💡 From 'Generating' Ideas to 'Recombining' Them

Decision Making / Ideation / AI / Generative AI

— An AI-Era Method of Ideation, Thought Through with the Decision Map —

This post is a sequel to the previous two posts in the series. Previous posts: “Maybe What’s Missing When You “Can’t Decide” Is Just One of Five Elements” / “Raising the Resolution of AI Agent Design”

Introduction

You ask AI, “give me 10 ideas.” What comes back is a pile of “plausible-sounding ideas” — ideas that sound about right but that you’ve heard somewhere before, ideas that don’t stray far from the average, ideas you could just as well have found with a search.

This isn’t the model’s fault. AI is, fundamentally, a device that returns the highest-likelihood output given the context it’s handed, so an unremarkable context gets you unremarkable ideas back. Ask in an ordinary way, and only ordinary proposals come out.

So what do you do?

This is where the decision map the series has been working with comes in. Ideation, at bottom, can be written as the problem of how to recombine an existing map. As scaffolding for switching AI from “a device that returns the average” to “a device for trying recombinations in parallel,” the decision map does real work.

This time I’ll cover the topic in four parts.

• Part 1: Why “give me ideas” only gets you mediocre proposals
• Part 2: Ten operations on the decision map — a toolbox for ideation
• Part 3: Scaling recombination with AI
• Part 4: Selecting ideas — quality assurance in the AI era

Part 1: Why “Give Me Ideas” Only Gets You Mediocre Proposals

If the problem setup is sloppy, the output falls to the average too

When you ask AI “give me ideas for a new business,” the structure of the output that comes back looks like this.

• Input: a vague question
• Setting: no particular constraints
• Evaluation criteria: an implicit “plausibility,” “safeness”
• Output set: anything goes

Of the five elements, only the input is properly specified. AI implicitly fills in all the rest. And the standard AI implicitly fills things in with is close to the average tendency of the training data. That’s why you only get back run-of-the-mill proposals.

This isn’t a limitation of AI — it’s just that the resolution of your own problem setup is low.

The unit of ideation isn’t “the idea itself”

From here comes a fundamental shift in the method of ideation.

Ordinary methods of ideation aim to “generate ideas directly.” SCAMPER, brainstorming, 5W1H — in the end, they all try to generate the idea itself.

The decision-map method of ideation is different. Rather than generating an idea directly, you decide which of the five elements to move and how, and you get the idea as the result of the recombination.

Once you make this shift, what you hand to AI changes. Instead of “give me ideas,” you hand over “what does this map become if you change this element like this.” That’s where Part 2 comes in.

Part 2: Ten Operations on the Decision Map — A Toolbox for Ideation

The big picture

Ideation based on the decision map decomposes into ten operations that move the five elements.

Think of this less as an exhaustive list and more as “the first ten moves to try when you’re stuck” — that’s the easiest way to use it. You don’t need to use all of them. Try 2–3, depending on the situation.

Of the main operations, let me dig into the four that tend to be especially effective.

Operation ②: Change the Setting — Doubt “the constraint that can’t be moved”

The move that tends to work best is moving an assumption you’ve never moved before.

To take a past example: for a long time, image-generation AI research pushed to raise quality while keeping the setting “text→image” fixed. The moment the setting linking “text” and “image” was loosened and diffusion models entered the picture, the quality of image generation changed by orders of magnitude.

In another example, the moment an LLM’s implicit setting of “return a single response” was changed to “it’s fine to include the thinking in the output,” Chain-of-Thought appeared. Move just one setting, and the quality of the output changes.

The form of the question for changing the setting:

• Is this assumption actually necessary?
• What happens if you loosen it? What happens if you tighten it?
• What happens if you swap in a different assumption?
• Can a constraint you thought couldn’t be moved be promoted to a design variable?

Operation ⑥: Synonymous Transformation — Swap the Constraint and the Goal

The duality between the goal and the setting (constraint), which I touched on in the previous post, is also an extremely powerful operation for ideation.

“Maximize sales with a budget under ¥1M” (constrained) and “maximize sales − λ×excess budget” (unconstrained) are mathematically equivalent, yet the starting point of your thinking is completely different. The former searches for measures that grow sales without moving the budget. The latter treats the trade-off between budget overrun and return directly. Even on the same problem, the visible solution space changes.

The form of the question for synonymous transformation:

• If you promote what you’re currently treating as a “constraint” to a “goal,” what do you see?
• If you demote what you’re currently treating as a “goal” to a “constraint,” does a different goal surface?
• Can what you think of as an “input” actually be rewritten as a “setting”?

Operation ⑦: Element Transplant — Borrow from a Different Field

Sometimes a map with the same mathematical structure shows up in a different field. Notice this, and you can bring a tool from one field straight into another.

For instance, channel coding in communications and compressed sensing in imaging share the same structure — “decoding from sparse observations.” The sum-product algorithm, developed in information theory, is known to be equivalent in structure to what’s called mean-field approximation in statistical mechanics, and to variational Bayes methods (see the references below).

The form of the question for cross-field element transplant:

• Is there a different field with the same structure of five elements as this task?
• Can you bring in the “setting” or “evaluation criteria” that work well in that field?
• Does a result “shown to be impossible” in a different field also apply here?

Element transplant is an operation AI is particularly bad at. AI is far better at returning within-field knowledge than at seeing through cross-field structural isomorphisms, so this is a place where the human’s decision map has a big role to play.

Operation ⑧: Backward Derivation — Assemble the Elements from the Desired Output

One more, at the end. The operation of fixing the desired output (the ideal) first, and deriving the five elements backward from there.

This is the reverse of ordinary problem-solving — instead of “start from the constraint and input and derive the output,” it’s the operation of “fix the output and, working backward, assemble the input, setting, and evaluation criteria necessary to produce it.”

