Where AI Meets Code
summary
Introduces a simple surfacing model for how LLMs pull concept clusters and shows the lost‑in‑the‑middle failure via a needle‑in‑haystack token test (primacy/recency bias). Demonstrates practical tactics: projecting Java classes into math, PlantUML, or other languages to reveal behavior; using generated JUnit tests to characterize running code and even deleting code then regenerating from tests as an experiment that exposes limits. Distinguishes low‑risk scaffolding and tools from high‑risk medical/financial use, and invents waywords for reusable prompt abstractions.
AI-written summary · about
chapters (18)
- 0:00Intro
- 2:26The surfacing model
- 5:01Context window
- 8:28Spectrum of use
- 10:58Techniques
- 14:33Lensing
- 16:44Generalization/Specialization in tooling
- 18:06Do "side by sides"
- 21:05Create pidgin languages
- 23:03Make waywords
- 29:53More techniques
- 31:18Cost/Risk
- 34:02Review fatigue
- 35:46Work in small chunks
- 36:50Automated testing
- 39:07Rigor
- 43:46Our trajectory
- 44:44Outro
related talks
15mUsing AI effectively in large Java codebases
26mAl and Software: The Good, The Bad, The Scary
21mLLM-Powered Unit Testing: Writing Reliable Java Tests with AI Assistance
58mAI Assistant: Don’t Code Alone
39mBack to Basics: Cutting Through the AI-Generated Code Noise
21mAI's impact on software architecture
59mIt AI n't What You Think!
30mUsing AI to Create Better Requirements and Designs