TubeReads

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

We are mobilizing capital, intelligence, and physical infrastructure on a scale not seen since the Great Wall of China—or perhaps the war effort. Andrew Feldman's Cerebras sits at the center of a $25 billion backlog, churning out inference chips as OpenAI, Anthropic, and SpaceX race to capture yesterday's demand. Meanwhile, Robin Rombach of Black Forest Labs is handing Martin Scorsese the tools to visualize dreams before they become scripts, collapsing the distance between imagination and production. But as reasoning models burn through tokens at unprecedented speed and sovereignty concerns resurface, two questions loom: Have we already crossed the threshold of AGI without noticing? And is open source about to eat the world—or will the frontier labs always hold the high ground?

Duración del vídeo: 1:03:57·Publicado 10 jul 2026·Idioma del vídeo: English
9–10 min de lectura·11,596 palabras habladas → resumido a 1,945 palabras (6x)·

1 —

Puntos clave

1

Cerebras holds a $25 billion backlog, racing to meet insatiable demand from OpenAI, Anthropic, SpaceX, and hyperscalers building data centers that consume more power than mid-size cities. Demand is booked; supply is the bottleneck.

2

By every definition posited 20 years ago—Turing tests, gold-medal mathematics, PhD-level reasoning—AGI has arrived. The models now check their work, ask what you haven't considered, and iterate recursively toward exponentially better answers.

3

Open-source models like Qwen, Kimmy, and Flux are closing the gap with frontier labs. Enterprises want sovereignty, cost control, and the option to customize—making open source a strategic imperative, not just an experiment.

4

Black Forest Labs is enabling creators—from Martin Scorsese to startup founders—to visualize ideas instantly. The technology is collapsing the cost and time required to produce launch videos, storyboards, and even high-end film production assets.

5

Loop-maxing and recursive learning are producing not incrementally better answers, but vastly superior ones. The exponential curve is steep, and we don't yet know where it ends—or whether the bottleneck shifts from intellectual problems to human coordination.

En resumen

AGI is no longer a future milestone—it's here, embedded in reasoning models that check their own work and propose questions we didn't know to ask. The race for superintelligence is now a race for deployment, sovereignty, and the right blend of open and proprietary infrastructure to power an exponential leap in human capability.


2 —

The Unprecedented AI Buildout: Data Centers as Modern Pyramids

Data centers now consume more power than mid-size cities, rivaling historical mobilizations.

The scale of AI infrastructure deployment is historically unparalleled. Andrew Feldman describes data centers the size of football fields drawing power loads previously associated with entire municipalities, being erected across the U.S., Canada, the Nordics, France, the Middle East, and even nations like Kazakhstan and Tajikistan. Every country and every state feels compelled to participate. Unlike typical technology cycles, this is not a case of «if you build it, they will come.» Demand is booked. OpenAI, Anthropic, SpaceX, Google, Microsoft, and AWS are not chasing speculative futures—they are scrambling to fulfill existing commitments.

Cerebras sits at the heart of this mobilization with a $25 billion backlog. Feldman notes that customers are ordering chips before they are finished, trying to capture yesterday's demand rather than tomorrow's potential. The irony is acute: in an era defined by exponential growth, supply chains are the constraint. The buildout is less about vision than about execution at a pace that recalls only the war effort or ancient megastructures.


3 —

Token-Maxing and the Evolution from Experimentation to Strategic Deployment

Early unconstrained token use is giving way to disciplined, enterprise-grade deployment strategies.

EARLY PHASE
The Costco Problem: Aisles of Unnecessary Tokens
When AWS first enabled developers to bypass IT departments, engineers signed up liberally, often spawning projects of dubious value. The same pattern emerged with AI tokens: enterprises gave teams unlimited access, resulting in massive experimentation—some invaluable, some wasteful. Feldman likens it to shopping every aisle at Costco and leaving with four $22 items you didn't need.
MATURE PHASE
Strategic Routing: Frontier Models vs. Open Source
Enterprises are now running AI «like a business,» deploying frontier models for hard, high-value problems and open-source or cheaper models for routine tasks. A new class of systems thinkers—often startup founders and VCs—has emerged, individuals who intuitively grasp when to apply which tool. The AI itself now coaches users to clarify goals, build requirements documents, and avoid token waste.

4 —

AGI Has Arrived—We Just Didn't Notice

By every historical benchmark, artificial general intelligence is already here.

💡

AGI Has Arrived—We Just Didn't Notice

Feldman and Calacanis agree: by any definition posited 10, 20, or 50 years ago—Turing tests, gold-medal mathematics, PhD-level reasoning—AGI has been achieved. Models now check their own work, propose follow-up questions the user didn't think to ask, and iterate recursively toward vastly superior answers. The models understand intent, not just syntax. Calacanis recounts asking an AI to identify global trends nobody else has spotted, then watching it debate with itself whether to scrape Hacker News, Reddit, or Instagram. «That's a reasoning model,» Feldman observes. The paradigm has shifted from summarization to strategic problem-solving, and the implications are only beginning to unfold.


