Gothenburg 2026 - Day 3 - RB and AI

Observed Benefits

• Extracts builds structures when vague
  • (Build steps, tool versions, metadata)
• Clone the developer for multiple defined solutions
  • Fuzzing adjacent (testing multiple spaces at once)
  • Developer defined rules (clear success criteria)
• Extracting relations from large codebases
• Translate (documentation) for better communication
• Rubber duck
• Accessibility (less typing: RSI)

Pitfalls

• Copyright / Licensing
• Information Overload / Quantity Expansion (Diff Explosion)
• Community perception
• Lazyness & False Confidence
• Lack of Etiquet
  • New users are less aware of it
  • Megaphone of bad
• Cheating (Not optimized solve your problem, but complete your check)
• Mostly Proprietary models (even in open weight models)
  • Can we get to custom trained models for projects?
• Specular language
• Harder to find “good” (finished) projects (lots more window dressing)
• Generational education pipeline

Open Source ??

• How many TBs of “source” material
• “Open Weights” is NOT open source. It’s binary, built output
• Train cost: Rebuild GCC - Minutes, Retrain model - Weeks and $$$

Other

• Speed
  • Task dependent (Good and bad depending on the situation)
• Human Element
  • Help avoid pitfalls and acheive benefits
  • More involvemnet is better (contrary to marketing)
• Skill customization over general “reproducibility” skill
• Build policies for acceptable practices
  • Instead of knee jerk reaction to outright ban