I’ve been in Chicago this week at the Performance Trust Fall event. The IT department has been running a hackathon. My contribution on the team was developing a trial MCP server to integrate database access into potential LLM agents. It’s pretty cool when it works.
I’ve also discovered that Anthropic, the makers of Claude, have defined a bunch of MCP servers which can be run on one’s computer to grant access to local files for the LLM. When I’m back on my own machine, I’d like to see if I could get it to consume the original Markdown files that comprise this site.
My reading for this Chicago trip has been Isabel Wilkerson’s Caste: The Origins of our Discontents. Not only ought the subject matter be required reading for any American or American-to-be; her writing is exceptional. She addresses some of the most horrific elements of American history with candor while developing empathy for both ends of the caste system. It’s taken me a few months to start reading this book because of anticipated grief, anger and despair. Yet Wilkerson has crafted a work so great that it pulls no punches yet doesn’t leave one in a puddle on the floor.
LLM chats are replacing my web search usage. When I’m solving a software problem a succinct chat response, complete with an example custom to my situation, is exactly what I used to look for in my search queue. When I’m looking for general information, my browser (Arc) will search 10+ sites and aggregate the answers. That’s how I discovered tickets to The Sound of Music two nights ago, by searching for things to do in Chicago today.
Quality information is a baseline necessity for an LLM to consume, but all that grows from that baseline - summaries, comparisons, compilations, syntheses - can be so quickly and thoroughly accomplished by an LLM that it’s hardly useful anymore. Developers of a public API need to produce accurate documentation, but now an LLM can generate the examples on-the-fly. Theatre events need to be published, but it’s not useful for a human to compile date night ideas.
For my own website, the most valuable content is that which offers a definition or sample of an idea. What I personally benefit from is not those building blocks but the ways that insights spark from synthesis. Ironically, that makes my site’s content mostly useful as a bare repository, with tools on top that bring out its value. I wonder if there’s a way to go backwards; to figure out what building blocks are missing from a more robust final result?