## Core Definition The **2000-Hour AI Trust Curve** is Steve Yegge's observation that it takes approximately 2,000 hours of LLM usage to develop genuine trust in AI systems. This trust is not about believing the AI is perfect, but about developing the ability to **predict its behavior and failure modes**. **Source**: Steve Yegge interview on Vibe Coding (January 2025) ## The Trust Paradox Trust in LLMs is counterintuitive: | Common Assumption | Reality | |-------------------|---------| | Trust = AI is accurate | Trust = I know when AI will fail | | Trust grows with success | Trust grows with predictable failure patterns | | Trust = confidence | Trust = calibrated expectations | | Less trust = more checking | Right trust = knowing what to check | ## Why 2,000 Hours? The number represents roughly: - **1 year** of full-time AI-assisted development - **2-3 years** of part-time usage - Enough exposure to encounter diverse failure modes - Sufficient pattern recognition for intuitive prediction ## What Trust Looks Like **Before trust (< 500 hours)**: - Over-rely or under-rely on AI suggestions - Surprised by failures - Check everything or check nothing - Treat AI as magic or as useless **Developing trust (500-2000 hours)**: - Start recognizing failure patterns - Develop intuition for "this feels wrong" - Learn which tasks AI excels at vs. struggles with - Build mental model of AI behavior **After trust (> 2000 hours)**: - Predict failures before they occur - Know when to trust vs. verify - Calibrated confidence in AI outputs - Can explain AI behavior to others ## Practical Implications 1. **Don't expect immediate proficiency** - Early struggles are normal 2. **Failure is learning** - Each AI mistake teaches behavior patterns 3. **Volume matters** - Can't shortcut the exposure requirement 4. **Context-specific** - Trust in one AI tool doesn't transfer fully to others ## Connection to Mechanical Sympathy The 2000-hour mark is when users develop [[LLM Mechanical Sympathy]] - the intuitive understanding of how the "machine" behaves at its limits, similar to an F1 driver's relationship with their car. ## Cross-References - **Source Document**: [[Steve Yegge on Vibe Coding]] - **Related Concepts**: [[LLM Mechanical Sympathy]], [[AI Development Mastery Stack]] - **Contrast**: [[Vibe Coding vs Mastery]] - Yegge embraces vibe coding; Karpathy warns against it - **Application**: [[AI Productivity Paradox]] - Early productivity drops before trust develops --- *Atomic concept extracted January 2026 from Steve Yegge on Vibe Coding*