1. You hired 100 geniuses. They are all in the break room because no work ticket arrived
Generative AI can write code, draft text, research, translate, compare alternatives, and draw from enormous bodies of knowledge.
That sounds like a world where everyone should be running ten one-person companies by tomorrow.
Yet many people stop at one question:
“So what do I build?”
A powerful model with no task is still idle capacity. One hundred Ferraris with no destination are just an unusually expensive parking lot.
A scarce resource in the AI era is not always model capability.
It is the person who gives capability its first job.
2. The real strength is often not a grand invention, but the first spark
“Idea power” sounds like the ability to suddenly invent something nobody has ever imagined.
In practice, a more ordinary ability can be more useful:
- this is annoying
- could these two things connect?
- can this existing feature serve another purpose?
- why does this step exist at all?
- what would we learn from a minimal prototype?
These are ignition points.
The first idea does not need to be complete. Its job is to create a surface that produces the next idea.
An empty table is hard to critique. Put even a rough prototype on it and people immediately see what is wrong, useful, missing, or unexpectedly promising.
The first move is not the answer.
It is a machine for generating more answers.
3. “Idea generation” is not one ability
It can be decomposed into several operations:
| Ability | What it does |
|---|---|
| Divergence | Produces multiple possibilities from one problem |
| Analogy | Imports structures from another field |
| Recombination | Turns existing A + B into a new C |
| Problem finding | Notices the inconvenience before the assigned problem |
| Reframing | Redefines what should actually be solved |
| Convergence | Selects what deserves implementation |
| Decomposition | Converts a goal into AI-manageable tasks |
| Initiation | Produces the first tangible artifact |
| Re-ideation | Uses the artifact to generate the next idea |
Divergent thinking is an important piece. The APA describes it as creative thinking that departs from commonly used or previously taught strategies.[1]
But divergent-thinking tests are not equivalent to real-world creativity as a whole.[2]
Producing many ideas is not enough.
Practical creativity also means reframing, selecting, building, observing, and thinking again.
4. Creativity often behaves like “generate → explore → generate again”
Creative-cognition research includes the Geneplore model, which separates generative processes from exploratory processes and treats creativity as a cycle between them.[3]
AI product work makes that cycle visible:
- think of something
- discuss it with AI
- make a rough specification
- build it
- run it
- notice that it is not quite right
- find a new angle
- build again
The idea does not need to be complete before implementation.
Implementation becomes part of thinking.
When a prototype takes three months, you try to solve much of the problem in your head first.
When a prototype can appear in thirty minutes, the prototype itself becomes a thinking note.
Ideation and execution stop living in separate departments.
5. “I cannot multitask” is correct only if the human is expected to do everything
Most humans are bad at holding ten projects, ten specifications, ten code paths, and ten verification states in working memory at once.
That is not a failure.
It is the wrong architecture.
The human does not need to multitask. The system can run work in parallel.
The human can retain:
- desired direction
- priorities
- acceptable boundaries
- final adoption decisions
Then research, specification, implementation, testing, translation, and auditing can be separated where dependencies allow and assigned outward.
If you are bad at multitasking, externalizing work into clear responsibilities can become even more valuable.
The goal is not to make the brain run more tabs.
It is to stop using the brain as the tab manager.
6. A giant corpus is a materials warehouse, not a building plan
A huge AI knowledge base is like an enormous home-improvement store containing lumber, steel, glass, tools, and heavy machinery.
Impressive.
But if nobody decides what to build, the store remains a store.
AI can retrieve information, compare options, write code, and suggest ideas.
Direction appears when someone says:
“I want to remove this friction.” “These two systems might become one.” “This experience could be taken one level further.”
That directional vector converts capability into work.
The person extracting value from AI does not need to know more than the AI.
They need to be able to tell it where to dig.
7. The ability to remake the problem can matter more than the ability to solve it
Design research has treated framing and frame creation as central practices for difficult, open-ended problems. Dorst emphasizes reframing as a way to make problematic situations more workable and desirable.[4]
This becomes more important when AI is a fast answer engine.
If you ask the wrong question, AI may produce a very polished answer to the wrong problem at remarkable speed.
Instead of only asking:
“How do we make this task faster?”
ask:
“Does this task need to exist?” “Does the user want faster work, or no work at all?” “Can we solve the underlying need from another entry point?”
As answer generation accelerates, problem framing gains leverage.
