Open Source Wins, AGI Is Here, and Scorsese’s AI Toolkit with CEOs of Cerebras & Black Forest Labs
In a Nutshell
Cerebras CEO Andrew Feldman reports that AI data center construction is consuming more power than all prior decades combined, with $25B in advance orders from OpenAI, Google, Microsoft, and others, while reasoning models running for hours to days on unlimited tokens are already delivering superhuman trend analysis and problem-solving. AGI has been achieved by any pre-2020 definition—every benchmark passed, the Turing test exceeded—yet the real constraint is now coordination and deployment rather than raw intelligence. Black Forest Labs CEO Robin Rombach shows that latent diffusion models are moving from unpredictable image generators to controllable multimodal systems already used by Scorsese for storyboarding, with the same models fine-tuned in hours to control robots, enabling both high-end film production and physical AI.
These notes were generated by AI and may contain inaccuracies.
Andrew Feldman, CEO and founder of Cerebras, returned to discuss the unprecedented scale of AI infrastructure development. The conversation compares the current mobilization to the Great Wall of China, the pyramids, and the war effort in terms of capital, time, and human resources dedicated to building AI systems.
Cerebras customers are constructing data centers that will consume more power in the coming years than the previous 50 years combined. Individual buildings the size of football fields are being built with power feeds exceeding those of midsize cities. These facilities are under construction across the United States, Canada, the Nordics, Paris and France, the Middle East, Kazakhstan, Tajikistan, Georgia, and Armenia. Every country and every U.S. state is participating in this buildout.
The primary buyers of this capacity include OpenAI, Anthropic, SpaceX, SpaceX AI, Google, Microsoft, and AWS. These organizations are placing orders for chips before manufacturing is complete, creating a $25 billion backlog at Cerebras. Demand is booked in advance rather than speculative, which is described as highly unusual compared to typical technology cycles.
The discussion addresses whether current token consumption creates genuine value. The position is that massive value is being generated alongside significant experimentation. The analogy is drawn to early AWS adoption, where engineers were encouraged to experiment freely with credit cards, resulting in both useful applications and wasted spend. Similarly, Costco shopping behavior evolved from browsing every aisle to strategic purchasing. Enterprises are moving from unconstrained token access to more disciplined allocation based on demonstrated productivity.
A new category of professionals is emerging who understand how to deploy AI effectively. These individuals demonstrate systems thinking. The technology is improving rapidly enough that users, including startup founders, venture capitalists, and associates, are being trained by the AI itself to define clearer goals and create requirements documents.
Prompting has evolved from precise instruction to intent understanding. Models now suggest improvements such as chart types that were not explicitly requested. This represents a shift from summarization and web research to understanding user goals and providing abstracted solutions.
An example is shared of using GLM 52 through a BitTensor distributed crypto project with effectively unlimited capacity from excess compute, possibly located in China with free energy. The model was tasked with identifying world trends that others had not yet recognized. The reasoning process was observed debating internally about information sources, considering Hacker News, Reddit, and Instagram as trend vectors. This demonstrated a reasoning model working through strategy in real time.
Unlimited tokens are equated with unlimited reasoning. Running models for 25 to 48 hours already produces significant results. With Cerebras hardware providing 15x speed improvement, equivalent compute time could yield weeks or months of thinking within a 24-hour period.
Reasoning is fundamentally inference and is computationally intensive. Fast compute makes extended reasoning tractable rather than requiring prohibitive time. The internal token consumption of reasoning processes is what enables high-performance inference systems to deliver value.
A physical Cerebras chip from the first eight or ten produced was displayed. The company has broken from traditional Moore's Law trajectories that doubled performance every 18 months. The new architecture is expected to exceed 2x improvement within the next 18 months. Newer architectures have more optimization headroom compared to mature designs like GPUs that rely primarily on process node shrinks.
Cerebras customers are among the fastest-moving organizations in history and are demanding. Some customers, including OpenAI and Amazon, are developing their own chips. This is interpreted primarily as a desire to avoid complete dependence on external suppliers, drawing lessons from the x86 era when hyperscalers were dependent on Intel. The goal is to control at least part of their own destiny rather than necessarily to compete directly in the chip market.
Open source models serve different use cases than frontier models. Frontier models handle hard problems while open source handles ordinary tasks such as data movement between systems. The "cutting and pasting economy" does not require gold-medal mathematics.
Sovereignty concerns are driving demand for domestic open source models. Regulated industries including finance and healthcare cite HIPAA and FINRA requirements for on-premises deployment and data control. OpenAI's release of OSS 12B is noted as a positive step, but additional U.S. domestic open source options are needed to provide alternatives to Chinese models. Nvidia is also advancing open source models.
