NegativeSpace

A map of the startups that exist, and a search for the ones that don't yet.

YC companies · dimensions of meaning · clusters, none labelled by hand

Good artists copy.

Great artists steal.

Master artists create embeddings, find the farthest distance, and then eat a sandwich about it.
Probably not Picasso

See the generated ideas How it works

Built by NH66.ai, an AI engineering studio.

// READ

A description becomes a location

An embedding model reads each company's description and turns it into numbers. Those numbers work like coordinates, so companies that sound alike land near each other.

FIRST 12 OF NUMBERS

SIMILARITY TO

// GROUP

Neighborhoods form on their own

Nobody told the model what an industry is. A clustering algorithm found dense groups, then named each one after its most distinctive words.

// RANK THE NEIGHBORHOODS

Stack the dots and compare

The same dots rearrange into bars, so you can see which types of company are most common. Pick a measure to re-rank them.

// COMPARE · NEAR AND FAR

Rivals score high. Strangers don't.

Any two companies get a score from their coordinates. Nobody told the model these are competitors. It worked that out from how they describe themselves. Pick a pair.

AVERAGE SCORE ACROSS ALL CLUSTERED COMPANIES

// COMPARE · THE CROWDED CORNERS

Some companies are almost twins

These pairs are among the closest on the map, close to what a rewrite of one company scores. If you're describing the same thing in nearly the same words, the map shows it.

// COMPARE · THE LONERS

And some have almost no neighbors

The score here is with each company's closest match. A low score means nothing else on the map sounds much like it. Those companies show that the empty parts of the map can hold real businesses.

// MARK

They're marked here in orange. They stay on the map, so a gap can't simply be an idea someone already tried. They never seed a new idea.

NEAR THE GAP WE'LL EXPLORE NEXT

// BLEND

Mix three related companies

Each of these music startups solves one part of getting a song to a fan. Blending unrelated companies gives mush, so the ingredients must already be close on the map.

Earbits lets bands and labels bid to have their music played to the right audience, like search ads for music.

Nasdisc is a marketplace for buying and selling vinyl, cassettes and CDs.

Bop.fm gives every song one link that plays on Spotify, YouTube, iTunes or any other service.

THE BLEND

The numbers are averaged by those percentages. The result is a single new location, marked in pink, that no company occupies.

// PUSH

Move away from everything

The point now drifts away from all companies, measured across all dimensions. It stops before it loses the meaning of its ingredients. The map's 3D view only approximates this move.

SIMILARITY TO THE CLOSEST OTHER COMPANY

WHAT'S AROUND THE GAP

Nearby are . Going by their descriptions, none connects hearing a song to buying the record, which is the space between the ingredients.

// TRANSLATE AND MEASURE

Claude describes what could live there

A location isn't an idea, so Claude gets the ingredients, the nearby tags, and the companies it must not copy.

Bop.fm's one link per song becomes one link per release.
Nasdisc's physical records become the stock it finds.
Earbits's bidding becomes how stores pay for placement.

We don't take the draft on trust. It goes through the same embedding model as every real company, which shows where it actually landed on the map. If it's too similar to an existing company, or too far from the gap, Claude gets those scores and tries again, up to three times.

// CHECK 1 OF 3 · WHAT DOES THE SAME MEAN?

Reword a company, then measure it

Before we can judge an idea, we need to know what a copy scores. So we take a real description, embed it, have an LLM rewrite it, and embed that too.

EMBED

REWRITTEN BY AN LLM

EMBED

SIMILARITY BETWEEN THE TWO · 1 MEANS IDENTICAL

Different words, same company. Anything near this score is effectively a copy. Across rewritten companies, the median was .

// CHECK 2 OF 3 · WHAT DOES NORMAL LOOK LIKE?

Compare real companies with their closest competitor

Next we score every company against its most similar company on the map. Half score above and half below. These two sit at that middle point.

SIMILARITY TO THE CLOSEST COMPETITOR

A real company with a typical closest competitor scores about here.

// CHECK 3 OF 3 · WHERE DOES THE NEW IDEA LAND?

Score the idea the same way

We embed the idea and find the most similar of all companies.

VS CLOSEST COMPANY

SIMILARITY TO ITS CLOSEST COMPANY

The six ideas shown below score . A low score shows an idea differs from YC companies. It doesn't show the idea is good, or that nobody outside this data is doing it.

Six ideas we picked

Hand-picked from drafts, of which passed the distance checks. They're unverified, and may already exist outside YC. Pick one to fly the map to its gap. Numbers show similarity to the closest company.

// LIMITS

What this doesn't tell you

Ideas are compared with YC companies only, so any of them may already exist elsewhere.
The score measures how different a description is. It doesn't measure demand or quality.
Empty space on the 3D map is only a lead. It's a flattened view, so gaps are measured in the full embeddings.
Market claims in the drafts haven't been checked, and the six ideas were picked by hand from three runs.