Adversarial Communication
19 points by wrl
19 points by wrl
The verification often ends up being just as expensive as doing the work in the first place
There are many valid, interesting use cases where this is not true, e.g. writing a script that aggregates and visualizes certain data (I can read pandas or SQL better than I can write it, or adding a small feature to a web app frontend if I'm not super familiar with the code base (but I can at least judge whether the code and the result are sensible), or figuring out the root cause of a bug in a legacy part of the code base (I don't have to become familiar with the code, but I have to fix this bug today).
Productive use comes down to building good heuristics for which case is which.
yeah, that was my first reaction too. agents are useful when you don't spam yourself (or worse, someone else) with high-review-burden work; this is common when coding. if your task is like "add three fields to these screens, following existing patterns" then it really doesn't take any more testing work than if you'd written the code out by hand.
@glyph Re: 'Spamming “For Good”?' perhaps how credit reporting agencies are allowed to ignore form letters is relevant prior art?
Thought-provoking article. I particularly enjoyed the observation in the first footnote about everything instantly being a brown-field project now.
There is a bitter irony that the ability to understand the inherent value of actually writing the essay on their own is the sort of thing that students can really only learn by writing a bunch of essays.
Forsooth!