How The #1 AI Engineer in The World Actually Uses AI

Summary

The video describes how AI engineer Andrej Karpathy reportedly uses AI agents, emphasizing delegation, concise instructions, and judging models by practical behavior rather than benchmark scores.

Tools and workflow

Karpathy reportedly uses Claude Code on a Mac with Opus 5.5 at the default medium-effort setting. He avoids reading or writing code and running terminal commands, instead asking Claude to handle the work behind the scenes.

Short, plain-language prompts

He favors brief requests and asks for explanations in simplified technical English. The narrator says Karpathy believes piling on requirements can backfire when they conflict with earlier instructions.

Delegating complex tasks

Examples include asking Claude to “check recent cyberthings” to review and strengthen security, or giving it a server IP along with the keywords “update, deploy, secure, harden, and optimize.” The video says Opus can delegate work to Sonnet subagents, including in data-processing workflows.

Visual explanations

When text is hard to follow, he asks for diagrams, HTML pages, or explanatory videos, on the premise that visual material can communicate faster than long explanations.

Evaluating models

He reportedly distrusts public benchmarks and recommends considering Epoch’s ECI, which assesses the quality of benchmarks themselves.

The narrator also describes a “land or water?” experiment: asking a model to classify geographic coordinates, then using the answers to build a world map. This is presented as a way to probe what knowledge models contain.

Building and sharing projects

The video describes HackerTube, a site that collects videos repeatedly recommended on Hacker News. For projects involving data, Claude can filter links, regular expressions, or spam, then pass the remaining material into a Claude Code workflow.

Rather than sharing an application, Karpathy reportedly prefers sharing its data source and prompt so others can generate their own customized version.

General takeaway

The narrator presents these examples as a practical guide to using AI agents: give them room to work, use visuals when helpful, and treat apps as interfaces that can be rebuilt around the underlying data. The video also notes Karpathy’s polite tone toward Claude, though it does not establish whether politeness improves results.

Reviews, guides, or tutorials

This is an analysis of reported AI-usage practices, not a hands-on product review or step-by-step tutorial.

Main speakers and sources

Nick Saraev narrates the video. The practices discussed are attributed to Andrej Karpathy, based on the narrator’s account of his recent work and comments. Claude, Opus, and Sonnet are the AI tools discussed.