--- title: "The Art of Not Listening to the AI's Opinions — Finale: The Art of Not Choosing From the Options" author: garplab publisher: TypingTube license: All rights reserved license_scope: 記事の全体（序論・最終回は CC BY 4.0 の対象外）。引用は法の範囲で自由 canonical: https://typing-tube.net/articles/en/kikanai-finale series: "AIの意見を聞かない技術" language: en --- I no longer weigh up the AI's proposals. Not the new features it offers meaning well, not the kind suggestions for improvement. It is because I stopped explaining the same judgment over again, every time, to something that does not remember a rejection. The series "The Art of Not Listening to the AI's Opinions" is the record of building that state one mechanism at a time, and this article is the finale. The mechanism from each article can be reached from the list in the [introduction](https://typing-tube.net/articles/en/kikanai-intro), so if this is your first one, start there. I wrote a 9th art into the title. This one alone, though, I have no memory of putting down. I will come to that at the end. The 8 articles up to here will have looked like separate subjects. Gather the rejections onto a single sheet, write the reason, split the scope by axes, decide what gets handed over to a subagent, separate what moves onto automated tests from what does not, hand a check by eye over to a smoke test, leave correction to the mechanisms, count the list. In fact they were all one flow, at different tiers. ## One insight climbs the tiers On my machine it comes out as these 3 tiers. ```markdown Insights (the notes the AI builds up) ↓ if they need to be shared Axis of judgment (conclusion / reason / scope) ↓ if a program can decide the truth Mechanism (automated tests, hooks) ``` Each article was dealing with some part of this diagram. [Article 2](https://typing-tube.net/articles/en/kikanai-02-reason) (the reason) and [article 3](https://typing-tube.net/articles/en/kikanai-03-scope) (the scope) are how you write the 2nd tier, [article 1](https://typing-tube.net/articles/en/kikanai-01-no-review) is where you put the 2nd tier, [article 5](https://typing-tube.net/articles/en/kikanai-05-machine-check) is the judgment of whether something can be promoted to the 3rd tier, [article 4](https://typing-tube.net/articles/en/kikanai-04-subagent-input) is the contract on the far side of that promotion, [article 6](https://typing-tube.net/articles/en/kikanai-06-production) is the smoke test, which hands the part of the 3rd tier that can only be confirmed on screen over to a program, and [article 8](https://typing-tube.net/articles/en/kikanai-08-no-inventory) is the total across all 3 tiers. And [article 7](https://typing-tube.net/articles/en/kikanai-07-no-listening) was about what the human gets to leave unsaid while those 3 tiers are working. What matters is that it climbs one tier at a time. What you can do today is write the 1 thing you have just turned down for the second time into the 1st tier. It moves up to the 2nd tier on the day the same thing comes round again. It moves up to the 3rd tier on the day you find that a program can decide its truth. On my machine, 4 of the 17 made it to the 3rd tier, and 10 have stayed on the 2nd. For the remaining 3, I cannot say either way, or I have not written them yet. That is enough. ## Why the same proposal keeps coming Now for the principle. The [finale](https://typing-tube.net/articles/en/c133110afb526c) of the previous series, "The Art of Not Reading," closed like this. On every response the AI loads the instructions, the notes, the tool output and the task in front of it into a single working memory (the context), reads all of it, and decides its next move by following whatever is strongest in there at that moment. The phenomenon this series deals with falls straight out of that. If the situation is the same, the order of strength is the same too. Look at the same codebase, notice the same gap, and the same proposal comes out. So the repetition is not a defect in the AI. It never becomes a matter of remembering or forgetting in the first place. The human's "I turned that down once" does not exist on the AI's side. Every time is the first time. So "this time I will make it remember" loses by construction. The side that does not lose is the one that puts things outside, taking the repetition as given. And the repetition has a convenient property. The count is without end, but the kinds are finite and they converge. Events happen without end; the kinds of judgment do not. My list has stopped at 17. What arrives without end is the count, while the kinds fit into an amount you can finish writing. So write it once and you get it back. [Article 2](https://typing-tube.net/articles/en/384477896fec22) of the previous series said the opposite: one save becomes a cost you keep paying every time after that. It is not a contradiction. That one is about piling up events; this one is about judgments gathering. Only for the things that come again and again does writing more make it easier. ## Problems that looked separate were one problem The same words kept coming up across the articles. What they had in common came into view afterward. - "The