Unexpected AI Use Cases

• Writing wedding speeches using AI
• Setting answering machines to AI-generated voices
• Using AI for daily coding assistance
  • Makes coding "infinitely faster"
  • Can describe UI changes (e.g., "build dark mode") and AI implements full feature
  • Humans becoming "narrators" describing desired outcomes
  • AI creates better results than humans doing it manually

AI's Counter-Intuitive Strengths

• Excels at creative/storytelling work (opposite of initial expectations)
• Creative applications:
  • Goal: enable anyone to "make South Park from their bedroom"
  • Generate photos (example: Eric and Rihanna playing volleyball on beach)
• Human-like AI voices:
  • Previously "sounded horrible"
  • Now "indistinguishable from human voice"
• Semantic search capabilities:
  • Previously didn't work well
  • Now excellent at finding relevant text from random pieces of data
  • Strong reading comprehension and question-answering abilities

Industry-Specific AI Implementation

• Fashion industry challenges:
  • New terms constantly emerging
  • Need continuous model fine-tuning
  • Example trends: "mermaid court" one month, "ballet court" next
  • Models require monthly updates for trend accuracy

AI Performance and Limitations

• Current capability: 85-90% solution accuracy
• Requires significant fine-tuning and additional "hacks" for genuine value
• Best performance with:
  • Structured tasks
  • Single, specific objectives
  • Clear process engineering
  • Iterative prompt debugging and tuning

Technical Challenges

• Marrying deterministic software with probabilistic models
• Variability in outputs with same inputs (vs. traditional programming)
• Benefits of randomness: better space exploration, improved model learning
• Reliability varies by use case:
  • Entertainment: variability acceptable if "funny"
  • Critical applications (cars): more problematic

Hallucination Problem

• Definition: AI generating non-existent content that appears plausible
• Poor fact vs. fiction distinction
• High-stakes verification challenges (medical diagnosis example)
• Industry anti-hallucination efforts created opposite problem:
  • AI now often claims ignorance of known information
  • Won't acknowledge articles definitely in training set
• Human-like limitation: can't provide consistent citations
• Trust and accuracy metrics insufficient alone

Human-AI Collaboration Requirements

• "Human in the loop" essential
• Humans needed for:
  • Initial assessment of AI outputs
  • Supervising corrections
  • Preventing hallucinations
• Technology should serve humans and deepen human connections
• All YC AI companies facing similar steering/control challenges