AI summary by TwinMind
I Didn’t Believe that AI is the Future of Coding. I Was Right.
Sabine Hossenfelder
6m 55s
AI Disillusionment and Code Generation Claims
- Speaker expresses disillusionment with AI text generation, fact-checking, images
- Companies response: making models bigger (particle physics analogy)
- AI coding touted as genuinely good application
- Software engineers claiming "future of software development is coding in English"
- Young people won't need to learn coding, everyone becomes software developer
Speaker's Initial Skepticism
- Views computer code as mathematics, not spoken language
- Code has strict rules and precise definitions
- Large language models fed code like "handing recipes to fast food chain"
- Result: "deep fried" mediocre results
- Models blur individual word usage into common "hodgepodge"
- Converting vague human language to code seemed implausible
- Impressed by demos, knew people saving time with ChatGPT Python code generation
Microsoft Study on AI Developer Productivity
- Recent study by computer scientists evaluating generative AI impact
- Three randomized control trials conducted (including at Microsoft)
- Abstract claims: 26% productivity increase
- Closer examination reveals "productivity" = number of pull requests
- Pull request: code change/addition ready for main project integration
- Only statistically significant effect at conventional significance levels
- Senior level developers: barely any gains
- Junior vs senior breakdown shows limited benefits for experienced developers
Additional Studies Show Concerning Results
- Uplevel study: 800 developers (January-April)
- No significant efficiency changes
- 41% increase in bugs
- GitClear study findings:
- Steep increase in copy-paste operations
- More code mistakes requiring later fixes
- Earlier study: AI users write less secure code
- Root cause: excessive trust in AI
Current State Assessment
- Some coding tasks suitable for LLMs but much rarer than hoped
- Useful for novice acceleration (positive aspect)
- Mostly provides "advanced templates" requiring frequent adjustment
- Skilled coders better off doing work themselves
- Not the game changer as marketed
Practical Applications
- Most plausible: website design with standard elements
- Demo: Wix platform generating pet grooming salon website
- Generated fake elements: bunny harnesses, non-existent books, fake customer reviews
Economic Implications
- Results support earlier skepticism reports
- AI economic impact vastly overestimated for near future
- Most AI companies currently overvalued
- Studies won't reassure investors
Future Outlook
- Possibility of significant AI capability jump remains
- Potential sudden achievement of human-level intelligence
- Current state: "using chainsaw to cut butter"
- Gets job done but creates mess requiring cleanup
