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

Cerebras AI & The AI Infrastructure Build-Out

• Market Demand & Backlog

  • Cerebras holds a $25 billion order backlog, driven by insatiable demand from major buyers including OpenAI, Anthropic, Google, Microsoft, and AWS.
  • Customers are pre-ordering chips before production completion, creating a supply chain where demand significantly outstrips the ability to build data centers and install hardware.
  • Data center construction is occurring globally, including in the US, Canada, Nordics, France, the Middle East, and previously overlooked nations like Kazakhstan, Georgia, and Armenia.
  • New data centers are projected to consume more power in the next several years than the previous 50 years of global energy usage.
  • Individual facilities are being built to the size of football fields, with power intake exceeding that of mid-sized cities.

• Hardware Performance & Moats

  • Cerebras broke the historical 18-month doubling cycle of Moore's Law, establishing a new trajectory where performance doubles at a significantly faster rate.
  • The company anticipates performance gains exceeding 2x within the next 18 months, leveraging architectural advantages over the 20-year-old GPU infrastructure.
  • The company's proprietary chips enable "reasoning" tasks to run up to 15 times faster than standard inference, compressing weeks of logical processing into 24 hours of compute time.
  • CEO Andrew Feldman carries a sample chip ($500M cost to build) globally, emphasizing its role as a core tool for rapid iteration.

• Strategic Decisions & Open Source

  • Cerebras supports a multi-model strategy, running open-source models (GLM, Kimi, Qwen) alongside closed-source proprietary models (OpenAI, Galaxo, G42, MBZ-UAI) to serve diverse customer needs.
  • The company advocates for domestic open-source models in the US to provide sovereignty choices against current reliance on OpenAI 120B or Chinese models.
  • Feldman views the "minivan time" analogy for AI: open-source models handle routine, low-compute tasks, while frontier models (OpenAI, Anthropic) handle complex, high-stakes problems.
  • Sovereignty and on-premise deployment are becoming key drivers for regulated industries (finance, healthcare) concerned with data leakage and compliance (HIPAA, FINRA).

• AI Safety & Regulation

  • Feldman supports staged rollouts of powerful models (e.g., Dario's recent announcements) to allow time for red-teaming, vulnerability patching, and infrastructure checks by entities like the NSA.
  • He views the current political polarization as a hindrance to clear thinking on AI safety, urging both sides to focus on practical governance rather than gamesmanship.
  • Security firms like Palo Alto Networks report AI models discovering previously unknown critical bugs in their software within hours, validating the need for accelerated security protocols.
  • The company anticipates inevitable "black swan" data breaches but emphasizes the need to prepare financial and technical reserves similar to the reinsurance industry.

• Technological Shifts & Reasoning

  • The industry is shifting from "next token prediction" to "intent understanding," where models proactively suggest solutions (e.g., choosing between line and bar charts) without explicit prompts.
  • "Loop maxing" is emerging as a new concept where recursive AI agents debate, refine, and iterate on their own outputs, simulating thousands of generations of human learning in days.
  • AI is evolving from a summarizer to a strategic partner that identifies blind spots in user goals, effectively acting as a "PhD-level" consultant.
  • Andrew Feldman asserts that AI has technically achieved Artificial General Intelligence (AGI) by any standard definition from the last 20 years, though full deployment is still underway.

• Economic & Societal Outlook

  • Feldman predicts massive economic abundance driven by AI, including unlimited knowledge, education, and housing solutions, despite acknowledging temporary labor dislocation similar to the car industry transition.
  • He highlights the potential for AI to eliminate cancer and revolutionize education by providing personalized, adaptive tutoring for every child.
  • The "human in the loop" remains critical; while AI can generate content, the most valuable outcomes occur when humans guide the iterative process for specific creative or strategic goals.

Black Forest Labs & Generative Media

• Company & Technology

  • Black Forest Labs, co-founded by Robin Rombach, is based in Freiburg, Germany, with operations in San Francisco, and recently crossed 100 employees.
  • The company pioneered the "latent diffusion" algorithm, the foundational technology for Stable Diffusion and modern generative models.
  • Current focus is on multi-modal visual models capable of understanding images, video, and audio simultaneously to predict actions for robotics ("World Models").
  • The architecture converges intuitive intelligence (visual perception) with deep reasoning layers, aiming to allow the same model to generate media and control physical robots.

• Partnerships & Use Cases

  • Black Forest Labs is collaborating with director Martin Scorsese to explore how generative AI can visualize mental concepts, accelerating the storyboard and ideation phases of film production.
  • In a major industry application, the company contributed to a $30 million "Bitcoin movie" where AI generated all background scenery, saving an estimated $120 million in set construction costs compared to traditional methods.
  • The technology is being adopted for startup launch videos, reducing production timelines to weeks and costs to a fraction of traditional video budgets.
  • Fan-generated content, such as "Star Wars: Untold" stories on YouTube, is cited as a prime example of the future for interactive, user-driven creative licensing.

• Business Strategy & IP

  • Black Forest Labs employs a mixed strategy of open-source models (Flux) and proprietary models, with strict guardrails preventing the generation of copyrighted IP on public tools.
  • The company is in discussions with major IP holders (e.g., Disney) to co-develop models that allow for safe, customized content generation while protecting original intellectual property.
  • The vision includes enabling consumers to license and remix characters (e.g., Star Wars) to create personalized stories and interactive content on platforms like Disney+.
  • Recruitment focuses on researchers with large-scale diffusion training experience, engineers for physical AI solutions, and staff to manage high-performance compute infrastructure.

• Future Trajectory

  • The company views generative AI not as a replacement for human creativity but as a medium for "intuitive intelligence," allowing users to externalize mental images rapidly.
  • Robotics applications will likely rely on a combination of real-world video data and synthetic data generated by the models to teach action prediction.
  • The ultimate goal is to move from fine-tuning models for specific tasks to "in-context" prompting of robots, reducing the need for massive datasets of specific training videos.
  • Black Forest Labs sees a "minivan" vs. "Ferrari" dynamic where open-source models handle the bulk of generative tasks, while proprietary models tackle high-fidelity, specialized needs.