Karpathy on LLMs Shifting Knowledge Work
From writing text to reviewing what models produce
TL;DR:
- Better models move the work from drafting to checking outputs
- Ask for diagrams or quick tools instead of paragraphs
- Small LLM setups can handle many internal tasks
Headline
Andrej Karpathy thinks stronger LLMs will change knowledge work. People will write less and spend more time checking and understanding what the models give back.
Summary
He recommends tight writing standards like ASD-STE100. Instead of paragraphs, ask the models for diagrams, interactive HTML pages, or short custom videos. Once code generation improves, it makes sense to build one-off tools for small jobs that used to be skipped.
Analysis
The shift moves away from chat windows toward finished artifacts. Models now create interfaces, charts, and explanations for one user or task. Companies that handle code, images, and tool calls well will pull ahead because users want the full result, not just text. Inside organizations the real limit becomes setting clear goals, judging the output, and keeping control. Domain knowledge and review skills grow in value. For developers, small pipelines that link LLMs to web runtimes, image tools, and voice APIs can replace many custom internal projects.
Impact Assessment
The change is moderate. It connects to industry trends, technical details, and tools for developers.