A 20-cent AI that can't write a sentence caught 13 requests our classifier missed
mikegiannulis · x · 2026-09-18
The author shares a counterintuitive experiment: a tiny model that can barely write a full sentence, tested for just 20 cents, caught 13 customer requests that their support classifier had missed, and raised the rate of ranking the correct source passage first from 60% to 90%. The takeaway: we may be using LLMs for the wrong jobs — many tasks don't need a fluent writer, and small specialized models can beat big ones on classification and retrieval ranking at a fraction of the cost.
Related event: TypeSafe Launches Jev, a Decision-Only Model That Never Writes a Word(44 posts)→
More from coding & agent
- Skip Scoping, Start Proving: A Five-Step Playbook for Shipping Enterprise AI at Startup Speed — alex_verem · 2026-09-18
- Berkeley study: the right agent harness cuts cost of the same result by 71% — MartinGTobias · 2026-09-18
- Meta Built an AI Agent as a Domain 'Secondary Expert' — With Auditable Knowledge Architecture — Meta_Engineers · 2026-09-18
- dotey on the AI code quality debate: black-box verification is replacing code review — dotey · 2026-09-18
- AgentSky launches as an agent marketplace: 40+ coding agents in browser, 44x cost gap between models — Scobleizer · 2026-09-18
- AWS Ships Six Open-Source Skills to Let Coding Agents Deploy Hugging Face Models on SageMaker — AWS ML Blog · 2026-09-18