Karpathy talks about AI interaction again: In the future, a problem may be worthy of making a temporary app

Dongcha Beating AI News, Andrej Karpathy, currently working on the Anthropic pre-training team, once again discussed how large models should output results to humans. He believes that as AI takes over more and more specific tasks, people will spend more time understanding the results, checking the process and making judgments. How to make people understand what AI has done faster will become increasingly important.

His advice escalates from text all the way to video. Text can be drawn from ASD-STE100 and explained in shorter, less ambiguous language; complex content can be drawn directly, and then further generated into web pages with animation and interaction. He is most optimistic about customized explanation videos, such as asking AI to make a 3Blue1Brown-style animation and dubbing on the spot for a question.

This continues Karpathy’s previous judgment on AI interaction. In May of this year, he already recommended using HTML to carry complex large model output, and also summarized the evolution of AI output into plain text, Markdown, HTML, and then to future interactive video and simulation environments. Images, animations, and interactive interfaces can transform complex information directly into a more understandable form than a long paragraph of text.

Now that code and model capabilities are getting cheaper, software that previously wouldn't have merited dedicated development can be made readily available. A web page, an interactive tool or even a video can only serve one person, explain one problem, and then be thrown away after use.