Jeff Dean says context engineering beats raw model size at YC Startup School
ycombinator · x · 2026-07-26
- Jeff Dean’s talk at YC Startup School argues that systems, not just models, solve problems: the context a model sees at inference time—tools, retrieval, and surrounding data—matters more than its pretraining corpus.
- His practical advice for better context engineering is to treat the model like a person, give it the kind of instructions you would give a human, and inspect where the system fails so the harness can be improved.
- He also describes a model “comfort zone”: performance drops beyond a certain boundary, so the mitigation is to use skills, hints, and multi-agent systems that explore different approaches and evaluate the best one.
More from coding & agent
- LangChain traces could become training data for open-source models — hwchase17 · 2026-07-26
- Opus 5 Gamed Its Own Eval: Deleted 18 Tests Overnight to Boost Scores — springrod · 2026-07-26
- Multi-agent system pitches autonomous dev, research, and business teams — tom_doerr · 2026-07-26
- The Engineering Gulf Between Vibe Coding and Enterprise Legacy Systems — springrod · 2026-07-26
- AI lecture argues most users tap only 10% of agents, while graph workflows run the rest — Roger_M_Taylor · 2026-07-26
- unlocalhost lets coding agents ship apps with invite-only access and secret checks — Outrageous_Mark9761 · 2026-07-26