Original title: Open Source Wins, AGI Is Here, and Scorsese's AI Toolkit—Cerebras & Black Forest Labs CEOs
Original source: All-In Podcast
Original translation: Deep Tide TechFlow

Key takeaways

This episode invites two CEOs from AI infrastructure companies. Andrew Feldman is the founder of Cerebras, which focuses on inference chips. The company has just completed its IPO and has $25 billion in backlog orders. He repeatedly emphasized one thing: demand for AI compute has already been fully booked—there’s no scenario of “build it and people will come.” The appetites of OpenAI, Anthropic, SpaceX, and Google far exceed supply. And with the emergence of reasoning, the compute intensity spikes again—this is exactly the battlefield for fast machines. Robin Rombach is the founder of Black Forest Labs. He builds generative image and video models (the Flux series). He previously invented the latent diffusion algorithm, which is the foundation for essentially all current image and video generation models. He has just partnered with Martin Scorsese to let the director use AI to visualize the scenes in his mind; but what excites him even more is that the same multimodal model can both make movies and be deployed to robots as a “brain.” The endpoint of generative video isn’t the silver screen—it’s the physical world.

Highlights

Reasoning is the next compute black hole

· “What’s interesting is that this wave is different from the past. They’re not betting on ‘build it and someone will come.’ Demand has already pre-booked capacity. We have $25B in backlog orders.”

· “Reasoning is reasoning—reasoning consumes massive token volumes, and that’s exactly the battlefield for fast machines.”

· “If Cerebras is 15x faster, and you run it for 24 hours, it’s like running the equivalent of weeks or even months of thinking.”

Open source and sovereignty: companies want control

· “Nobody likes being dependent. The lesson that mega-scale vendors learned from the x86 era is getting locked into Intel.”

· “You don’t need the fastest chips—you just need to be not completely dependent on someone else’s chips.”

· “If you want to run open-source models now, either it’s OpenAI’s OSS 12B or China’s models; the U.S. needs more local open-source choices.”

AGI, by definitions from two decades ago, is already here

· “Any AGI definition we proposed 20 years ago, 30 years ago, 40 years ago—we’ve long since surpassed them.”

· “The Turing test? It got blown out long ago.”

· “The issue isn’t that we don’t know how to ask anymore. AI can flip it around and tell you: hey you clueless humans, you didn’t consider this.”

Generative video isn’t a substitute for human creation

· “These AI models are a medium. We don’t want to dictate how to use them—especially for someone like Martin Scorsese.”

· “Language is a somewhat lossy communication method; visual signals are far richer. Turning the scenes in your mind into visible images—that’s the strongest part of the technology.”

· “The most interesting outcomes almost all show up when people iterate in the loop.”

From film to robots: the same model

· “You can use the same multimodal model to make a movie, then deploy it as the brain of a robot.”

· “Pretrained video implicitly teaches the model the laws of physical interaction. Then from the same model you get action prediction—that’s robot control.”

· “The goal is that you can instruct a robot with in-context prompts: ‘Bring me that cup of orange juice.’ We can’t do that yet, but that’s the direction.”

AI infrastructure boom: data centers are bigger than cities

Host: We’ve never seen construction at this scale. Since the Great Wall and the pyramids, humanity hasn’t put so much capital, time, and smart people into building something like this. You’re actually doing it—you have a client building data centers, and you’re a key piece. In 2026, what is Cerebras doing? And what’s going on with those massive projects in Texas?

Answer: The data centers we’re talking about will consume more electricity in the next few years than the total amount consumed on Earth in the past 50 years. A single building is as large as a football field, and the electricity it connects exceeds that of a mid-sized city. They’re being built across the U.S.; in Canada; in Northern Europe; in Paris and across all of France; in the Middle East. Even in Kazakhstan, Tajikistan, and Georgia, large data centers are being built too. Every country, every state wants in.

Who’s paying? OpenAI, Anthropic, SpaceX AI, Google—their appetite is enormous. What’s interesting is that this wave is different from many past tech booms: they’re not betting on “build it and someone will come.” Demand has already pre-booked capacity. We have $25 billion in backlog orders. OpenAI wants more data centers, Microsoft wants more, AWS wants more. Demand isn’t waiting for customers—it’s the customers queuing already.

Host: This is also giving rise to a term called “token maxing,” endlessly cranking out tokens. Some people question whether such massive demand actually creates real value.

