The 5 Levels of AI Coding (Why Most of You Won’t Make It Past Level 2)
Source: YouTube — Nate B Jones, published 2026-02-18 Link: https://www.youtube.com/watch?v=bDcgHzCBgmQ
Summary
The most important source in this wiki for understanding where AI-assisted software development actually stands and where it’s going. Uses Dan Shapiro’s five-level framework to map the gap between frontier teams (StrongDM’s 3-person dark factory) and the industry average (stuck at L2, measurably slower with AI). The distance is not a technology gap — it’s a people, culture, and organizational gap that no tool can close.
See Five Levels of AI Coding for the full concept page with all levels, examples, and data.
Key Data Points
- METR RCT: Experienced devs 19% slower with AI tools; believed they were 24% faster
- StrongDM: 3 engineers, zero human code, Attractor agent, external scenarios, digital twin universe
- Claude Code: 90% self-authored; Boris Cherny hasn’t written code in months; 4% of GitHub public commits
- Junior pipeline: US postings down 67%, UK grad roles down 46%
- AI-native economics: ~600K SaaS average
- J-curve: Most orgs stuck at the bottom, interpreting the dip as evidence AI doesn’t work
Pages Created or Updated
- Five Levels of AI Coding — new
- Attractor — new
- Nate B Jones — updated
See Also
- Four Prompting Disciplines — spec engineering is the L4/L5 skill
- Frontier Operations — the human skills that persist across levels
- AutoResearch and Evals — scenarios are evals