| ▲ | Nzen 3 days ago | |
I considered the section about programs interrogating one another to accomplish some goal [0] the most evocative idea. I'll admit that my limited fantasies resembled something that resembled swagger's openapi or hateoas. When I heard about Anthropic's model context protocol [1], it reminded me of this talk. I feel pretty skeptical that llm based systems are apt to craft a pigdin with a tool, as that seems like the kind of interaction that would use up lots of the context window. I'll grant that I've heard of people working around that by having their llm leave a summary of their session [2], to bootstrap the next, fresh session. I guess one could leave a pigdin dictionary that suited the llm traning data, as well. [0] intro starts at 13:13, regarding arpanet, but description starts at 13:53 [1] https://modelcontextprotocol.io/docs/learn/architecture [2] https://news.ycombinator.com/item?id=44661223 <edit to add> Also, Bret Victor's team was able to involve the light pens mentioned in his dynamicland research group / lab. | ||