Language Model Agents Enable Semi-Formal Programming
summary
Argues that large language models let programs live between bricklike code and unfired clay by enabling semi-formal programming, where some AST leaves are formal types and others are informal instructions. Demonstrations include llm_parser extracting structured job fields, AutoGPT generating and looping on subgoals (sometimes looping minutes), and the Figma Genius chat API creating in-situ UI actions. Describes building ephemeral interfaces that scrape sites and send actions back, treating LMAs as executors, compilers, and co-designers; prototypes mostly work but sometimes loop or misgeneralize.
AI-written summary · about
chapters (13)
- 0:00Semi-formal programming intro
- 2:30LLMs as compilers for informal tasks
- 5:57Direct execution vs structured LLM interaction
- 8:12Agents bridging formal and informal code
- 11:45Representing semi-formal programs
- 13:33Semi-formal languages and jargon
- 15:55Q&A: LLMs as compilers vs interpreters
- 17:42Exploring semi-formal programming
- 19:49Teal draw and software choices
- 21:15Medical diagnostics jargon
- 23:25Formalism vs informalism in UIs
- 24:49Audience views on LLMs
- 25:31Challenges and future of semi-formal programming
related talks
34mProgramming Languages for AI Programing Agents
36mProduction-Ready AI Agents: From LLMs to Small Language Models
01h 43mBuilding LLM Agents (Hackathon)
36mLLMs and AI agents demystified
51mTowards a Pragmatic Understanding of Large Language Models
53mSemantics-based program manipulation in the age of LLMs
57mArtificial Languages are Dead. Long Live Artificial Languages!
01h 00mFrom LLM Theory to Practical Agentic Implementations