50,000 Lines of LLM Haskell: What Broke, What Held, and Letting Go
summary
Claims that 50,000 lines by Claude produced a mostly working Haskell codebase, including a smartwatch app running Beam and Serpent, and several libraries: a conversion library called Witch and Post bindings for a DB. Workflow uses containerized agents with heavy testing, emulator-based CI, Nyx for build grinding, Playwright screenshots to give models UI sense, and a dedicated GitHub bot (containerized agents with dedicated GitHub bot Yepy Sloth). Results speed prototyping but require strict test/PR review and distrust of hallucinations.
AI-written summary · about
chapters (24)
- 0:00Introduction
- 1:08Why this talk
- 2:52Why LLMs lie
- 8:15Haskell and strong typing
- 10:44Testing strategies
- 12:52AI agents flaws
- 13:36Security constraints
- 15:45Trust model
- 18:13AI testing benefits
- 23:13AI cheating decisions
- 26:04Prompting techniques
- 28:52Smaller models for code
- 31:42AI weirdness
- 34:51Letting go of code
- 35:54AI energy use
- 37:18Job security
- 38:41Open questions
- 40:48Marketing verification
- 42:11AI energy impact
- 43:54Summary
- 44:17Intrinsic motivation
- 55:14Distrusting AI
- 1:03:25Concerns about intellectual property and workers
- 1:06:00Q&A
related talks
54mHaskell Project Showcase and Community Contributions
01h 26mReflections on Haskell@Meta
46mPainless software development with Haskell
29mHow to roll your own LLM
52mProject Presentations - Haskell
22mHaskell Tools, Compilers, Libraries and Community Projects
01h 11mHaskell in Production: Adventures in Blockchain Building
49mCoding Tomorrow