Decentralized AI Training: Can It Dismantle Centralized Powerhouses? | Crypto x AI Event
Episode publishedPodcast: The Delphi PodcastThe Delphi Podcast
Transcript preview
Preview only
Before we get started, we'd like to thank our sponsors for making this event possible. Special thanks to our platinum sponsor, Olas. Olas enables everyone to own a share of AI, specifically autonomous agent economies. We're also excited to highlight our silver sponsors, NEA, empowering decentralized applications and blockchain ecosystems. Venice AI, a private and uncensored alternative to popular AI apps. MIRROR, unified AI infrastructure secured by crypto. And Therick, the first on chain multi agent system. To learn more about all of our sponsors, check the description below and dive in. Enjo
Everyone, welcome back to Delphi's crypto AI conference. I'm Tommy. I'm a founding partner at Delphi Ventures, and I'm honored to host one of our keynote panels on decentralized for centralized training. The training of AI models is a foundational bedrock for AI itself. And today, we're gonna discuss if decentralized training can actually beat the centrally trained models of OpenAI, Anthropic, Perplexity, and others. Today, I have the best group of people I could find to be on this panel. I'm I'm honored to have them. I have Dylan from News Research. I have Ben Fielding from Jensen. I have Tra
It's great. Good.
Great. Great.
Yeah. That was good to have you guys. I'm gonna start off with, with an easy one. The first question I have is, will we have an open source AI trained model through a decentralized network that is actually competitive with centralized models on a near enough timeline that's actually useful? So will we have something to centralize an open source model that's competitive with a centralized model on any near term timeline, or is that not doable?
You are asking these are for people that are dedicating their work right now to doing just that. So I think you're not gonna have a very contentious answer here, but, like, resounding yes. I think if you look at, you know, the scale with which Nvidia has shipped h 100, if you know with which GPT 4 was trained something like $500,000,000, like, open RD Marshall resources to that scale. So there's no doubt that, you know, you could harness that and just like we're doing for years in the past.
Ben, Travis, what do you guys think? Obviously,
also a bit of a biased view, obviously, working on that stuff. Right? But I take maybe a bit of the contrarian perspective that we are maybe not completely there yet. Right? It's not, too contrarian to to say, hey. Maybe, there's still extra work needed obviously on the distributed training side. From the research perspective to actually scale to, like, the largest model sizes, obviously. Right? Assemble on the engineering side, getting the fault tolerance to work, to onramping and offramping of compute resources on a decentralized network and stuff like this. So, yeah, I don't think we are ne
Travis, I remember you also doing some work on the inference side. I know that's not exactly what you were, you know, dedicated to at this time, but what do you think about, the sort of progress that's been made lately towards maybe making a new trade off towards Inference time compute versus, you know, where we thought maybe spend all your courses to the, pretraining and training side. Seems like that's a good setup for a decentralized network considering it's like a lesser load on, interconnect.
Keep exploring with a free account.
A free account preserves the short preview. Upgrade to Pro for the full transcript and episode chat.