What is an AI moat? 11 wedges into markets incumbents already own.

Jeff Dean built MapReduce, BigTable, TensorFlow, the TPU and Gemini, so his rule for picking an AI problem is worth taking literally. Run today's general models against your domain and read the score. If they are completely failing, he says, that is probably a good sign. If they are kind of able to do some of it but not very well, that is not a great sign, because the capability is already present and more data or a bigger model finishes it.

His instruction is exact. Look for something where the model succeeds 0% or 1% of the time, not 20%.

Dean's second filter is the one founders skip. He asks whether the thing you are working on is a durable thing, or whether the models at the forefront get better at it in the next six months or 12 months, or whether it is something they will not manage for a couple of years. Six months means you are building a feature the labs have not reached. Three years means you have room to become a company.

General models keep getting better at a broader and broader range of things, so the window closes on its own schedule, not yours.

So where does the two or three people in a room advantage come from? Dean names two shapes. The first is a product with access to a particular kind of data the underlying general model does not have, like a user's own personal information, so your product sees what the frontier cannot. The second is training the narrow model yourself. He points at AlphaFold, a very specific model for protein folding that beat general models in its domain, and says the same shape works in material science or chip design.

A niche model may not take much compute, so this is judgment, not capital.

Scale AI started as something else. Alexandr Wang came into Y Combinator building an AI agent to help people get medical care, worked on it a month or two, and had Jared Friedman pull the team aside to say he did not know if it was going anywhere. What Wang did next is the wedge. He went back to first principles on what training a model required, and found three inputs.

Compute you could get by pressing a button. Code you could get by pressing a button. For data, he says, there was no effective way to get it at all.

Wang's read is that the startup story changed shape. Ten years ago, when he started Scale, it was David versus Goliath, and you had to be clever, find an angle into the market, and compete with far fewer resources. Now, with the power of agents, he thinks it is much closer to Goliath versus Goliath, with the startup as a mecha Goliath vastly enhanced by AI. His conclusion is that startups which properly embrace agents can easily outcompete incumbents.

What founders skip is that the multiplier is available to both sides. If agents make four people enterprise-sized, they do the same for the enterprise.

Altman's general observation is that startups tend to win when the technology landscape is moving very quickly, when costs are coming down and cycle times are short, and he thinks all of that is happening now. He walks the pattern back through the internet boom, the wave of people building on Facebook, and the iPhone App Store. The line to keep is that great startups tend to cluster when the ecosystem shifts and incumbents lose a lot of their advantage. You are not hunting a market where the incumbent is bad, but one where what the incumbent was good at stopped mattering.

OpenAI began as a Y Combinator research project at a moment when, in Altman's telling, people said the team was not only wrong that AGI might be possible but was going to single-handedly cause another AI winter. His highest bit of advice is to find things you can develop reasonable conviction in that the conventional wisdom says are just wrong. For years it felt like they knew the biggest secret in the world while everybody called them an idiot, and that dismissal bought them time without massive competitors. Start the same startup as everybody else and you get hype and money, rarely the big outcomes.

Matt Murphy led Anthropic's seed round at Menlo Ventures, and gets the Anthropic-will-eat-your-app question constantly. His answer opens with a concession founders will not enjoy. If the model just does something and your application is not distinctive enough, the model takes that market away, and it probably was not that defensible anyway. Then he names what distinctive looks like at Legora, his legal AI investment.

Its lawyers get in and understand workflows that cross organizational boundaries. On one case there are corporate lawyers, outside firms and a client, plus context inside each firm. Not quite an n-squared problem, as he puts it, but you need workflows that understand it.

The obvious objection to a vertical wedge is that the vertical is too small to carry the ambition. Murphy's answer is that the wedge was never the market. Legora's strategy, he says, is absolutely to expand out of legal into tax, compliance and accounting, because what the company built is a base platform that happens to be really good at understanding sophisticated service teams. Legal is where that got proven, not the ceiling, and he expects another leg of the stool nobody has named yet.

You earn the adjacent vertical by being undeniable in the first one.

Do margins still matter? Murphy says they do a lot, then describes a market that looks nothing like the SaaS one. A lot of great companies right now sit around 20 to 30 percent gross margins, and because of the cost of inference it is harder to say you are going to be an 80 or 90 percent gross margin company anymore. What he underwrites is the path. Great companies land at 60 to 70 percent, and the credible plan is some optimization, not being completely tied to inference, and often building your own model on your own data.

Boris Cherny built Claude Code. Product overhang is the gap between what the model can already do with today's models, not a future model, and what any shipped product lets it do. When the product gets in the way, he calls that hobbling. Sonnet 3.5 was the best coding model available and the coding products of the day did single-line autocomplete, sometimes multi-line, plus read-only chat.

So the team got rid of the scaffolding and gave the model the simplest possible harness, enough to write an entire file and build an entire feature. Cherny says there is now so much product overhang that he is not seeing startups capture.