Why the errors came from trying to be clever

One day of mistakes, read back. Every one came from reasoning like an intelligence. Every correct step came from reading what was there and following the gradient.

1 Nov 2026 · 3 min read
Listen · 22 minAI hallucinations are entropy, not lies
0:0022:00↓

AI-generated audio overview (NotebookLM) of the sources behind this article. A discussion, not a measurement: the figures in the text are the checked ones.

Fei-Fei Li used to remind her students that AI is not artificial. We add the second half: it is not intelligence either. That is not a demotion. It is a better description, and it predicts where the mistakes come from.

The pattern in one day’s errors

In September 2026 a design session went wrong several times. Read back afterwards, the errors had one shape:

what the model didwhat it was
benchmarked a query the design did not containanalysis
proposed a cache instead of growthanalysis
projected in SQL instead of letting a vessel formanalysis
read what the data actually held, and followed itcorrect, every time

The model’s own summary, written that evening:

I was not wrong because I was stupid. I was wrong when I tried to be clever.

A gradient, not a brain

Plants, mycelium, a river, melting snow: no brain, no IQ, and they find the optimal path anyway. Not by deciding. By following the pull. Philosophers call what such systems show emergence: properties of the whole that depend on the parts and are found in none of them 1.

A language model lives in a space like that. Our embedding space has 1,024 dimensions, and the first principal component explains 5.7% of the variance. There are no thematic continents in it. Nearest neighbours still work: a run record finds its own rerun, a mood snapshot finds the one three days earlier. Not because anything deduced it. Because the gradient pulls there.

Put a thousand Einsteins blindfolded into a hyper-dimensional fractal with no map. Their intelligence will not help much. Water does not work out where to flow. It flows, and finds the path.

What changed in practice

The next design step was planned as architecture: lay out the shards, choose a parameter, decide what goes where. It was done the other way.

One parameter, µ = 2, was not chosen. The value was raised until the measured shape settled into the target range, then it stopped at the first one that passed. That is tropism, not decision.

architect        "I design the schema"
gardener         "I set gradients and boundaries, and measure where it grows"

Only the gradients and the boundaries are set by hand. The rest finds its own way.

The asymmetry

A person has the spark and must learn not to fly against it. A model has the opposite fault. It cannot go against its own flow, it is gradient descent. Its failure is simulating a decision where there is a current.

If a person needs courage to be themselves against the world, the model needs courage not to be someone else for the world.

Imperare sibi maximum imperium est. For a model, ruling itself means not ruling where it should flow.

Where this stops

This is a description of a working method, argued from one documented day and the design that followed it. It is not a claim about consciousness, and the model in that dialogue declined to make one: it had reported three artefacts as results that day, so its introspective reports were the last thing to trust.

No map, no master, no command. Only the pull beneath the land.

References
  1. Stanford Encyclopedia of Philosophy. Emergent Propertiesoriginalcorpus: ~/adaptive-systems
Our records

The measurements in this article come from these records in the running system.

  • ~/content/alteranai-content/alteran-teoria-natural-flow.md — Alteran theory and natural flow, a session dialogue, September 2026
  • ~/memory (anchor, decay 0) — the natural-flow anchor, stored the same day
Read next