When 'Be More Creative' Moves 0.04 But Collision Moves 0.28

Cédric Lion released an open-source tool this month called Open Collider. It's a small piece of software with a precise function: it takes a prompt asking an LLM for ideas, and it makes those ideas measurably better.

The reason this matters is in the title number above. Lion's actual measurement, validated against 4,320 blind LLM-judge verdicts and 12 different real-world ideation problems, is this:

• The prompt "be more creative" moves the LLM's output approximately 0.04 units in embedding distance from its default response.
• A prompt that forces structurally distant domain collisions moves the LLM's output approximately 0.28 units.

That's seven times more. Same model. Same brief. Same number of words in the prompt. Just a structurally different approach.

This is the kind of measurement that should change how every business using AI prompts for ideas, brainstorms, names, headlines, content, or anything else that benefits from variation. We've been refactoring our own research synthesis pipeline and content-ideation workflows around the finding this week. The post below is the practitioner's version: what's happening, why it works, and how to apply it tomorrow.

The "artificial hivemind"

The underlying problem Open Collider solves is what researchers have started calling the artificial hivemind — the phenomenon where, if you ask an LLM the same question 100 times, the responses cluster around the same general answer. Different wording. Different framing. Same underlying substance. The model has been trained to produce the most-likely-correct response, which is the same as the response most like all the other responses it's seen.

This is a feature for most use cases. You want consistent, predictable output. You want the model to give you the answer most people would call right.

It's a bug for any use case that involves generating novel ideas — where the goal is not convergence, but divergence.

Here are the use cases where the bug is expensive:

• Content marketing teams generating blog headlines, social posts, email subject lines
• Designers asking for visual concepts, branding directions, naming options
• Product teams brainstorming feature ideas, UX patterns, use cases
• Strategy work asking for "what could we do here?" or "what are we missing?"
• R&D teams looking for novel approaches to known problems

In all of these cases, the customer wants 10 meaningfully different options, not 10 variations on the same option. The standard LLM workflow gives them the latter.

Asking the LLM to "be more creative" is the obvious instinct. Lion's measurement shows it's the wrong move. Being more creative moves the output a small amount in the same general direction. The output is slightly less typical. It's not meaningfully different.

The right move is to force the model to collide the brief with structurally distant domains.

Bisociations, since 1964

The technique Open Collider operationalizes isn't new. It's been a creative-thinking framework since Arthur Koestler's 1964 book The Act of Creation. Koestler's claim: novel ideas don't come from "thinking harder" or "thinking more creatively." They come from the collision of two unrelated mental frames.

Koestler called these collisions bisociations. The classic example: Einstein collided "elevator" with "free-falling object" and produced the equivalence principle of general relativity. The two frames were already in his head; the collision is what produced the novel idea.

Most human creative work, in any field, operates on this principle without naming it. The novelist collides domestic life with science fiction. The product designer collides agriculture with consumer electronics. The marketing strategist collides high fashion with home services.

Open Collider implements the same move programmatically: given a prompt, it first generates a set of structurally distant domains — semantic clusters from far parts of embedding space — then asks the model to produce ideas that combine the original prompt with each distant domain. The result: 10 ideas that aren't variations on one theme; 10 ideas that come from 10 different conceptual neighborhoods.

The numbers we cited at the top — 0.04 vs 0.28 in embedding distance, 7x improvement — are the measured result of this technique. The 12-of-12 sign-test wins (statistically significant at p=0.0002) confirm it's not an artifact of one model or one prompt. The 60%-plus originality preference from blind LLM judges across 4,320 evaluations confirms that humans (or LLM-judges acting as proxies for human readers) actually prefer the bisociated output.

This is one of the cleanest empirical results we've seen in the prompt-engineering literature this year. The technique works. The measurement is rigorous. The cost is essentially zero — Open Collider is MIT licensed, the implementation is a few hundred lines of Python, and the API cost per ideation run is on the order of pennies.

How we're using it

We don't use Open Collider directly through the published Anthropic API for one practical reason: our research synthesis pipeline already runs through Claude Code, and the Open Collider workflow integrates cleanly there. We've added a step in our ideation skill that calls Open Collider's collision-generation logic, then feeds the bisociated output back into our standard analysis pipeline.

