the un# podcast

The decentralized AI training section is genuinely packed with useful framing - the open-weight vs. publicly-owned model distinction, cost economics of swarm training, and the parameter-count progression are non-obvious and actionable. However, roughly the first third of the episode is standard crypto-history and DeFi-to-on-chain-finance territory that offers little new to anyone following the space.

“the big AI companies by the end of 26 will have spent about $600 billion on AI infrastructure”

“in about a week, the agents had reconstructed the same algorithm that Google kept private. And then in another week, they improved it by over 40%”

The framing of decentralized training as producing 'publicly owned' tokenized models - distinct from both proprietary and open-weight models - is a genuinely fresh construct that reframes the AI ownership debate. The broader crypto-AI convergence narrative, however, is now widely circulated, and the DeFi evolution commentary recycles familiar takes.

“we will separate the weights among different, um, kind of participants who help train the model will tokenize them. Right? And that token will give you, um, the rights to the API revenue for that model”

“there's no one entity who can like, censor the whole thing. There's no one entity who can take it offline”

Brukhman is a genuine early-stage practitioner with a real 10-year track record in crypto investing, not a career podcast guest, and his portfolio company references (Prime Intellect, Pluralis, Bagel) reflect direct investment conviction. The score is tempered because he is a VC fund manager discussing his own book throughout the AI training section without that conflict being acknowledged or challenged.

“In September of 22 we published a thesis at Coin Fund on decentralized trading”

“I remember in 2018 we became one of the largest token holders of the live peer protocol. And that's not because we bought a safe note or an equity note or even tokens from the team, but rather we helped the team mine their token supply”

The episode is notably specific by podcast standards: named companies, parameter-count timelines, ARR estimates with source caveats, and concrete dollar figures like Kirkland & Ellis's $500M model training budget all add real texture. Some figures (Anthropic at $100B ARR by year-end) are cited from analysts without scrutiny and seem aggressive, and the Google quantum-algorithm agent story lacks sourcing detail.

“it went from 10 billion parameters last year, company called Prime Intellect in our portfolio. Then it went to 40 billion. Then a couple of months ago a company called Templar on Bittensor treated a 72 billion parameter model. Last week a 100 billion parameter model was trained”

“Kirkland and Ellis, big law firm has put aside $500 million, train their own model, their own proprietary data”

The host sets up topics competently and asks useful connective questions that advance the narrative, but never once pushes back on bold or self-interested claims - Brukhman promotes multiple portfolio companies without challenge, cites aggressive revenue projections without scrutiny, and makes strong feasibility claims about decentralized training that go uncontested. The result is an informative monologue with light prompting rather than a probing interview.

“And um, you mentioned about the origins and the philosophy of um, how crypto spoke to you as a distributed, decentralized, uh, system”

“Right. It's almost like this long arc from network investing to um, to decentralized intelligence investing”