Taking Reinforcement Learning Cross Datacenter — Nan Jiang, Modal

Nan Jiang from Modal presents a technique for running reinforcement learning (AI training through trial-and-error) on GPU servers distributed across multiple continents.

The problem is that checkpoints — saved versions of a trained AI model — weigh around 500 GB and take hours to transfer between data centers, which prevents rapid model updates. His solution is to send only around 500 MB instead: a small patch containing only the most important changes. This works because less than 1% of the weights actually change between versions — not because gradients are sparse, but because updates are very small, often a thousand times smaller than the precision allows. Modal calls the implementation Stitch.

More from AI Engineer

Inside Arca's Long-Horizon Agents That Run the Business - Aman Singh, Arca

AI Engineer · 23 Oct

Real AI Agent Attacks and What Would Have Stopped Them — Andrey Rybka, Bloomberg

AI Engineer · 23 Oct

Outside of tech there has been no financial payoff to US companies for using AI, now viewed as a margin-hurting cost center, as VC-backed firms burn more tokens

Techmeme · 4 h ago

Source: Replit recently moved most AI spend from Anthropic to OpenAI, which gave 30%+ off in a deal worth $100M+/year, after Anthropic discontinued its discount

Techmeme · 6 h ago