5 —

«Computers are really dumb. They do exactly what you tell them.»

AI has evolved from rigid prompt interpreters to intent-understanding collaborators.

“One of my colleagues 20 years ago, a really smart computer scientist said, «Computers are really dumb. They do exactly what you tell them.» And at first, prompting was like that. You modified your prompt a little bit and it changed the answer dramatically. And increasingly it's understanding what your intent was.”

— Andrew Feldman


6 —

The $25 Billion Question: Where Does Demand End?

Cerebras and peers face insatiable, pre-booked demand with no ceiling in sight.

Cerebras Backlog
$25 billion
Pre-booked orders from OpenAI, Anthropic, SpaceX, Google, Microsoft, and AWS, reflecting demand outstripping supply across the AI infrastructure sector.
Moore's Law Trajectory for Cerebras
Over 2x in 18 months
Feldman expects Cerebras to exceed traditional Moore's Law doubling every 18 months, leveraging architectural innovation rather than just fab node shrinks.
Cost of a Single Cerebras Wafer
$500 million
Feldman jokes he carries a wafer everywhere—when one costs half a billion to make, it earns a special place.

7 —

Open Source and Sovereignty: The New Strategic Imperative

Enterprises and nations demand control, cost efficiency, and customization—fueling open-source adoption.

Open-source models—Qwen, Kimmy, OSS-12B—are closing the performance gap with frontier labs. Calacanis describes «smart routing» his own tokens between Claude and Kimmy, finding the open-source model indistinguishable for many tasks. Feldman observes that regulated industries—finance, healthcare—are demanding on-prem, domestically hosted solutions to avoid data leakage and maintain sovereignty of intelligence. OpenAI's release of OSS-12B was a strategic concession, but Feldman argues the U.S. needs more domestic open-source champions to give the world a choice beyond Chinese models.

Nvidia has also quietly developed open-source models, though Jensen Huang has been reluctant to promote them for fear of alienating customers like OpenAI and Anthropic. Cerebras, by contrast, positions itself as neutral infrastructure: it runs OpenAI's closed models, Chinese open-source models, and bespoke enterprise models like those developed by GlaxoSmithKline and G42. Sovereignty is not just a geopolitical concern—it's a business model. Feldman predicts this trend will accelerate as governments reassess dependencies exposed by incidents like the Fable/56 release controversy.


8 —

Recursive Learning and the Exponential Curve

🔄
Loop-Maxing
AI models now iterate recursively: ask a question, learn from results, refine the query, and repeat. Each loop adds information and improves output exponentially, not linearly.
📈
Exponential Gains
Feldman and Calacanis discuss how recursive loops produce answers that are «not a little bit better but vastly better.» The slope is steep, and we don't yet know where the curve flattens—or if it does.
⏱️
Time Compression
Running reasoning models for 24–48 hours with unlimited compute yields weeks or months of human-equivalent thinking. Fast inference chips like Cerebras make this tractable and scalable.
🧬
Generational Learning
Feldman likens AI iteration to studying fruit flies in genetics: thousands of generations in days, versus the 15–20 year human cycle. Paradigms no longer wait for people to die—they iterate hourly.

9 —

Black Forest Labs: From Stable Diffusion to Multimodal World Models

Robin Rombach's team is building models that generate images, video, audio—and predict real-world actions.

Robin Rombach co-founded Black Forest Labs after inventing latent diffusion, the algorithm underlying Stable Diffusion and most modern generative models. The core insight: compress natural data—images, video, audio—into efficient representations, then train transformers on them. This is why JPEG and MP3 work; latent diffusion translates that principle into neural networks. Black Forest Labs is now building multimodal models pre-trained on images, audio, and video simultaneously, with a roadmap that includes action prediction for robotics.

Rombach describes the convergence: a model that generates a video inherently understands physics and object interactions. By extending that model to predict actions, the same architecture can be deployed on a robot in the real world. The company splits its efforts between open-source models like Flux and proprietary offerings for high-end clients. Rombach is adamant: these are tools, not dictates. The goal is to expose as many «manipulation layers» as possible, letting users and developers iterate and customize. The company has raised significant capital, crossed 100 employees, and is hiring researchers, engineers, and customer-facing roles in Germany and San Francisco.


10 —

Scorsese's Vision: AI as a Medium, Not a Replacement

Martin Scorsese used Black Forest's tools to visualize Eastern European village scenes for pre-production.

