Thinking can be outsourced to AI, but understanding cannot—AI makes thinking cheap, while the value of understanding keeps rising.
Raymond’s explanation
This line comes from Andrej Karpathy: “You can outsource your thinking, but you can’t outsource your understanding.” AI can already find information, write reports, run processes, and even think at a higher quality than people. But AI cannot answer questions such as “Why do this?”, “Where should the next step go?”, or “What is good?” on a person’s behalf.
Raymond’s plain-language translation consists of four contrasts: AI can help find information, but “what to look up and why” is up to the person; AI can help write a report, but “why write it and who will read it” is up to the person; AI can automate a process, but “what real problem this process solves” is up to the person; AI can generate 100 versions in a second, but “which one is good and why” is up to the person. Karpathy’s line “The agent fills in the blanks; you decide where the blanks go” is the engineering version of the same idea.
Three things that AI cannot take away support understanding: comprehension (grasping what is fundamentally happening), taste (the ability to judge what is good or right), and curiosity (the drive to ask “why” one level deeper). This is why, the stronger AI becomes, the greater the gap between people who know a field and those who do not. Distinguishing whether AI is flying or stuck in a given task—the jagged intelligence Karpathy describes—depends on how deeply a person understands that task.
When to use it
- When deciding whether AI should handle something: execution and information processing can be outsourced; defining the problem and judging acceptance cannot.
- When someone asks “Now that we have AI, what do we still need to learn?” The answer is understanding—journal, read slowly and reorganize what is read, talk with people outside one’s field, turn “this version is better” into a sentence beginning “because…,” and preserve inefficient time to cultivate curiosity.
- When creating content: if AI-generated content has none of the creator’s point of view, it is useless because anyone can produce the same thing.
Counterexamples and boundaries
- “Outsource thinking” does not mean stop thinking: a person remains part of the system, and information still needs to enter the brain; only the rough work of processing information is handed off.
- Do not overcorrect by refusing AI: protecting understanding does not mean doing everything oneself. Execution costs have approached zero, so insisting on doing the work manually wastes the increasing value of “direction, taste, and judgment.”
- This boundary shifts with the depth of one’s understanding: the more one knows about a field, the greater the range that can be safely outsourced.
Origin
Andrej Karpathy introduced the idea in a May 2026 interview with Sequoia Capital (in the same interview, he also discussed jagged intelligence, the MenuGen example, and “the agent fills in the blanks; you decide where the blanks go”). In a long article on 2026-05-10, Raymond broke it down into the boundary between thinking and understanding, three assets AI cannot take away, and five daily practices for cultivating understanding.
Where it has been discussed
Articles and newsletters
- 2026-05-10_You Can Outsource Thinking but Not Understanding: Key Points from Karpathy’s Latest Interview and the Real Competitive Advantage in the AI Era — full interview highlights and Raymond’s reflection.
Social-media posts
- 2026-05-10_You Can Outsource Your Thinking to AI but Not Your Understanding_27662409376680584 — the long Facebook version, with five concrete practices for cultivating understanding.
Related concept pages
- AI Tool Applications — the IPO principle (Input is your irreplaceable point of view) applies this boundary to writing.
- Learning Methods — slow reading, reorganization, and Feynman-style output all build understanding.
Related concepts
Super AI Individual, Externalizing Judgment, Train AI, Not Learners, Tools Extend Thinking