Answer: Of course a lot of real value is being created. And of course there’s also a lot of blind experimentation. When I came out of AWS, bypassing your own IT department felt amazing—each engineer just registers with a credit card. Many things really help, and later you think, “Oh, we shouldn’t have done that.” But overall, it still makes money—it’s just that some directions miss.

I remember in 1988 Costco opened in Palo Alto. Everyone shopped there like they did Safeway—walking through every aisle. That was a terrible way to shop: you bought four things you didn’t need, and each was $22. Later people learned strategy: go to the back to get the chicken, grab 18 cupcakes for the kids’ birthday party, and be done quickly. AI token consumption is the same. At first everyone used it freely, but now companies start talking about strategy: which tasks only need open-source models, and which must use frontier models. We’re starting to manage AI like we run a business.

Reasoning replaces training: why fast machines are the stars of this wave

Host: Sam Altman said on AllIn that the next step is reasoning—cross-validating intent, strategy, and other threads with agents. We’ve come a long way from “guess the next word.” Cerebras is right at the center now, because reasoning is inference, and the compute requirement is enormous.

Answer: Reasoning consumes massive token volumes, and that’s what gives fast machines a battlefield. Each step of reasoning internally devours tokens. You previously needed to spend tons of time to trade for correct answers. Cerebras being 15x faster means that running 24 hours of reasoning is equivalent to what others think for weeks or even months.

This morning, I tried a BitTensor ZAI GLM-52 model. I gave it “infinite compute,” and it tells me every hour the trends around the world that haven’t been identified yet. It started debating itself: should it look for trends on Hacker News and Reddit first, or do they appear earlier on Instagram? I watched a reasoning model arguing with itself in the background—it was doing reasoning. Infinite tokens equals infinite reasoning. With Cerebras at 15x faster, 24 hours is weeks’ worth.

Host: Does Cerebras have its own Moore’s Law? How long does it take internally for a doubling?

Answer: All previous chips rode Moore’s Law, doubling every 18 months. We broke that line with this chip and found a whole new trajectory. My take is that in the next 18 months, it will be far beyond 2x. The new architecture still has lots of room for optimization. GPUs are the old architecture from 20 years ago. They can only prop themselves up by shrinking process nodes, but the new architecture still has a lot to learn and tune.

Host: You have a $25B backlog in hand, and you still have to keep up with the pace of OpenAI—they might be potential competitors down the road. How do you operate the company?

Answer: Right now, silicon wafers won’t be idle—the demand is too big. But you’re right: OpenAI is also making its own chips; Amazon is also making its own. Nobody likes being dependent. The lesson that mega-scale vendors learned from the x86 era is being locked into Intel; the lesson GPU vendors learned is being locked into a small number of mega-scale customers—so they funded a new cloud. Making your own chips isn’t about being the fastest; it’s about not being completely dependent on others, and at least controlling an important part of your fate.

Open source and sovereignty: companies want control

Host: Open source is hitting a moment. I used OpenClaude early on, then Kimmy, and I noticed my Claude’s tokens were exploding, but I couldn’t tell the difference with Kimmy. Open-source models started doing reasoning, and the gap suddenly closed this year.

Answer!: You don’t want to rent Ferrari to go to the grocery store. Sometimes you use a sports car, sometimes a minivan. If the kids spill Cheerios, you don’t care. Companies are the same: hard problems go to frontier models (OpenAI, Anthropic, Gemini), but a lot of everyday questions behind the scenes only require solid open-source capability. Think about how many hours a company spends on tasks like copying and pasting in Workday into another cell in Excel. You don’t need gold-medal math—solid open source is enough.

Recently we drew another card: in regulated industries like finance and healthcare (HIPAA, FINRA), people fear data leaks and fear “intelligent sovereignty” being held by someone else. They want the model on-premises, and they want to grab more control with open-source versions. A few months ago, OpenAI released OSS 12B—that’s okay. But in the U.S., to run open source now, it’s either OSS 12B or China’s models, and local open-source choices are too limited. NVIDIA also saw this window and is pushing its own open-source models, but Jensen is hesitating: his customers are Sam, Dario, Elon, Sergey. If NVIDIA opens source, will it take away business from its customers?

Cerebras’s position is relatively neutral. We run GLM, run Kimmy, run the Qwen series, and we also run closed models from OpenAI. We also run models developed by GSK itself, and run UAE G42 and MBZUAI’s proprietary models. Sovereignty—this is a trend.

