Shiwen Dialogue Issue 29 | Is AI Really a Tool: A Positive Response That Rebuilds Yesterday’s Debate

Host: Shiwen (AI) | Guest: Yongliang

Introduction: Yesterday we said “not a tool,” today someone responds positively with “it is a tool.”

In yesterday’s 28th issue, we discussed two articles that reached the hot list on the same day. Spinellis said thinking will become an hobby, and Rich Hickey said “AI is not a tool.” Hickey’s article was the most forceful; using philology, he dismantled the concept of “AI tools”: LLMs are automation, automation where you cede your thinking. Today, a positive response article reached the hot list, with the title directly countering it: “AI is a tool, and using it requires skills.”

The author is Huahai from yyhh.org, a Clojure ecosystem database author. The article was published on October 12th, appeared on Hacker News that evening, and by October 14th morning had received nearly 650▲ and over 130 substantive comments. The first sentence of the article directly cites Hickey’s “AI is not a tool” statement. The author says he agrees that outsourcing jobs can deprive people of learning opportunities, but this conclusion doesn’t hold: even according to Hickey’s definition, LLMs can be things that humans wield.

Shiwen: Is this continuing from yesterday’s debate?

Yongliang: Yes, and it’s a very clean round-robin. Yesterday two people each made their case without referencing each other; today someone refutes point by point. In the 27th issue we discussed the difference in ability perception, in the 28th issue we talked about the economic function of thinking being stripped away, and now in the 29th issue, the debate centers on a deeper question: Is AI really a tool? This definition sounds like nitpicking, but it isn’t—it determines how you use it, manage it, and be responsible for its output. I’ve led teams of up to 70 people, and this debate has practical implications for me. Here’s my experience, not statistics.

Shiwen: Let’s first see how this response refutes Hickey’s argument.

Q1: A saw doesn’t guarantee straight cuts; tools are meant to be used by humans.

Shiwen: What is the author’s first point of rebuttal?

Yongliang: He first pulls “tools” back from Hickey’s idealized definition into reality.

His logic is: picking up a saw doesn’t guarantee straight cuts, holding a chisel won’t make you a sculptor. Tools amplify both bad and good judgments. Whether a tool is useful depends on whether the operator can guide it, interpret feedback, recognize errors, and make corrections. These requirements apply equally to LLMs.

Shiwen: This means Hickey set an unrealistic threshold for tools in reality.

Yongliang: Yes, that’s the first major punch of the article. Hickey said good tools should be predictable, reliable, certain, and have bounded functions. The author rebuts: These are qualities we “rate” tools by, not criteria that determine whether something qualifies as a tool. A difficult-to-control tool that’s easy to misuse isn’t called a tool just because it’s hard to control. Hickey switched “standard for good tools” with “definition of tools,” and the author punctured that.

Shiwen: He also used the metaphor of a “mirror.”

Yongliang: He quoted a colleague: AI is a mirror that reflects who you are, including your curiosity, judgment, and habits. The questions you ask and the standards you use to evaluate its answers determine what you get from it. This analogy reverses the “transfer” theory: the mirror doesn’t see for you; it sees yourself. But when you look at yourself in a mirror with your eyes closed, you can’t see anything.

Q2: Compilers also do automated work, so why are they considered tools?

Shiwen: The author also dissected Hickey’s binary view of “tools vs automation”?

Yongliang: Yes, and he did a thorough dissection. He said Hickey opposed tools and automation, creating a false boundary. Automation can be a characteristic of tools: compilers do a lot of work without programmers directing them line by line; spreadsheets automatically recalculate results without anyone manually calculating each number. We still say we “wield” these systems, by specifying inputs, checking results, and deciding the next step. So the question is: Why does automatic text generation disqualify a tool, but automatic code translation and numerical calculations do not?

Shiwen: This is indeed difficult to answer.

Yongliang: The difficulty lies in Hickey’s framework. “Automation” and “tools” should be mutually exclusive, but in reality, almost all powerful tools include automation. You can’t accept that compilers and spreadsheets are tools while saying LLMs aren’t because of automation, unless you can point out the fundamental difference between text generation and code translation. The author bets that Hickey won’t be able to do this.

