Karpathy's AI Engineering Ladder: Prompt Only Gets You to 30%
A five-stage progress bar from LLM to Graph, explaining why changing the prompt alone is not enough. Includes the original post's video and several key takeaways.
Andrej Karpathy gave a one-hour talk on AI engineering at Stanford. Kshitij Mishra condensed the lecture into a progress bar:
10% → LLM
30% → Prompt
50% → Agent
70% → Loop
100% → Graph
This progress bar is not precise, but it is a useful reference. It lays out the current state of many teams: they are stuck at the Prompt layer and understand AI engineering as tweaking prompts.
As you go further down, the questions change. The model needs to get context, have memory, be able to call tools, and continue running based on results. Agents let the model break down tasks and use tools. Loops let the system keep executing after receiving feedback. Graphs organize multiple steps, tools, and data flows into a structure. What the nodes are, how the edges connect, and how state is passed matter more to whether the system can work reliably than a single beautifully written prompt.
The original post includes a strong line: “Delete everything, keep Graph.” Don’t take it literally as “delete the database.” It means that once a system reaches a certain level of complexity, scattered prompts become hard to maintain, and the graph structure is more worth preserving. Who calls whom, what each step produces, and where to return after a failure all need to be visible.
For AI practitioners with a development background, this progress bar can serve as a checklist: which layer are you stuck at now, and what is missing for the next layer? For non-technical readers, it offers a way to judge: an AI product is not just a model plus an input box; the system layer around it is the engineering.
The original post's video is on X. Someone in the replies asked about a YouTube version, but the original post did not provide a link.
发布时间: 2026-10-28 22:45