Skip to content
outpace.
Video · Making AI useful

The 5 Levels of AI Coding (Why Most of You Won't Make It Past Level 2)

Nate B Jones on the five levels of AI coding, and why most teams sit at level two while believing they are further along.

By Nate B Jones · 42:15

Nate B Jones pointing at a stack of coloured bars numbered level 0 to level 5, with the level 2 bar labelled "you are here"Watch on YouTube ↗ (opens in a new tab)

Nate B Jones walks through Dan Shapiro's five levels of AI coding, from autocomplete up to what Shapiro calls the dark factory, where a specification goes in and working software comes out with nobody reading the code. His example is StrongDM, three people running that way since 2024, whose agents are checked against behaviour scenarios kept outside the codebase so they can't be gamed. Against that he sets a 2025 trial where experienced developers using AI took 19% longer while believing they were 24% faster.

A few weeks ago my team needed five campaign messages. Five. AI gave us context, structure and a genuinely good flowchart. What it didn't give us was five messages we could send. The work hadn't gone away. It had moved, to deciding what matters to the person reading, what each message is for, and what we want them to do next.

That's the gap in this video, in a different setting. The 19% isn't a story about slow tools. It's the producing getting cheap while the deciding gets harder, and nobody naming the new work or learning to do it.

What got us out of it wasn't more output. We went back to the reader's situation, gave every message one job, and made sure the thing we promised could actually be opened. That became a campaign system we can use again, not a template repeated five times.

Take one job your team now does with AI. Ask the people doing it what judgement matters now the producing is cheap, then practise that part together on real work.

Highlights

  • Most teams who call themselves AI native are at level two: the AI writes across files, and a person still reads every line.
  • In the trial he cites, developers using AI took 19% longer and thought they were 24% faster. Measure the work rather than asking people how it felt.
  • StrongDM keeps its behaviour checks outside the codebase, where the agent can't read them, so it can't write code that only passes the test.

From the transcript

The distance between them isn't a technology gap. It's a people gap. It's a culture gap. It's an organizational gap. It's a willingness to change gap that no tool and no vendor can close.

TopicsAI adoptionWork design

Continue reading

All articles →
Illustration of a woman seen from behind, turning her head, as white wave lines rise across her shoulders against a deep blue sky.

What AI Can't Fake

Generative AI hasn't killed the expert. It has changed what expertise means: how fast you learn now matters more than what you already know.

Chris Rosato ·
Illustration of a woman in profile resting her chin on her hand, against a yellow background of pipes and line drawings.

Beyond Transactions

Faster answers are the smallest thing AI offers. The bigger opportunity is using it to think, create and lead better.

Chris Rosato ·