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

2026-07-10 - 64 min - source - Read full transcript

Jason Calacanis (host)Andrew FeldmanRobin Rombach

Key insights

Cerebras has a $25 billion order backlog, and demand for AI compute is now outstripping the industry's ability to build data centers and fill them with hardware.

Feldman says this is not the usual 'build it and they will come' pattern - customers like OpenAI, Anthropic, Google, and Microsoft are pre-ordering chips before Cerebras finishes building them, effectively trying to capture demand that already existed before supply did.

ai-compute-buildout

Cerebras claims to have broken the traditional Moore's Law doubling cadence and expects to more than 2x performance again in the next 18 months.

Feldman argues newer chip architectures have far more headroom for optimization than a 20-year-old design like the GPU, which is now mostly reliant on shrinking to the next fab node; a young architecture can still find algorithmic and design gains that compound faster than the historical 18-month doubling rate.

ai-compute-buildout

Jason personally shifted spend toward open-source models after finding Moonshot's Kimi indistinguishable in quality from Claude for his own use cases, and began smart-routing between the two.

He frames this as the moment open-source reasoning models closed the gap with frontier closed models in 2026, comparing frontier models to a Ferrari you save for hard problems and open models to a minivan for everyday work.

open-source-ai

Enterprises in regulated industries increasingly want open-source models they can run on-prem and domestically, driven by data sovereignty and data-leakage concerns, not just cost.

Feldman says finance and healthcare customers (subject to HIPAA, FINRA, etc.) are specifically asking for domestic open-source deployments they control, rather than trusting data to closed frontier labs.

open-source-ai

Feldman argues the US currently offers a thin choice in open-source models - OpenAI's OSS-120B or Chinese models - and needs more domestic open-source champions, including from NVIDIA.

He says NVIDIA has been reluctant to promote its own open-source models because doing so would put it in direct competition with hyperscaler customers like OpenAI and Anthropic, leaving a domestic open-source gap that Chinese labs are filling.

open-source-ai

Feldman argues that by any AGI definition proposed 20-plus years ago, including the Turing test, we have already achieved AGI - the open question now is deployment, not existence.

He frames this against the show's recurring point that humans struggle to know what questions to ask of a technology moving this fast, comparing today's AI skeptics to the position science fiction authors would be in if shown current models.

reasoning-and-agi

Recursive self-checking loops - where a model checks its own work, asks what wasn't considered, and re-runs with the new information - produce compounding rather than incremental quality gains, with no observed plateau yet.

Feldman calls this 'loop maxing' and says he has personally watched a reasoning model debate itself over where to search for information; both he and Jason describe not knowing where the exponential curve of quality-per-loop eventually stops.

reasoning-and-agi

Feldman frames reasoning itself as fundamentally an inference-compute problem, which is why fast inference chips like Cerebras's matter more as reasoning models scale.

The more tokens a model consumes internally to reason and verify itself, the more a chip's raw inference speed becomes the bottleneck; running a model for 24-48 hours of internal reasoning at higher speed compresses what would otherwise take much longer into a usable timeframe.

reasoning-and-agi

Black Forest Labs is converging image, video, audio, and robotic action-prediction into a single multimodal model, because generating a realistic video of the world requires implicitly modeling that world's physics.

Rombach says the same architecture that makes images and video can be fine-tuned with a relatively small amount of task-specific data to predict physical actions, which is why the company sees a direct line from generative media to real-world robotics.

generative-video-and-ip

Rombach frames Black Forest Labs' partnership with Martin Scorsese as proving the near-term value of generative video is pre-visualization, not fully automated filmmaking.

He describes sitting with Scorsese while the director explored a scene concept (a village in Eastern Europe), iterating on generated images until the mental picture in Scorsese's head was on screen; Rombach is explicit he doesn't see full generated feature films as the near-term goal, but a human-in-the-loop creative medium.

generative-video-and-ip

Black Forest Labs' IP strategy is to block certain protected characters and franchises from public generation by default, while offering rightsholders licensed or custom co-developed models built on their own content.

