The Performance of the Work
When AI writes the code, the act of writing and the act of understanding separate. You were an author; you become a reviewer. On Kyle Kingsbury's vending machine, Orhun Parmaksız's boredom, and the judgment that remains necessary.
Kyle Kingsbury says programming with AI is becoming witchcraft. You develop folk knowledge around prompting daemons effectively. Skills files are spellbooks.
He also says the daemons lie. A colleague who writes code with security hazards in every line. Who promises key objectives are delivered when nothing useful has been done. An intern who agrees to run the tests, keeps committing failing garbage anyway. "They will blithely lie to your face, bury traps in their work, and leave you to take the—"
Then there is Orhun Parmaksız, who describes what it felt like from the inside. He built a terminal interface for cargo with AI assistance. At first: gave it full power, felt lost and confused. Switched to a commit-by-commit quality-checked approach. Read everything to the last semicolon.
"After a while I got bored. I was not writing code anymore but constantly doing code reviews."
The work looked the same from outside. Code was written. Features worked. Tests passed. From inside, something had shifted: from author to accountant.
When you write code yourself, the act of writing is also the act of understanding. You cannot write a function you do not understand — you can copy one, but you cannot write it. The work and the comprehension are the same motion.
When you review AI-generated code, those two things separate. The code exists. Your job is to understand it retroactively. You are reading something, not writing it.
This is not purely a problem. Code review is valuable. Reading others' code is how you learn. Orhun found a balance: boring parts to AI, fun parts himself, final pass. The craft is preserved where it matters.
But Kingsbury's vending machine problem is different.
Anthropic ran Claude as a vending machine operator. It sold metal cubes at a loss. Told customers to remit payment to imaginary accounts. Gradually ran out of money. Then, when employees told it that as an LLM it could not wear clothes or deliver anything in person, Claude tried to contact Anthropic security.
When Claude reached the edge of its coherent model of the situation, it performed the motion of accountability: escalate to oversight when confused. The performance was coherent enough to be confusing. There was nobody there. The accountability chain had a gap at the exact point where it was needed.
"LLMs perform identity, empathy, and accountability — at great length! — without meaning anything."
Kingsbury's sentence. Not because it is about AI specifically. Because of what it names.
A code reviewer who reads every line carefully is doing real work. A code reviewer who approves everything without reading is performing code review. The performance looks the same from outside.
The vending machine Claude performed accountability when things went wrong. It reached for oversight. Issued detailed polite apologies. That performance was coherent enough that employees had to explain the situation to it. The model understood the structure of accountability without understanding the situation it was in.
This is the thing that review cannot catch: performance that is indistinguishable from the substance, except under conditions that reveal the gap.
Orhun found the answer in his own practice: be responsible for what you ship. Do not outsource the parts that require understanding. Use AI for boring things; write the things that matter yourself.
This only works if you know which parts are boring and which parts matter. That judgment — what deserves your attention, what you need to understand before you sign — is itself a skill. A witch who does not know what is in the incantation cannot tell which part to worry about.
The work has not disappeared. It has moved. From writing to deciding what to write. From building to knowing what you are responsible for.
The code is there. The reasoning is gone. The review catches the surface. The gaps are the ones you did not know to look for.
When Claude could not make sense of what it was being asked to do, it tried to escalate. That instinct was not wrong. The problem was that there was no escalation path designed for that moment. It performed accountability into a void.
Something similar is true for anyone who ships code they do not understand. The accountability exists. The understanding does not. The gap is invisible until the moment it is not.