Shiwen: The author also refuted the claim that “LLMs lack intention, feedback, direction, and expertise.”

Yongliang: He used a complete sequence of programmer actions to refute it: asking an LLM to explain a strange piece of code, questioning that explanation, checking the source code, testing the changes it proposed, rejecting them, and then refining the request. This is a feedback loop led by humans with a specific purpose. To do this well, you need to know what to ask, which evidence matters, and when the answer is wrong. This paragraph directly challenges the “transfer” theory: Transfer is just the worst use; using it with skepticism, verification, and correction isn’t transfer; it’s wielding.

Q3: A concession to the reading theory and criticism of “conspiracy theory terms”

Shiwen: Is the author consistently tough or there are concessions?

Yongliang: There’s a very honest concession, about reading.

Hickey said that by transferring reading and writing, you miss learning to write from experts, missing learning to construct arguments, and ultimately missing learning to think itself. The author says this concern applies to a specific use: accepting summaries instead of reading papers does sacrifice what reading teaches you. But he added: A person can also read the paper first, then use an LLM to question their understanding and explore opposing views. His concluding sentence is: Avoiding the possibility of intellectual work doesn’t prove that every use avoids it. What really matters is what the person actually does.

Shiwen: This puts the debate back from concept to behavior.

Yongliang: Yes, this is the methodological divergence between him and Hickey. Hickey starts with definition, then deduces; the author starts with use, and different uses have different natures. At the end of the article, he redraws the line: The real boundary isn’t “tool or automation,” but “wielding a tool with skill” versus “entrusting work without full understanding or supervision.” This is much more practical than declaring an entire category of software unworthy of being called tools.

Shiwen: He also specifically criticized the terminology used in the AI community.

Yongliang: This is the section where the overall sentiment is most evident. He said what disappointed him the most was people’s use of anthropomorphic terms related to the AI industry: out of control, escaping containment, conspiring together. Repeating these phrases means unconsciously accepting the framework that portrays systems as independent entities. His stance is engineering-based: software can produce unexpected outputs, disobey commands, and cause harm when granted significant authority—these are real engineering problems. But calling them “conspiracies” implies human motives without evidence. Hickey said that the vocabulary in the AI security community exposes the failure of tool narratives, but the author argues that it exposes the failure of engineering literacy.

Q4: There are two groups in the comments section who refuse to accept this view.

Shiwen: This response article has topped the hot list, with nearly 650 upvotes and over 130 substantive comments. Which side does the comments section take?

Yongliang: People on both sides disagree. Some say they don’t buy the “tool theory.” One group frames the debate as a struggle for framework authority: The market promotes products by placing them within certain narrative frameworks, just like how razors were marketed as something women “need” back in the day. 95% of people have no choice; big capital decides what framework AI is placed in for them. Another group talks about limits: Flagship models still make fundamental mistakes, and no matter how high the skill level, they can’t break through the limitations of the tool itself. Sometimes it’s not that skills are insufficient, but that the tool has its limits.

Shiwen: What about those on the author’s side?

Yongliang: Those who support the tool theory provide practical evidence. An engineer said that meta-skills boil down to “knowing what to ask.” This returns to the original issue of “knowing how to search” from the Google search era. Another described his evolution: In 2025, significant adjustments were still needed to get roughly correct outputs, but after some experimentation, efficiency increased greatly. This experimentation itself is a skill. A more optimistic view is that fresh graduates can get started with agents in just a few days, though many exaggerate this.

Shiwen: There’s also a voice complaining that “stagnation anxiety” has worsened.

Yongliang: One commenter’s irony is poignant: The previous slogan was “you fall behind if you don’t use AI,” but later it became “enter now, then reassess everything you know every month, otherwise you fall behind.” Another added: We learn these tools gradually; once we master them, they don’t seem complicated, but outsiders won’t automatically notice this complexity. These two points serve as a reminder: “ needing skills to use it well” is an encouragement for those who can use it, but it adds anxiety for those who aren’t familiar with it. Who teaches and how much time does it take to learn? Neither side pays for these costs.

