Prompts, files, and drafts can stay on the desk. That is not anonymity, and it is not a substitute for good operational hygiene. It is a different default than shipping every token through someone else’s log.
An essay in public · One page
The intelligence moves home.
AI used to live in the cloud. Now increasingly capable intelligence can live on your desk, in your laptop, and eventually everywhere you do.
That shift is not a single invention. It is a convergence: open weights you can download, models that do more with fewer parameters, quantization that shrinks a checkpoint without erasing it, inference stacks that fit on a desk, consumer machines with enough unified memory to hold them, and a growing preference for privacy and ownership as the metered cost of remote inference keeps falling.1
The old assumption
For a while, intelligence had a location.
Frontier AI became synonymous with enormous centralized infrastructure: rooms of accelerators, specialized networking, and teams whose job was to keep a model answering through an API. That was not a plot against the personal computer. Scale created centralization. Training runs of that size needed clusters. Serving them at product latency was expensive. The rational interface was a remote call.
A laptop could write the prompt. It could not, for a stretch of years, credibly hold the model. Compute concentrated. Inference was billed by the token. Control sat with whoever owned the weights, the cluster, and the terms of service. The cloud was not the enemy of local machines. It was where the capability actually lived.
The rest of this page is the evidence. Chapters stay here: open a headline to read the argument, play with the figures, then keep scrolling. Nothing navigates away.
The Gap
Open to load this chapter.
Small Models
Open to load this chapter.
The Machine
Open to load this chapter.
Why this matters
When intelligence fits in the machine you own, the relationship changes.
Five reasons. Each card opens in place — no new page.
Open weights can be stored, versioned, forked, and kept after an API disappears or a price sheet moves. Licenses still bind. Hardware still depreciates. You own a checkpoint, not an oracle.
Local decode has no round trip and no queue. It can still be slow. A 2-bit 284B MoE thinking at max effort is not a snappy autocomplete. Speed is a property of model, quant, and memory bandwidth — not of “local” as a slogan.
Once the machine is paid for, marginal inference is electricity. That can be cheaper than metered APIs, or more expensive if the hardware sits idle, or if a hosted frontier model is simply better at the task. Open Economics for the cited prices; this card will not invent a payback month.
You can pin a version, inspect outputs, and decide when to upgrade. You also inherit the work: drivers, quant recipes, context budgets, evals. Control is a workload.
Local is not free. It costs silicon, power, time, and a willingness to accept that some jobs still belong on a cluster.
A necessary pause
Local isn’t automatically better.
The thesis is not local or cloud. It is local and cloud: a personal machine for the work that should never leave the room, and remote infrastructure for the work that still needs a building full of accelerators.
The Stack
Open to load this chapter.
Economics
Open to load this chapter.
Future
Open to load this chapter.