Rombach cites this as the model he'd recommend to a major IP holder like Disney, contrasting it with OpenAI's more public dispute with rightsholders over Sora-generated characters.

generative-video-and-ip

Feldman defends the government-mandated staged rollout of Anthropic's frontier model as reasonable red-teaming, not partisan overreach, citing Palo Alto Networks CEO Nikesh Arora's account that early testing exposed critical unpatched security vulnerabilities.

He compares it to phased approval for powerful new pharmaceuticals, arguing that once a model is capable enough to pose a plausible cyberattack risk, giving defenders weeks to patch known holes before broad release is not unreasonable - while acknowledging political polarization makes it hard to evaluate such moves on the merits alone.

ai-safety-regulation

Media referenced

Goodfellas - movie - Jason's benchmark for what a finished AI-made film would need to match; used to contrast today's five-to-ten-second AI clips with a real feature.
Star Wars Stories Untold - other - Jason cites this fan-made YouTube series, built with generative AI on unofficial Star Wars stories, as already pulling millions of views per video.
Star Wars franchise (original trilogy era storyboards/miniatures) - movie - Jason references George Lucas's use of storyboard artists and miniatures, and Lucas's permission for non-commercial fan films, as a precedent for AI-assisted fan creativity.
Aliens / Gladiator - movie - Cited as films where director Ridley Scott was known for drawing his own storyboards, an analog precedent for AI-assisted pre-visualization.
Raiders of the Lost Ark - movie - Cited as a film Spielberg reportedly sketched himself, another storyboard precedent for AI pre-visualization.
Untitled Bitcoin movie (Gal Gadot) - movie - Jason relays Gal Gadot telling him the production shot on a soundstage with AI-generated backgrounds instead of green screen, cutting the budget from a reported $150M to $30M.

Companies

Cerebras - Andrew Feldman's company; makes wafer-scale inference chips, has a $25B customer backlog, and claims to be outpacing traditional Moore's Law doubling.
Black Forest Labs - Robin Rombach's company; makers of the open-source Flux image model, now building multimodal image/video/audio/action-prediction models, based in Freiburg, Germany and San Francisco.
OpenAI - Referenced for its OSS-120B open-weight model release, its Sora video tool's licensing dispute with rightsholders, and as a frontier-model competitor to Anthropic.
Anthropic - Referenced via Dario Amodei and the staged government-mandated rollout of its 'Fable' model, discussed as a case study in AI safety regulation.
NVIDIA - Jensen Huang's reluctance to publicize NVIDIA's own open-source models discussed as a case of not wanting to compete with hyperscaler customers.
GlaxoSmithKline - Runs its own proprietary models on Cerebras's infrastructure, cited as an example of enterprise-owned model deployment.
G42 / MBZUAI - Cerebras's UAE partners, running their own custom-designed models on Cerebras hardware.
Palo Alto Networks - CEO Nikesh Arora told Feldman that testing an early Anthropic-class model against their security software surfaced critical bugs the team had to patch for six weeks.
Disney - Used as the example major IP holder Jason asks Rombach to advise on open-source vs. custom model training strategy for protecting franchise characters.
AppLovin - Sponsor read: ad-tech platform for mobile game and e-commerce advertising.
Nasdaq - Sponsor read: positioned as the technology platform powering global capital markets infrastructure.