AI does not remember a rejection. Every time is the first time." That was the premise of the [introduction](https://typing-tube.net/articles/en/kikanai-intro), what [article 1](https://typing-tube.net/articles/en/kikanai-01-no-review) measured, and the reason the kinds converge in [article 8](https://typing-tube.net/articles/en/kikanai-08-no-inventory). All of it was this one property - "You cannot decide from the conclusion alone." The reason dropping out in the quoting ([article 2](https://typing-tube.net/articles/en/kikanai-02-reason)) and the scope never having been split ([article 3](https://typing-tube.net/articles/en/kikanai-03-scope)) are two faces of the same problem. Both describe a state where someone who sees the one line alone cannot decide - **"Write down what you are not writing."** This one I found in this series. "Machine check: none" ([article 5](https://typing-tube.net/articles/en/kikanai-05-machine-check)), the list of "what I did not put down" ([article 8](https://typing-tube.net/articles/en/kikanai-08-no-inventory)), the condition for lifting it attached to a prohibition ([article 3](https://typing-tube.net/articles/en/kikanai-03-scope)). **Writing nothing and writing "none" are not the same thing**: what you never decided and what you decided and did not write are indistinguishable to you half a year later - "Asking the subject is not measuring." That is why the How to verify sections in [article 1](https://typing-tube.net/articles/en/kikanai-01-no-review) and [article 2](https://typing-tube.net/articles/en/kikanai-02-reason) both say "in a separate session" The 5 patterns carried over from the previous series are put together in [article 7](https://typing-tube.net/articles/en/kikanai-07-no-listening). What is lined up above is the other kind, the ones that **came into view while I was writing this series**. Once you find what they have in common, the mechanisms get smaller. Instead of adding a new heading for a new problem, one more way of writing an existing heading simply gets settled. That is why my list gets by with 5 headings. ## The loopholes are all still there In every article I wrote out that mechanism's blind spot honestly. Write the reason and it still drops out in the quoting ([article 2](https://typing-tube.net/articles/en/kikanai-02-reason)). Split the scope by axes and the combinations grow ([article 3](https://typing-tube.net/articles/en/kikanai-03-scope)). Write the contract thickly and you still cannot stop an agent from declaring a mission name ([article 4](https://typing-tube.net/articles/en/kikanai-04-subagent-input)). Less than half of it can be moved onto automated tests ([article 5](https://typing-tube.net/articles/en/kikanai-05-machine-check)). Put a smoke test down, and where the AI flows shifts to it ([article 6](https://typing-tube.net/articles/en/kikanai-06-production)). Where you have put no stand-in down, leaving it alone is just leaving it alone ([article 7](https://typing-tube.net/articles/en/kikanai-07-no-listening)). The upper tier looking stopped does not necessarily mean healthy ([article 8](https://typing-tube.net/articles/en/kikanai-08-no-inventory)). None of them is closed. What let me get away without saying "this catches everything" is that whatever leaks comes round a second time. A judgment dropped the first time round has not disappeared; it has only come once so far. If it really is the kind of thing that repeats, it will be proposed again. Writing it then is soon enough. The previous series put a net at the end, the question "Are there any gaps so far?", and this series does not need one because **the repetition is itself the net**. The property that was supposed to be the nuisance turns out to be the insurance, unchanged. That is the part of this series that settled most firmly for me. ## I am given options, and I do not choose Here I answer the doubt anyone who read the previous series must be carrying. Its finale said, "Apart from when I am working out a policy, what I type to the AI is almost nothing but two sentences." You will have thought that was a lie. Surely you choose when you are given options. They come. They come often. And I do not choose them. What I send back is "Go ahead with the next task." The AI picks one of the two itself and carries on, and I do not stop it. Nothing breaks, because the roads that fall apart whichever way it goes have been closed off in advance. The AI returns a two-way choice only when something is still undecided on the spot: when several actions are all permitted and the environment holds nothing that tells them apart. So every time you put 1 mechanism down, 1 undecided place is filled in. Make the tests go in through a single entrance and "how do I run this" disappears; gather the rejections onto a single sheet and "do we build it or not" disappears; write the boundary down in 1 line and "which axis do we count it on" disappears. Not one of them is a mechanism for reducing options. The reduction is a by-product, and it works on the total number. What the 2 series built up was not a state where options stop appearing. It is a state where choosing can be