Podcast Atlas
Source note Episode guide Original audio Tags: All-In, Ai, Ai-Infrastructure, Open-Source-Ai, Agi, Multimodal-Ai, Generative-Media, Robotics

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

Summary

This All-In episode pairs Andrew Feldman of Cerebras with Robin Rombach of Black Forest Labs to connect AI infrastructure, inference economics, open models, release safety, multimodal generation, and robotics. Feldman frames reasoning as token-heavy inference, argues demand is already ahead of data-center and chip supply, and presents open-source model availability as part of model sovereignty. Rombach explains latent diffusion, Stable Diffusion, Flux, visual control layers, and why the same multimodal systems used for images and video could eventually support world models and robots. The episode is highly optimistic and founder-facing, so claims about backlog, AGI thresholds, film budgets, and cyber testing are kept source-scoped.

Key Claims

  • Feldman says AI demand is already exceeding the industry’s ability to build power-hungry data centers and fill them with accelerators, naming OpenAI, Anthropic, Google, Microsoft, AWS, and xAI as major capacity seekers.
  • Feldman says Cerebras has a $25 billion backlog; the wiki preserves that as a source-scoped company claim.
  • The episode treats reasoning as inference because long-running reasoning systems consume many visible and internal tokens, making speed and latency part of answer quality.
  • Jason’s “unlimited tokens as unlimited reasoning” frame extends Token Maxxing toward Unlimited Token Workflow and loop maxxing, where repeated AI passes can improve answers but also increase review burden.
  • Feldman argues that enterprises will route workloads across frontier, cheaper, and open-source models, extending Model Routing Cost Control from user behavior into enterprise procurement.
  • Custom chips are presented as both performance tools and bargaining/sovereignty tools: AI companies and hyperscalers do not want total dependence on one hardware supplier.
  • The open-model discussion names Kimi, Qwen, GLM 5.2, Llama, and customer-specific models, and argues that the United States needs more domestic open-source models so users outside China have more choice.
  • Feldman says staged rollout and government red teaming can be reasonable when a model becomes powerful enough to create meaningful cyber risk, qualifying Frontier Model Release Governance without calling for a blanket pause.
  • The Palo Alto Networks example is used to argue that AI-enabled vulnerability discovery can be useful defensively while still proving that powerful models create real security risk.
  • Feldman says AI has passed older AGI benchmarks such as the Turing test, while the conversation implies that AGI definitions move as each prior benchmark is met.
  • The first half links AI to abundance through cancer, energy, calories, education, housing, and tutoring, while acknowledging job dislocation.
  • Rombach describes Black Forest Labs as a company rooted in latent diffusion, Stable Diffusion, and Flux.
  • Rombach says future visual models are moving from image/video/audio generation toward action prediction, where one model can generate media and eventually help robots choose actions.
  • The discussion of generative media control layers frames better tools as moving beyond slot-machine outputs into text, image, multi-reference, and video manipulation.
  • Martin Scorsese’s use of Black Forest Labs tools is framed as visual ideation: getting a mental image out of a director’s head, not replacing the director.
  • The episode treats AI video as useful for storyboarding, startup launch videos, and some production scenery, but Rombach says high-end film remains one of the hardest delivery environments.
  • IP-controlled generative models are presented as a possible route for studios and rights holders: public tools can block certain IP while partner models can enable licensed creation.
  • Rombach says Black Forest Labs has crossed 100 people and is hiring for large-scale model training, diffusion, flow matching, customer engineering, physical AI, IP partnerships, and infrastructure.

Key Quotes

“token maxing” - the episode’s phrase for heavy AI inference consumption.

“unlimited reasoning” - Jason’s shorthand for abundant tokens in long reasoning workflows.

“AI models are a medium” - Rombach’s framing of filmmaker use.

Connections

Contradictions

  • No direct contradiction found.
  • The episode’s “AGI is here” claim is definitional rather than a settled wiki position. It qualifies AGI Narrative and AGI Three Acts by saying older benchmarks may have been met, while other pages still reserve stronger AGI claims for coding-agent, automated-research, physical-world, or broad autonomy milestones.
  • The source’s backlog, cyber-testing, film-budget, and model-performance claims are kept source-scoped because the episode is an interview format rather than an independently verified report.