"A job's shape is two numbers: how much it must read before the first token, and how long the person will wait. The edge serves one shape well and one badly, and the industry is building for the bad one."
Shape, Priced by the Curve
The curve lesson gave the two numbers that decide a job on a Mac: time to first token is prompt tokens over the prefill rate, and decode is output tokens over the decode rate at that context. The code block prices seven jobs on office's 27B at the rates the lab measured. Dictation cleanup — three hundred tokens in, three hundred out — is ten seconds. A mail draft, fifteen. A page of translation, under a minute. A transcript summary, thirty-six seconds with an eight-thousand-token prompt. A photo tagged by a vision model, three seconds in the background. Those five are short-context and latency-tolerant: the person is not watching a cursor, or is watching one for the length of a breath. Then the other two. A chat with the household's vault loaded — a hundred thousand tokens — waits five minutes before the first word, nearly all of it the prefill wall; a whole-repository coding turn waits seven, and a coding session is dozens of turns, each replaying the context the machine already read, because the door the agent uses keeps no prompt cache across turns — the hub's MLX server and mlx-lm's own API do have one, as the journey and fleet tracks showed, but the Ollama rung does not — and there is no batch to hide the replay behind.
The 95% and the 5%
The household's doctrine names the populations. The first five jobs are what most people would use a model for — summary, translation, mail drafts, dictation, photo tagging — and they are already running on the fleet: transcripts by the thousand, on-device dictation on every Mac, an offline model behind the prose editor, a local model tagging images. That is the 95%'s work, and the edge serves it well on a 24 GB laptop, let alone a Studio. The last two jobs are the industry's work: coding, with long context, whole-repo reads and dozens of loop steps, which is what makes a 256K context window unusable on a Mac in practice and what a pool whose common doors keep no prefix cache across turns serves worst. The doctrine's sentence is exact: the industry is building models for the workload the edge cannot serve and productizing nothing for the workload it serves well, because coding is where the revenue is and the 95% does not yet pay. The rivals track's checklist put the fix in software — a prefix cache and batching in the platform — and this lesson's arithmetic shows how much it would buy: a vault chat's five-minute wall paid once instead of every turn.
Why the Household Runs Coding in the Cloud and Dictation at Home
The fleet's records are the two columns of the code block sorted by tier. The coding agent's frontier legs are cloud; its bottom rung is local and slim, because the slim harness is what makes a local coding turn short enough to be worth taking. Dictation, tagging, summaries and search are local or one setting away from it. That is not an ideology about where models should run; it is the job's shape read off the curve — a wall of seconds goes to the room, a wall of minutes replayed dozens of times goes to a rack with a prefix cache and thousands of streams to hide it behind. The household's own reading goes one step further and this quest carries it as judgment: the day coding is quiet enough for the edge is the day its revenue cools, and the model that survives that day is not the one that scores on coding benchmarks. The next lesson is the arithmetic of the machine that has to wait for that day.