8. Faster implementation creates more ideas, not merely more output
In a preregistered experiment on professional writing tasks, Noy and Zhang found that ChatGPT access reduced average completion time by 40% and raised evaluated quality by 18% in the tasks studied.[5]
That does not mean every job becomes 40% faster.
The deeper implication is simpler:
Lower experimentation cost means more experiments.
More experiments produce more information from reality:
- is this actually useful?
- does it look worse than imagined?
- is another use case stronger?
- can it connect to a neighboring feature?
- where do users hesitate?
That information is difficult to manufacture purely in your head.
So when the cycle becomes short—
idea → implementation → observation → new idea
—ideas stop behaving like isolated sparks and start behaving like a chain reaction.
Implementation speed becomes the RPM of the idea engine.
9. But if AI generates every idea, everyone may climb the same mountain
Generative AI can improve idea support. It can also create convergence.
Doshi and Hauser found that access to generative-AI story ideas improved individual story evaluations on average, while AI-assisted stories became more similar to one another, reducing collective diversity.[6]
That matters.
If everyone repeatedly asks the same model for “ten creative ideas,” everyone may board the same tour bus.
Human contribution therefore remains valuable:
- friction personally observed
- odd field details
- constraints of an existing system
- analogies imported from unrelated fields
- strange combinations
- the question, “Everyone does it this way, but why?”
Do not ask AI to originate everything.
Throw one unusual stone into the water and let AI amplify the ripples.
10. The practical loop is simple: spark → sentence → split → delegate → inspect → spark again
A useful operating loop can be extremely small:
1. Spark — “Wouldn’t it be useful if this worked like that?”
2. One sentence — State who should move from what state to what state.
3. Decompose — Separate research, design, implementation, and verification.
4. Delegate — Give responsibilities to AI instead of carrying all of them mentally.
5. Produce an artifact — Text, screen, code, data, anything visible.
6. Read back — Verify that the target state actually exists.
7. Re-ideate — Ask what becomes possible now that this exists.
Step seven is why the loop compounds.
A enables B. A + B suggests C. C changes how the whole service can be defined.
A tiny spark eventually produces a power plant.
11. The danger is when idea speed exceeds organization speed
Fast ideation plus fast implementation creates architectural sediment.
One day you open the wiring and ask:
“Who built this?”
Yesterday-you did.
The fix is not to suppress ideas.
It is to strengthen structure:
- avoid multiplying sources of truth
- do not build a second mechanism for the same responsibility
- integrate into existing systems where possible
- parallelize only work that is truly independent
- read back after implementation
- place gates around important changes
- separate “we can build it” from “we should build it now”
Do not weaken the accelerator.
Upgrade the steering and brakes.
12. Conclusion: the scarce skill is shifting from “using AI” to “creating work for AI”
Models will keep improving.
Knowledge access will expand. Coding and writing will accelerate. Running multiple agents will become easier.
That shifts the valuable human layer.
Not merely:
“I can build everything myself.”
But:
I can notice what should exist.
I can create the first spark.
I can reframe the problem.
I can decompose the work.
I can delegate it.
I can inspect the artifact.
I can generate the next idea from reality.
AI is enormous capability.
Capability waits until somebody gives it a job.
So the strongest person in an AI-heavy system is not always the person who codes faster than the model.
Sometimes it is the person who says:
“Wait. Couldn’t we build this?”
In a company containing one hundred genius AIs, the first missing hire may not be genius number 101.
It may be the person who writes the first line on the work ticket.
Sources
- APA Dictionary of Psychology, “divergent thinking. dictionary.apa.org
- American Psychological Association, “The science behind creativity” — notes that divergent-thinking measures are useful but do not map cleanly onto real-world creativity as a whole apa.org
- Ward, T. B., Smith, S. M., & Finke, R. A., “Creative Cognition,” in Handbook of Creativity; see also Geneplore descriptions of iterative generative and exploratory processes cambridge.org
- Dorst, K. (2011). “The core of ‘design thinking’ and its application.” Design Studies, 32(6), 521–532 doi.org
- Noy, S., & Zhang, W. (2023). “Experimental evidence on the productivity effects of generative artificial intelligence.” Science, 381(6654), 187–192 doi.org
- Doshi, A. R., & Hauser, O. P. (2024). “Generative AI enhances individual creativity but reduces the collective diversity of novel content.” Science Advances, 10(28), eadn5290 doi.org
