Open Source Wins, AGI Is Here, and Scorsese's AI Toolkit with CEOs of Cerebras & Black Forest Labs

Episode Summary

Executive Summary: The transcript spans two major conversations: first, Andrew Feldman of Cerebras describes the unprecedented global AI infrastructure buildout, the shift from token-maxing experimentation to productive reasoning, the importance of fast inference chips, and the case for open source, sovereignty, and thoughtful AI governance. Second, Robin Rombach of Black Forest Labs explains how multimodal generative models are evolving from image generation toward video, action prediction, robotics, and interactive creative workflows, especially for film, IP holders, and fan-created content.

Main Topics: Global AI infrastructure buildout (Priority: 5/5): Feldman argues AI is driving a historically unprecedented physical and financial mobilization, with data centers, power demand, and chip purchases scaling across the U.S., Europe, and other regions. Inference, reasoning, and recursive intelligence (Priority: 5/5): The discussion centers on the transition from simple prompt-response models to reasoning systems that spend internal compute on better answers, making fast inference chips strategically important. Open source, sovereignty, and model choice (Priority: 4/5): The speakers emphasize that enterprises and governments increasingly want domestic, controllable, and regulated AI options alongside frontier closed models. AI governance, safety, and staged deployment (Priority: 4/5): They debate whether powerful models should be rolled out in steps, with red-teaming and security review, and whether polarization is obscuring reasonable policy. Superintelligence, AGI, and recursive learning (Priority: 4/5): Feldman frames AGI as already effectively reached by older definitions, and both speakers discuss recursive loops, exponential improvement, and the road to superintelligence. Black Forest Labs’ multimodal and creative AI roadmap (Priority: 5/5): Rombach describes the company’s work on image, video, audio, and action-prediction models that can support filmmaking, storyboarding, and eventually robotics. IP, film production, and fan creativity (Priority: 4/5): The conversation covers how AI is changing content creation economics for studios and enabling new fan-driven and licensed uses of copyrighted worlds like Disney and Star Wars.

Key Arguments: AI demand is far ahead of supply; customers are ordering compute before hardware is finished, making the buildout demand-led rather than speculative. Cerebras’ speed advantage matters because reasoning models consume large internal token budgets, so faster inference turns longer reasoning into practical products. Some AI experimentation will be wasteful, but the net value is still enormous, similar to the early cloud and Costco adoption curves. Enterprises do not want unlimited open-loop token spending; they are learning to run AI like a business with model routing and task-specific allocation. Open source is becoming essential for ordinary, regulated, and sovereignty-sensitive workloads, while frontier closed models remain best for the hardest tasks. Governments asking for staged releases, red-teaming, and patch time for unusually powerful models is presented as reasonable rather than anti-innovation. AGI is effectively already here by many older definitions, but full deployment and societal adaptation are still incomplete. Recursive loops in AI create compounding gains; asking, checking, and re-running jobs can move results from incremental improvement to dramatically better outputs. Black Forest Labs sees generative models as a medium, not a fixed product, and wants humans in the loop for iterative creative workflows. Multimodal models are converging toward world models that can generate, understand, and eventually act in the real world via robotics or computer use. AI will significantly lower the cost of high-end video production, storyboarding, launch videos, and fan films while opening new licensing models for IP owners. The most valuable consumer and enterprise products may combine open source flexibility with proprietary frontier capabilities depending on the task and sensitivity of the data.

Data Points: Cerebras backlog: $25 billion - Feldman says the company has a very large demand backlog for compute capacity. AI infrastructure power demand: More power in a few years than the previous 50 years on Earth took - Used to illustrate the scale of upcoming data-center electricity usage. Building size: Football-field-sized individual buildings - Description of the physical scale of AI data centers. Team size at Black Forest Labs: 100+ people - Rombach says the company just crossed this hiring milestone. Black Forest Labs founding timeline: 2 years ago - The company was started by Rombach and his co-founders about two years prior. Early model resolution: 64 x 64 pixels - Rombach contrasts the earliest image models with current high-resolution multimodal systems. Video duration capability: Multi-minute videos - Current generation can now produce much longer videos than the earliest models. Expected chip performance trajectory: Way over 2x in the next 18 months - Feldman says Cerebras expects to improve beyond a simple doubling trend. Historical processor trend: Doubling about every 18 months - Feldman references traditional Moore’s-law-like scaling before Cerebras broke it. Generation time horizon: 50 to 70 to 100 years - Used in the example of cathedral-like construction projects spanning multiple generations. Human learning cadence: 15 to 20 years - Feldman compares human generational learning speed to longer historical cycles. Startup launch video cost reduction: $100,000 to $250,000 down to weeks of work - Rombach and the interviewer discuss AI lowering production costs dramatically. Film budget example: $30 million vs $150 million - A Bitcoin movie reportedly used AI-generated scenery and would have cost much more traditionally.

Pivotal Quotes: "There is clearly massive value happening." — Andrew Feldman: On whether the current AI boom is real value creation or just token-maxing experimentation. "Unlimited tokens, I believe, means unlimited reasoning." — Interviewee/host: During a discussion about how more compute enables deeper model reasoning over long runs. "AGI, I think, I suspect you'll agree with me that we've hit it." — Andrew Feldman: On the claim that by older definitions, artificial general intelligence has effectively already arrived.

Implications: The AI race is shifting from model demos to industrial-scale infrastructure, governance, and productization. Winners will combine fast inference, model choice, safety, and workflow integration while creative industries rapidly adopt AI as a production medium.