For instance, fix “the moment the user is maximally satisfied” first. From there, work backward on:

• What must have been given as the input at that moment
• What must have held as the setting at that moment
• How you’d write it as an evaluation criterion

When the “build up from the constraint” way of thinking has gone rigid, reassembling the elements by working backward gets you out into a different solution space.

Part 3: Scaling Recombination with AI

AI is “a parallel recombination device”

Trying all of the operations above by hand is a lot of work for a human. But AI can try the same work in large-scale parallel. This becomes the essential advantage of an AI-era method of ideation.

The point is to make the prompt you hand to AI not “give me ideas,” but an explicit form of the five elements and the operation.

[Current map]
Goal:    Reduce LLM hallucination
Setting:    Model is fixed; only prompt-level countermeasures
Evaluation Criteria: Factual accuracy rate
Input:    The user's question
Output Set: A natural-language response

[Request]
Try the following operations on this map in order, and give me 3 proposals for each.
- Operation ②, change the setting: what happens if you loosen one assumption in the setting?
- Operation ③, change the evaluation criteria: what different solution appears if you change the evaluation criteria?
- Operation ⑦, element transplant: what other field has the same structure of "managing output under uncertainty"?

This prompt switches AI from an average ideation device into a structured, parallel search device. The quality of the proposals that come out is completely different.

Hand over structure, and AI’s output space widens

Why does this work?

AI returns the highest-likelihood output given the context it’s handed. If the context is only “give me ideas,” it converges on an average idea across the whole training data. But when the context includes the structure of the five elements and the operation, the output is narrowed to what’s possible under that structure — that is, to the highest-likelihood proposal under this particular combination of the five elements.

Just because a human decided “move this one setting,” the space AI searches changes dramatically.

This is the same thing I wrote in the previous post — “the gap between people who can hand over the five elements and people who can’t” — showing up under the same structure not just in agent design, but in the scene of ideation too.

Part 4: Selecting Ideas — Quality Assurance in the AI Era

Even if you generate 100 proposals, 90% are sophistry

Have AI generate ideas in parallel, and the count easily passes 100. What happens here is that most of them turn into “proposals that sound plausible but are thin on substance.”

Recombine the five elements, and the shape comes together. Once the shape is there, it looks about right. But a proposal that has shape but no substance — one where the structure alone was built while the density of the relevant internal model stayed thin — is, more often than not, close to sophistry.

Say AI produces the proposal “bring financial-engineering risk management into education” via “element transplant.” As a shape, it holds together. But unless you deeply understand both the reality of the educational field and the financial risk models, that transplant stays a superficial analogy.

The criteria for selection are “consistency” and “internal-model density”

This is where human judgment remains. There are mainly two criteria for selecting ideas.

1. Consistency with the upper-level and adjacent maps

Does the proposed idea contradict the upper-level goal it would be executed under (Operation ⑨’s consistency check)? Does it collide with a different, adjacent map? Run it through this question, and proposals carrying a structural contradiction fall out.

2. Density of the internal model

Do you (or a trusted expert) have a sufficiently dense internal model of that field? Move the structure alone in territory you don’t have a model for, and the output turns into sophistry. Run it through this check, and analogies with no real grounding fall out.

AI cannot make this judgment. Whether you can ask “do I actually understand this domain” is, in principle, something only a human can ask.

That’s exactly why the person who hands over the five elements becomes “the one who takes on responsibility”

Under the decision-map method of ideation, it works out like this:

• The combinatorial explosion of ideation — AI scales this in parallel
• Handing over the structure of the five elements and the operation — that’s the human
• Judging consistency and internal-model density — that’s the human

If this three-way division of labor breaks down, quality drops somewhere. Hand it to AI without structure and you get mediocre proposals; skip the judgment and you get a mountain of sophistry.

Put the other way: only someone who holds the five elements as a structure can use AI as a device for recombining ideas. This is a conclusion that follows directly from the previous two posts — a point that has come up again and again throughout the series.

Conclusion — Dividing the Roles of Ideation and Judgment

An AI-era method of ideation, thought through with the decision map, compresses to one line like this:

Let AI scale the recombination of ideas. Leave the judgment to humans.

What draws that boundary line is the decision map’s five elements — goal, setting, evaluation criteria, input, output set.

How you design the recombination operations, which combination you choose, whether a given proposal is sophistry or the real thing — only someone who can treat the five elements as a structure can ask these questions.

Put the other way, for someone who can handle the five elements, AI becomes a powerful tool that raises the ceiling of ideation by a full level. A human decides the direction — “I want to move this” — AI tries recombinations in parallel, and the human picks. Through this back-and-forth, you gain reach into a space of ideas a single human could never have arrived at alone.

At the close of the series, let me leave this line here once more.

Having a decision map means you can replace “struggling” with “designing.”

And in the AI era, that also means you can replace “generating ideas” with “recombining” them.

The people who can handle both design and recombination will stand at the center of creation in the era to come.

References

• Toshiyasu Matsushima, “How Should We Grasp and Teach Data Science? — Drawing on Waseda University’s Efforts,” Operations Research, Vol. 65, No. 10, 2020, pp. 544–550. https://orsj.org/wp-content/corsj/or65-10/or65_10_544.pdf
• Waseda University Data Science Education Team (author), supervised by Toshiyasu Matsushima, Introduction to Data Science I — Foundations of Data-Based Decision Making (Data Science Library 1), Science-sha, 2022.
• As an example of the same mathematical structure appearing in different fields: Matsushima (2020) points out that the sum-product algorithm in information theory, mean-field approximation in statistical mechanics, and variational Bayes methods share an equivalent structure.