All-In: Open Source Wins, AGI Is Here, and Scorsese’s AI Toolkit

概览

This episode is split between two AI infrastructure and model-building conversations: first with Cerebras CEO Andrew Feldman, then with Black Forest Labs CEO Robin Rombach. The first half focuses on the massive physical buildout behind AI, the economics of inference, open-source models, AI safety, and whether current reasoning systems already meet older definitions of AGI.

Feldman argues that demand for compute is already far ahead of supply, with AI companies and hyperscalers racing to build data centers and secure silicon. The discussion frames reasoning as inference-heavy work, where faster chips and more tokens can turn long-running AI agents into systems that produce dramatically better answers.

The second half turns to visual and multimodal AI. Rombach explains Black Forest Labs’ roots in latent diffusion and Stable Diffusion, then describes a future where image, video, audio, action prediction, robotics, and world models converge into shared multimodal systems. The conversation also covers Martin Scorsese using generative tools for visual ideation, production workflows, IP licensing, fan films, and hiring.

分段落总结

[00:00] AI Infrastructure Buildout

[事实] The episode opens by describing the AI buildout as a historic mobilization of capital, talent, data centers, chips, and power. [事实] Andrew Feldman says data centers being built now can be the size of football fields and draw more power than midsize cities. [事实] He says major AI companies and hyperscalers, including OpenAI, Anthropic, Google, Microsoft, AWS, and xAI, are demanding more capacity.

[03:20] Demand Is Already Ahead Of Supply

[事实] Feldman says Cerebras has a $25 billion backlog and that demand is outstripping the industry’s ability to build data centers and fill them with hardware. [事实] He frames the market as chasing “yesterday’s demand,” not speculative future demand. [推测] The conversation presents AI infrastructure less as a normal tech cycle and more as a capacity-constrained industrial race.

[04:27] Token Maxing And Enterprise Value

[事实] The hosts discuss “token maxing,” where users consume large amounts of AI inference. [事实] Feldman compares current AI experimentation to the early AWS era, when companies allowed engineers broad access before learning where the value really was. [事实] He says some AI usage will be wasteful, but that does not mean the net value is not enormous. [事实] He expects enterprises to route workloads by need, using frontier models for hard problems and cheaper or open-source models for ordinary tasks.

[06:24] AI Rewards Systems Thinking

[事实] Jason says AI tools are exposing whether users have clearly defined goals, requirements, and systems. [事实] Feldman says older computers did exactly what users told them, while newer models increasingly infer user intent. [事实] They discuss newer reasoning models that can suggest better charts, strategies, or follow-up questions even when the user did not specify them exactly. [推测] The discussion implies that AI adoption may reward users who can define goals and workflows clearly, not just users who write clever prompts.

[08:34] Reasoning Agents And Unlimited Tokens

[事实] Jason describes using a BitTensor-related project with GLM 5.2 to run a trend-hunting task where the system debated sources such as Hacker News, Reddit, social media, and Instagram. [事实] Feldman identifies this as a reasoning model working through a problem. [事实] Jason says the experience made him think of unlimited tokens as unlimited reasoning. [事实] Feldman says running reasoning systems for 24 to 48 hours can produce amazing results, and that faster Cerebras compute could compress weeks or months of thinking into shorter elapsed time.

[11:08] Inference And Cerebras’ Hardware Bet

[事实] Feldman says reasoning is inference and is computationally intensive because it consumes many internal tokens. [事实] He says fast compute makes reasoning workloads more practical by reducing the time needed to get good answers. [事实] Feldman says traditional processors followed Moore’s law, while Cerebras’ newer architecture has room to improve faster than a mature GPU architecture. [事实] He expects Cerebras to exceed 2x improvement over the next 18 months.

