Using AI effectively in large Java codebases
summary
Shows concrete wins on two-million-line Java: a curated corpus of programming and incident triage tasks, automated evaluations that compare Opus vs Haiku and Claude, and a measurable jump from ~30% to ~90% correctness when a written skills-as-context layer is applied. Explains token cost tradeoffs (older models + skills can be cheaper), using an LLM-as-judge to run weekly regressions, and when to reserve humans for complex designs versus letting AI handle routine fixes.
AI-written summary · about
chapters (9)
- 0:00Introduction to AI in Java
- 1:08Using GitHub Copilot effectively
- 3:17Evaluating AI skill performance
- 6:12Automating skill assessment
- 8:44Cost savings with AI skills
- 10:11Handling model updates
- 11:36Weekly AI feedback loops
- 11:59AI for incident debugging
- 13:49Future of AI in development
related talks
49mClean Code in the AI era
52mAI for Busy Java Developers
26mAl and Software: The Good, The Bad, The Scary
50mPanel Software Engineering in the age of AI
45mThe AI Revolution in Software Development: From Code Writing to Quality Assurance
44mAI and Code Quality: Building a Synergy with Human Intelligence
24mHow AI is increasing the surface area for serendipity
39mBack to Basics: Cutting Through the AI-Generated Code Noise