Five Levels from Spicy Autocomplete to the Dark Software Factory
Dan Shapiro maps AI-assisted development onto a five-level framework (echoing NHTSA's driving automation levels). The progression runs from manual coding (Level 0) through task offloading, collaborative partnership, human-in-the-loop review, autonomous specification, to full black-box automation (Level 5). He places himself at Level 4 and says small teams under five people are already reaching Level 5 with "nearly unbelievable" results.
Key Quotes
No direct quotes extracted -- the fetch captured the framework but not the original voice.
Key Themes
The framework's real value is naming the Level 3 malaise: the point where the AI becomes the senior developer and the human becomes a code reviewer. Many find this unsatisfying -- you're not coding anymore, but you're not really designing either. You're just reading. This maps precisely to the "code janitor" fear in The Next Two Years of Software Engineering and the dehumanization concern in AI Zealotry.
The "illusion of completion" at Level 2 is a sharp observation. It explains why so many people feel productive with AI but don't see it in the output metrics.
Critical Analysis
The driving automation analogy is useful but imperfect. Software development isn't a single axis from manual to autonomous -- it's more like a constellation of activities (design, implementation, testing, review, deployment) each at different automation levels simultaneously. Shapiro's framework collapses that into one dimension, which is good for communication but misleading for planning.
The Level 5 "Dark Factory" framing is provocative. Comparing software development to Fanuc's unmanned manufacturing facility is a deliberate choice -- it strips out the craft narrative entirely. That's either honest or reductive depending on what kind of software we're talking about. CRUD apps? Sure. Novel distributed systems? Not yet.
The piece would be stronger with more specifics about what those Level 5 small teams are building. "Nearly unbelievable" without examples is just hype. See Building low-level software with only coding agents for the closest thing to an actual Level 4-5 case study, and Fresh Eyes for a tool that tries to solve the Level 3 review problem.
Sources: Five Levels from Spicy Autocomplete to the Dark Software Factory Last updated: 2026-05-14