DeepMind-led paper makes design docs the source of truth, code disposable — SMART regenerates in 1.5-3h for ~$100

Roger_M_Taylor · x · 2026-09-11

A DeepMind/MIT/Stanford/Google paper, "Design Docs Are All You Need," proposes treating natural-language design docs as the durable artifact and code as a disposable build product. Their SMART ML performance-modeling library's main branch is almost code-free — a DAG of self-contained docs from which coding sub-agents regenerate the implementation on every spec change. Two ingredients make this reliable: docs built around step-by-step worked examples as in-context demonstrations, and a minimal recursive operator IR with symbolic SymPy cost expressions. Regenerated code reproduces hand-audited reference models (including DeepSeek-V3 serving on a TPU pod slice) to round-off precision; full regeneration takes 1.5–3 hours and $100.

Original post →

More from coding & agent

coding & agent channel →