From Bin: "We learned the hard way that agents, when working with JSONs and XMLs, have no visual intelligence. When an agent writes a JSON blob, it can be accurate, verifiable, and correct structurally.
But the agent has no idea whether this JSON is going to be good looking or not. That is actually the biggest problem we ran into.
That's when we turned to code. Code, especially HTML, is LLM's native language."
Full interview on @PeterYangYT!
