FILE 00117KSomeone on X pointed out that Covenant-72b can't count the R's in strawberry. They're right. But so were the people who laughed at GPT-4 for the same mistake two years before it started passing the bar exam. The interesting question was never whether the model fails. It's why, what that reveals about intelligence, and what happens next.
FILE 00229KPre-training gives AI knowledge. Post-training teaches it judgment: what to refuse, how to reason, what to value. This is the phase where alignment happens, and while decentralized efforts existed, weight sync over public internet made them impractically slow. This week, a research paper from Grail demonstrated that the bandwidth barrier keeping RL post-training centralized was 99% redundant, an artifact of how we were moving data rather than a physical constraint. The implications extend far beyond compression ratios.
FILE 00329KFor eighty years, the most powerful technologies have required concentration: co-located machines in fortress datacenters, tightly controlled by those who could afford the infrastructure. This week's research breakthrough from Templar marks something different, a technical path toward intelligence as genuinely distributed public infrastructure, where your home GPU can train frontier models alongside Google's datacenters.
FILE 00410KWhy a technical breakthrough in gradient compression could reshape who controls the future of artificial intelligence
FILE 0057KMeet the cryptopunk miners of Templar training AI models on their gaming rigs, challenging the assumption that only trillion-dollar companies can build artificial intelligence.