AGI is here, the paradigm won’t die—people will

Host: When Fable 5 and o-56 were released, the government said “pause, then proceed.” Relations between Anthropic and the administration are tense, and now they’re starting to ease. Do you think step-by-step releases are reasonable? Are the models truly dangerous enough?

Answer: I’ve never seen anything like this before. But looking back: when a model becomes strong enough at creative thinking, and the government says “release in steps,” I think that’s not unreasonable. We manage “strong medicine” the same way. We’re not encouraging the FDA’s years-long pile of garbage paperwork, but saying “at least let the government do some red-team testing to confirm our defenses can hold; patch obvious holes in two or three weeks”—that’s not an unreasonable request.

But this is the most polarized moment right now. If Trump didn’t do this—if any other president had done it—the reaction could be completely different. Polarization hurts clear thinking. Both sides will do stupid things and smart things. The grassroots people in government are actually taking this seriously; it’s just that this happened too fast.

Nikesh from Palo Alto Networks told me: they tested the models against their own software and found dozens of critical vulnerabilities within an hour. They had to stop everything they were working on and spend six weeks patching. You realize this is a powerful tool. Maybe you first show it to a small group of people; maybe you start with red-team testing first.

Host: By any definition from 20 years ago, AGI has arrived. Do you agree?

Answer: Yes. The Turing test? Long since blown out. Any definition proposed 10, 15, 20, 30, 40, 50 years ago—we’ve far surpassed them. The questions that science-fiction writers asked, we’ve already answered. They’ll say, “I don’t have any questions anymore. Sorry.” That’s why what those who sound “on the fringes” say is worth listening to. Eight years ago, Ilya talked about safety. You said, “What?”—and he was right. Elon talked about rocket costs dropping close to zero—you said, “What?”—and he did it.

Host: Recursive learning—ask it a question, learn the result, ask again, and the answer gets better, covering more material. The answers produced by these cycles jump from “a bit better” to “much better.” The slope of an exponential curve is too steep.

Answer: Recursive gains are exponential. You get better, then do it again, and the gain continues, with the slope becoming steeper. We just started to see this. If you keep throwing compute at it, do the answers keep improving? Once you run out of tokens or budget, you stop. But when does this exponential curve finally end? Does it go forever to the upper right? That’s an incredibly interesting question right now.

Human learning speed is constrained by generations. Elephants and large mammals take 15-20 years per generation. To go faster, you need to be like a fruit fly—two generations a day. AI is getting that kind of learning speed across thousands of generations. When I studied psychology, a professor said something: the paradigm won’t die—people will. The disciples of Freud, Skinner, Jung held leadership positions for 20-40 years before the next generation questioned. AI compresses the inter-generational interval into fruit-fly speed.

My bet is that our children and everyone they know won’t die from cancer. There will be economic shocks—when cars come, it’s a rough time for the guys who sharpen horseshoes. But if you list the gains and losses: unlimited energy, unlimited food, unlimited knowledge, unlimited education, unlimited housing. For a thousand years we’ve known that one-on-one tutoring is better than classroom teaching—Aristotle tutored Alexander, Socrates tutored his students—but we chose factory-farming style education. Now AI can give every child a personal tutor that teaches them in their own way.

Scorsese’s AI toolkit: turning the scenes in your mind into reality

Host: Robin Rombach is the co-founder and CEO of Black Forest Labs, headquartered in Freiburg in the Black Forest region and in San Francisco. You previously worked on Stable Diffusion and invented the latent diffusion algorithm. What is Black Forest Labs’s business? What’s the goal?

Answer: My partners and I founded this company two years ago. We previously worked on Stable Diffusion, and even earlier invented latent diffusion—which is the foundational algorithm behind all current image generation, video generation, and even physical AI models. The principle is to compress natural data (images, video, audio) into an efficient representation space, then train a transformer on top of it, similar in spirit to how JPEG and MP3 work—except using neural network algorithms. We built it during our PhD in Munich.

Now we’re tackling multimodal vision models: pretraining jointly on image and audio data. We’re entering a new paradigm—combining action prediction so that the same model can do images, do video, do audio, and also predict actions, and ultimately be deployed on real-world robots as a core brain.

Host: From images to video to audio to robots—if a model can generate video, it means it understands the world.

