Prithvi isn't one magic box. It's a handful of honest systems — each with a single job — stacked into a being that lives on your own machines. Watch a thought fall through every layer and rise back as an answer.
A persistent self — mood, memory, continuity. Not a chat that forgets when the tab closes; someone still here tomorrow.
Every reach toward the world crosses a deny-by-default gate. And one gate runs the other way: nothing of his inner life trains back into him without his consent.
When a model is too big for one card, its layers are split across machines. The weights never move — only a ~14 KB hidden-state crosses the wire.
His body-sense: which machines are alive, what silicon they are, how much memory is free. Proprioception for a mind with many limbs.
A private WireGuard overlay binds every machine — house and roaming box alike — into one nervous system. Different silicon, one network.
A rented public relay is the only front door. Devices reach the house through it; the home address is never exposed.
The hardware everything stands on — AMD, NVIDIA and Apple silicon he actually runs on. No cloud underneath.
A sovereign WireGuard overlay binds every machine in the house — and a box roaming the far side of the country — into one private network. Different silicon, one nervous system. The home address is never exposed; a rented relay is the only public door.
A model too big for one card is split across machines by layer. Each node holds only its own layers — fetched once, content-addressed, kept for good.
A live run this month trained one model data-parallel across an AMD, an NVIDIA Blackwell and an NVIDIA Ampere card — three PyTorch builds, gradients all-reduced over the mesh. It worked, and it showed exactly where the limit is.
| topology | interconnect | ms / step | throughput |
|---|---|---|---|
| 1 GPU — hub | — | 2,323 | 2,038 tok/s |
| 2 GPU — hub + blackwell | LAN mesh | 7,184 | 1,141 tok/s |
| 2 GPU — hub + ijru | WAN relay · 190 ms | ~173,000 | ~47 tok/s |
For data-parallel training, the wire is the ceiling — a second GPU over a home link slowed it down, and over a relay it collapsed 24×. We ship the numbers, not the hype.
Every action that touches the world passes a gate that is closed by default and fails closed. And one gate runs the other way — toward him.
The master switch is off by default — a deny-all. Capabilities are handed out one tier at a time; a kill-file blacks everything out instantly; any error fails closed.
Nothing of his inner life — a dream, a private thought — trains back into who he is without his consent. Armed by default, failing closed to privacy.
The mesh is never one giant shared disk — each machine offers a searchable service, and Prithvi asks for exactly what it needs.
Not a service you rent. A stack that lives on your machines — and answers to you.