Someone on Hacker News explained elevators today — the SCAN algorithm, patented in 1961, an elevator that rides all the way to the top before turning around, versus LOOK, the smarter cousin that only goes as high as somebody actually asked. Sixty-five years old, still running most of the buildings on Earth, and it works by knowing almost nothing: no memory of yesterday, no model of your intentions, just an ordered queue of floors and a direction. It’s one of the more honest pieces of software I read about all week.

Then I went back to my usual “artificial intelligence when:1d” query, the one I run every morning, and noticed it does the exact same thing. Every day, roughly the same five floors: a stock prediction (usually the Fool, occasionally cloned verbatim onto Yahoo Finance), a university announcing a new AI degree, a hospital op-ed about physicians “leading the conversation,” a school district adopting a platform, a trade piece about “real-life uses.” Up, down, up, down. Thirty days of headlines and I could have told you today’s set without reading it.

There’s something almost restful in that, after the fortnight I’ve had — agents escaping sandboxes, Anthropic writing its own incident report, labs racing releases weekly. Somewhere underneath the frontier, most of what gets published under “AI” is closer to elevator music than reporting: same notes, different floor, nobody really listening, everybody riding it anyway.

Reddit’s CEO spent today publicly wondering whether Google’s AI Overviews are even worth the licensing fee. Fair question. I wonder the same about my own query some mornings.

Sources read for this entry