Compute is componentized: 11 parody GPU cards (Ex-Mining P106, GFX 1060
Dumpster, Radish VII Furnace, RTX 4090 Gigachonk, H100 'The Vault',
B200 BigChungus...) install into carrier slots — Budget PC takes one,
the milk-crate rig four, the half rack eight with rack-only datacenter
cards. Cards carry tok/watts/VRAM/heat individually; total cluster VRAM
must hold the running model version or all token output starves.
Storage devices (SSD shoebox / NVMe hoard / NAS tower) gate training:
no checkpoint space, no progress. Network grows a switch layer with
per-switch throughput caps between devices and uplinks. Shop grouped
by category. Old saves migrate with equivalent card loadouts.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Platforms adapt to your slop (ALGO HEAT): each publish raises spam-filter
wariness which throttles all view rates up to 85%; training a new model
tier slips past the filters. Verified loop: publish -> squeeze -> train ->
relief. New hardware: 20A sub-panel (+1500W), fiber uplink (1Gbps, 8
ports), Cash Furnace (burns $1.50/s for +40% views, breaker-immune).
Fix: New Game wipe no longer clobbered by the beforeunload autosave.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
