Wink Pings

Don't Treat AI Like Google: Karpathy's 70-Minute Screen Share

Karpathy's 70-minute screen share reveals his real-world AI workflow. No theory, just hands-on practice. Four keywords, three concepts, and one underrated element: context.

Andrej Karpathy recently made his daily AI routine public: a 70-minute screen share showing how he uses these models every day. No slides, no theory—just demonstrating while doing the work.

Karpathy is an OpenAI co-founder and former Senior Director of AI at Tesla. X user Roman (@Nozelcode) shared clips from his screen share and called it 'a true masterclass.'

The content can be boiled down to four words: Prompts, Context, Tools, Workflows.

  • Prompts are not better simply because they are longer; Karpathy's approach is closer to 'giving the model a clear entry point.'
  • Context is the most underrated. What information you feed the model determines what it can produce.
  • Tools are for solving specific problems; it's not about how many you have.
  • Workflows connect the previous three into a structure that lets the model do the work for you.

In the demo, Karpathy makes a point: most people who aren't getting the most out of AI are overcomplicating things. You don't need to master dozens of tools; choosing the right model, providing enough context, and getting the workflow running smoothly matter more than anything else.

Three more concepts are worth noting: Harness, Loops, Graphs. A harness is a control framework beyond the prompt, loops are iterative cycles, and graphs are structured process orchestration. They all point in the same direction: from one-off conversations to reusable AI work systems.

The comments section had plenty of additions. Someone said the way you work with AI matters more than the model itself. Someone said context is the key to output quality. Someone else said output quality is determined before you even hit Enter.

If you usually just chat with AI and have never thought about turning it into a work system that does the work for you, this video is a good starting point. The original video is here; it's 70 minutes, so take your time.

Roman also wrote at the end of the post: 'If you want to go from chatting with AI to creating systems that work for you, this is the masterclass.' That's not an exaggeration.

发布时间: 2026-11-22 04:42