PowerUp Coaching — Living Knowledge Base

Insight: Modal Subtraction — Generate, Strip the Median, Skeptic-Test the Outliers

6 min read

“Generate 20 candidates, get rid of the stuff that everybody else already knows or talks about, get rid of the middle, then put a skeptic gauntlet on what’s left — see whether the things on the fringes actually pass muster.” — Lou, 2026-05-21

Session context: 2026-05-21_Mastermind — Lou walked the group through a multi-day conversation that started from Meta’s Large Concept Model (LCM) paper and ended with a concrete algorithm. The technique was extracted into a standalone skill (Sakana LCM / OutsideTheModal) and ships with the session’s R&D folder.

Core Idea

Insight - Latent Terrain Cartography — Navigating Off-Modal AI Responses to Find Non-Obvious Ideas named the discipline of leaving the model’s modal response on purpose. Modal Subtraction is the algorithm that operationalises that discipline. It runs in three explicit passes:

  1. Generate ~20 candidates for a question, problem, or idea.
  2. Subtract the median. Have the model surface what a typical LLM response to this prompt would say, then remove every candidate from the list that overlaps substantially with that median. What remains is the off-modal set — the ideas that would not have surfaced if you had asked the question once and accepted the first answer.
  3. Run the survivors through a skeptic gauntlet. Off-modal does not mean good — it just means not consensus. The skeptic pass asks: is this actually defensible? Does it survive an adversarial reading? Most off-modal candidates will fail this test, and that is the point — the algorithm is designed to over-generate at step 1 and aggressively filter at steps 2 and 3.

The mental model is statistical, not poetic. An LLM’s response distribution has a heavy mode (the consensus answer) and long tails (rare but possible answers). Naive prompting samples from near the mode. Cartography says “sample from the tails.” Modal Subtraction adds the missing step: don’t trust the tail samples either — most of them are tail because they are wrong, not because they are insightful. The skeptic gauntlet is the discriminator that separates rare-and-true from rare-and-noise.

This is conceptually adjacent to Insight - Isolation Outperforms Debate When the Goal Is Discovery, Not Refinement (generate in isolation, then converge) and Insight - Run Your Prompt Through Multiple Models and Synthesize at the Top (sample broadly, synthesise selectively). What is new is the explicit median-strip step — most multi-candidate workflows assume more candidates = better. Modal Subtraction inverts the assumption: more candidates only helps if you then remove the ones that look like everyone else’s. Otherwise you have just generated 20 versions of the same modal answer.

Why This Matters for Knowledge Entrepreneurs

The knowledge entrepreneur’s defensible asset is the non-obvious take. If your audience can get the consensus answer by asking ChatGPT directly, they did not need you. Modal Subtraction is the production method for non-obvious takes: it gives you a repeatable way to find positions that the typical LLM consensus would not surface, then verify those positions are defensible before you publish them.

The trap most people fall into is using AI to generate ideas and then publishing the first thing that “feels good.” That move maximises modal output — you are essentially publishing what the model would have published if it had your byline. Modal Subtraction is the structural answer: build the median-strip into your ideation process so the only ideas that survive into your content are the ones that are not what the model would have said by default. The cost is one extra inference call and a willingness to throw away 90% of your candidates. The payoff is content that compounds your authority instead of dissolving into the global feed of AI-assisted output.

This also reframes what “AI helps with ideation” actually means. The naive read is “AI generates ideas, I pick the good ones.” The Modal Subtraction read is “AI generates many ideas, AI also tells me which of those ideas are statistically boring, I pick from what remains.” The selection step is no longer pure human judgment — the model can identify its own modal output and disqualify it, leaving you to make the higher-leverage call between the survivors.

Practical Application

Run the Modal Subtraction Pass any time you need a non-obvious angle:

  1. Generate: ask the model for 20 candidate ideas, positions, or angles on the question. Push for variety explicitly.
  2. Surface the median: ask the same model, separately, “what would a typical LLM say in response to this prompt?” Save that answer.
  3. Subtract: feed the candidate list and the median back to the model and ask it to remove anything that overlaps meaningfully with the median.
  4. Skeptic-test the survivors: for each remaining candidate, ask the model to argue against it as a skeptic would. Keep only the ones that survive the argument.
  5. Pick from the survivors — that is your starting material.

Codify this as a slash command or skill so it runs in one invocation. The Sakana LCM skill in Lou’s R&D folder (shipped with this session) is one implementation; the OutsideTheModal command is another.

Coaching question: “If I removed every idea on this list that any AI would generate by default, what would I have left — and is what I have left actually defensible?”

Evolution Across Sessions

Builds on Insight - Latent Terrain Cartography — Navigating Off-Modal AI Responses to Find Non-Obvious Ideas (2026-02-19), which established the discipline of leaving the modal response on purpose, and Insight - EigenThinking — Turn Your Cognitive Fingerprint Into Intellectual Property (2026-02-19), which applied that discipline to self-modelling. The new development is the explicit median-subtraction step as a separate inference call — turning the cartography discipline into a three-stage algorithm that can be wrapped as a skill and run repeatably. Cartography said “sample the tails.” Modal Subtraction adds “and prove the tail samples are not noise before you keep them.”

Source

Graph View

Backlinks