AI agents 4
- Spending compute at inference time to search over plausible solutions is the general fix for unreliable long agent runs.
- Agents can already be run for days or weeks on a single hard problem, and almost nobody has internalised that.
Jeff Dean
Chief Scientist at Google DeepMind and Google Research, at Google since 1999, where he co-created MapReduce, Bigtable, TensorFlow and the TPU.
Jeff Dean did not write this page.
We collected these quotes from things they published elsewhere, and every quote links to where it was said. They have no account here and have not endorsed this site. Quotes are word for word; the short line under each one is our own restatement, not their wording. Their own site. Is this you? Claim it or ask us to remove it. Or tell us what is wrong here.
What Jeff Dean says most about
coding agents 3
- Specifying what you want matters more now that agents write the code than it did when the work went to a human who could ask follow-up questions.
- AI models have reached the level of a junior engineer on longer-running, agent-based coding tasks.
energy 3
- The numbers an AI engineer should know by heart are memory bandwidths and the energy cost of a single operation.
- Inference, not training, is now the constraint, and specialized low-energy hardware will beat general-purpose GPUs and TPUs on latency.
LLMs 3
- What sits in a model's context is far clearer to it than its training data, which is trillions of tokens stirred into a soup of parameters.
- A startup's durable advantage over a general model is having visibility into data the general model will never see.
scaling laws 3
- Spending compute at inference time to search over plausible solutions is the general fix for unreliable long agent runs.
- Before building on a gap in the general models, work out whether that gap survives six months or three years.
semiconductors 2
- Inference, not training, is now the constraint, and specialized low-energy hardware will beat general-purpose GPUs and TPUs on latency.
- A chip that can do nothing but low-precision dense linear algebra is enough for machine learning, even though it cannot run Chrome.
taste 2
- When agents do the building, the scarce skill is taste: choosing which problem is worth the time at all.
- Taste can be trained deliberately: write down what you think will matter in twelve months, then go back in twelve months and score yourself.