14 Macs across 4 countries trained an AI agent via home internet
Pluralis Research ran a session for an AI agent using only 14 consumer Mac computers spread across 4 countries. These Macs handled all the 'rollouts' — the step where the agent tries actions and generates results — running low-precision int8 inference via MLX. Meanwhile, a single B200 GPU on a different continent performed the actual model weight updates in higher-precision bf16.
The two sides synced only through over ordinary home internet, with no dedicated datacenter network connection. This matters because rollout generation makes up roughly 80% of the total in agentic RL, so offloading it to cheap consumer hardware is a big deal. The hardest technical problem was that the Macs' rollouts came from a slightly outdated, int8-quantized version of the model, while the real training used the latest bf16 weights — creating a mismatch called an 'off-policy gap'.
To fix this, a technique called PULSE sent only the roughly 0.5% of weight values that changed between versions, shrinking transfers from 9GB down to about 82MB. A DPPO-style filter also discarded around 0.3% of low-confidence results to keep training stable.
Key points
- 14 consumer Macs across 4 countries generated all RL rollouts; one B200 GPU elsewhere did the actual training updates
- Synced only via over home internet, no datacenter interconnect
- PULSE technique sent only changed weight deltas (~0.5%), cutting transfer size from 9GB to ~82MB
- A DPPO-style filter dropped ~0.3% of low-confidence tokens to keep training stable
- Code is ; posted by an employee of the lab that built it (Pluralis Research)