Ghost in the Haskell
summary
A GHC plugin called Alama holes plugs a local Llama server to synthesize typed holes by sending the hole type, in-scope bindings, and first-sentence docs; suggestions are parsed and GHC typechecks them so only syntactically valid and type-correct fills are offered. Models tested range from ~1B (slow/poor) to 24B running on a 4070 GPU (2–10s latency) and improve weekly; harder examples succeed ~50% of the time. A tiny guide variable passes intent strings to steer generation.
AI-written summary · about
chapters (17)
- 0:00LLMs and Haskell intro
- 1:34Programming languages and LLMs
- 3:01LLM challenges for code
- 8:56Haskell ideal for LLMs
- 11:26Haskell code generation
- 15:16Typed hole plugins history
- 17:05Alama LLM plugin intro
- 20:16Context for LLMs
- 24:51Plugin performance
- 33:42Guiding LLMs with types
- 38:37Plugin demo
- 41:10Lens and traversal example
- 44:04Debugging LLM suggestions
- 47:13Cloud LLMs with plugin
- 50:27Local vs cloud models
- 50:48Why Haskell is uniquely positioned for LLMs
- 51:13Q&A
related talks
21mProduction Haskell Demystified
01h 26mHaskell Language Server
46mPainless software development with Haskell
52mProject Presentations - Haskell
26mghc-specter: a GHC plugin that inspects the GHC state on live
01h 07m50,000 Lines of LLM Haskell: What Broke, What Held, and Letting Go
47mIndustrial Strength Laziness: What's Next?
54mHaskell Project Showcase and Community Contributions