Concrete places where we've started using it:

Blog ideation. When we have a topic in mind, we run a quick collision pass to generate three or four angles we wouldn't have thought of. About a third of the time, one of those angles is the actual post we end up writing. The default "what's the obvious angle?" output gets boring fast; the bisociation pass introduces friction that's productive.

Headline rewrites. We draft a working headline, then ask the collision tool to remix it against three distant domains (e.g., a technical post gets remixed against domains like "musical performance," "boxing," "knitting"). About one in five passes produces a meaningfully better headline. The others are filed for future reference.

Service-page positioning. When we're writing the "what we do" page for a new service offering, the collision pass helps us avoid the standard consultancy-page template. Instead of describing the service the way every other consultancy does, we describe it through a frame the reader hasn't seen on a consultancy page before.

Pitch generation for forward-deployment work. When we're preparing a discovery conversation with a new customer, we use the collision pass to think about their industry from three or four different conceptual frames. The customer often appreciates a question they haven't been asked before more than they appreciate a standard discovery process.

When not to use it

This isn't a universal upgrade. Bisociation is productive friction — useful when you want unexpected output. It's the wrong tool for tasks where you want consistent, predictable, conformant output.

Don't use Open Collider for:

• Customer support responses (consistency wins)
• Technical documentation (clarity wins)
• Compliance / legal / regulatory copy (precision wins)
• Code generation (correctness wins)
• Summarization tasks (faithfulness wins)
• Translation (accuracy wins)

The pattern: any task where the default LLM response is the desired LLM response, don't introduce friction. The default-is-best tasks are exactly the ones where the artificial hivemind is the feature, not the bug.

The use cases where it works are the use cases where you actively want variation — ideation, brainstorming, naming, headlining, positioning, framing. We use it more for ideation than for production output. Production output gets the consistent model; ideation gets the colliding model.

A small Easter egg

We discovered Open Collider through an X bookmark sweep — the same sweep that produced the Gracia PGA Championship 3DGS story, the Gemini 3.5 Flash demonstration, and several other research findings we've published this month.

The bookmark itself was buried — a single post from Lion announcing the release, with a screenshot of the tool's output on a sample brief. It would have been easy to miss. Our research processing skill caught it, classified it, and surfaced it as a high-priority signal worth deep-diving.

The same kind of pattern — small signal that turns out to be load-bearing — happens all the time in our work. Most of the practical wins in this practice come from catching the small useful signals before they're obvious to the market. This is part of why we run our own intelligence pipeline; the larger AI news cycle is dominated by headline products, not by the small tools that genuinely change a working consultant's daily work.

Open Collider falls in the second bucket. We're betting it'll be a standard tool in serious AI-using consultancies by the end of the year.

What to do this week

If you use an LLM for any kind of ideation, brainstorming, or creative-output task:

Read Lion's full release thread. It's worth understanding the technique end to end. The repository at github.com/CL-ML/open-collider has the working code.

Try it on one task you've been doing manually. Whatever you've been prompting "be more creative" for, run the same brief through Open Collider's collision-generation step instead. Compare the output side by side. (We expect you'll feel the difference within one example.)

Decide whether to integrate it or just borrow the pattern. The MIT license means you can integrate the tool directly. Or you can borrow the technique — explicitly ask your LLM to combine your brief with three structurally distant domains — and apply it without installing anything. Both work; the integrated version is faster.

Don't apply it to your production output. Apply it to your ideation, headlining, framing, and positioning work. Production output should stay on the default model.

If you're a working business owner or consultant who uses LLMs for any of the ideation tasks above and you want to talk about how to integrate techniques like this into your own workflow — not "tell me about AI strategy" but "help me build a workflow that produces better ideas this week" — get in touch.

---

Related reading:

• Gemini 3.5 Flash Reads a Paper, Codes the Demo, Ships the Site — the other piece of the May AI news cycle, and the broader argument about what changes for consulting
• Why Your AI Agent's Biggest Problem Isn't the Model — It's the Context — the underlying architectural reason that how you ask the model matters more than which model you use
• Forward Deployed Engineering in 2026 — What the Role Actually Does — the consulting model that creates demand for techniques like this
• Business Intelligence Is Pattern, Not Bookmarks — how we caught this signal in the first place