AI or Not AI? We Don't (Yet) Believe the AI That Replaces Devs!
summary
A year-long set of experiments at SopraSteria evaluated Copilot/ChatGPT-style tools on real development work. Copilot autocompletion speeds trivial coding but often generates buggy snippets needing substantial fixes; generated unit tests help explore and onboard legacy code but don't replace test-driven development; attempts to auto-migrate COBOL to Java failed without human-led migration. Using retrieval-augmented generation (RAG) over project documentation improved query accuracy; overall the tools speed simple tasks, increase code duplication and some bugs, require human review, and a two-hour training session reduced unrealistic expectations.
AI-written summary · about
chapters (7)
- 0:00Introduction
- 7:20LLM theory and limits
- 12:32Autocompletion and chat IA
- 17:32Legacy code with IA
- 25:10IA as creative tool
- 33:35Craft and Green IT impact
- 39:54Conclusion
related talks
01h 16mAI in Software Development: Boosting Efficiency
46mAI Integration for Devs: ChatGPT and Copilot in Your Projects
33mAI assisted software delivery
57mOne Team’s Journey to Using AI Without the Hype
48mWill AI change our developer profession?
01h 02mPANEL - Unleashing the Future: The Impact of Gen AI on Development Productivity
59mAI is Eating the Developer Career! Wait, what? Maybe it's the other way around?
20mWill AI Replace Coding or Improve It?