Jul 10, 2026 · 1h 3m · allin
Open Source Wins, AGI Is Here, and Scorsese’s AI Toolkit with CEOs of Cerebras & Black Forest Labs
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of the All-In Podcast, host Jason Calacanis interviews Cerebras CEO Andrew Feldman and Black Forest Labs CEO Robin Rombach to explore the massive global AI infrastructure buildout, the emergence of AGI and reasoning models, and the evolution of multimodal visual AI into robotics.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. The hosts hold 43.8% of the talking time here. How this is scored →
The hosts as informed peer 4.7 Guest teaching 4.8 Guest disagreement 1.2 The hosts pushing back 1.4
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 19:05 Feldman cuts off Jason on Nvidia strategy
Feldman interrupts Jason mid-sentence as Jason brings up Nvidia, redirecting the debate toward the national security imperative of domestic open source models over Chinese alternatives.
Hardest push from the hosts ▶ 14:50 Jason challenges hyperscaler custom chip motivesJason refuses to accept the standard PR framing around hyperscaler silicon, directly asking whether projects like Amazon's or OpenAI's custom chips are genuine hardware ventures or calculated flexes to force price concessions from Nvidia.
Biggest teaching moment ▶ 35:57 Feldman presents Drosophila fruit fly learning speed modelFeldman re-educates Jason on the speed of technological evolution by pointing out that human knowledge transfers across decades-long human generations, whereas AI models iterate across thousands of generational equivalent cycles at the speed of fruit flies.
The host holds their own ▶ 25:21 Jason brings specific enterprise security proof from Palo Alto NetworksJason demonstrates high domain expertise by citing direct conversations with Palo Alto Networks CEO Nikesh Arora regarding AI agents discovering previously unknown critical vulnerabilities during automated code auditing.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| Physical Scale of Data Centers and Cerebras's $25B Backlog | 3 | 5 | 1 | 1 | Jason introduces the massive scale of AI infrastructure buildout, prompting Feldman to explain the physical footprint and power requirements of modern data centers. Feldman reveals Cerebras's $25B backlog, highlighting unprecedented demand that outstrips hardware capacity. | |
| Token Maxing, Costco Analogy, and Enterprise Cost Control | 4 | 5 | 1 | 2 | Jason questions whether token maxing generates real enterprise value or wasteful experimentation. Feldman reframes enterprise AI adoption using a Costco shopping analogy, showing how chaotic initial usage evolves into strategic cost management. | |
| Reasoning Models in Action: BitTensor Autonomous Trend Hunting | 5 | 4 | 1 | 1 | Jason shares his personal testing of autonomous agents via BitTensor that debated trend hunting internally. Feldman identifies this behavior as a reasoning model actively evaluating intent in real time. | |
| Cerebras Wafer-Scale Engine Hardware and Long-Horizon Thinking | 4 | 4 | 1 | 1 | Jason highlights that reasoning requires high inference compute, leading Feldman to showcase physical wafer-scale silicon. Feldman emphasizes how high-speed compute makes long-horizon reasoning tractable. | |
| Crushing Moore's Law and Custom Hyperscaler Silicon | 6 | 5 | 2 | 4 | Jason probes whether hyperscalers building custom chips are just flexing for lower Nvidia pricing or genuinely competing. Feldman explains historical platform dependencies on Intel and GPUs that drive tech giants to control their own destiny. | |
| Open Source AI Progress, Data Sovereignty, and National Security | 6 | 5 | 3 | 3 | Jason advocates for open source models while Feldman draws a distinction between daily routine tasks and frontier model tasks. Feldman interrupts Jason's points about Nvidia to argue that America needs domestic open source champions to counter foreign models. | |
| Frontier Model Governance, Political Polarization, and Red Teaming | 5 | 5 | 2 | 3 | Jason asks whether government red teaming and delayed model releases are political gamesmanship or necessary safety measures. Feldman defends government red teaming by comparing AI safety checks to pharmaceutical trials. | |
| Cybersecurity Vulnerabilities, Black Swans, and AI-Driven Inquiries | 5 | 4 | 1 | 1 | Jason cites Palo Alto Networks discovering unrecorded vulnerabilities using AI reasoning. Feldman warns of inevitable black swan data breaches that the industry must steel itself for in advance. | |
| Reaching AGI, Passing the Turing Test, and Exponential Loops | 5 | 4 | 1 | 1 | Jason argues that AGI has effectively been achieved by definitions established decades ago. Feldman agrees, noting the Turing test has been surpassed and explaining how recursive loop maxing drives exponential model performance. | |
| Organizational Friction, World Models, and Generational Learning Speeds | 5 | 6 | 1 | 1 | Jason explores future limits of AI problem solving while Feldman points out that human organizational friction remains the primary bottleneck. Feldman explains how AI iteration operates at Drosophila fruit-fly speeds compared to human generational timeframes. | |
| Technological Abundance, Curing Cancer, and Personal AI Tutors | 4 | 5 | 1 | 1 | Jason contrasts doom-laden AI narratives with Feldman's optimistic outlook. Feldman argues AI offers concrete opportunities to eradicate cancer deaths and provide Socratic tutors to every child. | |
| Sponsor Read: Nasdaq Global Marketplace Platform | 2 | 5 | 0 | 0 | Jason reads a sponsor message for Nasdaq before introducing Robin Rombach of Black Forest Labs. Rombach explains his background inventing latent diffusion and neural compression algorithms during his PhD. | |
