Andrej Karpathy just shared a complete GPT in 243 lines of python.

A first-year CS student can read the whole file in an hour. This is the fifth iteration in a six-year compression arc: micrograd, minGPT, nanoGPT, llm.c, now microgpt.

This shows the algorithm is a commodity. The hard part is shaping what goes into the model and what you do with what comes out.

Context engineering is the clearest example. Two teams using the same foundation model will get wildly different results depending on how they construct prompts, manage retrieval, and structure agent orchestration. The model weights are identical. The product differentiation is entirely in the surrounding system.

Same story with evals. When the algorithm is a commodity, measuring whether your specific implementation actually works becomes the moat. Every AI PM needs a framework for this.

The best AI PMs I talk to on the podcast all land in the same place. The model is a component, not a product.

So PMs need to master these 5 skills:

1. Evals: news.aakashg.com/p/ai-e…

2. Observability: news.aakashg.com/p/ai-p…

3. AI foundations: news.aakashg.com/p/ai-f…

4. AI product strategy: news.aakashg.com/p/ai-p…

5. Context engineering: news.aakashg.com/p/cont…

243 lines is the engine. Everything above it is the product. And that's where PMs should contribute.