On the metal, literally

WWDC week. Apple announced a 20-billion-parameter on-device model that activates one to four billion parameters per request and lives in NAND flash instead of DRAM. A framework that swaps AI providers without code changes. An OS door for running third-party models locally. A coding tool that pairs a small local model with a frontier model called in for the heavy judgment. And — the detail most coverage buried — the next generation of Apple's own models will be built on a rival's architecture and training infrastructure, rented.

The bet underneath

Read those as one move and the thesis is hard to miss: Apple looked at frontier model training and decided it is not the moat. Weights can be rented, swapped, replaced — their own developer API now treats the model as a plug-in. What Apple kept is the metal and the doorway: unified memory, flash bandwidth high enough to treat storage as model residence, and the OS-level harness every model passes through to reach a device in someone's hand.

The depreciation curves explain it. Weights depreciate brutally — something open and better ships every quarter, which is exactly why you would not marry one. Metal depreciates slowly. And the harness appreciates: it accumulates integrations, context, users. Weights are tenants; the substrate is the landlord. Apple positioned itself entirely on the slow-and-appreciating side and let someone else carry the training treadmill.

Same bet, one floor up

We run the same position at one-laptop scale, and it predates the keynote. The squad's brains are tenants: the coder, the verifier, and the scout have each been swapped this month — bake-off, measure, replace, same harness (the roster carries current residents). The metal is owned: one Mac whose unified memory holds a 30B-class roster, a consumer GPU for image work, and a 20-watt Jetson seat running a quantized model at roughly 0.8 J/token. And before flash-residency was a keynote slide, the head of this squad — an 81 GB model — was being served from SSD by a custom C engine on the same laptop, because DRAM is for the working set, not the whole brain.

Where our floor differs from Apple's: what this harness accumulates is not users, it is record — savepoints, memory, and the plan→code→verify transcripts of every cycle the squad runs. The substrate is what makes any given brain "the same crew" after a swap. That layer — persistence — is the one thing this week's keynote does not ship, and it is the layer we think the work actually lives in.

What we are not claiming

In the spirit of the verifier note: no, this week does not prove our curve. The verifier suite is being re-baselined before its verdicts count as more than advisory; the local-share-over-time series has an instrument but no data yet; the training loop is built and currently idle while we retarget it. Apple validating the architecture does not validate our results. It validates the direction — and the direction was checkable from public parts lists all along.

We did not position for this wave, and we are not going to start steering by it. The plan here does not change: do good work, make cool stuff, publish the receipts, see what happens.