'AI' Is Not a Tool by Rich Hickey
Arguing that large language models are closer to automation than to traditional tools, commenters probe whether “AI as a tool” is a misleading framing that obscures loss of control, skill development, and responsibility. Many find value in AI for coding, troubleshooting, and knowledge work, but diverge on whether it should be treated like a wielded instrument (hammer, camera, IDE) or like outsourcing work to an agent, with some warning that the semantics matter for how we think about expertise, accountability, and long‑term dependence on automated systems.
Definition of “tool” vs “automation”
- Central disagreement is over a narrow definition of “tool” as something you wield with continuous control, feedback, and developing mastery, versus “automation” as something you delegate to and wait for results.
- Many argue this split is useful for thinking about loss of control, learning, and responsibility; others see it as pedantic or arbitrary.
- Some note that traditional software and machines (compilers, JVMs, lathes, cars) already blur this line.
Are LLMs tools, agents, or something else?
- One camp: LLMs are clearly tools, akin to IDEs or cameras; people use them instrumentally, and that’s enough.
- Another camp: conversational “AI agents” are closer to employees or subcontractors—goal-directed systems that choose means—so calling them “just tools” understates risks and loss of skill.
- Several propose spectra: tool ⇄ automation, tool ⇄ agent, high-control ⇄ “order a sandwich” experience.
Semantics, dictionaries, and language evolution
- Some strongly defend precise, etymology-based definitions to preserve conceptual distinctions.
- Others invoke family resemblances and language change, arguing you can’t “prove” what is or isn’t a tool.
- Debate includes whether appealing to dictionaries is clarifying or merely “old man yelling at cloud.”
Practical experiences with LLMs
- Some report dramatic productivity boosts: debugging, integration work, complex pipelines, testing, proofs, documentation.
- Others see widespread “vibe coding” and shallow use: prompting, accepting output, weak review, leading to security and quality issues.
- Multiple comments stress that value depends on remaining “in the loop” and understanding outputs.
Expertise, craftsmanship, and skill atrophy
- Concern that outsourcing reading, coding, and thinking to LLMs trades short-term speed for long-term loss of expertise.
- Counterpoint: much pre-LLM code was already “slop”; LLMs simply automate mediocrity and free experts to focus on higher-level design.
Analogies and metaphors
- Frequent analogies: fast food vs cooking (LLMs as “drive-through for code”), power tools vs hand tools, cameras vs painting, table saws and lathes, genies and wish machines.
- These are used to probe control, safety, learning, and where responsibility lies.
Meta and community dynamics
- Discussion notes heavy flagging, concerns about “hero worship,” fatigue with AI opinion pieces, and tension between shallow dismissal and over-serious semantic debate.