[14:03] Why AI Companies Build Their Own Chips

[事实] Jason asks whether OpenAI, Amazon, and others are building chips to gain pricing leverage or because they want to enter the chip business. [事实] Feldman says nobody likes being dependent and points to lessons hyperscalers learned from dependence on Intel and GPU makers learned from dependence on a small number of hyperscalers. [事实] He says companies may not need to build the fastest chip, but they do not want to be entirely dependent on others’ chips. [推测] The segment frames custom silicon as a sovereignty and bargaining strategy as much as a pure performance strategy.

[16:11] Open Source Models And Sovereignty

[事实] Jason says open-source models such as Kimi narrowed the gap with frontier systems enough that he could not always tell the difference for some tasks. [事实] Feldman says sophisticated users will choose frontier models for hard problems and open-source models for many ordinary enterprise workflows. [事实] The discussion names GLM, Kimi, Qwen, OpenAI models, GlaxoSmithKline models, and UAE partner models as examples Cerebras runs. [事实] Feldman says sovereignty is a trend and that the United States needs more domestic open-source models to give the world a choice.

[20:09] Model Release Safety And Government Red Teaming

[事实] Jason asks about whether powerful model releases should be staged because of cyber risk. [事实] Feldman says it does not seem unreasonable for government to ask for staged rollout and red teaming when a model becomes powerful enough to pose a meaningful threat. [事实] He says political polarization hurts clear thinking about AI safety. [事实] Feldman says guardrails can add latency, and fast chips can make those guardrails less painful.

[25:22] Cyber Risk And Inevitable Breaches

[事实] Jason says Palo Alto Networks found unknown bugs when testing advanced AI against its software, requiring weeks of patching. [事实] Feldman says when a model finds critical openings quickly, it shows the tool is powerful. [事实] They discuss beta testing, red teaming, and the likelihood that a major data breach will happen at some point. [事实] Jason says he now asks AI systems to check their work and tell him what goals or questions he has not considered.

[28:27] AGI And Better Questions

[事实] Jason asks whether AGI has already been reached but not fully deployed. [事实] Feldman says that by definitions from 20 years ago, AI has already passed AGI benchmarks such as the Turing test. [事实] They say older science fiction questions have largely been answered by current systems. [推测] The segment suggests that the definition of AGI keeps moving because each new capability changes what people consider meaningful intelligence.

[31:00] Recursive Learning And Loop Maxing

[事实] Jason introduces “loop maxing” as the idea of repeatedly feeding outputs back into AI workflows. [事实] Feldman says recursive gains can be exponential: ask a question, learn from the result, ask again, and get a vastly better answer. [事实] He says it is unknown whether better answers eventually plateau or keep improving with more compute and tokens. [事实] They also discuss that some future challenges may become people and coordination problems rather than intellectual problems.

[34:12] AI In The Physical World

[事实] The speakers discuss AI moving from screens into the real world through robots and world models. [事实] Jason imagines physical construction tasks, such as building a new version of Versailles, becoming possible with many autonomous machines. [事实] Feldman compares AI learning speed to biological generation cycles and says AI compresses learning across the equivalent of thousands of generations. [推测] The discussion treats physical AI as a future stage where recursive digital learning meets real-world execution.

[37:44] Abundance Versus Dislocation

[事实] Feldman says AI could help create a future where children and people they know do not die of cancer. [事实] He acknowledges economic dislocation and compares it to earlier technological changes such as cars replacing horse-related work. [事实] He lists potential benefits including energy, calories, knowledge, education, housing, and personalized tutoring. [事实] He argues that AI could allow every child to learn through a personalized tutor rather than a one-size-fits-all classroom.

[40:52] Black Forest Labs And Latent Diffusion

[事实] Robin Rombach introduces Black Forest Labs as based in Freiburg in the Black Forest and San Francisco. [事实] He says the company started two years ago and that he and his co-founders worked on Stable Diffusion. [事实] He says they invented latent diffusion, which compresses natural data such as images, video, and audio into more efficient representations before training models. [事实] He describes Flux as one of the company’s known open-source models.

