Y Combinator · Startup School 2026
Jeff Dean: The 1% Rule for Building in AI
Google's Chief Scientist explains why the best startup ideas in 2026 are the ones where today's AI fails almost completely — and how two napkin calculations changed computing history.
Tap a timestamp pill below to jump the video to that moment.
The problem
The foundation models are eating every obvious niche — so where do you build?
The dilemma facing AI-native founders in 2026 is a peculiar kind of abundance problem. The base models keep getting better, broader, and cheaper. Every quarter another category of application that looked like a startup opportunity six months ago gets absorbed into what the general model does for free. Jeff Dean, Google's Chief Scientist, sat down with YC's Diana Hu at Startup School to name this problem directly: how do you find a durable wedge when the thing you're building on keeps expanding?
His answer inverts the usual framing. Instead of looking for places where AI is strong and riding that wave, Dean argues founders should go hunting for places where the best models today still fail almost entirely. If a state-of-the-art model solves something 0% or 1% of the time — not 20%, not 50% — that gap is probably structural, not just a matter of needing a bigger model. It points to missing training data, missing tools, or a domain so specialized that the general training distribution will never cover it well. That's where a two- or three-person team can build something a large organization has no reason to prioritize.
“look for something where the model succeeds 0% or 1% of the time not 20%”
“Are they able to do this thing very well? And if they're completely failing, that's probably a good sign. If they're kind of able to do some of it but not very well, that's maybe not a great sign — because that's probably a sign that the capability is starting to be present in those models and with more training data or larger scale models or whatever it's likely to get better. So look for something where the model succeeds 0% or 1% of the time, not 20%.”
The flip side of this rule is equally important: avoid the 20% opportunities. If a model is already doing a task one-fifth of the time, capability is clearly on the way. You're racing against the roadmap of a company with hundreds of billions in compute investment. But at 1%? The gap is wide enough to build a moat before the general models close it — if ever.
How they solve it
Napkin math, hardware intuition, and the context layer nobody has built yet
Dean's most persuasive evidence is his own track record. He walked through two napkin calculations that each triggered platform-level infrastructure shifts at Google. The first: in 2001, he and Sanjay Ghemawat simply added up the size of Google's search index and the amount of RAM across their server fleet, noticed they were about to cross a threshold, and shipped a RAM-based search system in a matter of days. Search got dramatically faster as a result. No new algorithm, no research breakthrough — just a clear-eyed look at where the numbers were heading.
“the whole search index would finally fit in all of the RAM … shipped in a few days”
“Back in 2001 Google search used to run on hard drives. You and Sanjay did the math and realized that at some point the whole search index would finally fit in all of the RAM of all the computers you had running and you made that radical realization and you basically in a few days with Sanjay shipped in production a whole new search version that worked in RAM rather than hard drive — and that was the thing that got Google to be so fast.”
The second calculation came in 2013, when neural-network speech recognition had just started working at Google. The models cut error rates by the equivalent of twenty years of prior progress in just a few months of scaling. Dean did the arithmetic: if every Google user spent three minutes a day talking to their phone, the compute requirement would double the entire server fleet. Running this on CPUs was impossible. The implication was hardware-level specialization — which became the Tensor Processing Unit. The TPU now underpins much of Google's and Alphabet's AI infrastructure.
“if every Google user talked to it three minutes a day, you would have to double the fleet”
“Back in 2013 when speech recognition started to work at Google, you did the napkin math: if every Google user used their phone and talked to it and used the speech recognition system for just three minutes a day, you found that the system would require a Google server — you would have to double the fleet, which would be really really expensive just to do speech translation. … We realized that we needed some better solution than running on CPUs. And so we came up with TPUs, which are very specialized for running neural network inference.”
Dean's third thread is what he calls context engineering: the increasingly important craft of everything that surrounds a model call. The raw model is one piece; around it sits retrieval, tool definitions, memory, agent scaffolding, and the structure of information flowing through the context window. Unlike model weights — which Dean describes as “trillions of tokens stirred together into a soup of parameters” — information in context is transparent to the model and immediately actionable. Getting that layer right is, he argues, where much of the near-term leverage sits for both large companies and two-person startups alike.
On long-running agents — a topic the audience of Startup School founders cared about visibly — Dean was candid about the current ceiling. Agents that need to run for hours or days rather than seconds still fail at compounding error: each step that goes slightly wrong compounds into the next. The fix isn't a new model; it's better checkpointing, verifier loops, and the ability to decompose work into independently evaluable sub-tasks so errors can be caught and corrected before they propagate. He predicts agents running for weeks will be routine within a few years.
“agents are great for up to step 10 or something, then gets shaky at step 50”
“Everyone here has probably seen your agent go off the rail at perhaps step 30 or 40. Like agents are great for like up to step 10 or something and then gets shaky at step 50. What do you think is the constraint today? Is it context evaluators or just errors that compound because it's basically an open-loop system? Yeah, I mean obviously we want agents to be able to run for very long periods of time — weeks — but the hard thing right now is that errors can accumulate.”
For founders specifically, Dean pointed to domain-specific depth as the structural advantage a small team still holds. Google co-designs everything from processors to products, which means attention gets spread across thousands of use cases simultaneously. A two-person team that cares obsessively about one narrow domain — and builds something genuinely delightful for it — is solving a problem a large organization has no reason to prioritize. The window, he cautions, is not unlimited: general models are getting better fast, and the 1% opportunities of 2026 may be 20% opportunities by 2027.
“a two or three person team building something they're really excited about can have an advantage”
“The thing about Google is you co-design everything on the system from the processors to the products — which are the layers that someone like Google would keep building and compounding. And in a lot of cases that means that we don't have a lot of attention on particular domains where perhaps a really well-designed surface, and maybe a model and set of skills, or a specialized model that isn't in a general mix of things, can actually have a significant advantage because you can build something delightful and really high-accuracy for a domain you're really passionate about. That's where a two or three person team building something they're really excited about can have an advantage.”
Takeaway
The quick version
- Look for domains where the best model today succeeds 0–1% of the time, not 20%. The 20% cases are already on the roadmap of a well-funded lab.
- Napkin math still finds platform shifts: add up the numbers, see where they cross a threshold, and build what the crossing requires. The TPU started on a napkin.
- Context engineering — retrieval, tools, memory, agent scaffolding — is where most near-term leverage lives, for startups and incumbents alike.
- Domain depth is a genuine moat right now. General models spread attention across everything; a passionate small team can be dramatically better at one specific thing.
“Look for something where the model succeeds 0% or 1% of the time, not 20%. The 20% cases are probably a sign that the capability is starting to be present in those models, and with more training data or larger scale it's likely to get better.”— Jeff Dean, Google Chief Scientist at Startup School 2026
The 1% rule is not a consolation prize. It's a map to the places the big models haven't gone yet — and where the math still works for founders who do it on a napkin.