AI Has No Wisdom, and the Second Half Is Aimed at You
A post on alexn.org argues that AI has no wisdom — and that the people leaning on it won't grow any either. The debate is really about which judgments you should never hand over.
The post at alexn.org is called “AI Has No Wisdom and Neither Will You.” It collected 383 points and 543 comments on Hacker News. That thread count says something about where the argument has moved. Almost nobody with a codebase and a deadline is still litigating whether models can write useful code. The open question is what happens to the person doing the asking.
The two halves carry different weight
The first claim is settled. Models have no wisdom. They have recall, pattern completion, and a very good sense of what a plausible answer looks like. What they lack is the ability to make a call when information is incomplete and the consequences land on them. That is closer to a definition than a criticism: judgment includes having something at stake, and there is nothing on the other side of the API with anything at stake.
The second half is the one that stings. It puts the same ruler against the user. Hand over the parts of the work that require judgment, and your own judgment does not grow in the meantime. That sounds like a complaint from someone’s former manager. It also has a mechanism behind it, which is why people keep arguing about it.
What gets outsourced
Some delegation is genuinely free. Looking up a function signature, writing a regex, grouping a pile of stack traces into something readable — these were always documentation lookups wearing a new interface. The time saved is real, and nothing is lost.
Other delegation leaves a mark. Should this abstraction exist? Should we ship this feature at all? Is the current module boundary worth defending, or is it time to cut it? None of those questions have answers in the corpus, because the answer depends on your team, your users, and how much runway you have. A model produces the most statistically common response to the phrasing of the question. That response is usually coherent, often well argued, and occasionally exactly wrong for your situation. Coherence is easy to mistake for correctness. It reads smoothly, so you skip the step where you check whether it applies to you.
Judgment is accumulated, not retrieved
Judgment comes from carrying outcomes. You ship the wrong design, get paged at 2 a.m., roll it back, and write the postmortem. That sequence is unpleasant in a way that makes it stick. Replace it with “ask, then do what it says,” and the short-term numbers look fine: velocity up, tickets closed. The cost shows up later and elsewhere, usually in a design review a year out, when nobody in the room can explain why the system is shaped the way it is or defend a boundary they never had to argue for.
This is not a case against the tools. It is a narrower claim about which parts of the job can be routed through them. Anything that can be verified cheaply — does the test pass, does the code compile, does the query return what you expected — is a reasonable candidate for delegation. Anything where verification is expensive or impossible, because the only real test is how the market reacts six months from now, is not. That line is not always obvious, but it is far more obvious than the current level of debate suggests.
A ruler with a definite answer
There is a way to check where you stand, and it does not require a study. Turn off the assistant for a day and count how many decisions you can still finish on your own. If the answer is most of them, the second half of that headline is about somebody else. If it isn’t, the headline is about you — and arguing with it is not the useful response.