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

What It Covers

Cerebras CEO Andrew Feldman and Black Forest Labs CEO Robin Rombach join the All-In podcast to discuss the unprecedented scale of AI infrastructure buildout, the closing gap between open-source and frontier models, AGI arrival, and how generative video tools are entering real film production with directors like Martin Scorsese.

Key Questions Answered

• •AI Infrastructure Demand: Cerebras carries a $25 billion backlog, and demand still outpaces supply. Data centers now being built across Kazakhstan, Armenia, The Nordics, and The Middle East consume more power than mid-sized cities. Buyers including OpenAI, Anthropic, Google, and AWS are booking capacity before construction finishes — a demand-first dynamic that is historically unprecedented in technology infrastructure deployment.
• •Token Efficiency Strategy: Enterprises are shifting from unlimited token consumption toward strategic model routing — using frontier models like Claude or GPT for complex reasoning tasks while deploying cheaper open-source models for routine operations like data formatting or document processing. This mirrors how early AWS users learned to stop expensing every workload to the cloud and instead match compute cost to task complexity.
• •Reasoning Models Change Prompting: Modern reasoning models now interpret user intent rather than requiring precise prompt engineering. Running a reasoning model for 24–48 hours on a single task produces outputs equivalent to weeks of human analysis. Cerebras hardware, running 15x faster than standard inference chips, multiplies this effect — making extended reasoning loops practically viable for enterprise and research applications.
• •Open-Source Sovereignty Trend: Regulated industries in finance and healthcare are selecting on-premise open-source models over frontier closed-source alternatives to avoid data leakage and maintain compliance. Feldman argues the US needs more domestic open-source models beyond Meta's OSS 120B, since current alternatives are primarily Chinese models — creating a geopolitical gap in AI sovereignty that no single vendor has filled.
• •Latent Diffusion as Universal Foundation: Black Forest Labs' core algorithm — latent diffusion, originally developed during Rombach's PhD — compresses images, video, and audio into efficient representations for transformer training. This same architecture now underpins physical AI: a model trained to generate video implicitly learns world physics, enabling action prediction for robotics with only a few hours of task-specific fine-tuning data.
• •Generative AI in Film Production: A Bitcoin-themed feature film starring Gal Gadot replaced physical sets with generative AI backgrounds on a soundstage, reducing production cost from an estimated $150 million to $30 million — making the project commercially viable. Rombach describes Scorsese using Black Forest Labs tools to externalize visual concepts from pre-production scenes, treating the model as a storyboarding and ideation medium rather than a final output generator.