Every brainstorming session I’ve ever run with an LLM produces the same list. Different topic, same list. Ask it to “think outside the box” about your product and you get the median of every product blog post ever written: gamification, community features, an AI assistant, maybe a freemium tier. The model isn’t being lazy. It’s doing exactly what it was built to do — give you the most probable answer. And the most probable brainstorm is, by definition, the one everybody already had.
The fix came from a man who died in 2017 and never graduated high school.
A periodic table of ideas, drawn by hand
Patrick Gunkel (1947–2017) was an independent scholar who spent roughly five decades building something he called ideonomy — “the science of ideas.” No diploma, no degree, and yet he worked in MIT’s computer science lab under Edward Fredkin from about 1971 to 1987, Marvin Minsky credited him with coining the word “telepresence,” and in 1984 the Richard Lounsbery Foundation handed him an unsolicited five-year grant to keep going. The archive — hundreds of hand-drawn charts, lists, and monographs spanning more than 200 divisions of ideas — still sits frozen at ideonomy.mit.edu, with Grace Kind’s digital preservation project providing text-searchable versions of his five self-published volumes.
Gunkel’s premise was that ideas aren’t magic. They have properties that vary along dimensions, and there are repeatable operations you can perform on them: negate a property, substitute a value, combine two ideas, lift one to a higher abstraction and drop it into a different domain. He described ideonomy as a kaleidoscope for exhibiting all possible combinations of ideas, and a prism for diffracting any single idea into its components. Chemistry for molecules, biology for organisms, ideonomy for ideas.
He thought it was a job for computers. He was about forty years early.
Roll the dice, show your work
Here’s where it gets interesting for anyone using LLMs today. Someone packaged Gunkel’s machinery into a skill — a set of instructions an AI agent loads before brainstorming. The design makes two moves that fix the median-brainstorm problem.
Move one: it rolls dice. Before each session, a picker script randomly draws a “method tuple” — a couple of operators (say, abstraction-lift and dimension-identification), an output structure Gunkel called an organon (a chart, a tree, a procedure), and a few dimension-prompts (how old is this idea? does it accumulate or decay? is it valued positively or negatively?). The model doesn’t get to choose its approach, which means it can’t default to its comfortable rut. If you know music, this is Brian Eno’s Oblique Strategies — the 1975 card deck he and Peter Schmidt used to jolt recording sessions out of local minima, famously deployed during David Bowie’s Berlin trilogy. Draw a card, obey the card. Except here the cards are Gunkel’s operators, and the musician is a language model.
Move two: the machinery stays visible. Every section of the output is labeled with the operator that produced it and the specific move it made. Every session ends with a “trail” — a footer recapping each move and, critically, a mandatory line listing what was not surfaced: the directions the dice didn’t reach.
If you can’t see which method produced which idea, you didn’t brainstorm — you got the average, formatted nicely.
The trail makes the whole thing falsifiable. No trail, no ideonomy — just a thoughtful answer wearing a costume.
It found the hole in my own spec
Theory is cheap, so here’s what happened when I pointed this at a real project. I’ve been developing a design for a brand-safe social media platform — hierarchical trust delegation modeled on SSL certificates, where the NFL certifies teams, teams certify players, and branded hashtags carry the league’s shield. I call it MaxOro, and I’ve been kicking it around for years — long enough to stop seeing its flaws.
The dice drew abstraction-lift, dimension-identification, the procedure organon, and a dimension-prompt called direction: does this idea accumulate, decay, oscillate, or hold steady?
That one prompt cracked the spec open. Every trust mechanism in my design accumulates — reputation points, vouches, certifications, endorsement reach. Nothing expires. Nothing decays. Nothing gets revoked. And here’s the embarrassing part: my own governing metaphor is SSL, and real PKI spends most of its operational effort on exactly the decay side — certificate expiry, revocation lists, live status checks, browsers nuking entire root authorities. I borrowed the issuance half of the metaphor and quietly dropped the half that does the hard work. Six design documents and not one of them answered the question: what happens to a player’s certified posts when he gets cut from the team?
The abstraction-lift operator then made it worse, in the best way. Strip the proper nouns from “NFL certifies teams certify players” and you get a provenance tree — the same shape as wine appellations, academic accreditation, organic food labeling, and ham radio licensing. Every one of those domains has already paid tuition on problems my spec hadn’t priced: counterfeit labels, diploma mills, captured accreditors, certifier fraud. Each of their solutions is a candidate to lift back down into my design.
I turned the whole analysis into an interactive explorer, including a simulator where you can revoke a player’s certification under three different propagation policies and watch what happens to everyone he vouched for. The revocation simulator is the punchline. Explore it below, or open it full-screen.
What “not surfaced” buys you
The trail’s final line listed what the session didn’t explore: killing the central-root model entirely, making credentials portable across rival platforms, and the economics of who pays for revocation. That line is the most valuable part of the output — it converts a finished brainstorm into a map of the next three.
The best thing an idea machine can tell you is what it didn’t tell you.
Even if you never touch this particular tool, steal the pattern. When you brainstorm with an LLM: force a method it didn’t choose, make it label which method produced which idea, and demand a list of the directions it skipped. Do those three things and you stop getting the median. A guy with no diploma worked this out by hand, fifty years before the machines showed up that could actually run it. The least we can do is stop asking those machines for the same shitty list.