Good at Coding ≠ Good At Everything
Spot the mistake:
Coding and math are hard
AI is good at coding and math
Therefore, AI is good at everything that's as hard as coding and math
Set aside the details in point 2 (how you get good results for coding and math out of an AI); the real fallacy is the leap in point 3. The recent successes of OpenAI's Codex and Anthropic's Claude Code owe less to better models (they've long been trained on just about every freely available codebase) and more to the bespoke, domain-specific work that has gone into these products. What makes Claude Code so good at coding isn't the underlying LLM; it's the harness around it.
Absent that harness, you've got an AI that still makes basic (and hilarious) mistakes, like the infamous Car Wash example. That's where step 3 breaks down. If you want an AI that's truly good at your particular white-collar job, it needs its own harness for that job. You can't assign each task a difficulty score and declare that the AI handles everything below the line.
The implication: I have serious doubts about the "job apocalypse", because there is no general-purpose hard. There are only specific jobs, and each one has its own unique challenges.
