PowerUp Coaching — Living Knowledge Base

Modal Subtraction — How to Find the Ideas an LLM Would Never Have Given You by Default

5 min read

Topic

A three-stage ideation algorithm (generate, subtract the median, skeptic-test the survivors) that systematically produces non-obvious takes — the kind of content that compounds authority instead of dissolving into the global feed of AI-assisted output.

Target Reader

Thought leaders, knowledge entrepreneurs, and content-driven coaches who have noticed their AI-assisted ideas read more and more like everyone else’s — because everyone else is using the same models with similar prompts.

The Fear / Frustration / Want / Aspiration

The fear: in an AI-saturated content economy, your “ideas” are converging with everyone else’s because you are all sampling near the same modal response. The want: a repeatable process that surfaces positions the typical LLM would not produce by default. The aspiration: content that is defensibly different — not by stylistic flourish, but by structural ideation.

Before State

The reader generates 5–20 ideas with AI, picks the one that “feels best,” and publishes it. They suspect the picked idea is statistically modal but cannot reliably identify which of their ideas would have shown up in someone else’s AI session too. They lack the algorithmic vocabulary to systematically leave the modal.

After State

The reader runs Modal Subtraction whenever they need a non-obvious angle. They explicitly surface the median, subtract it, and skeptic-test the survivors. They publish only what survives both filters. They have the algorithm codified as a slash command or skill so it costs one invocation, not three.

Narrative Arc

Open with the convergence problem — why everyone’s AI-assisted content is starting to sound the same, and why “just be more creative” is not the answer. Frame the modal-response geometry: every LLM has a heavy mode and long tails; naive prompting samples near the mode. Walk through the three stages: generate broadly, subtract the median as its own inference call, then skeptic-test the survivors so off-modal doesn’t mean wrong. Close with the codification — wrap it as a skill, run it before every piece of content where novelty matters.

Core Argument

The non-obvious take is the knowledge entrepreneur’s defensible asset. Modal Subtraction is the production algorithm for non-obvious takes: a repeatable three-stage process that produces ideas the typical LLM consensus would not have surfaced, then verifies those ideas are defensible before you publish them. The key move that distinguishes it from naive multi-candidate ideation is the explicit median-subtraction step — most workflows assume more candidates = better; this one inverts the assumption.

Key Evidence / Examples

  • Lou’s multi-day exploration of Meta’s LCM paper that produced the algorithm, captured in his shipped LCM-vs-LLM/ folder.
  • The Sakana LCM skill that operationalises the algorithm — generate 20, surface the median, subtract, skeptic-gauntlet survivors.
  • Cross-reference to Insight - Latent Terrain Cartography — Navigating Off-Modal AI Responses to Find Non-Obvious Ideas — the discipline that this insight algorithmises.
  • The Expert Mind skill that uses Modal Subtraction internally for the 5-component business architecture audit.
  • The statistical framing: an LLM’s response distribution has a heavy mode and long tails; naive sampling pulls from the mode, naive tail-sampling pulls noise.

Proposed Structure (5–7 beats)

  1. The convergence problem. Why everyone’s AI-assisted content is starting to sound the same. The modal-response geometry.
  2. What “off-modal” actually means. Distinguish rare-and-true from rare-and-noise. Why naive cartography fails without a filter.
  3. The algorithm: generate. Push for ~20 candidates with explicit variety pressure.
  4. The algorithm: subtract the median. The pivotal step. Surface what a typical LLM would say, then remove anything in the candidate list that overlaps. Most workflows skip this.
  5. The algorithm: skeptic-test the survivors. Off-modal does not mean defensible. The gauntlet that separates insight from noise.
  6. Why this compounds your authority. Content that survives Modal Subtraction is content the average reader cannot get from their own AI session. That gap is your moat.
  7. How to codify. Wrap as a slash command or skill. Run before any piece where novelty matters.

Editorial Notes

Voice should be precise and slightly technical — readers should feel they are learning an algorithm, not a vibe. Resist softening the language (“a way to think about it…” weakens the piece; “the three stages are…” is what lands).

The framing tension to surface: most readers will assume their existing ideation process already produces non-obvious output. The article needs to prove — in the second beat — that it does not. The proof: their last 5 AI-assisted ideas, fed to any other LLM, would have surfaced the same 5 ideas. That is the moment the reader recognises the problem.

Next Step

  • Approved for drafting
  • Needs revision
  • Deprioritised

Graph View