← Back to codingwithaibook.com

From the book · Chapter 20.5

Escaping the Hivemind: Colliding Your Problem with Distant Domains

The agent is very good at executing an idea. It is quietly terrible at giving you a good one — and the reason is more interesting than it first appears.

The thirty-second idea

Ask Claude for ten ways to redesign your onboarding flow. Read them. Then open a fresh window and ask again. The wording will be different and the framing will be different, but the substance will rhyme. You will get progressive disclosure, a checklist with a progress bar, contextual tooltips, a "skip for now" escape hatch, and a celebratory confetti moment at the end. These are not bad ideas. They are the ideas anyone in the room would have surfaced in thirty seconds, which is precisely the problem. You did not need a model trained on a meaningful fraction of human writing to tell you to add a progress bar.

Researchers have started calling this the Artificial Hivemind. Jiang and colleagues, writing in 2025, sampled the same prompt hundreds of times across a range of models and found that the outputs cluster in a narrow zone, returning again and again to the same handful of obvious answers. The model is not malfunctioning when it does this. Converging on the common answer is what a system trained to predict the most probable next token is built to do. Ask it for the expected thing and it will hand you the expected thing, polished and confident, every time.

Why "be more original" doesn't work

The instinctive fix is to argue with the model. You tell it to be original. You tell it the obvious answers are off the table. You paste in three more paragraphs of context about your users and your constraints and your competitors, on the theory that a better-informed model will think better thoughts. Both moves feel productive. Neither does much.

The reason is worth sitting with, because it is counterintuitive. Adding context does pull the response away from the generic center, and toward the territory you described. But you can only pull the answer toward what you already put in the prompt. The richer your brief, the more tightly the model converges on your framing of the problem, your known reference points, the solution space you were already standing in. You escape the average and land squarely back in your own head. The output is more original from the perspective of a stranger and not at all original from yours. Telling the model to "be creative" is worse still: it has no idea where the unexplored territory is, so it gives you its statistical impression of what creative-sounding output looks like, which is itself a cliché.

Koestler's collision

There is a move that does work, and almost nobody makes it by instinct. Instead of adding more material from your own domain, you inject material from a domain that has nothing to do with your problem.

The idea is older than language models. In 1964 Arthur Koestler argued in The Act of Creation that genuine creativity is born when two ideas from unrelated frames of reference are forced to collide. He called it bisociation, and his examples were the great conceptual leaps of science and art, the moments when someone looked at one thing through the lens of a completely different thing and saw what no one inside either field had seen. Hofstadter and Sander made a related case in Surfaces and Essences in 2013: thought runs in grooves carved by everything you have already thought, and a mind left to its own devices will follow the path of least resistance straight back to the familiar. Drift, on its own, is conservative. It takes a foreign element to knock you out of the groove.

Picture the space of all possible ideas as a landscape, with similar ideas sitting near each other and unrelated ideas far apart. A normal prompt drops the model into one neighborhood and the answer lands somewhere nearby. A collision does something different. Take two genuinely distant subjects, say parasitology and platform economics. On the surface they share nothing: different vocabulary, different domain, different century of study. But the space of ideas has an enormous number of dimensions, and along some hidden dimension the two might sit much closer than they look. Parasitology studies organisms that hijack a host's behavior for their own reproductive benefit; the parasite that makes an infected rat lose its fear of cats is the canonical case. Platform economics studies systems where users generate value for the platform while feeling they are working for themselves. Underneath, both are describing the same mechanism: behavioral redirection for asymmetric benefit. Bisociation is the discovery of that hidden dimension, the one in which two distant things turn out to be the same shape.

This is also why the move is a gamble rather than a guarantee. The hidden connection has to actually exist. Some pairs of subjects are distant in every dimension, and forcing them together produces nothing but noise. You cannot know in advance which collisions will land, which is exactly why the technique demands volume and curation: throw many distant domains at the problem, accept that most will miss, and keep the few that strike something. Novelty, it is worth remembering, is not the same as value. A collision can be genuinely original and completely useless. The point of generating a lot of them is that you only need the good ones to survive the filter.

