Global Decentralized RL: Pluralis Research Pioneers Post-Training via 14-Mac Cluster Across 4 Countries
Event Summary
Pluralis Research has unveiled the first successful reinforcement learning (RL) post-training experiment conducted entirely over the public internet using a distributed cluster of consumer-grade Macs. By deploying 14 Macs across 4 countries for sampling (utilizing the MLX framework and int8 quantization) and a single B200 GPU on a different continent for centralized training, the project demonstrates a viable path for large-scale RL using heterogeneous, geographically dispersed hardware.
- ▶ Decoupling Sampling from Gradients: The experiment proves that the rollout phase of RL, which is notoriously inference-heavy, can be effectively offloaded to edge devices, reserving high-end GPUs for the compute-intensive gradient updates.
- ▶ MLX as a Production Catalyst: Apple’s MLX framework is no longer just for local experimentation; its high memory bandwidth makes Mac silicon a formidable competitor for distributed inference tasks in a production RL pipeline.
- ▶ Infrastructure Democratization: This setup lowers the barrier to entry for advanced RL research, shifting the focus from “GPU hoarding” to “intelligent orchestration” of existing consumer assets.
Bagua Insight
This is a paradigm shift from “GPU-Rich” vs. “GPU-Poor” to “Orchestration-Rich.” The real breakthrough here isn’t just the hardware, but the validation of asynchronous distributed sampling over high-latency public networks. In the RL loop, sampling is embarrassingly parallel; Pluralis exploited this by treating global Macs as a massive, elastic buffer for the B200. This architecture effectively bypasses the physical constraints of a single data center. It signals the rise of a “DePIN” (Decentralized Physical Infrastructure Networks) approach to AI training, where the bottleneck is no longer the number of H100s you own, but how efficiently you can harvest idle TFLOPS from the edge.
Actionable Advice
1. Hybrid Compute Strategy: Startups should pivot toward a hybrid model—leveraging localized “sampling farms” (Mac Studios/Mac Minis) to feed centralized training nodes, significantly cutting cloud OpEx. 2. Optimize for Quantized Rollouts: Implement int8 or lower precision for sampling phases to maximize throughput without sacrificing the final model’s convergence stability. 3. Monitor Distributed Orchestration Tools: Keep a close watch on frameworks that manage state synchronization across high-latency nodes, as this will be the critical “glue” for the next generation of decentralized GenAI development.