Grounded Copilot: How Programmers Interact with Code-Generating Models
summary
Study watches 20 programmers across four tasks and languages and builds a grounded-theory claim that interactions with Copilot are bimodal: acceleration mode where programmers have a plan and accept short, low-edit completions, and exploration mode where uncertain programmers use long, multi-line suggestions as jumping-off points. Experts explore less; novice or new Copilot users explore more. Tasks include unfamiliar APIs (custom crypto, networking, chat server) and a programming contest. Recommendation: models detect mode, surface intermediate values, integrate LSPs, and help distinguish multiple suggestions.
AI-written summary · about
chapters (13)
- 0:00Welcome to Grounded Copilot
- 0:21Code-generating models overview
- 1:23Study method and tasks
- 2:05Acceleration vs exploration theory
- 4:13Theoretical findings
- 4:55Acceleration mode examples
- 6:00Exploration mode examples
- 7:03Acceleration vs exploration
- 11:43Quantitative analysis and behavior
- 12:26Future tools recommendations
- 14:31Validation challenges
- 16:16Acknowledgments
- 16:42Q&A
related talks
01h 12mAI-Assisted Programming Today and Tomorrow
45mSupercharge your coding: Integrating ChatGPT and Copilot in long-term projects
31mGithub Co-Pilot: Addictive or Effective?
05mco_pilot
30mFrom COBOL to Copilots: Why “No-Code” Never Kills the Coder
25mImproving developer happiness with AI
01h 00mLevel up with GitHub Copilot: using AI to learn, code, and build
40mLevel Up with Copilot