Wink Pings

Karpathy’s 70-Minute Video: The Most Valuable Part Isn’t the Prompts

Karpathy didn’t use a single slide and demonstrated 70 minutes of hands-on AI practice. Prompts are only the surface; what really sets people apart is the three-layer system design of Harness, Loop, and Graph. The video may be dated, but the methodology is not.

OpenAI co-founder Andrej Karpathy simply shared his screen. For a full 70 minutes, without a single slide, he demonstrated how he works with large models every day. The video has been shared many times, and the comments are almost unanimous: a top player’s hands-on practice is more useful than ten secondhand summaries.

Some people pointed out that the model in the video is GPT-4o, so it may not have been recorded recently. But the method is not outdated.

Karpathy’s core chain has four steps: Prompts → Context → Tools → Workflows. Pick the right model, feed it the right context, and let the tools run the rest. It sounds simple, but most people get stuck; the problem is often in the system design, not so much in the model itself.

A viewer said after watching: top players also keep revising and starting over when working with models. The difference is that when they start over once, they get stronger; when ordinary people start over ten times, they just get tired. Why? Because they change the system around the model, while ordinary people only change the prompts.

This system finally got a name recently. Whitemon published a long article (link) that breaks agents into three layers: Harness, Loop, and Graph. In one sentence: environment, feedback, and flow.

Harness: Giving the Model Hands, Memory, and Walls

Harness is everything outside the model that lets it actually do work. The simplest definition: remove the model from the architecture diagram, and everything left on the page is Harness.

Specifically, it includes: a contract (goal, done_when), a map (AGENTS.md, not an entire manual), tools (ten narrow tools beat forty overlapping ones), memory (PROGRESS.md, not the chat log), permissions (walls instead of requests), and tracing (a trace of every run).

An agent without done_when will quietly solve a simpler task and then tell you it’s “fixed.”

Loop: Make the Model Prove Itself

Every agent has its own small loop internally: call the model, run a tool, feed the result back. But most people’s loop is just “try again.” A real loop needs to be designed: goal, evidence, and stopping condition.

One key point: don’t loop on confidence; loop on evidence. The model saying “should pass” doesn’t count as done; the test passing counts.

Before writing code, write the check first. A test that fails before the fix is more useful than ten prompts. Feedback should say “expected vs. actual,” not just “red.”

Graph: Deciding What to Do Next

A loop can only do one thing. When multiple things need handoffs, Graph comes into play. Nodes are clearly bounded pieces of work, edges are real dependencies, and state is the information that survives between nodes.

One easily overlooked point: four agents sharing one context is essentially one agent wearing four hats, not a team. To run in parallel, each worker needs its own context.

Don’t rush to Graph for things a loop can handle. Wait until several loops need handoffs, then split out nodes.

Back to Karpathy’s video. What he demonstrated—Prompts → Context → Tools → Workflows—is actually a manual version of Harness, Loop, and Graph. The only difference is that this loop runs in his own head and hasn’t yet been externalized into a system. Ordinary people can copy his prompts, but they can’t copy this skeleton.

A viewer put it well: tools are just the entry point; how to fix things after mistakes is the truly valuable part. Watching how he pulls himself back after the model leads him astray is ten times more useful than watching the smooth parts.

So the next time your agent says “done,” don’t rush to change the prompt. First ask: which layer said that?

  • Repeatedly retrying things that have already been tried—Harness’s memory has a problem
  • “done” isn’t trustworthy—Loop is missing a real check
  • Every step is correct but the whole thing is slow—Graph’s structure needs to be split up

Whitemon’s article includes a complete checklist that’s worth saving. Models change every quarter, but your contracts, tests, state, and gates remain. They are the foundation that keeps the system from crashing.

The video is here. It’s 70 minutes; I recommend setting aside a full block of time to watch it.

发布时间: 2026-11-21 16:40