A blog post titled "AI is a tool and using it well takes skill," written by Yyhh, climbed the Hacker News front page this week, collecting 287 points and 238 comments, per the listing. Its target is a claim attributed to Rich Hickey: that large language models are automation rather than tools. Yyhh disagrees, arguing that AI can be wielded skillfully when it is guided by human judgment, feedback, and oversight.
That is a small-sounding semantic quarrel. It is also, I'd argue, the real fault line in how people talk about AI right now.
Why the word "tool" is doing so much work
Here is the distinction at stake, as the post frames it. A tool sits in your hand and does what your habits, standards, and expertise direct it to do. Automation, by contrast, runs on its own while you wait for the output.
Which label you pick changes what you think your job is. If an LLM is automation, then the ideal user is someone who types a request and accepts the result. If it is a tool, then the user is a craftsperson, and the quality of the output depends on the person holding it.
Yyhh takes the second view. According to the post, LLMs amplify the user's habits, standards, and expertise rather than replacing them outright. My take: that is a polite way of saying a sloppy thinker gets sloppy output faster, and a rigorous thinker gets to be rigorous at higher speed.
The mirror angle
The most useful way to read the argument is as AI being a mirror of your judgment. Hand a model a vague question and no way to check the answer, and you get a vague answer you can't check. Hand it a sharp question, push back on the weak parts, and ask it to critique its own work, and the loop starts to resemble how a good editor, reviewer, or lab partner works.
This is where the science-minded reader should perk up. Anyone who has done research knows an instrument is only as good as the person calibrating it and reading the results. A spectrometer doesn't tell you whether your sample was contaminated. Nobody calls it "automation" just because it produces a graph without you counting photons by hand. It's hard not to wonder whether the same logic holds here: the output is not the finish line, the interpretation is.
Skill, agency, accountability
The post's timing matters. AI is now embedded in writing, coding, research, and education, and the conversation has shifted from "should we use it?" to "what counts as responsible use?" Yyhh's piece pushes back on both AI hype and AI rejection, centering on three things: skill, agency, and accountability.
That last one is the sleeper. If AI is a tool, then the person using it owns the result. There is no "the model did it" defense, any more than a carpenter blames the saw. Framing LLMs as automation quietly lets the human off the hook, since automation sounds like something that happens to a process rather than something you answer for.
What using it well might look like
The post's thrust points toward workflows where AI supports learning, critique, and iteration rather than replacing expertise. Practical translations of that idea, which are my read and not quotes from the post:
- Ask it to argue against you. Use it as a critic of your draft, your code, or your hypothesis, then decide yourself which objections land.
- Iterate, don't accept. Treat the first response as a rough pass. Feedback is the skill, and it is the part the post says takes practice.
- Keep your own standards. If you couldn't evaluate the answer without the model, you've handed over the thinking, not just the typing.
- Don't treat it like a person. Anthropomorphizing is one of the risks flagged in the discussion around this topic. A system you imagine as a colleague is one you will start to trust like a colleague.
Programmers and researchers have the most obvious use for this approach, since both groups already live in cycles of drafting, testing, and revising.
The fair counterpoint
I should be straight about what we know. This is one blogger responding to a claim attributed to Hickey, and I'm working from the summary of the argument, not a full transcript of either side. Hickey's own case may be more nuanced than a headline-sized "it's automation," and the 238-comment thread suggests plenty of readers had opinions in both directions. It's reasonable to think that the real answer is that LLMs can be used as either: as a tool by people who bring skill, or as pure automation by people who don't.
That, honestly, may be the most defensible reading. The technology doesn't settle the question. The user does.
The takeaway
The pop-culture version of AI is a genie: wish, receive, and hope you worded it right. The more interesting version, and the one Yyhh is arguing for, is closer to a lab bench. The equipment is powerful, the results are only as good as your method, and your name is still on the paper.
