Harness

Chapter 1 · Plain words first: the gear it wearsStart with a horse. How fast a horse runs is half the horse, half the gear: reins, saddle, shoes — every piece has to fit. An AI works the same way. The brain in the middle is the large model. The kit fitted around it is the Harness. In plain words, it is everything at its side while it works: are the rules written down, are the tools all there, what may it touch and what may it not, does it remember what it did, and does anyone check when it finishes. This episode is only about that kit.

Chapter 2 · The scene: new horse, same old storyYou have seen this scene. It botches the job again, and somebody immediately says: swap in a smarter one. So you swap. You even try the pricier one. Same failures — down to the way it fails. Hold off on blaming the horse. In the song, a legendary fast horse has slack reins, a crooked saddle, and one shoe lying on the ground. With gear like that, running steady would be the surprise. Most of the time the model is not the worst part; nobody has crouched down and gone through its gear piece by piece.

Chapter 3 · The working loop: think — do — look — think againGear alone is not enough; it has to turn. The work goes in a loop: think first, then do; once done, take a look; having looked, think again, and around it goes. You have met this loop — Episode 1, “Loop After Loop,” was about exactly this loop. That episode was about how the loop makes it smarter with use. This one steps back a pace: every step of the loop needs gear under it. Thinking needs the rules and the material at hand. Doing needs tools in hand. Looking needs the rules set earlier, used as the ruler. The loop is how it turns; the gear is what it turns on. With both, the turning stays steady.

Chapter 4 · Mind both ends: rules before, checks afterThe trick sits at the two ends. Before the work starts, set the rules — the trade calls this feedforward: the ruler is fixed before a hand moves. After the work, check the result against those rules — feedback: right or wrong, measured, not guessed. And keep the most valuable move for last. When a check turns up a mistake, do not let it go to waste: write the lesson back into the rules as one more line, and next time it steers around that pit by itself. That is how a rulebook grows — one line at a time. Same horse, new gear, and the work looks like a different animal did it. The brain never changed.

Five pieces of gear · One step at a time

1. Rules. What they can do: write down, up front, how the job is done, in what order, and what counts as good enough — so it follows the page instead of guessing. What they can’t: do the work for it. Write a rule wrong, and it will follow the wrong rule faithfully to the end.
2. Tools. What they can do: let its hands reach out — to look things up, to calculate, to fetch real material — instead of answering from memory alone. What they can’t: make it use them well. A tool left on the bench, or the wrong one picked up, and the job still goes wrong.
3. Boundaries. What they can do: draw the line between what it may touch and what it may not; for anything that matters, it asks a person first, and a dangerous move gets stopped on the spot. What they can’t: make the answer better. Boundaries guard safety, not quality.
4. Memory. What they can do: lay out the material before the work starts, and write what was done into a ledger, so next time it opens the ledger and picks up where it left off. What they can’t: guarantee the ledger is right. Write an entry wrong, and next time it follows the wrong entry with complete confidence.
5. Checks. What they can do: after the work, run it back against the rules and pull the mistakes out on the spot, before they slip through. What they can’t: invent the ruler. Anything the rules never mentioned, a check cannot catch — not in one pass, not in a hundred.

Sources for this episode

The concept frame comes from Wang Haobai’s “Harness Engineering: An Introduction,” shared at the Osaka AI Meetup on 2026-10-25, and from a public video topic a viewer recommended. Only the topic and the frame were borrowed; the lyrics and these notes are original to this site.

Closing question

Take the closing question with you. Next time it lets you down, before you swap anything — which one do you check first? Are the rules stated clearly? Are all the tools there? Are the boundaries drawn? Did it bring its memory along? Did anyone check the work when it finished? Of the five, pick the loosest one, and tighten that one first.

Process note

The captions in this episode were not split evenly by duration. We listened to the song line by line and measured when each line actually starts: in the chorus and the bridge, a few lines sit almost two seconds away from where an even split would put them, and an even split would leave the words off the singer’s mouth. The horse on screen is drawn the folk New-Year-print way, in a handful of lines — not a lifelike horse, just one you recognize as a horse at a glance. The gear diagram in the chorus never changes the picture: same horse, same drawing, with the five pieces lighting up one by one as the chorus names them. When all five are lit, the horse runs steady.