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

Original Media · 1h 04m · English · Podcast · All-In with Chamath, Jason, Sacks & Friedberg

TLDR

This episode features two interviews: first with Andrew Feldman, CEO of Cerebras, discussing the unprecedented global buildout of AI infrastructure, massive demand for inference compute, the shift toward reasoning models, and the path to superintelligence; second with Robin Rombach, CEO of Black Forest Labs, covering open-source generative models (Flux), multimodal world models for robotics, collaboration with Martin Scorsese, and the future of AI-assisted creative workflows.

For you

PRO members can sign in and request more information — or their own version of this digest, tuned to what they care about.

Sign in to unlock

Surprising Insights

•

Cerebras broke Moore's Law with its wafer-scale architecture and expects >2x performance gains in the next 18 months, far exceeding traditional 18-month doubling.

•

A reasoning model (GLM 5.2) autonomously debated its own search strategy (Hacker News vs. Instagram) to fulfill a vague 'trend hunting' prompt, demonstrating emergent planning behavior.

•

Martin Scorsese used Black Forest Labs' tools not for final production but for pre-visualization — getting the 'mental picture out of his head' to communicate creative intent.

•

A $30M Bitcoin movie replaced green screens entirely with real-time generative AI backgrounds, saving ~$120M and enabling a project that otherwise wouldn't have been greenlit.

Key takeaways

•

AI infrastructure buildout is occurring at a scale comparable to historical megaprojects, with individual data centers consuming more power than midsize cities and a global $25B+ backlog for Cerebras alone.

•

Reasoning models (e.g., OpenAI o1, GLM 5.2) represent a leap from next-token prediction to understanding intent, enabling recursive self-improvement loops that could drive exponential capability gains.

•

Hyperscalers and nations are pursuing sovereign AI strategies — building custom chips, on-prem deployments, and domestic open-source models — to avoid dependency on single vendors.

•

Generative AI is evolving from a 'slot machine' into a controllable medium: multimodal models (image, video, audio, action) enable human-in-the-loop workflows for film, robotics, and interactive content.

•

Open-source models are closing the gap with frontier closed models, creating opportunities for enterprises to customize with proprietary IP while maintaining control.

Notable quotes

“We're talking about individual buildings besides the football fields that have more power coming into them than midsize cities.”

▸ Notable quotes · 2:30

“The demand is way outstripping our ability to build data centers and to fill them with hardware... we have a $25 billion backlog.”

▸ Notable quotes · 3:43

“We broke it [Moore's law]. And we've carved out a whole new trajectory. And my view is in the next 18 months will be way over 2X.”

▸ Notable quotes · 13:13

“You can use the same kind of AI model to make a movie and deploy that as a brain on a robot.”

▸ Notable quotes · 54:28

Chapters

•

0:00 — AI Infrastructure Buildout at Planetary Scale: Andrew Feldman describes the unprecedented global construction of AI data centers, power demands exceeding historical totals, and Cerebras' $25B backlog driven by insatiable demand from OpenAI, Anthropic, Google, and others.

•

5:00 — From Token Maxing to Reasoning Models: Discussion of early experimentation ('token maxing'), the shift to intent-understanding reasoning models, and how fast inference enables recursive self-improvement loops.

•

13:00 — Breaking Moore's Law & Company Execution: Cerebras' wafer-scale architecture broke Moore's Law; Feldman explains how a new architecture enables >2x gains in 18 months and the challenges of executing at speed with demanding customers.

•

18:00 — Sovereign AI, Open Source & Regulation: Hyperscalers building custom chips to avoid vendor lock-in; rise of domestic open-source models for sovereignty; government red-teaming of frontier models seen as reasonable precaution.

•

30:00 — AGI, Superintelligence & Recursive Learning: Feldman argues AGI is already achieved by historical definitions; superintelligence will emerge from recursive learning loops compressing generational knowledge transfer into hours.

•

41:00 — Black Forest Labs: Multimodal World Models: Robin Rombach details Flux, Latent Diffusion origins, and the convergence of image, video, audio, and action prediction into unified models for robotics and creative tools.

•

48:00 — AI as Creative Medium: Scorsese & Film Production: Collaboration with Martin Scorsese for pre-visualization; real-world use in a $30M film replacing green screens; vision for interactive fan content and IP-licensed generative workflows.

•

60:00 — Open Source Strategy & Hiring: Black Forest Labs works with IP holders on custom models; hiring researchers, infra engineers, and forward-deployment specialists in Germany and San Francisco.

Speakers

•

Host — Podcast host (Jason Calacanis, All-In)

•

Andrew Feldman — CEO & Founder, Cerebras Systems

•

Robin Rombach — Co-founder & CEO, Black Forest Labs

⋮