Keynote: The dangers of probably-working software
summary
Huffman compression code copied from Wikipedia fails in production because padding zeros collide with prefix codes and TIFFs already-compressed images explode storage; an early GitHub Copilot user notes AI can produce “probably-working” code but black-boxes behavior; a .NET JIT bug in Sept 2015 makes optimized builds return different results; a tiny Express hello-world pulls 123 dependencies for Express hello world (65 unique), Node is ~10 million LOC, agents can brick machines by auto-running updates, and tools like agents.md and spec kit plus CI/tests and David Fowler’s bug-repro agent are what actually make AI-written code trustworthy.
AI-written summary · about
chapters (20)
- 0:00Welcome to NDC London 2026
- 1:50Quiz on untested code
- 3:13Huffman compression failure
- 8:10AI tools and probably-working code
- 10:58Limits of trusting unknown code
- 13:28Understanding .NET compiler quirks
- 15:40Hidden complexity in tools
- 17:07Dependency chains in projects
- 19:35Trusting software in critical systems
- 22:47AI code and risk assessment
- 27:23Balancing AI scope and quality
- 33:01AI improving but risks remain
- 37:13Faster coding vs better code
- 40:23Guardrails for AI development
- 41:07Testing failures on special days
- 44:18AI in software lifecycle
- 46:47Specification-driven AI development
- 47:52Documenting standards for AI agents
- 50:43AI tools for bug reproduction
- 52:51Summary and call to responsible AI use
related talks
36mKeynote: The Future of Software 
39mBack to Basics: Cutting Through the AI-Generated Code Noise
56mFundamentals of Software Engineering In the age of AI
32mEscape the Try-Again Loop: Verifying AI-Generated Code
32mBeyond the Hype: A Pragmatic Look at How Developers Actually Use AI Tools
15mKeynote: The Common Pathway of AI and Software
48mSacrificing the golden calf of “coding”
46mThe AI-Powered Developer