In a two-part interview episode, Chamath Palihapitiya sits down separately with Cerebras founder Andrew Feldman and Black Forest Labs founder Robin Rombach. Feldman describes an unprecedented AI infrastructure build-out — data centers that will soon draw more power than the previous 50 years combined and a $25B Cerebras backlog — and argues the demand reflects real, enormous value despite experimentation; he champions more domestic open-source models, defends the government's staged-rollout ask on Anthropic's Fable model as reasonable, and contends that by any older definition AGI has already been reached, with recursive 'loop-maxing' gains compounding exponentially toward an abundant future (curing cancer, personalized AI tutors). Rombach explains how latent diffusion and Black Forest Labs' Flux models are converging into multimodal 'world models' that generate images, video, and audio and predict actions — the same model that can make a movie can serve as a robot's brain — and recounts working with Martin Scorsese, framing the tools as a creative medium best used with a human in the loop. Both guests land on the episode's thesis that open source has caught up to frontier models for most tasks and that AI sovereignty and control are now a defining enterprise and national trend. The conversation closes on IP and content, with Rombach favoring new creation over replicating existing IP and Chamath predicting a future where studios license their characters so fans can tell their own stories.
Feldman describes the physical scale of the AI infrastructure build-out — data centers the size of football fields drawing more power than mid-sized cities, being built worldwide — and a $25B Cerebras backlog chasing demand that is already booked, not speculative.
Chamath and Feldman debate whether the compute spend is creating value; Feldman argues the net value is enormous despite waste, likening early unconstrained token use to AWS/Costco, and describes enterprises maturing into tiered deployment that routes ordinary work to cheaper open-source models.
Chamath recounts running the Hermes agent on GLM 5.2 with unlimited compute and watching a reasoning model debate itself; the two discuss the leap from next-word prediction to understanding intent, and how unlimited tokens implies unlimited reasoning.
Feldman explains why fast inference makes token-heavy reasoning tractable, and claims Cerebras broke the traditional 18-month Moore's-law doubling, carving a new trajectory he expects to exceed 2x in the next 18 months.
Feldman argues hyperscalers build their own chips to avoid dependency and control their destiny; he says open source is having a moment, the US needs more domestic open-source models, and Cerebras already runs GLM, Kimi, Qwen, OpenAI, and sovereign customer models.
On the government's action against Anthropic's Fable model, Feldman calls a staged rollout with government red-teaming reasonable (analogous to powerful pharmaceuticals), cites Palo Alto Networks finding critical vulnerabilities, and laments how partisanship clouds clear thinking.
Chamath and Feldman agree AGI has effectively been reached by any older definition and discuss 'loop maxing' / recursive learning producing exponential, vastly-better answers, using multi-generational stonemason and Drosophila-genetics analogies for compressed learning.
Feldman argues that weighing pros and cons honestly, AI is a massive net benefit — a real shot at eliminating cancer deaths, personalized AI tutors, and abundance — with some economic dislocation comparable to cars displacing carriage makers.
Rombach introduces Black Forest Labs (Freiburg + San Francisco), recounts inventing latent diffusion behind stable diffusion, and describes converging image/video/audio generation with action prediction so a single multimodal model can also serve as a robot's brain.
Rombach explains exposing manipulation layers to defeat the 'slot machine' critique, recounts Martin Scorsese exploring the Flux models in person, frames generative AI as a creative medium best used with a human in the loop, and discusses AI-scenery film production (e.g., a Bitcoin movie).
Rombach says the same multimodal model that makes a movie can be deployed as a robot brain, discusses combining perception, prediction, and simulation, and the goal of prompting robots in context rather than task-specific fine-tuning.
Rombach advises IP owners like Disney that new creation beats replicating existing IP and that blocking certain IP generation on public tools is sensible; Chamath predicts a future of licensed, fan-created stories, citing AI Star Wars fan films.
Dated and falsifiable — full record at /api/episode/<id>.
Every position taken in the episode, attributed to its holder, with conviction (0–1) and the verbatim receipt.
# full structured record curl 'https://aipod-two.vercel.app/api/episode/all-in-2026-07-10-open-source-wins-agi-is-here-and-scorsese-s-ai-toolkit-with-' # or over MCP aipod_episode { "id": "all-in-2026-07-10-open-source-wins-agi-is-here-and-scorsese-s-ai-toolkit-with-" }