Techniques and frameworks

Latent diffusion - The core algorithm Rombach and his co-founders invented (as PhD students, later used in Stable Diffusion) that compresses images/video/audio into an efficient representation for transformer training.
Loop maxing / recursive self-improvement - Feldman's term (coined live on the show) for models that ask a question, check their own work, and re-run with more information, producing compounding rather than incremental gains.
Token maxing - Feldman's term for early, undisciplined over-consumption of AI compute inside enterprises, likened to shopping every aisle at Costco before learning to shop strategically.
Action prediction / world models - Black Forest Labs' approach of training a single multimodal model to generate images/video and predict physical actions, on the premise that realistic video generation requires implicitly modeling real-world physics.
Red teaming / staged model rollout - Feldman's defense of the government asking an AI lab to release a powerful model in controlled stages so defenders can patch vulnerabilities first, compared to phased pharmaceutical approval.

Summary

Jason Calacanis runs a two-part solo interview with Andrew Feldman, CEO of Cerebras, and Robin Rombach, CEO of Black Forest Labs, covering the infrastructure and creative-media sides of the current AI buildout.

With Feldman, the conversation opens on scale: Cerebras is sitting on a $25 billion order backlog, and demand for inference compute is outstripping the industry's ability to build and fill data centers, an inversion of the usual "build it and they will come" pattern. Feldman claims Cerebras has broken the traditional Moore's Law cadence for chip performance and expects to more than double again in the next 18 months, attributing the headroom to running a young architecture rather than a 20-year-old GPU design. The two dig into why hyperscalers are building their own chips (control over destiny, not just cost) and pivot to open source, where Jason recounts discovering that Kimi matched Claude's output quality closely enough to start smart-routing between them - part of a broader theme that regulated-industry customers now want domestically controlled, on-prem open models, and that the US currently only offers OpenAI's OSS-120B against a wave of capable Chinese alternatives.

The discussion turns philosophical on AGI and recursive self-improvement. Feldman argues that by any pre-2020s definition, AGI has already arrived, and frames "loop maxing" - models checking their own work and re-running with new information - as producing compounding rather than incremental gains whose ceiling is not yet visible. He connects this directly back to hardware: since reasoning consumes enormous internal token volume, fast inference chips become more valuable, not less, as models get smarter. The segment closes on AI safety and regulation, where Feldman defends the White House's staged rollout of Anthropic's frontier model as reasonable red-teaming rather than partisan interference, citing Palo Alto Networks CEO Nikesh Arora's account that early access testing surfaced critical unpatched vulnerabilities. Both agree a major AI-enabled security breach is inevitable and the right posture is preparation, not prevention.

The second half shifts to Robin Rombach and Black Forest Labs, the open-source image lab behind Flux, now working on multimodal models that unify image, video, audio, and physical action prediction under one architecture - the throughline being that generating realistic video requires implicitly modeling real-world physics, which is the same capability needed for robotics. Rombach describes Black Forest Labs' partnership with Martin Scorsese, framing the near-term value of generative video as pre-visualization: getting a director's mental image onto the screen for iteration, not replacing full film production. Jason presses on cost disruption already visible in production, citing Gal Gadot's account of an AI-background Bitcoin film that reportedly dropped from a $150 million to $30 million budget by skipping physical sets, and the "Star Wars Stories Untold" fan-film series racking up millions of views using AI-recreated unofficial Star Wars stories.

The episode closes on IP strategy: Rombach's approach is to block certain protected characters from public generation by default while offering rightsholders licensed or jointly developed custom models, which Jason positions as the template he'd recommend to a major library owner like Disney. Both guests come across as bullish but not hype-driven, repeatedly grounding claims in specific customer behavior (backlog size, enterprise sovereignty demands, a real director's working process) rather than abstract futurism.

Notable Quotes

"The intelligence is getting so much better every step along the way that I'm watching individuals... they start playing with the tool, and then the tool starts playing with them." - Andrew Feldman

"By any definition we had 20 years ago, we've hit it. There was a Turing test, blew it away." - Andrew Feldman

"You don't want to take your Ferrari to the grocery store... there's minivan time." - Andrew Feldman

"Language ultimately is a little bit of a lossy communication medium... visual information is so rich, there's so much signal in it, and it's just another way of communicating." - Robin Rombach