Karpathy's 2-Hour Masterclass: The Comments Section Checks the Video Source First

A tweet calls Andrej Karpathy's 2-hour video a masterclass, covering Agents, Harnesses, Loops, and Graphs, and even labels it a must-watch. The tweet includes the video, but the comments are not unanimously cheering; instead, three kinds of reactions appear first: asking for the link, asking about the date, and checking the source.

Original post link: https://x.com/ajitcodes/status/2116364535997006310

Some users point out that the video in the original post is Karpathy's February 2025 “How I Use LLMs,” which is 2 hours and 11 minutes long; a comment says he was working at OpenAI at the time. This claim comes from the comments, and the material does not provide the original YouTube link, so it cannot be independently verified. Another user directly asks @grok whether this is a misleading post.

This does not mean the video itself may not be valuable. The issue is the packaging. An old course with a new title, paired with the four buzzwords Agents, Harnesses, Loops, and Graphs, can easily make people think it is a new tutorial on those topics. Bookmarking is fast; verifying the source is slow.

The more specific questions in the comments come from @saeid: how the loops section is explained, and how to use a loop to check coursework. Mian Maaz Ullah Khan is concerned about harnesses: what to do when a tool fails halfway through, a retry runs twice, or the agent loses state. For people who have built agents, these questions are closer to real work than “2 hours” and “masterclass.” Demos usually show the happy path; it is the failure path that determines whether a system is usable.

A little background for readers unfamiliar with these terms: Agents can be understood as models calling tools to complete tasks; a harness is an execution framework wrapped around the model; a loop is repeated checking or retrying; a graph is organizing steps into a directed flow. Different teams do not use exactly the same terminology, but the discussion usually focuses on state, retries, and failure recovery.

If you only want to judge whether it is worth watching, first confirm three things: the video's original source, its publication date, and its table of contents or subtitles. The original post does not provide a course title page, and the comments do not provide a clickable original link. What can be confirmed from the material is the tweet and the video file; it cannot be confirmed that this is truly a new masterclass covering all four topics.

Add one more verification step before bookmarking. For “masterclasses” in your feed, this step is more useful than bookmarking.