“He has clearly a vision in his head of a scene or a scenery where maybe a new movie will be shot and he's trying to explore that. We basically looked at the scenery of a village in Eastern Europe somewhere and he was describing it, we saw some outputs, we iterated on the outputs. And ultimately I think what he said in the end is getting the mental picture of something out of your head and communicating it in a visual way by making these images is something that just makes it easier to communicate and convey an idea of what is actually in your head.”

— Robin Rombach


11 —

From Storyboards to Production: The New Economics of Filmmaking

Generative AI is collapsing costs and timelines for launch videos, storyboards, and even feature-film production.

1

Ideation and Storyboarding Directors like Ridley Scott and George Lucas used to sketch storyboards or hire full-time artists. Now, tools like Flux enable rapid visualization of scenes, parallelizing the brainstorming process.

2

Startup Launch Videos Founders previously spent $100,000–$250,000 on launch videos. Now, they collaborate with a director for a week or two, using AI to generate assets, cutting costs by an order of magnitude.

3

High-End Production Gal Gadot described filming a $30 million Bitcoin movie on a sound stage without green screens. All scenery was generated by AI—what would have been a $150 million production became financially viable.

4

Fan Films and IP Licensing Fans are using generative AI to create «Star Wars Stories Untold,» recreating untold narratives from the universe. Millions of views suggest a future where IP holders license tools for creative customization.


12 —

The Path to Physical AI: Perception, Prediction, and Action

👁️
Perception
The model must understand and describe what it sees in a scene—object recognition, spatial relationships, physics.
🎬
Prediction / Simulation
Generating video is a form of prediction: what happens next? Pre-training on video gives implicit understanding of physics and interactions.
🤖
Action Prediction
The same model can be fine-tuned to predict the next action a robot should take, enabling deployment in factories or real-world environments with minimal task-specific data.
🔗
In-Context Prompting
The goal: prompt a robot in natural language—«pick up the glass with orange juice»—just as you would a language model, with no hardware-specific fine-tuning required.

13 —

The Ledger of Abundance: Cancer, Education, and the Human Coordination Problem

AI promises unprecedented gains in health, education, and resource abundance—if we solve for human coordination.

Feldman offers a stark optimistic framing: «We have a shot with this technology so not our children nor anyone they know dies of cancer.» He acknowledges economic dislocation—there will be winners and losers, just as cars displaced blacksmiths—but insists the ledger is heavily weighted toward abundance. Unlimited energy, calories, knowledge, education, housing: the promise is real. He cites the ancient knowledge that one-on-one tutoring—Aristotle to Alexander, Socrates to Plato—produces vastly better outcomes than factory-model classrooms. AI agents can finally deliver personalized education at scale.

Yet Feldman also observes that once hard intellectual problems are solved, the bottleneck shifts to «people working together problems»—motivation, coordination, friction. AI can help there too, he suggests, particularly as world models begin to learn from observing human behavior. Calacanis notes that reasoning models are already coaching users to think more clearly, define goals, and build requirements documents. The recursive loop is tightening: AI helps humans help AI help humans. The question is no longer whether we can solve problems, but whether we can organize ourselves to deploy the solutions.


14 —

Valores mencionados

CBRSCerebras Systems (implied; post-IPO)

15 —

Personas

Jason Calacanis
Podcast Host
host
Andrew Feldman
CEO and Founder, Cerebras
guest
Robin Rombach
Co-Founder and CEO, Black Forest Labs
guest
Martin Scorsese
Film Director
mentioned
Sam Altman
CEO, OpenAI
mentioned
Dario Amodei
CEO, Anthropic
mentioned
Ilya Sutskever
Co-Founder, OpenAI (former)
mentioned
Jensen Huang
CEO, Nvidia
mentioned

Glosario
Latent DiffusionAn algorithm that compresses natural data (images, video, audio) into efficient neural representations, enabling transformers to generate high-quality outputs; the foundation of Stable Diffusion and most modern generative models.
Token-MaxingThe practice of consuming large volumes of AI inference tokens, often through recursive or loop-based reasoning, to achieve exponentially better outputs.
Loop-MaxingA term coined during the conversation to describe iterative, recursive prompting where the AI refines its own outputs through repeated cycles, compounding learning and quality.
World ModelAn AI model pre-trained on diverse real-world data (images, video, audio) that develops implicit understanding of physics, objects, and interactions, enabling both generation and action prediction.
MFU (Model Flops Utilization)A metric measuring how efficiently a compute cluster is being used during model training; maximizing MFU is critical for cost-effective large-scale AI development.

Aviso legal: Este es un resumen generado por IA de un vídeo de YouTube con fines educativos y de referencia. No constituye asesoramiento de inversión, financiero o legal. Verifique siempre la información con las fuentes originales antes de tomar decisiones. TubeReads no está afiliado con el creador de contenido.