Decentralized training's unlock is open biases
Why Decentralized AI Needs Cosmos: Greg Osuri of Akash Explains (The Interop)
“The main thing for me for decentralized training is open biases… open biases means that everybody gets to contribute to reinforcing the model, and so therefore the biases are, in theory, supposed to be more balanced.” — 00:29:54
Context: Citing early ChatGPT and Google Gemini bias failures from RLHF done by ideologically homogeneous SF teams; “at least [it’s] a free market that determines… right now we don’t know who trained them” (the definition continues at [00:31:18]-[00:32:00]).