Cerebras currently runs GLM, Kimi, the Qwen model family, OpenAI closed-source models, custom models developed by partners such as GlaxoSmithKline, and models from G42 and MBZ UAI in the UAE.
The government's request that certain model releases be staged is viewed as reasonable when a model is sufficiently capable that it could pose meaningful threats. The comparison is made to pharmaceutical regulation. A short period for red teaming and infrastructure patching is considered prudent. Polarization around the specific administration complicates clear assessment of the underlying security concerns.
The rank-and-file government personnel are described as working diligently despite political noise. Competition between frontier labs is viewed as beneficial for technological progress. Self-regulation by the industry is encouraged, though guardrails must be designed while racing against both external competitors and internal standards of excellence.
When GPT-3.5 was tested against Palo Alto Networks software, it identified previously unknown vulnerabilities that required six weeks of patching. This demonstrated the offensive capability of the technology and reinforced the need for controlled release processes. Beta testing programs for operating systems and security patches are cited as precedents for staged deployment of powerful new technologies.
Warren Buffett discussed the reinsurance industry as an analogy for preparing for catastrophic events. Organizations must set aside resources in advance even though they cannot predict exactly when disasters will strike. The principle applies to AI systems and broader technological development: plan for the possibility of major breaches or disruptions while recognizing that specific triggers remain unknown. Black swan events by definition involve scenarios that were never considered or questions that were never formulated in advance.
Even recognizing the existence of unknown unknowns provides a useful starting point for planning. The discussion explored how reasoning capabilities in AI can surface blind spots by prompting models to identify what users have not considered. One effective prompting technique involves explicitly instructing the model to review its work for errors, identify unconsidered goals, and generate follow-up questions. Tools such as Perplexity already surface the next three suggested prompts automatically, and explicit instructions amplify this capability.
A memorable question encountered while raising capital was: "What was the smartest question you heard that wasn't covered by what I asked?" This approach reveals curiosity, humility, and a desire to map the full landscape. Applying the same method to AI expands perspective by asking what additional questions would help achieve expertise in a domain, what a PhD-level questioner would ask, or what a gold medal mathematician might inquire.
Definitions of AGI and superintelligence function as waypoints even if they remain imprecise. The conversation concluded that artificial general intelligence has already been achieved by any definition from 20 years ago. The Turing test has been surpassed, and every prior benchmark has been exceeded. Science fiction authors from previous decades would find their questions answered and would have no remaining questions to pose.
Early warnings about AI safety from Ilya Sutskever eight or ten years ago initially seemed extreme yet proved prescient. Similarly, Elon Musk's early statements about driving launch costs to near zero appeared unrealistic at the time. These examples illustrate why it remains valuable to attend to voices that sound fringe. The current moment allows technologists to observe recursive loops in which tools build upon themselves. The term "loop maxing" emerged during trend scouting and subsequently entered wider usage, demonstrating how models can anticipate emerging concepts.
Sam Altman, Ilya Sutskever, Dario Amodei, and others recognized five to six years ago that powerful recursive gains produce exponential improvement. Each iteration yields better results, incorporates more information, and covers additional material. The slope of improvement is steep, and throwing additional compute at the process raises the question of how far the exponential continues before token or budget limits intervene.
The discussion distinguished intellectual problems from people problems. Many organizational challenges involve coordination, motivation, and reducing friction rather than pure intellectual difficulty. World models that observe human behavior may eventually provide behavioral insights. The conversation extended to physical-world applications where recursive loops operate outside screens, such as constructing large-scale environments or deploying robots.
Historical large-scale projects often involved multi-generational participation by specialist families who apprenticed and passed on innovations. AI learning now compresses equivalent generational cycles into much shorter periods. Human learning typically advances at generational pace, whereas AI achieves learning across thousands of equivalent generations rapidly. Paradigm shifts in human knowledge historically required the passing of influential figures before new approaches gained acceptance; AI iteration shortens these intergenerational gaps dramatically.
The technology offers a credible path where neither our children nor anyone they know dies of cancer. Economic dislocation will occur, analogous to the displacement of carriage makers when automobiles arrived, yet the ledger includes massive abundance on the positive side: unlimited energy, unlimited calories, unlimited knowledge, unlimited education, and unlimited housing. Personalized AI tutors that adapt to individual learning styles represent one concrete application, replacing the factory-model approach of teaching to a mid-level standard.
Industries, capital, and intelligence are converging into a single interconnected system. NASDAQ powers more than 135 marketplaces and regulators globally and serves as the leading technology platform connecting capital to companies shaping the future.
