Karpathy talks about AI interaction again: in the future, even a single question may be worth making a temporary app for.

动察 Beating AI News Flash: Andrej Karpathy, currently working on Anthropic's pretraining team, has once again discussed how large models should output results to people. He believes that as AI takes over more and more concrete work, people will spend more time understanding results, checking processes, and making judgments. How to help people understand what AI has done more quickly will become increasingly important.

His suggestions escalate all the way from text to video. For text, ASD-STE100 can be referenced to explain in shorter, less ambiguous language; complex content can be directly illustrated, and further developed into webpages with animation and interaction. He is most optimistic about customized explanatory videos, such as having AI create, on the spot, a 3Blue1Brown-style animation with voiceover for a specific question.

This continues Karpathy's previous judgments about AI interaction. In May of this year, he already recommended using HTML to carry complex large model outputs, and summarized the evolution of AI output as plain text, Markdown, HTML, and then future interactive videos and simulated environments. Compared with a long passage of text, images, animations, and interactive interfaces can directly turn complex information into a form that is easier to understand.

Now that code and model capabilities are becoming cheaper and cheaper, software that was previously not worth developing specifically can also be made on demand. A webpage, an interactive tool, or even a video can serve just one person, explain one question, and be discarded after use.