Jeff Dean on Context Engineering

"Trillions of tokens stirred together into a soup" — that's how Jeff Dean describes a model's training data, contrasting it with the context for a specific task, which is far clearer and sharper. Google's former Chief Scientist, who helped build its AI from the ground up, made the core point at a Y Combinator interview with Diana Hu: which model you pick is secondary; what matters more is how you build the system around it.

Dean calls this context engineering. The model, he says, is "really only one piece" of a larger system. What shapes the result far more is tooling, access to information retrieval, multi-agent orchestration, and clear context relevant to your specific problem. His full argument is laid out by Search Engine Journal.

What Dean advises:

• learn from failure — watch where the model stumbles on real tasks;
• adjust guidelines and skills, not parameters: improve the context instead of retraining;
• iterate — every failure shows what information the model was missing.

The key upside is accessibility. Training a model takes massive GPU resources, while context engineering is within reach of anyone with an API and their own retrieval-and-tools setup. For SEO and content teams that's a direct signal: don't wait for a "smarter model" — structure your information so AI can find and use it correctly. Clear facts, clean sources, and thoughtful presentation matter more today than which neural network happens to be answering the user.