Escaping the idea basin

The basin is real, and deeper than you think

Every aligned language model has favorite ideas. Ask for 40 earned-media campaign concepts and 18% of them are the same idea — "humans vs. AI, blind test" — wearing different headlines. 57% sit in one emotional register. Clustered strictly for "same idea," a 40-idea batch holds roughly 14 actual ideas.

This isn't randomness failing you. Preference training teaches models that typical answers are good answers — researchers call it typicality bias — so the model learns to stay near the middle of everything it knows. The 2026 literature is close to unanimous that the collapse lives in the weights, and our measurements agree:

• Temperature does nothing. Max versus low temperature on the same model: 17 vs 14 distinct ideas. Noise.
• Verbalized sampling barely moves a frontier model. The paper-faithful version scored within noise of a random-constraints control, three replicates.
• Self-rewording helps the count, not the reader. Spec-first decomposition — have the model write 40 maximally-different one-line specs, then render each — reliably doubled distinct ideas. It led our scoreboard for two full rounds. Then a human read the output.

Wider isn't better: the viability trap

We had a strategist blind-read whole 40-idea batches and tick every idea he'd actually take to a client. Plain prompt: 33 of 40 viable. The spec-first "winner": 22 of 40. Cluster the survivors and both land near 12 usable, distinct ideas. The diversity machinery bought its spread almost entirely with ideas a human throws out.

Almost every method that widened the set did it by escaping the brief instead of escaping the basin.

A "provocation" instruction — render each idea as the strangest version of itself — made the trap vivid: machine-scored originality soared while human-scored viability collapsed to a 0.28 win rate. Surprising and un-runnable is not a campaign.

Update (September 2026): seven more readers, one conclusion moves

Because this result carried so much weight, we re-ran it as a panel: seven Fractl strategists blind-read the same plain and spec-first batches on both briefs — same sheets, same tick-every-viable instruction, no knowledge of which set was which. The viable-rate gap replicated cleanly: plain wins on usable ideas everywhere (readers averaged 23 vs 20 of 40 on the main brief, 29 vs 22 on the second). The yield conclusion moved. Keep only the ideas a majority of the panel ticked, cluster the survivors:

Majority = ticked by ≥4 of 7 readers (main brief) / ≥4 of 6 (second brief). Distinctness = strict same-idea clustering of the surviving ideas.

On the main brief, spec-first's survivors span 15 distinct ideas against plain's 12 — a modest real edge, not the wash a single reader saw, and it holds for six of the seven readers individually; on the second brief it's a wash. The mechanism is arithmetic, not taste: plain's 40 ideas only contain 14 distinct ideas to begin with, so its 25 survivors pile into the same clusters, while spec-first's 20 starting clusters survive a harsher cut with more of them still standing. Spec-first is still the wasteful way to widen a set — a quarter of its ideas die at review — but the spread that survives review is real spread. "Flat or worse" becomes "flat to modestly better."

The judges have their own basin

We tried to automate the human read. Four frontier models from three companies judged ~21,000 blind head-to-head pairs. Checked against 139 decisive blind human picks across two briefs:

Agreement measured on decisive (non-tie) pairs only. 50% = coin flip.

The failure has a mechanism. Judges score pairs in isolation and never saturate: a human reading 60 pairs is tired of "the AI writes its own obituary" by pair 12; the judges rewarded it every single time. In one experiment, 13% of all judge-favored ideas on a moving-company brief were the same divorce-data premise. And the judges share the generators' training, so they share the generators' taste — a judge can't see its own blind spot.

One technique helped: showing the judge the candidate pool first and declaring its premises used up lifted originality agreement from ~50% to 63% — and to 77% on the pairs where two judge families agreed. Useful for triage; not good enough to replace a human. Every ranking in this study that matters was made by a person.

The $110 brain surgery

If collapse lives in the weights, edit the weights. Open-weight models ship in pairs — the raw pretrained "base" model (the whole distribution, unaligned) and the aligned "instruct" model — with identical architecture. So you can blend them, parameter by parameter. We interpolated Qwen2.5-72B at 30 / 50 / 70% base on a rented H200 and measured how many regions of idea-space each blend reached:

The dial is real. Every step toward base widened the ideas, monotonically, and the model still followed instructions at 70% base. It's a lever no API will ever expose. And it wasn't enough. The open model's basin was so much deeper than the frontier model's that maximum blend never caught a plain Sonnet prompt on the main brief.

Two more findings from the open-weights lane:

• Division of labor beats either model alone. Let the blend emit raw one-line premises and a frontier model render each into a finished idea. This harvest-and-render pattern beat direct generation for every configuration — and on the second brief the 70%-base harvest (27.9) beat the best pure-frontier method (22.4).
• You cannot predict which models are wide. One community model family is wide at 405B and collapsed at 70B; its successor is the exact reverse. Same lineage, opposite outcomes. Screening each candidate empirically costs about $1 with this instrument — guessing costs you the campaign.

The lever nobody was testing: the brief

Halfway through, our strategist said the wide-model ideas were "cool, but I'm not sure what they're trying to be about." Our brief said ideas "must be something the company could produce and pitch to journalists" — and every generator read that differently. So we rewrote it the way he actually judges: name the goal, license tangents explicitly, but require a through-line — a CTA or mention of the product must not feel out of place in the finished project.

No new model. No technique. A clearer assignment nearly doubled the usable yield — more than anything else we tested, machinery included. A vague brief collapses the model into its safest cluster; a brief that names the goal and licenses tangents gives it permission to spread and a tether that keeps the spread useful.

No single generator wins. The portfolio does.

The final surprise. Under the rewritten brief, four very different generators each produced roughly 12–21 distinct viable ideas — call it a tie. Then we pooled everything the human had ticked viable and clustered the pool:

Different model lineages mine almost completely different veins of good ideas — same brief, same bar, near-zero overlap. Three times the swath of the best single method. The machinery earns its place not by beating the frontier model but by being uncorrelated with it.

The playbook

• Fix the brief first. Name the goal. License tangents. Require a through-line. The highest-leverage twenty minutes in the pipeline.
• Run a portfolio, not a prompt. Plain frontier prompt + spec-first + one or two unrelated model lineages as premise harvesters. The API-only version costs about $2 per brief.
• Anchor everything. Render every raw premise for the brief with a frontier model — keep what's surprising, draw the through-line.
• Dedupe the pool, then let a human pick. Clustering is the one job AI judges do reliably. Ranking originality is not — don't automate the last step.
• Skip: temperature, verbalized sampling, personas, provocation instructions, and any judge that claims to score creativity.

First step, today: take your next ideation brief and add three sentences — the goal, permission to be tangential, and the through-line rule. Then ask two unrelated models, not one model twice.