Jul 30, 2026 · 57m · y-combinator
Jeff Dean: The 1% Rule for Building in AI · Y Combinator
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this Y Combinator Startup School interview, Google Chief Scientist Jeff Dean and YC Managing Partner Diana Hu discuss the future of AI engineering, covering long-running autonomous agents, specialized inference hardware, context engineering, and strategic advice for startup founders.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. The partners hold 25.1% of the talking time here. How this is scored →
The partners as informed peer 6.0 Guest teaching 2.4 Guest disagreement 0.1 The partners pushing back 0.1
speaking balance: gold is the partners, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 14:43 Dean redirects focus away from batch size one training
When the host playfully asks if Dean could invent batch size one training over a weekend, Dean immediately pushes past the premise to state he is focused on inference optimization rather than training latency.
Hardest push from the partners ▶ 14:26 Host pushes on whether batch size one training can be solvedThe host challenges Dean directly on whether the fundamental systems bottleneck of training batch sizes could be eliminated by him in a dedicated sprint.
Biggest teaching moment ▶ 37:38 Dean outlines fault-tolerant computing with faulty transistorsDean educates the audience on why standard silicon fabrication enforces strict near-zero defect rates and demonstrates how applying distributed storage redundancy concepts at the transistor level could redefine hardware design.
The partners hold their own ▶ 12:06 Host details 1 picojoule compute versus 1000x IO energy bottleneckThe host articulates the exact physical energy constraints of high-bandwidth memory IO versus floating-point computation, showing how model training limitations are fundamentally systems and energy bottlenecks.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The partners as informed peer | Guest teaching | Guest disagreement | The partners pushing back | Why |
|---|---|---|---|---|---|---|
| Title Card and Opening Sequence | 5 | 1 | 0 | 0 | Diana opens by precisely citing Jeff Dean's previous public prediction from AI Ascent regarding AI reaching junior engineer capability and asks for his 2027 prediction. Dean responds cooperatively, forecasting automated machine learning research loops. | |
| The Next 'Fits in RAM' Moment: Specialized Inference Hardware | 6 | 2 | 0 | 0 | Diana contextualizes the discussion with Google's 2001 transition from disk to RAM indexing, prompting Dean to identify modern parallels. Dean identifies low-power, specialized inference hardware and multi-day agent execution as the modern equivalent. | |
| The Origin of the TPU and First-Principles Napkin Math | 6 | 3 | 0 | 0 | Diana recalls the origin of the TPU based on speech recognition scaling math and observes how TPUs pre-dated transformers. Dean elaborates on why designing flexible dense linear algebra accelerators proved durable across algorithm shifts. | |
| Updated System Latency Numbers and Energy IO Bottlenecks | 7 | 3 | 0 | 0 | Diana displays deep technical knowledge by citing Dean's famous systems latency numbers and highlighting the 1 picojoule compute versus 1,000x data IO energy disparity. Dean confirms this physical constraint dictates why batching is necessary across machine learning workloads. | |
| Inference Optimization and the Shift to Context Engineering | 6 | 2 | 1 | 1 | Diana pushes on whether batch size one training could be solved over a weekend, but Dean redirects the focus toward inference latency and aggressive low-precision specialization. Diana connects the discussion to compression theory. | |
| Mastering Context Engineering with Practical Code Optimization | 6 | 3 | 0 | 0 | Diana frames the industry shift from pure parameter scaling to context engineering, tool usage, and retrieval. Dean explains how Sanjay Ghemawat and he codified internal microbenchmarking workflows into agent skills and published the Performance Hints paper. | |
| Preventing Agent Failure and Leveraging Multi-Agent Search | 6 | 2 | 0 | 0 | Diana asks why multi-step agents degrade at step 30 or 40. Dean explains distribution drift when agents stray from training data and advocates for inference-time search with multi-agent evaluators. | |
| Where Small Startup Teams Can Win Against Tech Giants | 6 | 2 | 0 | 0 | Diana asks how early-stage startup teams can compete against vertically integrated frontier labs. Dean advises targeting domains where frontier models fail completely (0-1% accuracy rather than 20%) or where proprietary local data is critical. | |
| AI-Native Engineering: Specifications, Taste, and Crazy Thought Experiments | 6 | 4 | 0 | 0 | Diana inquires about developing taste in an agentic coding world. Dean shares a radical thought experiment about designing systems using unreliable transistors with high daily error rates by applying distributed systems fault tolerance, which Diana relates to neuromorphic computing. | |
| MapReduce Origins, Automated Research Loops, and Handling Rejection | 7 | 3 | 0 | 0 | Diana prompts Dean on past contrarian bets like MapReduce, automated scientific evaluators, and the famous 2014 NeurIPS rejection of knowledge distillation. Dean explains how fast neural surrogates speed up experimental loops by 300,000x. | |
