See there are two statements in your comment.
We need decentralized AI, there is no doubt about that. It is an important way to keep the power of AI democratized. But I don't believe that decentralized AI has potential to out compete the corporate models.
Let's for example look at the one of the companies mentioned in the tweet; NousResearch.
Their models are just fine-tunes of corporate models, they do not train models from scratch which is the most expensive part of training a model and that is hardly decentralized. I would say that you can even call these corporate models as well.
I believe there were one project that trained a model decentralized way from scratch, I can't recall its name right now. But the model was not impressive to say the least.
The best open-weight models are still corporate models. Alibaba, Xiaomi, ByteDance, DeepSeek is backed by a Chinese fund called High-Flyer.
Decentralized AI, is a pipe-dream. Sure there are lots of GPUs that people have, that people can contribute to the training. Problem is, how do you convince them to pay for the electricity. Let me put it this way, if we yeeted all the market cap of Hive into training a decentralized AI. We might have an AI with ~20-30B parameters and a middling quality.
RE: ArcHive: Jake Brukhman — Creator Content Thread (Volume 1)