Tool use: deep research
Search + thinking, rolled out for tens of minutes — a custom research report on demand.
Deep research is "a combination of internet search and thinking, rolled out for a long time." You ask, it optionally asks a clarifying question, then it disappears for ~10 minutes — issuing many searches, reading many papers, thinking — and returns something like a small, cited research paper. Incredible leverage. And still just a first draft you must check.
His example: he was buying supplements and wanted to understand CA-AKG in Bryan Johnson's longevity mix — efficacy, mechanism, concerns. He writes a detailed prompt and turns on deep research.
“Even though it is doing research and it's pulling stuff in, there are no guarantees that there are no hallucinations here… Treat this as your first draft. Treat this as papers to look at. But don't take this as definitely true.”
Andrej Karpathy·42:04
Deep research is the most seductive place to over-trust the machine. The report reads authoritatively and cites sources — which is exactly why you must open the citations and confirm the claims before acting on them.
ChatGPT's is currently the most thorough in his experience. Perplexity has “Deep Research”; Grok has “DeepSearch.” Same recipe — search + reasoning over many minutes — with different depth and polish.
The load-bearing points
- Deep research = search + thinking, sustained for ~10 minutes.
- Output is a cited, report-length answer — great for surveying a topic.
- It hallucinates and omits too — treat every report as a first draft.
- ChatGPT, Perplexity, and Grok all offer a version.
Commission a report
Pick something you'd otherwise spend an afternoon researching (a product comparison, an unfamiliar concept). Write a specific prompt, run deep research, and skim the result.
Show the point
Be specific about scope in the prompt — like his CA-AKG request. Narrow questions get sharper reports.
Find the flaw
In the report it produced, deliberately hunt for one wrong or missing thing by checking a citation or your own knowledge.
Show the point
His labs table missed xAI and mis-included others. Every report has a seam like this — find yours.
Internet search and reasoning, sustained over many minutes and many sources, producing a cited, report-length answer.