Daybook, 21 July 2026

Working on an interesting idea today — kept progressing on the evaluations and improving the agent architecture. And now I'm following something very similar to how an autoencoder works. So: I've got one system that discovers the latent semantics from the different artifacts, tables, and data sources — it builds a kind of ontology, how the different artifacts connect and how you can drive them. And that semantics powers the other agent. The way I'm validating it is through reconstruction — once the semantic system is in place, I trigger a task, and I measure how well, with those semantics, the agent can reconstruct the analytics workflows or the artifacts.
And there should be a dual feedback. From the traces, there has to be some analysis and feedback going back — one, to the system that discovers the semantics; two, to improve the agent itself. So I can figure out where the gap actually is: whether the semantics has gaps, or the agent does. I'm building it as a kind of ongoing cycle — a workflow that can keep running, figure out where the issues are, and come up with feedback on what to revise. That's the shape of it.
One honest annoyance: I realized today, kind of badly, how much Claude Code forgets over a session — whatever we did earlier, a lot of it just slips. That's fine, that's how it is right now.
I was at it from about 8:30 in the morning to 6, 7:30 in the evening — different tasks, reading some papers on agents, ideating. And after that I stepped out to a nearby Barnes & Noble, just to see what's there, maybe find something new. And I did — a book called The Simulation Hypothesis, which felt genuinely interesting, kind of resonating with me. So I bought it. But then, when I got home, I realized I'd left it in the cab. Still trying to figure out how to get it back. And somehow the day ends on a bit of a low. I'll be honest — I'm feeling pretty down as I write this. It's a small thing, a lost book, but it's sitting heavier than it should.
I'm doing what I can to get it back — reached out to Uber support, and I'm trying to get to the driver, and even the riders who took the cab after me, in case someone picked it up by mistake. And honestly, I can't quite shake it. It's not really about the book itself; it's something more than that, and that's why it's sitting on me so much.
There's a small lesson in how it even happened, too. During the ride we'd been having a really good conversation — cars, robots, AI, all of it — and I think I got so carried away in it that I just stepped out and left the book behind. The good thing led straight to the slip. So: be more careful. Even the conversations I love shouldn't make me lose track of the small things sitting right in front of me.
Settled · 21 Jul 2026
Disagree, or know something I missed? Write to mepc3197@columbia.edu · the title goes in the subject line
Related
- Perspectives on Problemssessionphilosophy · science
- The Art of Problem Solving in the Age of Superintelligencecourseunder wayphilosophy · science
- AI Safety, Alignment, Human & AI Systems, Evals, Diffusion and Futurescourseplannedphilosophy · science