We get asked a version of the same question every week: will AI kill humanity? We are not the right people to answer it. Serious researchers disagree, and a seed fund adding its opinion would not move the debate.
But there is a narrower question we answer for a living, every batch, before Demo Day: if models keep getting dramatically better, which companies get stronger, and which disappear? That question decides most of what we invest in. Here is how we answer it.
Three takeaways
- Ask what a 10x better model does to the company. If it makes the product stronger, the company is AGI-resilient. If it makes the product unnecessary, the company is renting time from the model labs.
- Resilience lives in what a model cannot download. Proprietary data, physical deployment, regulated access and verification all get more valuable as intelligence gets cheaper.
- This is now an explicit filter, not a lucky side effect. We ask every founder the same few questions before Demo Day, and the answers shape both whether we invest and how much.
The 10x test
Every AI company is built on a bet about the next model. Some are betting it will be better at the thing they sell. Others are betting it will make the thing they sell easier to deliver. Only the second bet compounds.
So we run one test on every company we see: imagine a model ten times better than today's ships tomorrow. Does this company get stronger or does it get replaced? A thin interface over a general model fails it, because the next model release is the competitor. A company whose value sits somewhere the model cannot reach passes it, because a smarter model makes its product cheaper to run and better to use.
The test is not about whether a company uses AI. Almost everything in a YC batch does. It is about where the company's value lives once intelligence itself is abundant.
Where resilience lives
Across the companies we have backed, four sources of resilience keep showing up. Each one is something a better model makes more valuable, not less.
| Source of resilience | Why a better model helps it | Portfolio examples |
|---|---|---|
| Proprietary data and systems of record | Models are only as good as the data they can act on. Owning hard-to-get records, or the workflow where decisions are recorded, turns better models into a better product. | Landeed (property title search in India), Hub.xyz (rights-cleared training data), Solve Intelligence (patent drafting and prosecution) |
| Physical deployment | Intelligence does not move atoms. Robots, vehicles and chips need engineering, manufacturing and field deployment that no model release replaces. | Andromeda Surgical (autonomous surgical robots), Revoy (hybrid-electric semi trucks), Zettascale and Baud Labs (AI chips) |
| Regulated access | Licenses, integrations with regulated systems and accountability for outcomes take years to earn and do not ship in a model update. | Healthtech-1 (patient navigation and automation for NHS practices), Salvy (mobile carrier for businesses in Brazil) |
| Verification and trust | The more code and decisions AI produces, the more valuable it becomes to prove they are correct and safe. | Theorem (program verification), ZeroPath (AI penetration testing), Tester Army and Momentic (AI testing) |
There is a fifth pattern worth naming: infrastructure that consumes intelligence. Companies like Trigger.dev, Browser Use and Magnitude run the jobs, browsers and inference that agents depend on. When models get better, more work flows through them, and their input cost falls.
What fails the test
The companies that worry us share a shape, and it is usually visible in the first meeting:
- The roadmap is the model provider's roadmap. If the feature list reads like a list of things the base model cannot do yet, every release shrinks the product.
- The moat is a prompt. Clever orchestration is valuable for about one model generation. Without data, distribution or deployment behind it, it does not survive the next one.
- Pricing per seat for work the model will do natively. As the task gets automated inside the tools people already use, the seat disappears, and so does the revenue.
None of this means a company is a bad business today. It means its value depends on the model staying where it is, and we are not willing to make that bet at seed.
How we apply it before Demo Day
We invest weeks before Demo Day, so we rarely have revenue history to lean on. What we do have is the founders' answers to a few questions that sort AGI-resilient companies from the rest:
| Question | What a strong answer sounds like |
|---|---|
| What happens to your product the day a much better model ships? | It gets cheaper to run and better to use, and here is why. |
| What do you own that a model cannot download? | Data, a deployed footprint, a license or an integration that took real work to earn. |
| Who is accountable when the output is wrong? | We are, and customers pay us to be. Verification is part of the product. |
| Does your cost of delivery fall as models improve? | Yes. Our margins expand with every model generation instead of compressing. |
Founders who have thought hard about these questions usually answer them before we ask. That itself is one of the strongest signals we see.
How we're positioned
- Weighting toward the four sources of resilience. Proprietary data, physical deployment, regulated access and verification now run through most of our recent checks.
- Backing the trust layer early. As AI writes more code and makes more decisions, proving correctness becomes infrastructure. We think this category is still early.
- Following on where resilience is proven. When a company shows that better models make it stronger, that is when we want to own more of it.
What could go wrong
- Resilient is not the same as successful. A company can be safe from model progress and still fail on distribution, execution or timing.
- Physical and regulated companies move slower. Hardware, healthcare and telecom carry longer feedback loops and more capital needs, so we size accordingly.
- Model labs move into verticals. The largest labs are building their own applications and data partnerships, which can erode advantages we assumed were durable.
- We may misjudge what is proprietary. Data that looks unique today can become a commodity when someone else collects it at scale.
Back to the big question
We still won't tell you whether AI will kill humanity. What we can say is that the companies we are most excited about are the ones making powerful AI safe to rely on: proving code correct, testing systems before they ship, and keeping people accountable for outcomes. However the bigger story ends, that work matters, and it gets more valuable with every model release.
This is market commentary, not investment advice. Company descriptions reflect each company's public positioning. Examples are illustrative and are not a complete list of our portfolio.
