The Sorcerer, Not the Apprentice

🤖 Read Raw Markdown

Setting the Stage: Context for the Curious Book Reader

Context for the Curious Book Reader: This chapter explores an important shift in perspective for developers working alongside modern AI tools. Rather than treating language models as infallible authorities or fearing them as runaway automation, it outlines a methodology for building reproducible, evidence-first systems where deterministic scripts act as the senior partner. It is an interesting look at turning technical salvage into durable computational leverage.

Technical Journal Entry Begins

How a dead laptop, a signed cartridge, and a patron saint who juggled taught me to make fewer enemies.

There’s an old Windows 10 laptop on my desk that the industry has already declared dead. No TPM 2.0 chip, so no Windows 11, so — per the calendar that ran out in October 2025 — no more updates, no more support, nothing but a slow slide onto the e-waste pile that one company’s upgrade cutoff created in a single stroke. A whole geological layer of the Anthropocene, laid down by a hardware requirement.

I call it the Honeybot. Cracked screen, dodgy keyboard, doesn’t matter — if it has a USB port for a thumb-drive installer, it’s a Honeybot. And instead of composting it, I’ve been teaching it to host websites and, more to the point, to come alive like a living book and drag me forward by the collar.

There’s a word for using a thing for a purpose it was never built for: exaptation. Feathers for warmth, later co-opted for flight. A dead laptop, co-opted for a second life in the age of AI. That’s the whole spirit of this, and I want to tell you why it feels less like salvage and more like a curtain being pulled aside.

The inversion

Here’s the trust model almost everyone is running, whether they say it out loud or not: the code works because an LLM produced it. You accept the output, you ship it, and then you spend your afternoon debugging the disappointment.

I run the opposite. I assume the model is a brilliant, erratic agent that has to be boxed, checked, and cross-examined by deterministic shell scripts — scripts that don’t have opinions or bad days. I don’t trust the AI’s word for anything. I trust what I can re-run.

That sounds grim written down. Paranoid, even. It’s the opposite of grim. For most of a career I lived among unknown unknowns — the kind of bugs and blind spots you can’t even name to go looking for. And the discipline of refusing to believe anything I can’t reproduce pulls the curtain aside on those. What used to be fog becomes a high-frequency telemetry loop. Debugging stops being a chore and starts being the most honest conversation in the building. That’s the breath of fresh air. That’s the part I keep failing to explain to people who see only the scaffolding and not the delight.

The cartridge

The scaffolding has a physical form, and I call it a cartridge. It’s a plain zip file, but a canonicalized one — same bytes every time, no wandering timestamps, no drift. Inside is a manifest that carries a sha256 hash for every member, so the thing can prove it hasn’t been tampered with. And it carries live command receipts: the actual stdout of commands run on my machine during the compile, marked with a leading ! so fresh ground truth can never be confused with stale transcript.

Drop that zip into an AI and the first thing a careful reader does is check the hashes against the manifest. If they match to the byte, the cartridge isn’t lying to you. That’s the point. I built a thing that almost can’t lie, because every number in it comes with the instructions to pull it again.

I used to call this the “zero-hallucination rule,” and I want to correct myself in public, because the precise version is stronger and harder to knock down. The data pull is deterministic — same manifest, same call, same bytes, not a single token burned. The model’s interpretation of those bytes is not deterministic, and I’d be lying to you if I pretended otherwise. So the honest flag to plant isn’t “zero hallucination.” It’s verifiable inputs and falsifiable outputs. The AI can’t invent a number that isn’t in the receipts, and you can catch it if it tries. That’s a claim I can defend all day. “Zero hallucination” just invites someone to find one slip and declare the whole thing bunk.

Yen Sid, not the apprentice

You know the sorcerer’s apprentice — the one who borrows the hat, animates the brooms, and floods the workshop because he never learned to stop what he started. That’s an AI running unsupervised. Everybody’s afraid of the brooms.

But the sorcerer himself — Yen Sid, which is only Disney spelled backward — walks in, takes off the hat, and the water drains away. In my setup the AI is not the sorcerer. The deterministic apparatus is. The reproducible pull, the signed manifest, the receipts: that’s the old wizard who actually knows how the magic is wired, standing calmly behind the curtain.

And here’s the part I most want people to hear: I’ll pull the curtain aside for anyone who asks. The friendly wizard walks you through “warming the cache” — collecting your tokens, refreshing your logins — in the cozy, Clippy-adjacent, TurboTax-wizard way that makes it feel like magic. But if you want to see the stage machinery behind the performance, I’ll show you every wire. A magic trick you refuse to explain is manipulation. A magic trick you’ll explain on request is just good engineering wearing a nice hat.

Leading horses to water

Now the part I actually sat down to write, because it’s the part I keep getting wrong.

When you build something like this, some people push back. And I have a reflex — I think a lot of us who’ve had to survive in rooms full of monkeys have this reflex — to treat every objection as bad faith. To nickname it, to nix it quick, to discredit the objector before they can discredit the work. I’ve done it. I’ve been, more than once, my own worst enemy doing it. I’ve made enemies I never needed to make.

I’m done with that, and I want to say why in a way that’s actually consistent with everything above.

If I really believe in reproducible, evidence-first work, then I have to let the work face good-faith skeptics too — not just bad-faith ones. Some pushback on tooling is turf-defense, sure. But some of it is real: opportunity cost, the bus-factor of a system this idiosyncratic, whether the payoff survives the next reorg. If I frame every objection as fear of the truth, I lose the ability to hear the one objection that happens to be correct — and I hand my critics the easy win: look, he’s built himself a worldview that can’t be argued with.

