Coding Was Easy Mode: Why AI's Best Win Won't Generalize
Part 2 of the Innovation Plateau series.
The strongest evidence for “AI will automate everything” is what it did to software engineering. But SE is not a preview of every white-collar job — it’s a structural anomaly, for two reasons.
The Training Data Nobody Else Has
First, the training data. The open-source ecosystem handed models billions of lines of peer-reviewed, documented, functional logic — GitHub and StackOverflow are essentially a curated corpus of problems and verified solutions. And the workflow itself generates telemetry: every commit is an atomic record of how a problem got solved, every code review a structured critique, every Copilot rejection a labeled data point. Law, medicine, and strategy have nothing like this. Nobody version-controls a market entry strategy, and the “correct” answer usually lives in off-screen human relationships.
The Compiler Is the Real Moat
Second — and this is the one people miss — coding has a compiler. LLMs are probabilistic engines that confidently hallucinate. Software is the rare domain with a deterministic ground truth bolted into the workflow: the code compiles and passes tests, or it throws an error you can feed straight back into the model. That tight retry loop is the entire bridge between probabilistic AI and reliable output. “Vibe coding” (Karpathy’s term) works because of it.
Hallucinate a legal argument or misread a compliance statute, and no error log appears. The physical world, like most strategic work, does not have a compiler.
| Software Engineering | Law, Finance, Strategy | |
|---|---|---|
| Training data | Massive, public, documented | Siloed, proprietary, subjective |
| Feedback | Deterministic — compilers, tests, error logs | Human interpretation, markets, rulings |
| Telemetry | Git, CI/CD, pull requests | Email threads and offline meetings |
Why This Matters for the Plateau
Every extrapolation from “look what Copilot did to programming” to “the same is coming for lawyers next year” quietly assumes these two advantages travel. They don’t. The domains AI is being sold into next have subjective ground truth, siloed data, and no automatic error signal — which means adoption there will be slower, more expensive, and require the kind of institutional scaffolding (typed outputs, reference simulators, human review gates) that coding got for free.
That’s one of the three legs of the plateau argument. The second — what agents do to inference economics — is part 3.