Dev recalls fighting for structured output in GPT-3 days as dynamic inputs now parse in <200ms
blixt · x · 2026-09-16
Developer @blixt recalls spending lots of time coaxing structured data out of GPT-3/3.5-era models, when constrained decoding was a lifesaver. He argues the new approach reinvents that concept at scale: good results for dynamic inputs in under 200ms. Responding to skepticism, he says the attention isn't just hype — it's one of those obvious-in-retrospect ideas that someone actually shipped.
More from coding & agent
- NoSpoon agent churns out "so bad it's good" AI slopdrama in ten minutes, site closing soon — Kyrannio · 2026-09-16
- Harness Engineering: the skill that replaced prompt engineering in 2026 — pstAsiatech · 2026-09-16
- Graphify, a 118K-star repo, turns codebases into queryable knowledge graphs for coding agents — techNmak · 2026-09-16
- Postgres pro tip: set application_name for human-readable query sources — DanielLockyer · 2026-09-16
- AI training at an energy company shows domain expertise doesn't teach agent engineering — NumbersProtocol · 2026-09-16
- Ministral 3 3B on a Galaxy S21 relays chats between Gemini and Z.ai across two browsers, 10/10 runs — Mean-Standard7390 · 2026-09-16