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

Cerebras CEO Andrew Feldman disclosed a $25 billion backlog for the company's inference chips, with customers including OpenAI, Anthropic, xAI, Google, Microsoft, and AWS placing orders before chips were finished. The scale of the buildout is such that data centers under construction over the next several years will consume more power than the previous 50 years of global energy use combined, with individual buildings the size of football fields drawing more power than mid-sized cities. Construction is active across the US, Canada, the Nordics, France, the Middle East, Kazakhstan, Tajikistan, Georgia, and Armenia.

A single Cerebras chip costs approximately half a billion dollars to produce, and Feldman claimed the company has broken Moore's law, predicting a better than two times improvement over the traditional 18-month doubling pace within the next 18 months. He attributed this headroom to newer architectures versus the 20-year-old GPU design. Cerebras chips run 15 times faster than alternatives, meaning 24 hours of compute could yield what would otherwise take weeks or months of model reasoning time. A finding disclosed within the last six weeks is that faster chips reduce the latency penalty imposed by AI safety guardrails, which can make AI feel slower and discourage adoption.

Enterprises have moved away from unconstrained token access toward strategic routing, sending ordinary workloads to open source or cheaper models while reserving frontier models for harder problems. Regulated industries in finance and healthcare are increasingly choosing on-premise open source deployments due to data sovereignty and compliance requirements. Feldman noted that the practical domestic open source choice in the US today is essentially Meta's 120 billion parameter model or Chinese models, and argued more domestic options are needed. Cerebras currently runs GLM, Kimi, the Qwen family, OpenAI closed source models, GlaxoSmithKline proprietary models, and models for G42 and NBCU AI.

Feldman described recursive learning as producing exponential rather than marginal gains, with each iteration yielding vastly better answers, and said the slope of improvement curves is steep enough that it remains unclear where gains end as more compute is added. He attributed early recognition of this dynamic to Sam Altman, Ilya Sutskever, and Dario Amodei, who he said understood five to six years ago that recursive gains are exponential. Feldman argued AGI has effectively been reached by any definition that would have been used 20 or 30 years ago and that the Turing test has been clearly surpassed. He predicted AI-driven medical research creates a real possibility that none of our children or anyone they know will die of cancer, and argued AI agents could deliver individualized tutoring at scale, something known since Aristotle and Socrates to produce superior outcomes but never implemented in mass education across the past thousand years.

Palo Alto Networks CEO Nikesh disclosed that when the company tested a powerful AI model against its own software, the model found previously unknown bugs requiring six weeks of emergency patching. Separately, speakers noted that a massive AI-related data breach is considered inevitable even if its specific timing and form are unknown, and that AI companies are currently inventing safety and governance processes without any existing playbook.

Black Forest Labs co-founder Robin Rombach previously worked on Stable Diffusion and co-invented the Latent Diffusion algorithm, which is the foundational algorithm behind generative models used for image generation, video generation, and physical AI. The company is based in Freiburg, Germany and San Francisco, was started two years ago, has recently crossed 100 employees, and is known for the open source model Flux. Black Forest Labs is developing multimodal visual models pre-trained simultaneously on images, video, and audio, and is combining that pre-training with action prediction so a single model can generate media while also predicting actions for robotics deployment. Rombach argues that pre-training on video gives a model implicit understanding of real-world physics, enabling robotics applications to emerge from the same architecture. Current robot deployments require only a few hours of fine-tuning data, though fully in-context prompting without task-specific fine-tuning is the longer-term research goal.

Rombach met with Martin Scorsese multiple times and demonstrated Black Forest Labs models to him directly. Scorsese used the models to explore and iterate on visual scenery for a potential new film, specifically a village in Eastern Europe, as a way to communicate a mental picture of a scene. Rombach frames this as a core value proposition, arguing that language is a lossy communication medium whereas visual information carries richer signal, making generative AI a tool for externalizing creative intent rather than replacing it. The podcast host noted that startup launch videos previously costing $100,000 to $250,000 can now be produced in a week or two, and cited a Bitcoin film with a $30 million budget that used generative AI for all background scenery with no green screens or built sets, with Gal Gadot quoted as saying the same film would have cost $150 million if sets had to be physically constructed.