Groundwork
Notes from building AI you own. Learned routing, on-device inference and training, a fail-closed datapath for decentralized training, and an honest account of what verification does and does not prove. Each note names its evidence state, the failures ship with the results, and links go only to code you can clone today.
Capability you own: the route, distill, train loop
Why the painkiller is no longer a better model but the open infrastructure that lets a person own the capability, run it on hardware they hold, and keep it working offline.
Democratized distributed training
Open-weights models trained by cohorts of hardware people already own: single Macs, Thunderbolt-linked islands, CUDA islands. The endgame, the trust plane that makes it open, and why the wire is the constraint.
Two Macs, one adapter
Two consumer machines trained, exchanged, and merged real LoRA adapters to byte-identical files across three rounds. What the run showed, what it cost, and the failures that ship with it.
A fail-closed datapath for decentralized training
Moving weight fragments over a peer-to-peer link with integrity asserted at every hop: what is proven on loopback, what a two-machine run has since measured, and what a node-loss demo still has to show.
What a training receipt does and does not prove
Reproducibility evidence and deterministic replay are real and useful. They are not hardware attestation and do not prove that no hidden step occurred. Drawing that line is the point.
Load-bearing claims, measured separately
The public ledger of the research program: what is measured, what is analyzed with measurement pending, what is pre-registered and not yet run, and the rules that keep it honest.
What we ship today
A maturity-labeled snapshot of the stack: what is shipped, what is a launch surface, and what is still in build.