Andrej Karpathy LLM output framework

by Vojtech: Andrej Karpathy

The real shift isn't just about asking LLMs for answers anymore; it's about demanding they build the actual thing. Andrej Karpathy laid out a framework where you stop accepting plain text as the default and start climbing up the ladder of output formats as the models get smarter. You can ask for ASD-STE100 standards to force clean, simple writing that hits that '80% of the way there' sweet spot if the strict rules feel too heavy. Or you skip the paragraphs entirely and request diagrams because parsing an image is just easier than wading through walls of text. If you really want to see what's possible, ask for HTML so the model generates a fully interactive webpage complete with animations instead of static words. But the one he is most excited about is custom explainer videos where you tell the AI to make a 3Blue1Brown-style breakdown on any topic X, then narrate it using your own voice API. The whole point here is that intelligence and code are becoming so cheap that we can finally commission massive, single-use artifacts that never made sense to build before. We are talking about apps designed to be used exactly once or videos explaining one specific concept without any fluff. Pushing this boundary is going to surprise everyone who thinks the best output is still just a block of text.