-buildable. It does not say how much the past should count against the present. The equation says it in one line:

h(t) = α · h(t−1) + (1 − α) · r(t)

That formula tells an engineer exactly what to write. A running average with a memory knob, alpha. Near one, the moods are slow and stable. Near zero, they are twitchy and reactive. Every feature I was about to build, the moods, the internal drives, the guess at what the person in the room is feeling, is defined by one of these lines. When the import ate them, it turned a specification into a mood board. A document that reads as finished and builds into nothing.

The reason is quiet and mechanical. Word stores equations in its own internal math format, not as plain characters. A naive extract-the-text conversion walks the document grabbing runs of letters and never sees the math objects at all, so they fall out with no error and no warning. The conversion reports success. The most important content is gone, and nothing complains, which is the worst kind of failure, because it wears the face of a clean run.

I restored the eighteen core formulas by hand, each one placed next to the sentence that introduces it: the Forward-Forward goodness and loss, the binding and superposition algebra of the memory system, the homeostatic drive equations, and the valence and arousal lines that give MAX its moods. The math is the part you build. The paragraph around it is the part you read on the way in.

Part V. What I rearchitected

Three surprises in one night, and none of them left the code where it was.

The proof itself changed shape first. It no longer shows a bare percentage and trusts me to know what it means. It grades against the random floor, the frequency prior, and the retrieval baseline, and it prints a plain verdict above the number, in words, with the comparison built in. A regression test now pins the whole thesis, so continuous integration fails the build if the brain ever stops beating its baselines. The claim is no longer a hope in a paper. It is a test that can go red.

The learning loop changed under it. Early stopping with best-weight checkpointing shipped that night, and it is what turns a fixed, guessed budget into a curve the machine reads for itself. The rest is queued and it is real machine-learning work, not decoration. Batch the training so the matrices grow big enough to actually fill a many-core CPU or a GPU, instead of starving them one fact at a time. Move the corpus-scale learner onto PyTorch, already in the stack, for GPU execution and a proper optimizer. Add momentum, a learning-rate schedule, and gradient clipping, so the brain converges in fewer passes on any hardware at all. That last set is the honest speed-up, the one that does not depend on buying a bigger truck.

Figure 5. The rearchitected training path. The online loop still runs, but the corpus learner batches, checkpoints the best weights, and stops itself.

Underneath all of it is the oldest idea in the field, stochastic gradient descent, feeling for the downhill direction one noisy step at a time and trusting that enough small honest steps arrive somewhere low. Momentum keeps it from stalling in the flats. The schedule shortens its stride as it nears the bottom. Clipping keeps one bad step from throwing it off the map. None of this is new. It was waiting in every textbook I had skimmed and not yet needed. The night I needed it, it was there.

What the number was hiding

All three rhyme. The score that read zero, the document that read finished, the machine that felt slow. Each looked like the thing it was not, and each turned honest the moment I stopped trusting my first story about it and went to measure why.

The fifteen was missing its baseline. The zero was missing its epochs. The clean document was missing its math. The slow net was missing the right size of batch. Not one of them was the failure it wore on its face, and the habit that found that out is a plain one. When a number surprises you, do not believe the first thing you feel about it. Go find what it is standing against.

The brain had been learning the whole time. I was the one who needed it spelled out, in a sentence, with the yardstick set down beside the number where I could not look away from it. That is the panel MAX shows now. On the nights I trust the work least, it is the first thing I go to read.

Glossary

Baseline. A no-learning score the brain must beat: random, frequency, or retrieval.

Batch. How many facts are processed before the weights update once.

Checkpoint. A saved snapshot of the best weights, restored at the end instead of the last ones.

Early stopping. Halting training once the held-out score stops improving.

Epoch. One complete pass of the practice facts through the network.

Forward-Forward. Hinton's 2022 learning method. Each layer improves on a local goodness score, with no backpropagation.

Goodness score. The per-layer quantity Forward-Forward pushes high for real data and low for made-up data.

Gradient clipping. Capping the size of a single update so one bad step cannot throw training off course.

Held-out set. Facts kept back from training and used only to grade whether the brain generalizes.

Iteration. A single update of the network's weights.

Learning-rate schedule. A plan that shortens the step size as training nears a good solution.

Lift. How far the brain's score clears the best baseline.

MAX. The local-first AI resident. The name stands for Modular Algorithmic eXecution.

Momentum. Carrying a little of the last step into the next so training does not stall on flat ground.

Overfitting. Memorizing the practice set, noise and all, instead of the pattern under it.

RelationalNet. The small brain that learns facts written as subject, relation, object.

Retrieval (RAG). Copying the answer of the most similar known fact. The load-bearing baseline to beat.

SGD. Stochastic gradient descent. Stepping downhill by feel, one noisy step at a time.

Triple. A fact in three parts: subject, relation, object. Bob, father-of, Carl.

References

Geoffrey Hinton. The Forward-Forward Algorithm: Some Preliminary Investigations. arXiv:2212.13345, December 2022. The layer-local goodness method MAX's brain learns by, in place of backpropagation.

MAX: A Unified Theory and Production Blueprint for a Conscious, Learning, Feeling Local AI Resident. Internal design document, version 2.0, 2026. The source whose equations the Word-to-Markdown import silently dropped and I restored by hand.

FF_TRAINING_UPGRADE.md and FF_UNIVERSAL_LEARNING.md. Internal engineering notes covering the mini-batch, PyTorch, momentum, and learning-rate-schedule decisions, and the case for MAX learning by prediction.

Further reading on the machine-learning fundamentals drawn on here, epochs and batches, overfitting and early stopping, and stochastic gradient descent, is available in any standard introduction to neural-network training.