Grown, not written
Nobody types the rules into an AI model. A simple routine adjusts its random numbers until its guesses get good.
The short answer. Training means guess, score the error, nudge every number a little, and repeat. We ran it on a tiny model, so these results are real, and small.
Who made this page. Claude, an AI made by Anthropic, whose CEO is quoted below.
Four steps, repeated
The point: a routine that repeats sets a model's numbers. Nobody picks them by hand.
An ordinary program is a recipe: if it knows that u follows q, somebody typed that line. A model has no such line. Its weights, the long list of numbers from Chapter 1, get set by training.
Picture a mixing desk with thousands of knobs, one for each number. After each round, a piece of maths called backpropagation works out which way to turn each knob to lower the error. Every knob moves a hair that way.
It's like walking downhill in fog, where height is the error. You can only feel which way the ground falls.
Our model has about 56,000 numbers. It reads up to five letters and guesses the next. It trained on words from a 1934 dictionary and is scored on words it never saw.
Tags: measured we ran it and counted. third-party someone else reports it. assessed a judgment call.
Replay the training
The point: the routine alone takes the model from gibberish to word-shaped guesses.
Press Train more to jump to a later saved moment of one real run.
It writes by rolling dice weighted by its guesses. measured
Its guess for the letter after q
Each bar is one letter's chance of coming next. Orange marks u. Nobody told it that u follows q. Between 300 and 1,000 nudges, the examples alone push that bar up.
Error score
High when it gave the right letter a small chance. Guessing blindly across 27 choices (26 letters and end of word) scores 3.3. The axis counts nudges, unevenly spaced, and the scale starts at 2.
A sample of its numbers
One square, one number. Orange above zero, teal below, stronger means further from zero. Changing these is all training does.
By the end it gets about 1 next letter in 3 right, nearly ten times better than a blind guess.
Is it just memorising? We built a phrasebook to check: it stores what came next for every run of letters in the training words. It won overall, because 95 in 100 test cases were runs it held. On new runs the network did far better, with about a seventh of the storage. That's training's edge: patterns squeezed into few numbers carry over to new cases. measured
The phrasebook, in numbers
Next letters right, all tests
Has the phrasebook seen these letters?
The phrasebook stores 376,477 counts, the network 56,097 numbers. On a new run, the phrasebook falls back on the commonest letter. The seven rules are one short attempt; better rules would score higher. measured
A real map of this run's error
Treat the error score as height. Each of the 56,097 numbers is one direction you can walk, so this map shows a thin shadow: 2 directions out of 56,097. We drew the two that best fit this run.
The start shows what a shadow loses. The map reads 2.36 there, and the real error was 3.31. So any early climb on the map comes from the flattening: the real error fell between every saved stage. From 250 nudges on, map and real error differ by 0.06 at most. measured
Same start, different start
The point: different random starts give different numbers and nearly equal skill. The same start gives an exact copy.
Two oak seeds in one garden cast similar shade from different branches. Each picture below is the same sample of numbers after a finished run. The percentage is next letters right.
Same start, trained twice
All 56,097 numbers match exactly.
Different starts
These began from different random numbers, with examples in a different order. Patterns differ, scores barely do. measured
More numbers: how different are they?
Any two different-start runs pick the same next letter about 7 times in 10. A correlation scores how closely two lists line up: 1 is identical, 0 is no relation. Runs from different starts scored 0.02. Changing only the starting numbers, once, gave 0.03 and still 36.0% right. Keeping the start and changing only the order gave 0.39 and 35.7% right, so the start leaves a lasting mark.
Are they the same sums, reordered? Partly. Inner sums have no fixed order, so two runs can grow similar ones in different positions. Matching each last-layer sum to its closest partner in another run gives an average similarity of 0.60. Best-of-200 matching flatters this, and it covers one layer.
Can you find a habit and delete it?
The point: no one dot holds the q-then-u habit. Removing it took a search, and it cost other habits.
A dot is one small sum inside the network. We switched off each of the 200 dots in the last inner layer, one at a time. u stayed the top guess after q every time. At worst its chance fell from 99% to 91%. So we switched off the most damaging dots together.
Chance it gives u after q
The cost of switching off 20
measured
Training changes habits too. A worn path through grass stays when people stop walking it, and fades when they walk elsewhere.
Stop walking it
Remove every word with a q and train 3,000 more nudges. u stays the top guess after q, every time.
Walk elsewhere: teach q then a. Picks a:
Teaching barely dented the overall score. measured
More numbers: a failed first try, and teaching
We first ranked dots by whether the top guess changed. None changed it, so that picked 40 dots nearly at random. q-then-u stayed at 99.6% while the overall score fell to 33.0%.
Overall score after teaching q then a: 35.8%. The same extra training with nothing changed: 35.9%.
Where "grown" stops being true
The point: "grown" is a useful metaphor, and it breaks if pushed.
Dario Amodei, Anthropic's CEO, wrote in April 2025, crediting a cofounder, that these systems are "grown more than they are built", and that "we set the high-level conditions". third-party
Against it: people wrote the network's shape, the routine and the score, and chose the words. You could fairly call the result compiled from data. assessed
Where it breaks: it doesn't eat, breed or live. And no two seeds grow into exactly the same tree, while this repeated down to the last number.
What a tiny letter-guesser can't tell you: how big models behave. Goals, deception or planning can't even be asked of this one. assessed
What you can fairly say
- Training is guess, score, nudge every number, repeat. Nobody writes the rules.
- In our tiny model, different starts gave different numbers and similar skill. The same start repeated exactly.
- No one dot held q-then-u. Removing it took a search and hurt other habits.
- That models are literally alive, or that people have no hand in them.
- That every bad habit can be removed, or that none can. We tested one.
- That training beats memory everywhere.
- Whether AI understands.