The same brief, run two ways

A concrete case makes the difference visible. Cédric Lion, who built and open-sourced a pipeline called Open Collider to operationalize this, used the following brief: structural redesigns of Spotify's Discover Weekly that break a user out of their taste bubble. Same brief, same transcript, same underlying model. The only thing that changed was whether the prompt contained a distant-domain collision.

The default ideas are reasonable mechanisms, all of them — Spotify's own product team could have proposed any one in a brainstorm, because they live in the same conceptual neighborhood. The collision ideas might be wrong, unworkable, or too clever by half. But none of them is anywhere near the dense, obvious center, and none would ever have come out of a prompt that only knew about music.

The part where someone tried to prove it wrong

It would be easy to stop at the Spotify example and call it a nice story. The more useful thing Lion did was try to break his own claim. He set up two honest falsifiers and ran them in parallel against the real technique, across twelve real projects, roughly twenty-three thousand generated ideas, and over four thousand blind pairwise judgments.

The logic of the test is clean. If the collisions only worked because the collision prompt happens to be longer and more detailed, then a long, in-domain "deep brief" with no distant material should reproduce the effect. It barely moved. And if the collisions only worked because they explicitly told the model to be original, then a prompt that simply instructed it to "be original, combine concepts from distant fields" should reproduce the effect too. It also barely moved. Measured as distance from the baseline cloud of ordinary answers, the genuine collision pulled the ideas away from the center on all twelve projects, by a wide margin, while the two imitators produced a small fraction of the shift. The instruction alone did not do it. The extra length alone did not do it. The actual collision of distant domains did.

Distance is not quality, of course; ideas can be far from the center because they are absurd. So the second half of the test asked three different models from three different vendors to judge anonymized pairs on which idea was more original and which was the better one to actually pursue. On originality the collision won roughly two times in three. On "better idea overall" it came out tied with the alternatives rather than ahead, which is the honest result: the technique buys you more genuinely original ideas without making them worse on average. Given that the entire game is to surface a few non-obvious ideas worth keeping, a wash on average quality with a clear gain in originality is exactly the trade you want. None of this is the last word — it ran on a single generation model, and the author is the first to list what he has not yet controlled for — but it is a good deal more than a vibe.

How to run a collision yourself

You do not need the full pipeline to use the mechanism. The portable version is a prompt structure, and it fits inside the explore step of the build loop, the moment when you are still asking what to build rather than how to build it. Start with your real brief, then — instead of enriching it with more of your own context — bolt on a few deliberately unrelated domains and force the model to build the bridge.

{your brief and any relevant transcript or context}

DOMAIN BANK (deliberately unrelated to the problem):
- Mycology: how fungal networks route resources to the parts of an
  organism under stress
- Naval logistics: how a fleet refuels and rearms without returning to port
- Restoration ecology: how a degraded ecosystem is seeded for recovery

For EACH domain above, find the hidden structural similarity to my
problem, then generate 15 ideas that bridge my brief to that domain's
core mechanism. Do not file the edges off to make them sound reasonable.

A few things make the difference between signal and noise. Pick domains that are genuinely far from your problem; the safe, adjacent ones ("how other apps do onboarding") collapse straight back into the Hivemind. Run each collision in its own context window rather than asking for everything at once, so one domain's framing does not bleed into the next. Generate in volume, because most collisions will miss and you are buying lottery tickets. And then curate hard — this is the step people skip and the one that makes the technique usable. The output of a collision run is raw ore, not finished product; your job, or a second pass with a model acting as a ruthless filter, is to throw out the eighty percent that is clever nonsense and keep the few that quietly reorganize how you see the problem.

This is the same skill the rest of the book keeps returning to: the agent generates, and you verify and select. Architecture decision records are a natural home for it, because the value of an ADR lives in the alternatives you seriously considered, and a collision pass is a cheap way to make sure the alternatives were not all the same idea wearing different hats. So is the non-coding work — RFPs, vendor reviews, incident retros — where the task is often to produce an angle no one else in the room will have thought of.

This is Chapter 20.5 of Software Engineering with AI. The pipeline and full dataset behind the experiment are open source at github.com/CL-ML/open-collider and open-collider-research (Cédric Lion, CC-BY-4.0).