I asked a model a question about database indexing last week and got a good answer. It gave me four strategies, when each one applies, and a note about write amplification I had not considered. I read it, nodded, and closed the tab.
An hour later it started bothering me and it took a while to work out why. The answer was fine. That was the problem. I had asked a question and received an answer, which is exactly the transaction I set up, and what came back was the most reasonable thing anybody could possibly say about database indexing.
I put money in. Something fell down the chute. I took it and walked off.
That is how most people use these things every single day, and it is the main reason they think the ceiling is a lot lower than it actually is.
Models Are the Generalization Agents
I said this in a working session a few months ago and I keep coming back to it:
Models are generalization agents. You’re not gonna find artisan or boutique products in them. They’re Walmart. They’ll have everything, but the most generalized version of it.
Walmart is not a criticism. It has everything, it is close by, it is open, and it is cheap. If you need a phone charger at nine at night, Walmart is the correct answer.
But nobody walks into a Walmart looking for the thing nobody else has.
A one-sentence prompt returns the centroid of everything ever written about the subject. Not the best answer. The middle of the distribution, which is the safest possible place for the machine to stand. And when the question is unremarkable, this is what you want. Most of my questions are unremarkable. Most of yours are too.
The trouble starts when you bring it something that actually matters to you and it hands you the middle of the distribution for that too, and it sounds so reasonable that you never notice you got handed the average.
The Flattening Happens Somewhere Surprising
The standard story is that RLHF ruined the models. Safety alignment sanded off the interesting edges, the alignment tax made everything bland, and if only we could get at the raw base model we would all be writing like Nabokov. It is a satisfying story because it has a villain.
It also appears to be mostly wrong about where the loss occurs.
A 2026 analysis of the OLMo 3 post-training lineages by Constantinos Karouzos, Xingwei Tan and Nikolaos Aletras (arXiv:2604.16027, a third-party analysis, not an Ai2 publication) goes looking for the exact stage where output diversity dies. The chain-of-thought distilled lineage keeps 38% of the base model’s diversity. The instruction-tuned lineage keeps 34%. Different roads to the same floor: the CoT line collapses at supervised fine-tuning, while the Instruct line bleeds most at DPO, the preference step most people mean when they say RLHF. The lineage trained with RL only, skipping both of those bottlenecks, keeps the most, at least 71% with a median of 94%, though it lands around half the distilled model’s score on grade-school math.
Read that again, because it inverts the folk explanation. The stage everybody blames is the stage that preserved the most.
I am not going to pretend this resolves into a clean recommendation about which model to pick, because it does not, and anybody selling you that certainty is ahead of the evidence. One paper and one model family does not settle. What it suggests is enough for this post. The flattening is real, it is measurable, and it is baked in upstream of anything you type. You are not going to prompt your way around a property of how the thing was trained.
Which means the answer was never a better sentence. It was never a magic phrase, never “act as a world-class expert,” never a jailbreak, and never the prompt template somebody is selling you on a newsletter. Those are all still questions. They are just longer questions.
A Question Gets You an Answer. A Move Gets You Something Else.
Here is the distinction the rest of this series runs on.
A question is a request. You describe what you want and the machine returns the most defensible version of it. The transaction is complete, the chute has delivered, and the quality of what you get is capped by how average your subject is.
A move is different. A move does something to the problem, or to the model, that makes the centroid unavailable as an answer. You force two unrelated things into the same frame and make it reconcile them. You impose an arbitrary rule that rules out the obvious response. You take the problem apart into parameters before you try to solve any of it. You ask for the answer from four different seats and pool them afterward. You let something genuinely random pick a direction. You invert the question and solve the opposite.
None of that is prompt engineering. It is closer to what a person does when they are stuck and finally stop staring at the thing.
That is the whole series. Eleven more posts after this one, each one a family of moves, each one with the mechanism explained well enough that you could build your own variants instead of copying mine.
Most of These Are Older Than Computing
Almost none of these techniques were invented for AI. Forced relationships, morphological analysis, the deliberately terrible idea, writing under an arbitrary constraint, brainwriting, defamiliarization. They are decades old at minimum, some of them a century. They were developed by novelists, engineers, design researchers and a few outright lunatics, all working on paper, and they have survived decades of use by people who were stuck.
It turns out a lot of them also work on models.
So every post in this series does two jobs. The human version comes first and stands on its own, which means you can run it today with a pen and no account anywhere. Then the translation, which is what the same mechanism looks like when there is a model in the loop. If you have no interest in AI whatsoever you should still be able to take something usable out of every one of these. If that ever stops being true, I have written the post wrong.
The Half You Cannot Hand Off
I want to be honest about the limit here, because the genre this post belongs to is usually dishonest about it.
I once spent a long working session designing a system with a model, and afterward I did the math and realized I had spent roughly as long as it would have taken me to just write the thing myself. The win was not speed. There was no speed. The win was that my attention went into judgment instead of getting burned down in the weeds of debugging, and I came out the other side understanding the design better than I would have.
The less comfortable part of that same session: the model made three confident predictions about how the system would behave and every single one was wrong. That is the mild version of a failure mode I have hit much harder, back when a self-hosted agent of mine hallucinated fake news for two days before I noticed. The two ideas that actually survived contact came from me. One of them came out of a completely unrelated domain that the model had no reason to connect to the problem and never would have.
So the moves in this series are not a way to get the machine to think for you. They are a way to stop getting the average. You still have to supply the taste, the judgment, and the weird connection from the unrelated thing you happened to know. What you get to outsource is the keystrokes, and honestly, the keystrokes were never the hard part.
Stop putting in a dollar and taking whatever falls out.
Next one: force two unrelated things together, and watch what the machine has to do to make them fit.
Top comments (0)