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

Summary

The episode opens with Andrew Feldman of Cerebras describing the AI buildout as an unprecedented physical and industrial mobilization, with data centers consuming city-scale power and demand already locked in via a reported $25 billion backlog. The discussion centers on reasoning models and why inference, not just training, is becoming the core bottleneck: longer internal token budgets and faster chips make deeper reasoning tractable, and Feldman argues Cerebras’ architecture is designed to accelerate that class of work. The hosts then pivot to open source, sovereignty, and regulated-industry deployment, making the case that many enterprises want domestic, on-prem, and customer-controlled model options. In the second half, Robin Rombach explains Black Forest Labs’ latent diffusion lineage, its move toward multimodal models that combine images, video, audio, and action prediction, and why human-in-the-loop creative iteration is the most important near-term use case. The episode ends by framing current systems as already meeting older AGI definitions, while emphasizing safety, staged rollout, red-teaming, and the likelihood of future data leaks as planning assumptions.

Notes

• Chamath frames Cerebras as an inference-chip company in the race for superintelligence.
• The AI buildout is likened to a historic mobilization of capital and talent.
• Andrew says AI infrastructure is becoming physically enormous.
• He says individual buildings will draw more power than mid-sized cities.
• Chamath says the biggest AI buyers are still insatiable.
• Andrew says demand is already booked and Cerebras has a $25 billion backlog.
• Google, Microsoft, AWS, OpenAI, and Anthropic are all asking for more data centers.
• He emphasizes that customers are chasing booked demand, not speculative demand.
• Early AI use is still experimental, but the net value is already substantial.
• Chamath says strong AI users are emerging with systems-thinking instincts.
• He says models now push users to clarify goals and requirements.
• Prompting has shifted from brittle text matching toward intent understanding.
• Newer models can infer structure the user did not explicitly ask for.
• Chamath describes a jump from summarization to solution generation.
• Chamath uses a trend-scouting workflow to show repeated model use.
• Andrew identifies that workflow as reasoning in action.
• Chamath’s premise is that unlimited tokens imply unlimited reasoning.
• Andrew says long inference runs can produce better answers.
• Reasoning is compute-heavy because it burns internal tokens.
• Feldman says Cerebras systems are so expensive he jokes about bringing one to interviews.
• He claims Cerebras broke from the old Moore’s-law-like processor curve.
• He expects the company to be well above 2x in the next 18 months.
• New architectures have room to outperform older GPU-era designs.
• The hosts focus on how hard it is to run a company with $25 billion of demand.
• Andrew says keeping up with OpenAI is especially difficult.
• Hyperscalers and chip companies are trying to reduce dependency through custom silicon and neoclouds.
• He says customers do not need the fastest chip, only less dependence on someone else’s chip.
• Chamath says open-source reasoning quality has caught up in practice.
• Andrew compares frontier models to Ferrari, and routine work to a minivan.
• Many enterprise tasks only need dependable open-source capability, not gold-medal intelligence.
• Regulated industries want on-prem deployment because of leakage and sovereignty concerns.
• Customers want domestic open-source options with more control.
• Andrew says the U.S. needs more domestic open-source models.
• He says the world’s open-source choices are currently limited.
• Cerebras runs open, closed, and customer-owned models.
• Sovereignty is becoming a major market trend.
• Andrew says staged rollout for powerful models is reasonable.
• He compares AI safety review to pharmaceutical regulation and red teaming.
• Fast chips make guardrails less painful.
• A security leader found tens of critical issues in an hour, underscoring AI’s power.
• Andrew says a massive future data leak is inevitable.
• He says even unknown unknowns are useful to acknowledge.
• Chamath says AI should help humans notice what they are not asking.
• He describes prompting AI to critique the prompt and ask follow-up questions.
• A good assistant should broaden the user’s question set.
• Chamath says AGI has already been reached under older definitions.
• Andrew agrees that earlier AGI definitions have already been surpassed.
• He says systems have blown past older Turing-test-era expectations.
• Andrew says technologists are now seeing old warnings validated.
• Chamath says the tools are starting to build themselves recursively.
• He says trend-scouting surfaced the term loop maxing.
• Andrew defines loop maxing as recursive learning with exponential gains.
• He says the endpoint of this exponential curve is unknown.
• He argues many future bottlenecks will be people and coordination problems.
• World models may yield behavioral insight by watching human behavior.
• He says AI may compress learning across the equivalent of thousands of generations.
• Andrew says AI could help prevent cancer, though there will be economic dislocation.
• Robin Rombach says latent diffusion underlies image, video, and physical AI.
• Black Forest Labs is moving toward multimodal models with action prediction.
• Video pretraining may help models learn real-world physics.
• The product strategy is to expose more manipulation layers for users and developers.
• Black Forest Labs views models as a medium, not a prescribed workflow.
• The Scorsese demo centered on iterating from a mental picture to visuals.
• Visual media is richer and less lossy than language.
• The best near-term use case is human-in-the-loop iteration, not autonomous movie generation.
• The same model family could support film, robotics, and possibly computer use.
• Action prediction is framed as prediction, simulation, and generation.
• Current robotics deployments may need only a few hours of fine-tuning data.
• Public tools block certain IP generation, while custom models are built with IP holders.
• Black Forest Labs has crossed 100 employees and is hiring in Germany and San Francisco.
• The company is hiring across research, customer deployment, and infrastructure.