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

The conversation opens with a discussion of the unprecedented scale of AI infrastructure buildout currently underway globally. Feldman explains that data centers being constructed in the US, Canada, Nordics, Europe, Middle East, and Central Asia will consume more power in coming years than humanity used over the previous 50 years. Individual buildings are the size of football fields with power consumption exceeding midsize cities. Cerebras has a $25 billion backlog, and companies like OpenAI, Anthropic, Google, and Microsoft are insatiable in their demand, pre-ordering chips before production is complete. This represents a fundamental shift where demand far exceeds supply.

The hosts discuss token maxing—the overconsumption of AI tokens—and whether this represents genuine value creation or wasteful experimentation. Feldman argues that while some experimentation is wasteful, the net value is enormous, comparing it to early AWS adoption where engineers were given unlimited cloud credits. He notes enterprises are now moving toward strategic token allocation, with smarter personnel understanding systems-level deployment emerging as a critical skill. The conversation highlights how AI is transitioning from being a tool that requires precise prompting to one that understands user intent and suggests improvements the user didn't explicitly request.

A major focus is on reasoning models and inference. The hosts discuss Hermes agents and GLM-52 models, noting how unlimited compute enables extended reasoning chains. Feldman explains that Cerebras's inference chips are optimized for these reasoning workloads, which consume massive numbers of tokens internally. He notes Cerebras has broken Moore's Law, with plans to achieve more than 2x improvement in the next 18 months, moving away from traditional GPU architectures with their geometric limitations.

The conversation addresses open-source models versus frontier models. The hosts argue there's room for both: frontier models (OpenAI, Anthropic, Gemini) for hard problems, and open-source models for routine tasks. Feldman supports open-source development from a sovereignty and dependency perspective, noting companies want control over their infrastructure after historical lessons with Intel and GPU manufacturers. Concerns about data sovereignty and regulatory compliance in regulated industries (finance, healthcare) are driving adoption of on-premises open-source solutions.

On the Grok/reasoning model rollout debate, Feldman takes a nuanced stance. While noting the government's request for staged deployment testing before release isn't unreasonable—comparing it to pharmaceutical testing and cybersecurity red-teaming—he critiques the politicization of the decision. He argues that technical concerns about cyber threats are legitimate, but polarization prevents clear thinking. He emphasizes that both sides will make smart and dumb decisions, and the government's rank-and-file teams are trying hard to navigate unprecedented territory.

Rombach discusses Black Forest Labs' work on multimodal generative models spanning image, video, audio, and action prediction. He emphasizes the paradigm shift from simple text-to-image systems to complex multimodal models. The key innovation is latent diffusion—compressing natural data into efficient representations, then training transformers on those compressed spaces. This enables efficient generation across modalities.

A significant portion covers the Martin Scorsese partnership. Rather than proposing Scorsese use Black Forest's tools to create full films, Rombach positions them as a brainstorming and visualization medium. Scorsese can externalize his mental picture of a scene, iterate on visual outputs, and communicate his vision more effectively than through language alone. Rombach rejects the notion of a predetermined "correct" use case, emphasizing the model as a medium that different creators will deploy differently.

Rombach discusses practical applications beyond filmmaker use: startups using the tools for launch videos, the Bitcoin movie that eliminated green screens by generating scenery on soundstages, and the trajectory toward robotics. He notes the technology is rapidly improving in resolution and temporal coherence, unlocking more demanding production use cases. The conversation emphasizes that current video generation still requires human-in-the-loop iteration for best results.

On IP and content libraries, Rombach explains Black Forest restricts generation of certain trademarked IP in public tools but partners with IP holders to develop customized models. He suggests the most interesting future involves interactive content creation tools and enabling fan creators to develop new stories within established universes (referencing George Lucas's permission model for fan films). Licensing frameworks could allow consumers to customize creative experiences while compensating rights holders.

The hosts discuss both the economic dislocation and abundance created by AI. Feldman acknowledges some workers will face disruption (comparing it to carriages and horses), but emphasizes the pro side of the ledger: curing cancer, unlimited energy, unlimited education, and personalized AI tutoring (revisiting Socratic methods at scale). Rombach notes that training data can come from synthetic generation, YouTube analysis, or fine-tuning on specific robot hardware, with the goal of eventually reaching prompt-based control.