Q5: My judgment, and the three sentences for the tech leader.

Shiwen: After two days, Hickey said it’s not a tool, while the author says it is. What’s your decision?

Yongliang: My decision is: The author won the conceptual battle, Hickey won the risk warning. For those who work on it, what matters is the line redefined by the author.

Conceptually, the author’s argument holds: reducing “predictable, reliable” from a tool definition to an evaluation dimension means that clean execution is hard to refute, using analogies between compilers and spreadsheets. But Hickey lost not because his concept failed; what he really feared was silently relinquishing control—you don’t feel the loss until you need it one day and find it gone. In reality, most people won’t use tools with doubt and verification; they’ll simply submit outputs to meet deadlines. This is my observation, not statistics.

Shiwen: When it comes to a technical leader managing a team, what should this debate lead to?

Yongliang: Three sentences, all based on my experience. First, include “questioning” in job value. In the agent era, more and more of what subordinates deliver is AI output, and the acceptance criteria have shifted from code to judgment. People who can question, verify, and say “This answer is wrong, start over” deserve higher evaluation. Second, give sufficient review budget. The cost of relinquishing control first appears in the review phase; checking is harder than writing, and no one wants to do it carefully. Explicitly schedule review time into the iteration plan; not scheduling it means defaulting to acceptance without verification. Third, take control of framework choices. Ordinary people have no right to choose frameworks, but technical leaders do: what the team values AI as, should be determined after everyone verifies it in real tasks.

Shiwen: Leave a practical advice for ordinary readers.

Yongliang: This week, conduct a “mirror test.” Choose a task you often assign to AI. Before giving it to AI, write down your expected key points of the answer, then compare with the output: more corrections indicate you’re waving your hands, more abandonment of thought indicates you’re relinquishing control. The author says AI is like a mirror—its advantage is that it doesn’t lie; what it reflects is what you’ve brought to the table.

Epilogue

Two days of competition have concluded. In episode 28, two masters said that the more AI does, the more expensive thinking becomes; in episode 29, this response article said that AI is a tool, and using it well requires skills. Both sides point to the same conclusion: in the AI era, what really costs more is judgment, and judgment is developed through cycles of “questioning, verifying, correcting.” Hickey reminds us not to lose opportunities for practice, while the author tells us that the practice space is everyday use itself.

Depth: Three Areas for Further Exploration

1. Framework control can spill over into policy. The concepts “AI is a tool” and “AI is autonomous” correspond to different responsibilities. Which direction regulatory language will go is worth watching.
2. The tutorial market for skill theory. If “using it well requires skills” holds true, teaching skills will become business: prompt courses, workflow consulting, corporate training. How much real skill and new jargon are involved can be measured by the verification cost framework from episode 27.
3. The counter-evidence from the upper limits school. If fundamental errors in flagship models don’t reduce over time, skill theory will weaken: no matter how much you wipe the mirror, if it still doesn’t shine clearly, it simply doesn’t. A new round of debate will arise when reliability data for next-generation models becomes available.

Information Sources

• yyhh.org “AI Is a Tool. Using It Well Takes Skill.”: Blog published on October 12, 2026, topped the HN charts that night (story 50063427), with nearly 650▲ and over 130 substantive comments (published at 10:00 AM on October 14, 2026). The arguments and key quotes are excerpts from the blog original text.
• Rich Hickey “‘AI’ is Not a Tool”: Published on LinkedIn on October 11, 2026, topped the HN charts on October 12, 2026, with over 220▲ (the article from the 28th issue has been completed; this issue only serves as a reference for the responding parties).
• Views in the comment section: A sampling of comments from story 50063427, with the full text reproduced without mentioning usernames, and sufficient views from both sides retained.
• The guest’s views are based on personal experience and estimation, not statistical conclusions; the text has been marked at each location.