Hi,

This is Plain Strata, the Thursday Layer.

In late 2025, the team with the best claim to dislike centralized AI did the centralized thing. Prime Intellect built its name training models across homes and offices on different continents, no single room, no single owner. Then, for its strongest model yet, a 106-billion-parameter system called INTELLECT-3, they put every chip in one building and trained it the old way.

They did not lose their nerve. They read a number off a page. Every training step, thousands of processors have to share their updates with each other, and the size of that update scales with the size of the model. Inside a cluster, a dedicated cable moves it in seconds. Across ordinary home internet, the same update can take more than five hours to cross, once per step, for a run that takes hundreds of thousands of steps. That ratio, sharing time over doing time, has a name: the communication wall.

There are four real tricks for holding that ratio down, and this episode walks through all of them: sync less often, send less each time, send a different thing, and move the slow part off the link entirely. Each one works, up to a point. The turn is what happens at the frontier, where the tricks and the wall collide.

One question to carry out of this: is the wall a permanent law, priced into cables, or a temporary price that bandwidth and better algorithms will eventually erase?

Listen: Spotify: https://open.spotify.com/episode/6D4gWUNLUpQ2hwyYyig3JZ?si=q_k2ZolVSkSVBAKNx8U3gQ Apple Podcasts: https://podcasts.apple.com/kg/podcast/plain-strata/id6783455764?i=1000775197926 YouTube: https://www.youtube.com/watch?v=4n4NGu4ip8o

The two voices are AI. The research and writing are mine.

Decentralized AI, layer by layer. Dastan