Jeff Dean: The rule 1% for AI products (AI column)
Video Watching Jeff Dean is always interesting and informative. He can start with a story about MapReduce or calculation on a napkin, and after a few minutes to bring the conversation to the architecture of the next generation of systems. I've told you about him before. TED Talk, major lecture at Rice University, Interview with Noam Shazier and A retrospective of AI development at Stanford AI Club. Now I've looked. conversation Jeff with Diane Hu at YC Startup School 2026 - «The 1% Rule for Building in AI”. If in previous lectures he explained how we came to modern models, then here the question is more practical: what else makes sense to build a small team when universal models quickly pick up more tasks.
The most useful idea of the release - the very rule 1%. According to Dean. It is dangerous to choose an area where the frontier model is already coping with approximately 20Percentage of cases: This is a sign that the ability has appeared and after six months or a year the base model can eat a significant portion of the product. It is better to look where today it is only successful. 0–1And the team has a way of turning that one percent into a working system.
This way can be closed subject data, a narrow specialized model, a good interface or engineering binding: tools, memory, skills and a verifiable execution cycle. This is not a guarantee of protection from large laboratories, but the discipline of task choice: to look not for a beautiful superstructure over the current model, but for a sustainable advantage around it.
Another important thought was that the model is only part of the system. Dean talks about contextual engineering as a space where a small team can still compete. Knowledge within scales is mixed with trillions of learning tokens, and information in context is given to a model right now and for a specific task. Therefore, the quality of the system is increasingly determined by what data it obtained, what tools it chose, how it decomposed the task and how it checked the result.
A good example is Dean’s work with Sanjay Gemawat. They described for the agent the usual cycle of optimization of low-level code: run microbenchmarks, change implementation, again measure performance, check a wider set of scenarios and cache size, then repeat. In fact, many years of engineering experience turned into skill. Their public document “Performance Hintscan be used as a starting material for a similar binding.
For long tasks, one instruction is not enough. Agents can already work for days and weeks, Dean said, but easily go beyond the boundaries of familiar distribution and start making mistakes. Reliability emerges from a clear specification, tests, skills, multiple parallel attempts, and a separate appraiser who discards failed branches. Transferring a program from one language to another works especially well: source code and tests become a detailed executable specification.
And, of course, the conversation with Jeff could not do without iron. He's talking again. TPU: 2013 According to the study, just three minutes of speech recognition per day would require Google to double its server fleet. The answer was a specialized chip for low-precision linear algebra, published by Google analysis of the first TPU showed in the 30–80 The best performance per watt compared to the then CPU and GPU.
Now, according to Dean’s prediction, the next big shift will occur in the specialized hardware for inference: fewer data moves, lower computational accuracy where it is permissible, and radically less latency. This is an important correction to talk about “model quality”: many of the limitations of an AI product turn out to be limitations of memory, power, bandwidth, and response costs.
Finally, Dean expects that in 2027 Automation of ML itself will increase significantly in the year: the system will lay out the task, run many experiments, evaluate the results and collect an improved version. AlphaEvolve It already shows the form of such a cycle, although Google’s results here should be taken as company data, not as an independent confirmation of the universality of the approach.
So it was interesting to hear Jeff’s thoughts on what the future holds for the industry.
#AI #Agents #Engineering #Architecture #Infrastructure #Product