| Founder Career Advice, Building High-Impact Teams, and Future AI Horizons | 5 | 1 | 0 | 0 | Diana asks Dean how a 25-year-old Jeff Dean would approach AI today, how to build high-leverage teams, and which grand challenges to tackle next. Dean outlines the trade-offs of frontier labs versus startups and emphasizes low-ego collaboration. |
Statements from this episode (15)
Jeff Dean: AI models have reached junior engineer capability level
“The models have been getting a lot better at sort of agent-based, longer-running coding tasks, and it seems pretty clear that they are now actually pretty capable, and depending on exactly your definition of junior engineer, it seems pretty spot on, I would sa…”
Dean: Specialized inference hardware will surpass general GPUs and TPUs
“I think, ah, you're gonna see more and more, ah high performance and low energy inference hardware systems, because I think everyone is now realizing that inference is the key to making, you know, these agent-based systems be available to more and more people,…”
Jeff Dean: AI agents can run autonomously for days or weeks
“Probably one thing is people don't quite realize how possible it is to have, you know, agent based systems that can run not just for an hour or two hours on a problem you care about, but for some problem domains and with highly capable models underlying them, …”
Dean: 1,000x energy penalty for moving data forces machine learning batching
“I mean, I think the example you raised of a thousand X difference in bringing moving data versus actually computing on it in, in terms of energy is, is a pretty significant one. And it shapes a lot of aspects of what we do in machine learning. Because if you d…”
Dean: In-context data is clearer to models than pretraining parameters
“And the nice thing about that is that information is really clear to the model. Unlike the training data, the model was trained on where it's all kind of like trillions of tokens stirred together into a soup of hundreds of billions or trillions of parameters, …”
Dean: Sanjay Ghemawat and I published a 30-page code optimization guide
“Oh we actually published the document maybe a few months ago called performance hints that Sanjay and I wrote. That's like a 30 page document about, you know, various kinds of performance tricks and Some people have taken that and then given it in summarized f…”
Jeff Dean: Inference-time compute search improves reliability in agent workflows
“Inference time compute to perform search over plausible ways of solving the problem that can get much, much higher performance or much more reliability in, Long-running agent flows.”
Jeff Dean: Google adds custom skill definitions for internal developer agents
“Particularly in the internal Google development environment, we have skills so that the agents can know how to use lots of our internal tooling for coding, or for code reviews, or for, you know, measuring performance, or, you know, fetching log files, and thos…”
Dean: AI startups should target tasks with 1% model success, not 20%
“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 tr…”
Dean: AI coding agents make crisp software specifications more important than ever
“The importance of specifying what you, what it is you want has actually gone up, because before you'd be handing it off to a very intelligent human who maybe has context or can ask you follow-up questions and agents can sometimes do that, but I think clear spe…”
Dean: Code translation is an ideal AI use case due to existing tests
“A use of a coding agent that works extremely well, is you can ask today's models to translate software from one computer language to another very effectively, because in that case, you actually have a incredibly detailed specification. You have the whole softw…”
Jeff Dean: AI models will not have good taste in problem selection
“I think models are not necessarily going to be that good at it. So you're going to have people steering a lot of AI assisted computation in order to accomplish great things and more quickly. But that essence of what it is you want your models to do is the key …”
Jeff Dean: Google built a neural simulator 300,000x faster than full DFT
“Take a bunch of output from those simulation runs, the input molecule configurations and the outputs of that. The extensive simulator and then use it to train a neural approximation to the simulator. So this is now a validation device, but instead of it taking…”
Jeff Dean: There is no impediment to fully automating ML model research
“There's no, ah, you know, real impediment to making that be a much more automated loop, where the model itself decides it's going to explore, or maybe with a nudge from some people, ah, at the various highest level, like, oh, why don't you try some new ideas a…”
Jeff Dean: Frontier AI models consume 1,000x more data than 18-year-old humans
“If you think about our large scale models today, they probably see a thousand times as much data as a human does by the age of 18, yet the human by the age of 18 is better in a lot of things and, you know, on par with those frontier models that have seen way m…”