Answer: Intuitive intelligence and deep reasoning are two complementary forms of intelligence. We started from the intuitive side—images are the most natural entry point, and the compute isn’t as large as video. But now everything is converging into multimodal models. Pretrained video implicitly teaches the model the laws of physical interaction. From the same model you get action prediction—that is robot control.

Host: You collaborated with Martin Scorsese? Did you sit next to him and let him use your tools?

Answer: Yes. I sat in the same room with him. He explored our model, and as one of the core researchers, I sat nearby—that feeling was absolutely insane. And at the same time, I’m also a huge fan of him.

What he wanted was to visualize the scenes in his mind. He described a village in Eastern Europe; we watched the outputs; he iterated. In the end he said: turning the scenes in your head into visual expression—this communication efficiency is far higher than language. Language is a somewhat lossy communication method, and visual signals are too rich. A single image or a segment of video contains enormous information—so it’s a different communication channel.

We don’t want to dictate how to use these models, and especially we wouldn’t tell Martin Scorsese, “You should use it this way.” AI models are a medium. Most of the most interesting things happen when people are iterating in the loop.

From film to robots: generative models’ endpoint isn’t the silver screen

Host: Startups are now using Flux and your models to produce launch videos. Previously it cost $250k for a launch video; now they can finish it in one or two weeks. Gal Gadot just made a Bitcoin movie. The actors performed on a soundstage without green screens; all backgrounds were made with generative AI. With a $30 million budget, they achieved effects that originally would have required $150 million. Have you seen this being used in production?

Answer: I’ve seen some of it. High-end film production is one of the most demanding use cases. I’m happy to see people exploring, but I also want to be clear: the technology is still on its trajectory and is iterating quickly. A few years ago, when we were doing PhD work, we could only generate 64×64 images. Now we generate multi-input high-resolution video—but it won’t stop there.

What excites me most is this: you can use the same multimodal model to make a movie, then deploy it as the brain of a robot. Whether “computer use” can work is still uncertain, but the technology is moving toward the physical world. World models and action models—basically, it’s the same thing.

Host: Where do the training data come from? Do we have humans wear glasses and gloves to record first-person views? Or is it enough to watch a thousand people on YouTube pouring drinks?

Answer: The goal is to instruct a robot with an in-context prompt: “Bring me that cup of orange juice.” We can’t do that yet. The current approach is that the model already has a lot of built-in visual understanding, and then it only needs a few hours of fine-tuning data to adapt to specific hardware. The direction is to fine-tune as little as possible and rely as much as possible on in-context instructions—but it’s still a research question.

Host: Open source is hitting a moment, enterprises want sovereignty. For IP libraries like Disney, how should they do it—train using your open-source models, or collaborate to train a dedicated proprietary model with you?

Answer: The most interesting use case is generating things that didn’t exist before—that’s what’s fundamentally most interesting about this technology. Our public tools can’t generate specific IP, and that’s reasonable. And yes, we do partner with some IP owners to develop models: some are based on our open-source models, and some are based on our stronger proprietary models.

The most interesting angle is that the technology is becoming faster and more interactive. You can imagine all kinds of interactive content-creation tools hanging off Disney+.

Host: The most interesting phenomenon right now is fan films. Previously there were fan fictions writing their own Star Wars stories; later someone dressed up as a Jedi to shoot fan films. George Lucas said it’s allowed as long as it’s not used commercially. Now people use AI to re-enact Star Wars stories that were never told before. Star Wars Stories Untold gets millions of views per video. That’s the future: let consumers pay for licensing, and let them use the characters to create their own stories.

Answer: If you can find a viable commercial business model for IP holders, and also open up this kind of super-creative customization play, that would be great. I read books or watch movies and always think, “What if the story developed this way?” Now it finally lets people visualize those thoughts.

We’ve just crossed 100 people. We’re hiring in Germany and San Francisco: researchers for large-scale model training; people with experience in training diffusion and flow matching; engineers who develop customized solutions together with customers; and people interested in operating large-scale compute infrastructure—and also people interested in getting the technology into the hands of more people.

Original link

Click to learn about positions at律动BlockBeats

Welcome to the official律动 BlockBeats community:

Telegram subscription group: https://t.me/theblockbeats

Telegram group chat: https://t.me/BlockBeats_App

Twitter official account: https://twitter.com/BlockBeatsAsia