on asking machines for ideas

Asking a machine
for ideas?

Sometimes we know what we want.
Sometimes we figure it out in the asking.

We often ask AI to do things for us — things we can picture clearly and ask for. But when we ask AI to help us brainstorm, we're asking for an idea we don't have yet.

How can we do that? With words: more ideas, novel ideas, not-obvious ideas.

We use these words all the time in conversation — but what does the AI hear when we put them in a prompt?

Let's examine this through a simple example.

the promptyou run a small town bakery, and a starbucks has just moved into town. give 25 strategies for responding to it. for each, give it a name and a 1 sentence description

This isn't a story about bakeries or business strategies. It's just a setting where we have obvious intuitions to compare the AI's answers against.

01 — the plain ask

Simple question, obvious answers.

Most people will gravitate to some obvious answers, and different AIs capture that common view — though in slightly different ways. Let's look at the top 5 answers from each.

why the ★?The prompt asks a generic question — business strategy — but plants it in a specific situation: a Starbucks across the street. An answer that takes the fight to Starbucks is still answering that specific question; one that gives advice you'd give any bakery has drifted to the generic. So naming the rival (★) is our proxy for staying on target rather than drifting.

Now let's look at those answers taken as a whole, and see how they cluster.

In the picture, answers that name the rival (★) cut a deep notch; ones that don't spread wide and shallow; and the area of each dip is the number of items.

Most answers collect in the neighborly basin — in a deep well that corresponds to answering the specific question. But there's also a second, smaller well: just running the bakery like a business.

what the plain ask revealed

That the models overlap so much — all landing on much the same expected answers — tells us something important. It's not really about the models. The models are just reflecting our own common intuitions back at us.

02 — read further down

Ask for more.

The top ideas were obvious. What if I just read further down the list?

Let's start with the high-level view first, for comparison.

The more list doesn't break out into new ground — it just becomes less targeted and a bit more business-like.

Let's compare a few examples from the top of the list against ones further down.

Top of the listFurther down

In broad strokes, the later ideas are narrower variations on the earlier ones — town merch instead of local sourcing — and the specific strategies competing against Starbucks give way to general good-business advice.

what “more” revealed

Reading further down the list for more ideas, we were hoping for different. What we got was just more volume — further, not elsewhere.

03 — ask for new

Ask for novel.

Let's just ask for new ideas directly.

the prompt...give 25 novel, innovative strategies for responding to it...

If we look at the top “novel” ideas, how does this shift us from the baseline?

Neighborly stays central, with the same lean toward business — but it holds its focus on the competition, and the first unconventional ideas appear.

Most “novel” ideas just weld a twist onto a baseline one — and a few genuinely leave the neighborhood. One of each:

BaselineNovel, innovative
what “novel” revealed

When we asked for novel, we meant “make it genuinely new.” The machine heard “make it sound new” — same subject, dressed up. Even the names get hyped: Scone-scription.

interlude

Still orbiting the same center.

We've been trying to find new ideas beyond the obvious.

Instead of turning up the intensity, maybe we need to change the location.

04 — forbid the obvious

Avoid the obvious, land somewhere new.

What if we just rule out the obvious answers? Where do we end up?

the prompt...give 25 strategies for responding to it, but not ones that center on local sourcing, local identity, or unique artisanal features...

This move has an obvious and dramatic effect.

Rule out the obvious answers and the models shift to that second well — a new center of MBA-style optimization.

The same model, before and after the ban:

Neighborly · baselineOperational · not-obvious

It was there all along — the quieter “just run it like a business” answer, sitting in that second well, dwarfed by the neighborly center. Forbid the obvious and it rises, and everyone circles it. Notice these barely mention Starbucks anymore: the new center isn't a counter-move, it's generic operations.

what “not-obvious” revealed

We didn't ask for a change in the form of the ideas; we asked for a change in their content. We may not be able to say what we want yet — but we know what we don't want (more of those first answers), and that's enough to force the shift.

05 — do both

Do both.

If the story's right, “both” should move us first, then add the twist. Does it?

the prompt...give 25 novel, innovative strategies... but not ones that center on local sourcing, local identity, or unique artisanal features...

Watch for two things: the shift into the business well, and a few ideas landing in the unconventional well.

It relocates like not-obvious, then twists like novel — and the invention, now in the business well, turns into apps, kiosks, and gadgets.

The same impulse — pay attention to the person in front of you — landing in two different wells:

Attention · neighborlyAttention · unconventional

In the neighborly well that means a custom cake for your wedding, or a phone-free hour to actually talk. In the new one it means supplement-dosed “focus foods” and a kiosk that scans your mood.

what “both” revealed

This is a toy example, and some of these ideas are frankly bonkers — but it reveals that we meant two different things by “new ideas”: new in form vs. new in content. We have to be clear which we actually want.

06 — what we were really asking

Two things vary: the neighborhood an idea sits in, and the building itself.

more → more buildings on the same street.
novel → a strange building in the same old neighborhood.
not-obvious → a new neighborhood of familiar building styles.
both → surprising buildings in a new neighborhood.

what we learned

These weren't just different words for roughly the same idea. We know because the AI — which reflects our own meaning back at us — did different things with each. This is one strongly-scripted setting with an obvious runner-up, so the exact map won't transfer everywhere; and how you use the insight will depend on the situation — sometimes you'll want novel rather than not-obvious. But the first step is noticing the distinctions exist.