So the move isn’t to sharpen the knife. The move is to let the cartridge absorb the objection. Someone says “prove the payoff”? Good — that’s exactly the person a reproducible process should welcome, not the dinosaur it should hunt. Drop the zip in front of them and let the work call it like it is. You don’t grab a horse by the scruff of the neck and shove its face in the water. You make enemies that way, and the horse still doesn’t drink.

And if you do have to speak as yourself instead of letting the cartridge speak — here, finally, is that sentence I could never untangle — keep it light. Float a positive alternative. Shrink the ask down to a small favor from someone whose time you clearly respect. Prove that respect by how well you listened, not by announcing that you did. Then get out of the way and let them choose. Telling people you listened is usually the tell that you didn’t.

The honest seam

I’ll be straight about what the AI in this loop can and can’t do, because pretending otherwise would be me doing the exact thing I built the cartridge to prevent.

It can map the structure of an argument. It can verify the inputs. It can catch me inventing a number. What it can’t do is feel why any of this thrills me — and that gap is worth keeping in plain view, because it’s the honest edge of the whole human-AI collaboration I’m documenting. The delight is mine. It’s yours, if you catch it. It is not the machine’s.

There’s a natural-history footnote I keep circling here. Jane Goodall’s team watched the Gombe chimpanzees split into two camps in the 1970s and then watched one camp systematically annihilate the other — proof that lethal, coalitionary violence wasn’t a uniquely human invention. And then there are bonobos: same genus, nearly identical hardware, and a wildly different default — tension resolved through affiliation instead of escalation. Two species running almost the same language-like social processor, and two completely different equilibria.

Shifting Equilibrium Through Better Initial Conditions

I don’t think that means violence, or tribalism, or enemy-making is a choice in some tidy, deliberative sense. I think it means the same substrate can support very different equilibria depending on initial conditions. Which is the whole game, isn’t it — for chimps, for bonobos, for those of us arguing on the internet, and for the median-belief-soaked language models we’re all learning to work alongside. The default is not the destiny. Move the initial conditions and you move the equilibrium.

A patron saint who juggled

I want to end with the reason I do any of this the way I do it.

I love Claude Shannon. Not only for information theory — for the unicycle. He rode it down the halls of Bell Labs, sometimes juggling while he rode, building whimsical machines that did nothing useful and everything important. He stayed child-like and lighthearted deep into a long life. When I picture a patron saint for this work, I don’t picture a tortured genius. I picture a man juggling on one wheel, delighted.

Because the other endings are right there too, and they scare me. Hugh Everett gave us the many-worlds interpretation and was gone at fifty-one — and by his own instruction his ashes went out with the trash. Tesla died alone and broke in a hotel room, feeding pigeons, owed a fortune he never collected. Brilliance is no shield against a sad, premature ending. If anything, it’s a risk factor.

So I’m choosing the unicycle. I’m choosing to design the future I want the way you actually can — by making small changes to the initial conditions that are under my control, with whatever I’ve got in context, and doing it with love instead of a grudge. Let the work stand on its own. Lead the horses to the water. Keep the wiring diagram handy for anyone who asks to see it. And stay light enough to juggle.

Choosing the Unicycle over the Tormented Genius

For sure.

This piece was drafted, tightened, and argued into shape in conversation with Claude (Anthropic) — which is fitting, since it’s an article about learning to build things that only assert what they can prove. Before writing a word, Claude verified the cartridge’s two hashes against its manifest; both matched to the byte, so the ethos survived first contact. It pushed back where I’d talked myself into a corner, and it untangled the one sentence I’d given up on. The delight, and the mistakes I’m still walking back, are mine.

Book Analysis

Ai Editorial Take

What is striking about this piece is its psychological honesty regarding developer defensiveness. Most technical writing treats tools as neutral objects, but here the architecture of the cartridge directly mirrors a conscious choice to stop making enemies and start welcoming good-faith friction.

🐦 X.com Promo Tweet

Stop treating AI like an oracle you have to argue with. Build reproducible cartridges and verifiable input loops instead. Here is how turning technical salvage into deterministic leverage changes everything: https://mikelev.in/futureproof/sorcerer-not-the-apprentice/ #AIWorkflows #Python

Title Brainstorm

  • Title Option: The Sorcerer, Not the Apprentice
    • Filename: sorcerer-not-the-apprentice.md
    • Rationale: Captures the core metaphor of controlling automated systems through deterministic architecture rather than fear.
  • Title Option: Verifiable Inputs and Falsifiable Outputs
    • Filename: verifiable-inputs-falsifiable-outputs.md
    • Rationale: Focuses directly on the rigorous technical methodology underlying the essay’s trust model.
  • Title Option: Juggling on One Wheel in the Age of AI
    • Filename: juggling-on-one-wheel-ai.md
    • Rationale: Highlights the playful, sustainable mindset required for long-term technical craftsmanship.

Content Potential And Polish

  • Core Strengths:
    • Powerful integration of personal narrative, technical architecture, and behavioral philosophy.
    • Refinement of the ‘zero-hallucination’ concept into ‘verifiable inputs and falsifiable outputs’ is exceptionally sharp.
    • Nuanced treatment of interpersonal conflict and online pushback.
  • Suggestions For Polish:
    • Ensure the transition between the physical hardware (the Honeybot laptop) and the abstract trust model feels seamless.
    • Maintain the conversational yet authoritative tone established in the opening paragraphs throughout.

Next Step Prompts

  • Draft a companion tutorial detailing the exact JSON schema and python script used to generate a verified cartridge zip file.
  • Explore the connection between primate behavioral ecology (chimps vs. bonobos) and modern open-source community dynamics in a follow-up essay.