The 1% Rule for Building in AI

Dean says AI models have become capable at longer coding tasks, roughly matching his earlier prediction that they were reaching junior-engineer ability. He expects more automated experimentation to improve machine-learning systems, and says specialized inference hardware could cut latency substantially; he asks readers to imagine a 50-fold improvement.

Dean recalls estimating that three minutes of speech recognition per Google user each day could require doubling the server fleet. That helped motivate TPUs, specialized for low-precision linear algebra; early versions were 30 to 80 times more energy-efficient than CPUs and GPUs of the day, with 20 to 30 times lower latency. He says data movement can cost about a thousand times as much energy as a calculation, making batching and memory bandwidth important design constraints.

Dean describes AI products as systems combining models with retrieval, tools, task breakdown, and evaluation. At Google, he and Sanjay wrote a skill that let an agent edit code, run microbenchmarks, and iterate on performance improvements. For long-running work, he recommends skills that keep agents on familiar paths and multiple agents that explore and evaluate different approaches.

Dean says small teams can win by building excellent products for domains they know well, especially where they have access to useful data or can train a specialized model. Founders should test what general models can already do and look for tasks where they fail almost completely, since partial capability may improve quickly. Clear specifications help agents succeed; translating Python software to Go is a strong example because existing code and tests define the expected behavior.

Dean argues that choosing worthwhile problems will remain a key human skill as agents write more code. He recommends building judgment through experience, revisiting predictions about what will matter, and questioning assumptions with thought experiments. He sees automated experiment loops as promising for machine learning, science, and engineering, and highlights data-efficient learning and continual learning as open challenges.