# Ideas - requests for startups, from this forecast

← [Startups](README.md) · [III.A](../README.md) · [Index](../../../README.md)

---

[Y Combinator's Requests for Startups](https://www.ycombinator.com/rfs) is a standing tradition: name a problem that is newly solvable, sketch what a company could be, and invite founders to build it. This page does the same job **through the lens of this corpus**. The filter is stricter than "cool AI product." Every idea below is something the rest of the document implies is **undersupplied relative to how much the equilibrium needs it** - usually a scarce complement ([Game 3](../../../02-games/3-firms.md), [compressed version](../../../00-overview/compressed.md)), not another layer of free cognition.

You do not need to build these exact companies. If one of them is wrong, the wrongness is informative: it means a claim elsewhere in the corpus is wrong too.

## How each idea is written

YC's format is: *why the world shifted, what to ship, why you*. Ours adds the economic spine so founders do not confuse a demo with a firm:

| Field | Meaning |
|---|---|
| **Why now** | The substrate or institutional change that makes this newly buildable or newly valuable |
| **Build this** | The concrete product or service - not a category slogan |
| **Own this** | The inelastic complement the company must hold (data, license, distribution, atoms, balance sheet, trust) |
| **Not this** | The Red Queen version - what dies when models improve |
| **Dead if** | The falsifier; if this happens, skip the idea |

An idea that cannot name **Own this** in one sentence is not ready. Re-read [selection](selection.md).

## The meta-request

> **Build companies that make verification, liability, physical capacity, or durable distribution cheaper - not companies that only make generation cheaper.**

Generation is already on a cost curve that will not leave margin ([inference economics](../../../01-substrate/inference-economics.md)). Verification, permits, powered land, licensed accountability, and human trust are not. The [compressed version](../../../00-overview/compressed.md) says the same thing in five bullets; the catalogs below turn those bullets into company shapes.

## What not to start (anti-RFS)

These recur in every batch deck and almost never clear the [selection](selection.md) map:

- **ChatGPT for X** with no exclusive data, license, or channel - a feature with a landing page
- **Another coding agent** without owning eval data, enterprise context, or the review bottleneck
- **Generic "AI employees"** sold as seats into competitive industries - pure Red Queen for the buyer, pure churn for the vendor
- **Thin wrappers on a single frontier API** - the model vendor's roadmap is your product plan
- **Vertical SaaS that only embeds workflow** with no data rights - rebuild cost is falling under the customer ([software](../software.md))

If your idea is one of these, either find the complement or expect a fast cheap death ([exits](exits.md)).

## Catalogs

| | Theme | Through-line |
|---|---|---|
| [Verification & trust](ideas-verification.md) | Provenance, human presence, review, evals, accountability | Making ground truth and trust cheap where generation flooded the channel |
| [Atoms & substrate](ideas-atoms.md) | Energy, robotics data, field ops, logistics, labs, aging | Physical clocks and rights that software cannot mint |
| [Institutions](ideas-institutions.md) | Insurance, compliance, apprenticeship, law, education | Where deployment is gated by licenses, balance sheets, and formation of experts |

## How this differs from a typical RFS list

YC optimizes for "we would like to fund this." This list optimizes for **"this forecast says value and social need concentrate here."** Overlap is large (trust layers, physical-world OS, defense, compliance, education) because the same scarcities are visible from both seats. Divergence: we de-weight pure consumer AI apps and model-adjacent wrappers unless they capture a complement, and we up-weight insurance, apprenticeship, and verification infrastructure that look boring in a demo day and decisive in 2032.

## Failure mode for the whole list

If [learned verification](../../../06-uncertainties/learned-verification.md) retires the master asymmetry, half the verification catalog softens. If [complement erosion](../../../06-uncertainties/complement-erosion.md) runs fast, the atoms and license ideas lose half-lives. If platforms intermediate all demand, distribution-native ideas win and independent complement owners become tenants. Score the list against those three pages, not against whether a given demo looks good this quarter.

---

**Related:** [Selection](selection.md) · [Formation](formation.md) · [Compressed](../../../00-overview/compressed.md) · [Game 3](../../../02-games/3-firms.md) · [YC RFS](https://www.ycombinator.com/rfs) (external format reference)

**Next:** [Verification & trust](ideas-verification.md)