Generative AI: Autocomplete for Everything
Generative AI: Autocomplete for Everything
Authors: Noah Smith and roon (a researcher at a prominent AI company) · Published: December 1, 2022 · URL: https://www.noahpinion.blog/p/generative-ai-autocomplete-for-everything
One-line: Written the week ChatGPT launched: generative AI will behave like every previous labor-saving tool — taking over tasks rather than jobs — and the resulting “sandwich workflow” (prompt, menu, edit) is the shape of most knowledge work to come.
Summary
The essay opens on the two emotions AI produces: dread among people who expect to be replaced, and guilt among the people building it. The moment is late 2022 — diffusion models had just matured, Midjourney and Stable Diffusion had landed, Stability AI had raised $101M and Jasper $125M. The authors’ position is blunt: “we think the fear, and the guilt, are probably mostly unwarranted.”
Tasks, not jobs. If AI causes mass unemployment it will be the first technology in history to do so. Industrial machinery, CNC machine tools, software and industrial robots each triggered the same panic and none delivered it; recent evidence associates robot adoption with higher employment at firm and industry level. The reason is that those technologies replaced tasks. Smith’s own ancestors were metalworkers who spent their days bashing metal by hand; two centuries later metalworkers direct machines to do the bashing, “and you can bash a lot more metal with a machine.” They cite Prediction Machines (Agrawal, Gans, Goldfarb) and Acemoglu–Restrepo’s task-level model, in which new technology can raise productivity at existing tasks, shift labor toward different tasks, or create new ones — and which effect dominates decides whether workers are helped or hurt. Compressed to Smith’s aphorism: “Dystopia is when robots take half your jobs. Utopia is when robots take half your job.”
Comparative advantage. The example is a venture capitalist (“let’s call him Marc”) who types inhumanly fast and still hires a secretary, because his hours are better spent elsewhere — “he ends up paying someone else to do something that he’s actually better at.” Applied at task level, this survives even general intelligence: AI resources are finite, and as roon puts it, every use of a frontier model is “lighting a pile of GPUs on fire.” AI businesses will expand and consume physical resources “until human workers themselves, and the work they do to complement AI, become the scarce resource.” Whether future wages rise depends on whether AI’s skills are similar to ours or complementary — and current generative AI is visibly different: hands with the wrong number of fingers, window panes that don’t line up, a Tesla that parallel-parks unlike any human.
How the apps actually work. GitHub Copilot is the exemplar. An LLM trained on public code suggests the next lines, so engineers describe in plain language what a snippet should do. This splits the value chain cleanly: humans reason about business logic and decomposition, the model remembers thousands of frameworks and incantations. “No software engineer will be running solo anymore” — everyone gets “a superhumanly knowledgeable, but idiosyncratic pair programmer.” The prediction that Copilot “will not only raise the productivity of individual programmers but create more programmers generally” is the one most worth revisiting. Midjourney is the second case: coherent composition, broken details, and a ceiling on narrative abstraction — “Midjourney isn’t dreaming up ‘Guernica’.” Prompting is an inexact science of communicating intent, “a problem not dissimilar to difficulties of delegating to other humans.”
The sandwich workflow. The essay’s most portable contribution. Three steps: a human has a creative impulse and writes a prompt; the AI generates a menu of options; the human picks one, edits it, and adds finishing touches. They walk it through op-ed writing (Smith imagines bullet points expanded into paragraphs in a facsimile of his own style, then edited back), illustration, industrial design (“55-inch TV stand with two cabinets”), architecture, graphic and interior design, legal briefs, memos, marketing copy, consulting decks and financial models.
They take the obvious objection seriously — that prompting and editing are less fun than generating, and that jobs get more rote. The answer has two halves, and the second is more honest than the first. Increased wealth should buy more leisure for creative hobbies; and anyway, “what we produce has always been dictated by the market, while our true expressive artisanal creation has always been done on our own time, and only a lucky few people have ever been able to combine the two.” CAD didn’t take the fun out of woodworking.
The close: “not something that replaces humans, but something that gives them superpowers. A proverbial bicycle for the mind.”
Full text
Archived privately against link rot: ../attachments/generative-ai-autocomplete-for-everything/generative-ai-autocomplete-for-everything.md
Connections
- Noah Smith — co-author with roon; the piece is the origin of the “sandwich workflow” term and of Smith’s task-versus-job framing that recurs across the Noahpinion run.
- The Openness of AI — Kyle’s essay on how AI capability diffuses; this piece is the complementary claim about where the value lands once it has diffused.
- Artificial Intelligence and Automation — the historical argument (machine tools, robots, word processors) that every later jobs-and-AI debate keeps relitigating.
- AI’s Big Messaging Pivot — the same author, four years on, watching the industry adopt in 2026 the argument he was making in 2022.
- Marc Andreessen — the comparative-advantage example is built on him by name.
- Anthropic — cited for the early experiment showing human-plus-model outperforming either alone.