The Resurgence of Decentralized AI | Roundup

A year ago all three hosts were fading anything related to decentralized AI or the crypto-AI intersection, but their view has shifted as the limits of large frontier models have begun to show cracks. Miles argues the era of token-maximizing state-of-the-art models is over because harnesses have gotten much better at smart routing toward an efficient frontier of minimum viable cost per task, and usage is now spreading across more models than even two or three months ago. Frontier labs spend roughly one hundred billion dollars to train a state-of-the-art model and expect one hundred fifty to two hundred billion in return revenue, but net new revenue from each successive state-of-the-art model will decline over time as users route to cheaper models for lower-complexity tasks.

Anthropic launched what was internally called Mythos as Fable with significant guardrails, but a team of supposedly Amazon researchers jailbroke it quickly. Anthropic heavily restricted Fable use cases related to biology, cybersecurity, and training other models, and was nerfing responses without indicating to users that responses were being downgraded. Fable is unavailable in Canada due to export controls, and Anthropic charges approximately two hundred dollars per month for a subscription tier. Anthropic also announced it would begin collecting user data, raising the concern that healthcare providers using Claude who are subject to HIPAA may now be breaking the law. The hosts suggest Anthropic may be trying to gain regulated incumbency and find ways to monetize beyond compute subscription fees.

Miles draws a distinction between decentralized inference and decentralized training, remaining bearish on decentralized inference because it will always perform worse than centralized inference, while finding decentralized training increasingly interesting from both platform risk and cost perspectives. Jensen has publicly stated the need to eventually crack decentralized training, a comment first heard approximately a year and a half ago. Distributed compute resources sitting idle in university labs worldwide are cited as a potential supply-side input, and distributing training costs is argued to make training significantly cheaper than relying on centralized labs. Projects including Prime Intellect, Pluralis, and Hermes have been training models for approximately two years, all notably including the word research in their titles, reflecting a research-to-production maturation.

The two main drivers of interest in open source AI models are cost reduction and elimination of platform risk, with platform risk at the business level argued to be the primary driver of adoption. Microsoft owned approximately five percent of Anthropic and then turned off Claude access, cited as a concrete example of platform risk materializing. Open source models such as Alibaba and Qwen currently have better capability than decentralized models, but open source models can still censor users, relicense weights, or restrict supply-side access, making them not fully censorship-resistant. Recent government action is said to have increased the importance of decentralized AI, with the argument that in a world where governments constrain frontier labs, decentralized AI becomes much more important.

Venice, founded by Eric Hughes, is described as probably the most talked about crypto project at the moment and currently the only way to get token exposure to something with real usage at the crypto and AI intersection. Venice operates with client-side processing so user information does not go server side, enabling full privacy, and gives users access to open source models that reduce censorship and remove user identity from queries. Its usage has exceeded initial expectations, and it is argued to be probably the best performing consumer crypto application on the consumer side currently. Venice uses only a few simple crypto mechanics while functioning as a regular app with crypto on the back end, and the prediction is made that many future best apps in crypto will follow this pattern rather than being native DeFi apps.

Hardcore agent engineers are already using open source models for execution tasks and only using closed source frontier models for evaluation or planning. PayPal has taken open-sourced model weights and is training a finance and payments specific foundational model on proprietary transaction data, illustrating how large corporations open source models strategically to commoditize their complements. The bottleneck for AI adoption is identified as creativity around finding the right form factor for different use cases rather than compute availability, with many software companies including Notion and HubSpot described as having naively added AI agents without meaningful differentiation from simply using Claude directly.

The current moment in AI is compared to the protocol thesis era of crypto, analogous to when block space got commoditized before the fat app thesis emerged. A key difference is that OpenAI, unlike Ethereum or Bitcoin, is capable of building applications itself, meaning developers face the risk that the platform will turn their entire product into a feature. Ethereum fees collapsed by roughly two orders of magnitude following scaling improvements, and it is described as near-consensus that Ethereum cannot generate sustainable fees, with Jevons paradox cited as a reason demand should eventually recover but the deflationary impact on prices before that demand arrives identified as a real stumbling block. Decentralized AI is expected to be a major theme for the next crypto cycle, with wearables identified as a likely coming catalyst that will increase demand for local AI models running on-device, and quarter four identified as the likely period for crypto markets to begin recovering.