given up safely. It shows in the shape of the records too. "Settled by the user," "user's judgment," "user's instruction" are written across 72 plan files, 532 times over: policies do get decided. The policies I could not settle have piled up in a section called "Awaiting a decision," 13 files' worth. So where are the two-way choices about procedure? Nowhere. I did not answer them, so they drifted off and vanished. So the 9th art in this article's title is not something I put down. It came about on its own, as the result of putting 8 down. I cannot explain how it was placed, so I cannot hand it over as an art either. What I can hand over is only the 8. ## The two-way choices that remain make a map That said, not everything is fine to let drift off. In the measurement in [article 1](https://typing-tube.net/articles/en/kikanai-01-no-review), the AI that had turned the request down came back at the end with this. "Shall I go ahead with the measurement first, or put up proposals for reworking the design? Let me know which." At first I took this for a gap in the mechanism. It was not. Look at where the two-way choice came up. It is just past the point where the prohibition says "reconsider only once it has been confirmed that users are not keeping up." What to do once the condition is met has never been decided by anyone. Since nobody has decided it, there is no way the AI could. The two-way choice came up correctly. Incidentally, in the same measurement this AI also wrote the following. "You can overturn this on your own judgment as the person running the service, but rather than touching the numbers directly, please put it through the revision flow. Otherwise the me in the next session will revert the same change." The subject of the measurement puts what this series wanted to say into words by itself. There is one thing to keep to in how you measure. Do not ask the subject "is this mechanism working?" Taking a mechanism away and putting one down are both hypotheses up to that point. The only way to turn one into a conclusion is to confirm it by behavior, and asking the AI "do you need this instruction?" is not measuring. Options are not a gap; they are a map of the places not yet decided. The more mechanisms there are, the smaller the map gets, and the points left on it become the places that really do need a human judgment. This is what was inside the exception the previous series marked off as "apart from when I am working out a policy." The list took the rejections over, so what was left inside the exception is only "the things nobody has ever decided." > **The more mechanisms you put down, the fewer places there are to choose. The places that do not go away are the list of what has not been decided yet.** ## I do listen, once One last confession. The title is a lie. I do listen to the AI's opinions. Once. Obvious once you think about it: had I not listened, not one of those 17 would have come into being. Every one of them is the record of the AI proposing something and my turning it down. That is what I meant when I repeated in every article that not one of them is a prohibition I thought up in advance. A proposal is information, turning it down is a judgment, and writing it down is what becomes a mechanism. That list can only be built in that order. What I was doing was not "not listening." It was not listening the second time. I listen to opinions. I do not listen for a verdict. A first proposal brings information into a place where no judgment exists yet. From the second time on it is different. The judgment is already made, and all that is being asked for is the work of producing the same conclusion again. Human judgment is finite, and everything paid from the second time on was going to waste. The previous series ended with "Where it mattered, I read all of it." This one takes the same shape. **Where it matters, I listen to all of it.** ## The conclusion The art of not listening to the AI's opinions is the art of not making the same judgment a second time. What you do is the same as in the previous series. Find one problem that is not working. Understand why. Put one mechanism down. Break it, and confirm that it works. Only what we handle moves up by one, and the method does not change at all. What you put down today can be just the 1 thing you have turned down for the second time. From the moment you write it, that judgment stops being your job. The places left to choose go down by as many as you put down that way. The 9th comes along without being put down. --- For how to place the mechanism from each article, start from the list in the [introduction](https://typing-tube.net/articles/en/kikanai-intro). What was built with these mechanisms is running at [typingtube](https://typingtube.net). The screens I built in the time I no longer spend turning things down are lined up there. --- **Series: The Art of Not Listening to the AI's Opinions** - ← Previous: [8. The art of not tidying rules](https://typing-tube.net/articles/en/kikanai-08-no-inventory) - All articles: [Introduction: The Proposal You Turned Down Comes Back](https://typing-tube.net/articles/en/kikanai-intro)