Ai

I told someone last week that his business couldn’t survive him. He’d called to pick my brain. Decades in the auction industry, multiple divisions, good revenue. But every decision in the company routed through him. Sales, operations, finance, client relationships. He’d spent his career being the best person on the microphone, and the rest of the business just happened around that skill. There were no documented processes. No decision-making frameworks anyone else could follow. If his potential successor took over tomorrow, they’d fail. Not because they’re incapable, but because the architecture underneath was never built for anyone else to run. ...

Dave Plummer is a retired Microsoft engineer. He created Windows Task Manager, ported Space Cadet Pinball to Windows, built the ZIP folder support, and wrote the Windows activation system for XP. He’s been coding since the TRS-80 days. Dropped out of high school. Worked at 7-Eleven. Went back to school, cold-emailed his way into Microsoft from Saskatchewan, and stayed for years building tools that billions of people still use daily. He now runs a YouTube channel called Dave’s Garage. Lex Fridman had him on episode #479, and four things from their conversation stuck with me. ...

Grafe Auction runs about 290 auctions a year across 48 states. Each campaign takes 10 to 14 days from start to finish. I’m the partner responsible for marketing, technology, and operations. That pace doesn’t leave much room for a system that’s clever but doesn’t hold up under pressure. It also doesn’t leave room for losing a thought from January that turns out to be the missing piece in February. The personal knowledge management crowd has an orthodoxy right now: tools don’t matter, just write. The AI crowd has its own: throw everything at the model and let it figure it out. Both are half right. Tools alone don’t create insight. Models alone can’t connect what they’ve never been given. ...

Stripe’s 2025 annual letter lays out five levels of agentic commerce, and what makes the framework stick is that each level isn’t defined by what the AI starts doing. It’s defined by what the human stops doing. Level 1: Eliminating web forms. You still research and decide. The agent just handles the typing, the checkout fields, the payment details. It’s a clerk, not a buyer. Level 2: Descriptive search. You stop searching for specific products and start describing situations. “I need back-to-school supplies for a third grader in Chicago, nothing too itchy or tight, he likes KPop and tennis.” The system reasons across weather, materials, sizes, taste, reviews, and delivery timelines. ...

The Observation I didn’t build structured data systems because I predicted WebMCP or the agentic web. I built them because “touch it once, make it findable by any system” is just good practice. “Computers do the mundane, humans critically think” is just a sound division of labor. These are principles, not predictions. But those principles, arrived at independently, produced exactly the infrastructure the agentic web needs. A straight line runs from a thought card I wrote in August 2025 to a platform that’s accidentally ready for a protocol Google announced six months later. ...

Last August, I had a conversation with a colleague that stuck with me. We were talking about whether auctions actually find true market value. My argument was simple: how can it be true market value when maybe 5-10% of potential buyers ever see our marketing? We’re calling it market value based on whoever happened to show up. I wrote about it at the time. My answer was to build communities, find your own lake, pursue smarter marketing instead of more marketing. I believed it. I still do. But six months later, I think I was thinking too small. ...