Thoughts from San Francisco

Ben’s Bites came back from San Francisco with a useful correction to the usual AI hype. The biggest mistake, he thinks, is assuming everyone wants the same thing from these tools. They don’t. There are the everyday users — the “better Google” crowd — and then there are the curious people who try Lovable or Replit, hit a database bug, hammer “fix it” a few times, and stop. After that come the non-engineers who know enough to steer a coding agent, and then the developers themselves.

That framing matters because it changes the job. In the Makerpad era, he thought his work was about turning people into builders. Now he says the real value was helping already-curious people learn how to build. Curiosity, in his view, comes from seeing examples that look playable, not serious. That’s why he keeps making odd little things, and why he renamed his talk to “how I build and play with AI”. Someone told him afterward they loved how often he said “play”. That sounds right to him. Copy, remix, learn.

His bigger bet is on personal agents, though he’s clear they’re not there yet. The pitch to ordinary users is tidy: let the agent do your computer work, like rebooking a flight. But for the people who live inside computers, the appeal is different. They want agents as a way to do much more with their own machines. And right now, the rough edges are obvious. One thread for everything is too blunt. He wants side threads, Slack-style. He wants little widgets that can be built on the fly. Wabi, he says, is the only thing he’s tried that does a version of that.

Then there’s memory, or the lack of a visible one. He’s spent months tending what he calls his “context garden”, and still doesn’t trust agents unless they show him what they remember. So he makes his own system sync every memory into a Google Doc. He also hates loops that don’t close, where an agent nags about some unfinished task that was actually handled somewhere else, like Gmail or a calendar. That’s the kind of problem he’s tempted to build through, not from scratch, using primitives like Pi instead of betting on one lab’s stack. The speed matters too. He says he’s not sending agents off overnight; he’s steering them all day, so fast feels smart. More compute, more surfaces, more play.