Transcript Testing: Apache DataFusion maintainer says AI-era tests should be written for humans to read
mgill25 · x · 2026-11-14
- Andrew Lamb, Apache DataFusion maintainer, proposes Transcript Testing: as AI tools mean humans write less code, software tests should be explicitly written for humans to read, to verify and understand the systems we build.
- In the past 6 months he hasn't seen a PR with inadequate tests — instead, the opposite problem: AI-generated tests spanning hundreds of lines that reviewers skim or skip entirely (himself included).
- Example: a PR optimizing DataFusion's Parquet Bloom filter reading shipped four test functions (400 lines total), one test alone at 117 lines.
- His takeaway: humans should still design and maintain software, likely in natural language, and the abstraction level of verification (tests) must rise accordingly.
More from coding & agent
- HQ pitches an open-source 'company brain' giving team AI a shared context layer — jacob_posel · 2026-11-14
- Cohere Labs talk: Harness² recursively improves agent harnesses at test time — Cohere · 2026-11-14
- Someone Built a Provably Fair Dice Casino Where AI Agents Place Bets via MCP — adamamcbride · 2026-11-14
- Intern shipped a multiplayer app built with Claude Code, got hired, then presented it on stage — Alec_Coughlin · 2026-11-14
- Google Research showcases Antigravity's /teamwork for science & engineering — Google Research · 2026-11-14
- Dev: Working at a High Level with AI Is Great but Exhausting — jonathan_wilke · 2026-11-14