Andrej Karpathy used an X post to make a practical point about working with language-model output: the best format may not always be plain text. He wrote that diagrams or images can be easier to process, parse, and understand than writing, and he also said asking for output "in HTML" can produce a beautiful, interactive webpage.
In the same post, Karpathy made a broader argument about where this is headed. He wrote that as LLMs get better, they will do more of the legwork autonomously, and more human work will rise toward oversight and understanding.
The formats he called out
Karpathy’s examples were straightforward. On the visual side, he said diagrams or images can be easier to process, parse, and understand than writing. On the interactive side, he said asking for output "in HTML" can produce a beautiful, interactive webpage, and that LLMs are getting really good at frontend and can create beautiful experiences and animations.
Taken together, those suggestions point to a workflow where the model is not just generating an answer, but also packaging that answer in a form that may be faster to inspect. Karpathy framed that as part of a larger shift in which improving models handle more of the routine work, leaving people to spend more time interpreting results and supervising the process.
The replies cited here
The replies in scope did not show a clear consensus, but they did show how other users picked up on Karpathy’s examples. Grady Booch replied to the diagram point by pointing to UML. Another user, elie, wrote that interactive HTML is "probably one of the most time-efficient ways to understand model output" while also saying models are still not good enough at building the right level of abstraction around plots.
That leaves the conversation in a fairly specific place. Karpathy was arguing for diagrams and HTML as useful ways to present model output, and the cited replies mostly engaged those concrete suggestions rather than treating them as a settled rule for how AI tools should explain themselves.
Comments