| Intuitive Intelligence, Controllable Generative Media, and Real-World Physics | 4 | 5 | 1 | 1 | Jason asks if video generation proves models understand physical world dynamics. Rombach distinguishes between intuitive visual intelligence and deep reasoning, showing how multimodal pre-training yields implicit physics models. | |
| AI Tools in Filmmaking and Collaborating with Martin Scorsese | 4 | 5 | 1 | 2 | Jason asks how close AI video generation is to creating full feature films like Goodfellas rather than five-second clips. Rombach shares details of sitting with Martin Scorsese to turn mental concepts into visual storyboards. | |
| Human-in-the-Loop Workflows and Historical Pre-Production Analogies | 6 | 4 | 1 | 1 | Jason connects generative AI workflows to classic movie pre-production methods used by Ridley Scott and George Lucas. Rombach agrees that human-in-the-loop creative iteration yields the most compelling results. | |
| AI in Movie Production and Physical AI Robot Integration | 5 | 5 | 1 | 1 | Jason shares an insider account from Gal Gadot using generative backgrounds instead of green screens to cut movie budgets by $120M. Rombach notes how rapidly quality scaled from 64x64 pixel images to multi-minute video models. | |
| World Action Models and Data Strategies for Robotics Training | 5 | 6 | 1 | 1 | Jason asks whether robotics training relies on synthetic generation or real-world video recording with sensory gloves. Rombach explains fine-tuning visual models for action prediction and striving for in-context physical prompting. | |
| Open Source Strategy, Corporate IP Rights, and Fan Content | 6 | 4 | 1 | 1 | Jason asks how IP holders like Disney should navigate open source versus proprietary models. Jason proposes licensing frameworks for fan-created content like Star Wars Untold Stories while Rombach highlights IP filtering safeguards. |
Statements from this episode (20)
Feldman: New data centers will consume more power than past 50 years
“What we're talking about now are data centers that are in the next several years going to use more power than the previous 50 years on earth took.”
Cerebras CEO discloses $25 billion order backlog
“And so, you know, we have a twenty five billion dollar backlog.”
Feldman: Enterprises are shifting from unconstrained token access to cost-optimized AI usage
“I think at first people opened up and said, everybody, as much tokens as you want. And in enterprises, there's no open loop. We don't give sort of any resource unconstrained to people. And now we're jumping on and saying, whoa, all right, these guys should hav…”
Feldman: Modern AI models no longer require precise prompt whispering
“And increasingly, it's understanding what your intent was. And if you have a chance to play with Fable or Five Six from OpenAI, increasingly, what, you don't have to get the prompt just right. You don't have to be a prompt whisperer.”
Calacanis: Unlimited tokens in AI models enable unlimited reasoning
“Unlimited tokens, I believe, means unlimited reasoning.”
Feldman: Cerebras wafer-scale chip cost half a billion dollars to develop
“When one costs half a billion to make,”
Feldman: Cerebras chip gains will exceed 2x over next 18 months
“And my view is in the next 18 months will be way over two X.”
Feldman: Extreme AI demand means no silicon compute goes unused
“Look, I think there is so much demand right now that, that, There is no silicon that will go unused.”
Feldman: Cerebras runs Qwen, Kimi, and custom models for GSK and G42
“We run today, we run GLM, we run Kimi, we run the Quen set of models, and we run OpenAI's models, the closed source ones, we run Models for, say, GlaxoSmithKline, which they wrote and developed. We run models for our partner in the UAE G-forty-two, and MBZ-UAI…”
Feldman: Government red teaming of powerful AI technologies is reasonable
“With a powerful new technology, it certainly doesn't seem unreasonable to say, hey guys let's at least do some red teaming at the government, so we know our defenses can block this.”
Feldman: A massive data leak is inevitable in the AI era
“I think we can also know that there will be a massive data leak.”
Feldman: AI has passed every historical definition of AGI
“I mean, if you think about, oh, there was a Turing test, blew it away. I mean, you think about that, that any period of time, sort of 1015, 20, 3040, 50 years ago, we, any definition we would have previously put forward. We've blown past it.”
Feldman: AI world models will provide behavioral insights by watching humans
“I think that's some of the things the world models are gonna bring us as they begin to watch human behavior.”
Feldman: AI gives humanity a shot at ending cancer deaths
“There's a shot that our children, none of them, nor their people they love will die of cancer.”
Rombach: Pre-training AI models on video gives implicit understanding of physics
“Pre-training on videos gives, like, implicit understanding of the physics of interactions with the real world, and then you can get stuff like action prediction, like robotics out of the same model.”
Robin Rombach: Most Valuable AI Video Uses Require Human-in-the-Loop
“But I think ultimately. Like the real interesting use cases, they come when you have like a human in the loop who iterates and uses it as a medium.”
Calacanis: Portfolio startups now create launch videos in two weeks
“I've seen with a lot of the startups I'm investing in now, They'll just spend a week or two working with you know, a director to make a launch video.”
Rombach: Same generative AI models making movies can power robot brains
“I think what really excites me is you can use the same kind of AI model to make a movie and deploy that as a brain on a robot.”