Programmer’s Apprentice Season 2: Future Directions in AI-assisted Coding
summary
Claims that developers are the last generation to write code by hand and that general LLMs replace years of custom models are backed by concrete work: deployed systems for code search, code recommendations and automated bug fixing at Meta, a prototype IDE copilot called CodCompose, and a Nov 30 2022 inflection when ChatGPT outperforms bespoke models. Experiments show LLMs synthesize and fix code, generate unit tests, and run agents using a Starlark scripting host, but results are brittle, hard to debug, stochastic and pose security and cost problems.
AI-written summary · about
chapters (12)
- 0:00Intro
- 5:23Non-monotonic logic
- 7:57Virtuous cycle
- 8:48Vicious cycle
- 11:44AI in software engineering at Facebook
- 12:44Code searching using natural language
- 14:58Code recommendations
- 15:59Automated bug fixing
- 20:48Let's automate ourselves away & have fun doing it
- 35:26LLM as scripting client
- 41:49LLM-based software is very powerful, but...
- 45:19Outro
related talks
01h 28mFuture of Software Programming with AI Bots at Hand
26mThe present and future of programming with AI (Keynote)
33mAI is coming for your job
43mFrom Lambda to LLMs: AI Coding and the Future of Programming
22mAdapting to the Future of Work in Tech
27mWill AI (Finally) Replace Developers?
01h 01mWill AI Replace My Job? A Town Hall for Software Developers
32mComputer Programming Is Dead; Long Live AI-First Programming