Chairman τao

🚨 NEW: The Green Party has drafted proposals to reduce the time police can detain terror suspects from 14 days to 4 [@genevieve_holl]

“The NHS is Failing because Billionaires are getting Richer and Richer” thanks for clearing that up Hannah !!! 🙄deary me… 🤡 @TheGreenParty

@MothinAli Thoughts on banning halal slaughter mate? Nobody from the 'Greens' seem keen to answer me on that?

🔥 Exactly. Templar changed how I think about AI infra. I didn’t expect much from decentralized AI, but seeing @tplr_ai train a 72B model on 1.1T tokens across ~70 permissionless nodes on Bittensor ( $TAO). That alone is already unusual, but what really changed my mind is how they made it work. - At this scale, training is limited by coordination. Normally you’re pushing ~280GB of data per synchronization step between nodes, which makes decentralized training basically dead on arrival. - @tplr_ai compressed that down to ~2.2GB and reduced sync frequency massively using SparseLoCo. When I look at that, I see them removing the core bottleneck that killed every previous attempt 🤯. That’s why I think calling this a DeepSeek moment is actually not exaggerated. DeepSeek showed models can be trained cheaper. Templar shows they can be trained without central coordination at all. -> Those are two very different directions, and this one feels structurally harder to compete with. Another signal I don’t ignore: when people like Anthropic’s Jack Clark publicly frame it as real infra: - In my experience, that kind of validation usually comes after something already works, not before. - This is still pre-training. The real edge in AI comes from post-training, RLHF, alignment loops, basically where models become actually useful. Templar is moving there next with Grail, and for me that’s the real test. If they can decentralize that layer too, then we’re no longer talking about decentralized compute, they’re talking about a fully permissionless AI production pipeline. What makes Templar stand out to me is the timing and direction they chose. 1/ They went after coordination when the entire AI industry is quietly hitting scaling limits. - That’s a very different bet, and usually the ones who attack constraints, not trends, are the ones that matter later. 2/ Another catalyst I see is the permissionless design. - Most decentralized AI systems still gate participation in some way, which kills network effects early. - Templar went fully open from the start, which means if this model works, it doesn’t just scale linearly, but compounds with more contributors, more experimentation, more edge cases being solved in parallel. Also, the fact they are building toward post-training (RL layer) tells me they understand where real value sits. Pre-training gets attention, but post-training is where models become usable, sticky, and monetizable. If they execute here, they start owning part of the intelligence layer itself. 3/ My prediction based on this: In the short term, most people will still underestimate it because model quality gap vs centralized labs will be the easy argument. But over time, I think Templar becomes: - a backend layer for open AI development. - a coordination network for distributed compute. - and eventually a marketplace for intelligence refinement. Not dominant overnight, but quietly embedded everywhere. And if that plays out, the upside comes from becoming the system that anyone can build on when they don’t want to rely on @OpenAI at all.

@GarethHowe95938 @vidaio_ Did I find it? 🗿🥚

On the @theallinpod this week, @chamath asked @nvidia CEO Jensen Huang about decentralized AI training, calling our Covenant-72B run "a pretty crazy technical accomplishment." One correction: it's 72 billion parameters, not four. Trained permissionlessly across 70+ contributors on commodity internet. The largest model ever pre-trained on fully decentralized infrastructure. Jensen's answer is worth hearing too.