Robin Rombach is co-founder and CEO of Black Forest Labs, based in the Black Forest town of Freudenstadt and with presence in San Francisco. The company develops open-source image and video models. Rombach previously worked on Stable Diffusion and co-invented the latent diffusion algorithm, the foundational technique behind current generative models for images, video, audio, and physical AI applications. Latent diffusion compresses natural data into efficient representations on which transformer models are trained.
Current work focuses on multimodal visual models pretrained on images and audio simultaneously, moving toward integration with action prediction. The same model can generate images, videos, and audio while predicting actions suitable for robotic deployment in the real world. Pretraining on video yields implicit understanding of physics and real-world interactions.
Early generative models functioned like a slot machine where prompts produced unpredictable outputs. Progress has expanded manipulation layers: text-to-image evolved into text-plus-image-to-image editing, then into multi-image semantic combination. The same progression now applies to video. Exposing additional controllable layers allows users and developers to iterate, modify style, color, and aesthetic.
Black Forest Labs is working with director Martin Scorsese. The collaboration centers on using models as a medium for visual exploration rather than prescribing specific usage. Scorsese explored generating imagery of an Eastern European village scene, iterating on outputs to externalize mental pictures that are difficult to convey through language alone. Visual information carries richer signal than language, which is inherently lossy and subject to interpretation.
The goal is not necessarily to generate complete feature-length films autonomously. More valuable use cases arise when humans remain in the loop, iterating as part of a workflow. Storyboarding represents an immediate high-value application; directors such as Ridley Scott, Steven Spielberg, and George Lucas historically created their own storyboards or collaborated with artists for this purpose.
Startups are replacing expensive launch videos that previously cost $100,000–$250,000 with shorter, lower-cost productions created in one to two weeks using generative tools. The same base models support diverse creative and commercial outputs, including a forthcoming Bitcoin movie exploring alternatives to green-screen techniques.
Black Forest Labs is seeing use cases in actual production workflows, though high-end film production remains one of the most demanding applications. The technology is on a clear trajectory of rapid improvement. A few years ago during PhD-level research, the only achievable output was 64x64 pixel images. Current capabilities include multi-input videos at high resolution, with continued improvements expected to unlock additional high-end use cases.
The most important consideration is maintaining a human-in-the-loop production workflow. Multimodal generative models excite interest because the same AI model can be used both to make a movie and to deploy as a brain on a robot. This versatility connects world models and world action models, which are fundamentally the same underlying technology.
The approach combines prediction, simulation, and generation. Predicting actions requires understanding visual inputs to determine reasonable next actions. This involves perception: understanding what is seen enables transformation into new content, prediction of actions, or description of scenes. No single element drives progress; it is the combination of these capabilities.
Current deployment involves many different hardware platforms and robots running in factories, each with different action representations that require model tuning. In practice, models leverage strong visual understanding capabilities and then require only a few hours of fine-tuning data to adjust for specific tasks. The goal is moving toward more in-context learning, though this remains a research problem.
Black Forest Labs implements safeguards on public-facing tools that prevent generation of certain IP. The company also works directly with IP holders to develop models together, sometimes based on open source models and sometimes on more powerful proprietary models. This collaborative approach provides attractive value for rights holders.
Disney has already established precedents by working with ChatGPT and licensing specific characters for output. OpenAI previously had a relationship with Disney for Sora that is no longer active, but the licensing model for characters demonstrates one path forward.
The technology is becoming faster and more interactive, enabling new forms of interactive content creation tools that could be hosted on platforms like Disney+. Fan films represent a particularly interesting development. Previously, fans wrote Star Wars fanfiction or created live-action films with costumes. George Lucas permitted non-commercial Jedi films and even released how-to resources including lightsaber sound files.
Now fans are creating AI-generated stories from the Star Wars universe that have not been officially told. Star Wars Stories Untold has become one of the largest channels, with videos receiving millions of views. This points toward a future where customers pay licensing fees, rent software, or pay based on output, then create their own stories with established characters.
Black Forest Labs has an office in San Francisco and recently crossed 100 employees. The company is hiring in both Germany and San Francisco after raising additional funding.
They seek researchers with experience in large-scale model training, particularly diffusion model training and flow matching training. Engineers are needed to work with customers developing customized physical AI solutions and to jointly develop models with IP owners. Additional roles include engineers experienced in large-scale compute infrastructure management to maximize MFU and ensure training runs smoothly. The company also seeks people interested in getting the technology into users' hands for deployment.
Corporations want additional control but also need frontier models and proprietary capabilities for refined features, suggesting a bright future for the open source approach combined with proprietary offerings.
Keep All-In Podcast in your library
Save the videos and channels worth coming back to, and find them again in one place.
