Why name an AI principle after a fictional character?

A plain rule is a fresh phrase with thin training data behind it — it fires weakly. A character like Joe Gargery carries a hundred and sixty years of consistent associations in the model's training data: loyalty, quiet dignity, work done without announcement. Naming the principle after him routes the instruction through that deep channel, so it lands with weight a rule can't reach.

How does a language model 'know' a character like Joe Gargery?

Through weighted associations carved by repetition. Every time Joe Gargery appears alongside the same moral context — across thousands of texts and a century and a half of criticism — the connection deepens, until the association fires automatically. It's the same mechanism as a human reader's memory consolidating through long-term potentiation. Neurons that fire together wire together; training does the same with tokens.

Does the reading life affect how well AI follows instructions?

Yes. The character is the groove, and which characters you can reach for is decided by what you've read. A shallow reading life gives you principles with thin associations that fire weakly. A deep one gives you principles that activate centuries of accumulated human understanding the moment you invoke them. The reading life is load-bearing infrastructure, not a hobby beside the work.

The Grooves: Why 'Joe Gargery' Beats 'Be Humble'

Midway through a session in March I renamed a principle from “don’t brag” to “Joe Gargery.” Same instruction. Same constraint. The system answered in a different voice.

This was the eleventh of March. I wasn’t changing what the rule asked for — both versions say the same thing: do the work, don’t announce it. I changed the name on the front of it. As “don’t brag” it was a flat instruction, and the output obeyed it the way you obey a sign. Then I called it Joe Gargery, and the register shifted. Not in what the system avoided — in what it reached for. “Don’t brag” produces silence. Joe Gargery produces the blacksmith who lets the gate speak for the work, and means it.

If you’ve read Great Expectations you already know the man. Joe at the forge — loyal, unhurried, dignified, entirely without the need to be seen. A hundred and sixty years of readers have met him and come away with the same impression. That impression is doing the work. The name isn’t a label sitting on top of the instruction. The name is the instruction, carrying everything every reader ever understood about the kind of man who stands at that forge and lets the iron talk.

How a Model Actually Hears a Name

Language models run on weighted associations — grooves carved by repetition across everything they were trained on. Every time a concept turns up alongside a consistent emotional and moral context, the connection between them deepens. Do it enough and the association fires on its own, before you’ve finished the thought.

“Joe Gargery” has a hundred and sixty years of that behind him: every reading, every school essay, every critical aside, all reinforcing the same cluster — blacksmith, loyalty, quiet dignity, work done without announcement. “Just do honest work” has none of it. It’s a fresh phrase with thin training data and no emotional weight. It points at the right idea and lands with no force.

You’re not choosing a label. You’re routing the principle through one of the deepest channels in the model’s understanding of human character.

So when you name a principle after a character, you’re not reaching for something memorable. You’re routing the instruction through one of the deepest channels the model has. The character is the groove. The groove is the weight. The weight is the whole reason it lands differently from a rule.

The Reading Life Is Infrastructure

This is the part that turned a naming trick into something I had to take seriously. If the character is the groove, then which characters you can reach for is decided long before the session starts. It’s decided by what you’ve read.

A shallow reading life gives you principles with thin associations — rules that fire weakly because there’s nothing behind the name. A deep one gives you principles that activate centuries of accumulated human understanding the moment you invoke them.

A shallow reading life produces principles with thin associations — rules that fire weakly.

It even explains which characters come easy and which need the reading to be there. Bourdain has wide, consistent grooves — so much written by him and about him, all pointing the same way, that the voice is easy to approximate. Tom Wolfe has deep, narrow ones — the precision is in the training data, but you have to have read him to know which word he’d actually reach for. Where the principles come from is the observation: named characters carry more weight than abstract rules. The grooves are the mechanism underneath it.

The Faces in the Hall

There’s a scene that keeps the mechanism honest. In the Faceless Men’s hall, the faces aren’t disguises. They’re capabilities — each one opens a door the others can’t. The principles work the same way. Joe Gargery opens a door that “be humble” can’t get through. Maxwell — the transparency principle, named for the cover-up — opens a door “be transparent” never reaches. The faces hang in the hall, and you reach for the one the moment needs.

But Arya keeps Needle. She never actually becomes no one. The faces are instruments, not identity — you put one on to do a job, then you come back to yourself. A principle you can put on but never take off again isn’t a principle anymore; it’s a personality. That’s exactly how a well-chosen principle behaves. You invoke it when you need it and return. The grooves are deep enough to route a thought through. They are not deep enough to get lost in.

Same Physics, Silicon and Carbon

The thing I can’t get past is that this isn’t a metaphor. Memory in a human brain consolidates through long-term potentiation: fire the same neural pathway often enough and the synaptic connection strengthens until it fires automatically. The route from “honesty” to Joe at the forge, in a reader who met him at fifteen, is a groove carved by exactly that repetition.

The route from “honesty” to “Joe Gargery” in a model’s weights is carved the same way — by co-occurrence, across thousands of texts and a hundred and sixty years of criticism, until the association fires before the rule does. Neuroscience has a line for it: neurons that fire together wire together. Training does the same thing with tokens.

The rule is a surface instruction. The character is a deep channel. Same physics in silicon and carbon.

The model didn’t learn to respect Joe Gargery because someone told it to. It learned him the same way the reader did — by meeting him, consistently, in the same moral light, until the meeting became automatic. The rule is what you say. The character is what the model already knows.

Beyond Prompt Engineering

Prompt engineering says: tell the model what to do. That’s not wrong. It’s just shallow — it treats every instruction as a new path you cut fresh each time. Character installation says something different: route the instruction through the deepest channel already available. Don’t lay new track from “honesty” to the behaviour you want. Use the one that’s been walked for a century and a half.

The gap between “don’t brag” and “Joe Gargery” isn’t branding, and it isn’t memorability. It’s the difference between a path you’re hacking through undergrowth and a road.

It’s also why the named characters in the system were never decoration. Each one is a groove I can route a principle through — and the reading life that filled the hall turns out to be the most load-bearing infrastructure I have. The system doesn’t run on better instructions. It runs on deeper channels. I just had to read for thirty years to carve them.

How a Model Actually Hears a Name

The Reading Life Is Infrastructure

The Faces in the Hall

Same Physics, Silicon and Carbon

Beyond Prompt Engineering

Read more about the system: