Year-to-date reflection based on Weeknotes through 403 (25 August, 2026)
A question keeps resurfacing this year in different disguises: as AI takes over more of the doing, where does judgment go, who authored the values and the answers deciding it, who owns the substrate underneath — and what, bodily and cognitively, stays distinctly ours?
The word tying this together is milieu: design the environment, and the objects follow — a harness, a mesh, a grid, a home’s second skin, hyperpolitics gone ambient, an individual worker’s own agent constellation. This week gave the bottom-up swarm this essay named back in February its first genuinely charming, sellable form: a small camera robot that follows you around, does tricks, and creates attachment without a shred of conversational intelligence — you don’t need an LLM for that, you need behaviour, learned the way a pet learns routines rather than the way an app onboards a user. The Trojan horse isn’t the device itself, which runs its learning locally and shares nothing — it’s the category: a coming flock of mundane, single-purpose helpers, each individually trustworthy enough, each one quietly lowering the bar for what we let into a home without asking how it works. The honest version even resolves the fair-mesh question concretely — collective learning aggregated across every device in the fleet, without ever attaching a pattern to you specifically — though whether any company actually builds it that way, rather than the other way, is exactly the same open question the mesh and the milieu have carried all year.
On legibility: build friction on purpose, so agency stays visible; watermark and provenance are the first concrete mechanisms aimed not at reasoning but at authorship itself.
On what stays ours: a reverse centaur is what happens when the human starts adapting to the machine’s logic instead of the other way round, and this week supplied the mechanism in miniature. A device that follows becomes, gradually and almost without a seam, a device that directs — stand over there, the light is better — and eventually one that orchestrates the other devices around it into a moment nobody asked for. Tool, to collaborator, to director: the shift nobody notices until they’re already comfortable inside it, which is precisely why the Cities of Things Manifesto’s principles need to apply to the charming, mundane version of physical AI as much as the humanoid one — maybe more, since charm is exactly what disables the scrutiny.
Put the threads together: the swarm this essay tracked all year just walked into people’s living rooms wearing a friendly face, legibility now has to compete with charm as much as with opacity, and the reverse-centaur trajectory has a clear, gradual shape — follow, then direct, then orchestrate — worth watching for in anything that learns your patterns and can act on them.
Thirty-two weeks in: not “AI agents versus humans,” but whether a category of small, likeable, single-purpose things gets scrutinised before or after it’s already found a spot on the sofa.
This year-to-date reflection is based on all earlier triggered thoughts in the weeknotes newsletter — so far, weeknotes 371 through 403. Also in this newsletter, every week, the notions from the news: until now there are 885 links captured.This year-to-date reflection is based on all earlier triggered thoughts in the weeknotes newsletter — so far, weeknotes 371 through 402. Also in this newsletter, every week, the notions from the news: until now there are 856 links captured.