Skip to content
C.W.K.
Stream
Lesson 02 of 06 · published

Local's Real Value Is Availability

~11 min · edge-era, availability, privacy, outage, our-judgment

Level 0Spec-Sheet Skimmer
0 XP0/91 lessons0/19 achievements
0/100 XP to next level100 XP to go0% complete
"Local means private only says where the inference happens. The data lives elsewhere already. What a local model actually buys is an answer on the day nothing else answers."

The Narrative and the Reason

The usual case for a model in the house is privacy. The household's doctrine sets it aside in one paragraph: "local means private" says only where inference happens; the data already lives in mail, in cloud storage, in code hosting, in databases, in photo libraries, and a chain is as private as its weakest link, which is where the data lives, not where the model runs. Bringing inference home while leaving the rest in place, the doctrine says, locks one door with twelve windows open; perfect privacy is an unplugged machine, which is useless. Privacy is a risk to manage — know what goes where, set the blast radius, keep the handful of true zeros written nowhere — not a reason to buy a pool. This quest carries that as the household's judgment, checks nothing against it but the household's own fleet, and finds the fleet consistent: the household runs local where local is good enough and buys the cloud where quality wins, and neither choice was made for privacy.

The Seat Local Actually Fills

The seat, in the doctrine's words, is availability: offline, or the vendor's eight-hour outage. The code block sorts the household's model-backed jobs by what keeps working when the cloud is unreachable and when the network is gone entirely. Memory search, image generation, on-device dictation and the coding agent's bottom rung keep working in both cases, because they run on a pool in the room. Voices, transcripts, summaries, frontier chat and the agent's frontier legs stop, because they were chosen for quality from a vendor's rack. The split is the fleet track's cloud-by-choice lesson seen from the other side: every job the household sent to the cloud for quality is a job that stops when the cloud does, and the local tier under each of them — the setting the tier lesson described — is what makes the stop a degradation rather than a silence. That is what the pool bought: not secrecy, but a floor under every surface.

Why This Is the Edge's Real Argument

The rivals track showed that for any single job a card or a rack does it faster, and the fleet track showed that for the household's best jobs the cloud does it better. What neither can offer is presence: a machine in the room, on, holding a model, needing nobody. The physics track priced that presence at 9 watts idle and 270 at full load on a Studio, silent, for years — the always-on row of the rival track's decision table, the one the rival never wins. The household's read is that this is the edge's whole value until the market's conditions change, and that it is enough: a family whose search, images, dictation and last-rung agent survive an outage has bought availability with the pool and quality with the subscription, and would give up neither for the other. The next lesson is which work belongs on the floor, by its shape.

Code

availability.py — the household's jobs under two kinds of outage·python
#!/usr/bin/env python3
"""Local's real value is availability. A year of a household's model-backed jobs against two
kinds of outage: the vendor's (a service down for hours) and the room's (no network at all).
Which jobs keep working, by tier, from the fleet track's census. Hours are illustrative."""
jobs = [  # job, tier today, keeps working when the cloud is down?, when the network is down?
    ("memory search (embed + rerank)",     "local hub",         True,  True),
    ("image generation",                    "local hub",         True,  True),
    ("dictation (on-device)",               "local, every Mac",  True,  True),
    ("coding agent, bottom rung",           "local, every Mac",  True,  True),
    ("coding agent, frontier leg",          "cloud",             False, False),
    ("voices (text to speech)",             "cloud",             False, False),
    ("video transcripts",                   "cloud",             False, False),
    ("transcript summaries",                "cloud via the hub", False, False),
    ("frontier chat",                       "cloud",             False, False),
]
print(f"{'job':36} {'tier':18} {'cloud down':>10} {'network down':>12}")
for j, t, c, n in jobs:
    print(f"{j:36} {t:18} {'works' if c else 'stops':>10} {'works' if n else 'stops':>12}")
print("\nthe household's rule, in its doctrine: privacy is a narrative, not a reason -- the data already lives in a dozen services;")
print("the seat a local model actually fills is availability: offline, or the vendor's eight-hour outage. The four rows that keep working are that seat.")

External links

Exercise

Sort your own model-backed jobs into the two outage columns and write the table on your card. Then write, honestly, the reason you own or want a pool — privacy, cost, speed, availability — and check it against the columns. If the reason is privacy, list the other links in your chain beside it.
Hint
Most people's tables have a floor they did not know they had — dictation, a local reranker, one agent leg. The exercise is to name it, because a floor you cannot name is not one you will be standing on when the outage comes.

Progress

Progress is local-only — sign in to sync across devices.
Spotted a bug or have feedback on this page?Report an Issue

Comments 0

🔔 Reply notifications (sign in)
Sign inPlease sign in to comment.

No comments yet — be the first.