Rich Hickey draws a sharp distinction between something you wield and something you ask to do the work for you. His target is the reassuring phrase “just a tool”: a language model can produce useful results while changing what you practice, control, and learn.

The argument turns on three costs:

  • A tool gives continuous control and feedback. Requesting a result from a model, Hickey argues, substitutes a different relationship.
  • Learning a model’s changing quirks is not necessarily learning the underlying craft. A software update can invalidate those techniques.
  • Replacing a paper with its summary skips more than information: argument structure, author credibility, unexpected details, and material for future ideas.

His drive-thru analogy makes the distinction plain: buying dinner is not using a tool to learn cooking. The useful question is not merely whether AI gets the job done, but which abilities the chosen workflow stops exercising. That concern is stronger than the essay’s sweeping claim that LLM use lacks feedback altogether.

The 112-comment thread on Hacker News complicates the categorical framing without dismissing the learning cost.

What the thread adds

  • didibus — reports a narrower clarification from Hickey in a comment elsewhere: “I’m talking about its embodiment in conversational systems purporting to be ’tools’ for doing homework, research, coding, writing etc.” This is a commenter’s report, not wording in the essay itself.
  • acjohnson55 — offers a counterexample from their own coding-agent practice: “There’s a huge amount of direction and feedback.” They describe a range from unsupervised work to structured processes.
  • munificent — separates output from apprenticeship: “I think a camera is an excellent tool for producing photographs. It is a shitty tool for becoming a better painter.”
  • rcxdude — challenges equating software work with writing code: “a useful piece of software is much more than just the effort of writing out the code”.
  • drivebyhooting — counters the assumption that delegation always displaces learning someone would otherwise undertake: “I was never going to learn webGL, AI or no.”

Handles are pseudonymous. HN publishes no per-comment scores; ordering is HN’s ranking. This is a slice of the thread, not a consensus.