Matt Holden: structured output is LLMs' real magic, now at 300ms and nearly free
holdenmatt · x · 2026-09-18
Replying to James Cowling, Matt Holden says function calling / tool use / structured output was always his favorite LLM feature: turning unstructured intent into strict, validated schemas is "the magic you can actually program with." He notes the rest can now be shed for 300ms latency at near-zero cost, adding that he "f'ing hates JSON Schema."
More from coding & agent
- Users beg Claude Code lead for a reset for weeks — total silence from Thibault Sottiaux — RileyRalmuto · 2026-09-18
- Vague prompts force AI to make endless calls: why good engineers still hand-write key code — bendee983 · 2026-09-18
- Codex compaction: the LLM requests a blank context and keeps its own notes — mitsuhiko · 2026-09-18
- Google's ToolGrad flips tool-use dataset generation: answers first, queries later — AxSaucedo · 2026-09-18
- Founder advice: add an MCP to your SaaS, hype or not — jonathan_wilke · 2026-09-18
- Full harness open-sourced: controller, planner, safeguards, and media verification on GitHub — imjustnewatai · 2026-09-18