[42:51] Multimodal Models And Action Prediction

[事实] Rombach says Black Forest Labs is working on multimodal visual models trained on images, videos, and audio. [事实] He says the next paradigm combines those models with action prediction so the same model can generate media and eventually be deployed on robots. [事实] He describes intuitive intelligence and deep reasoning as complementary forms of intelligence. [事实] He says video pretraining can give models implicit understanding of physics and real-world interactions.

[44:50] Control Layers For Generative Media

[事实] Jason asks how generative image and video systems move beyond slot-machine-style outputs. [事实] Rombach says the solution is exposing more manipulation layers to users and developers. [事实] He describes the progression from text-to-image to image-plus-text editing, then to combining multiple images and prompts. [事实] He says the same principle is now applying to video and becomes more interesting when many modalities are inputs and outputs of the same model.

[46:31] Scorsese And AI As A Creative Medium

[事实] Jason asks about Black Forest Labs’ work with Martin Scorsese. [事实] Rombach says AI models are a medium and that Black Forest Labs does not want to prescribe how filmmakers use them. [事实] He says Scorsese used the tools to explore a visual idea, including a village in Eastern Europe, and iterate on outputs. [事实] Rombach says a key value is getting a mental picture out of someone’s head and communicating it visually. [推测] The partnership is framed more as professional ideation and visual communication than as replacing a director with an automated movie generator.

[50:29] Storyboarding And Startup Videos

[事实] Jason connects generative AI to storyboarding and cites directors such as Ridley Scott, Spielberg, and George Lucas as examples of visual planning traditions. [事实] He says startups that once might have spent $100,000 to $250,000 on launch videos can now spend a week or two working with a director and AI tools. [事实] Rombach says Black Forest Labs supports exploration of many kinds of launch videos and products built on the same base models. [推测] The segment presents AI video as an immediate cost and speed advantage for early-stage storytelling.

[52:25] Production Use And High-End Film

[事实] Jason describes a Bitcoin movie where actors worked on a sound stage while generative AI created scenery, saying the budget was $30 million rather than a possible $150 million with physical sets. [事实] Rombach says Black Forest Labs sees some production use cases, though high-end film is one of the most demanding areas. [事实] He says the technology has progressed from 64-by-64-pixel images during his PhD years to high-resolution, multi-minute videos. [事实] He declines to predict exactly when high-end film use will fully arrive, but says the technology will keep improving.

[54:23] One Model For Movies And Robots

[事实] Rombach says the same kind of AI model could be used to make a movie and serve as the brain of a robot. [事实] He connects world models, world action models, computer use, generation, prediction, and perception as related parts of the same technology direction. [事实] He says action prediction requires understanding visual input well enough to predict a reasonable next action. [推测] This suggests Black Forest Labs sees media generation as part of a broader path toward embodied and interactive AI, not just content tools.

[56:32] Robotics Data And Fine-Tuning

[事实] Jason asks whether robots will learn from first-person human data, YouTube videos, synthetic data, or some combination. [事实] Rombach says the goal is to prompt robots in context, similar to prompting a language model, but that systems are not there yet. [事实] He says current robots use different hardware and action representations, so models need task-specific adjustment. [事实] He says a model with broad visual understanding may only need a few hours of fine-tuning data for a specific task, with the longer-term goal of more in-context operation.

[58:01] Open Source, IP, And Content Libraries

[事实] Jason asks how major IP holders such as Disney should think about open-source models, owned content libraries, and model control. [事实] Rombach says the most interesting content-creation use case is making something new that has not existed before. [事实] He says Black Forest Labs’ public tools prevent generation of certain IP. [事实] He says the company works with some IP holders to develop models, based either on open-source models or more powerful provider models.

[60:27] Fan Films And Licensed Creativity

[事实] Jason discusses fan fiction, Star Wars fan films, and AI-generated untold Star Wars stories gaining millions of YouTube views. [事实] He suggests IP owners could let fans pay licensing or software fees to create their own stories with protected characters. [事实] Rombach agrees that a model working for IP owners while enabling creative customization would be valuable. [事实] He says generative tools can help people visualize alternate ideas they have after reading a book or watching a movie.

[62:04] Black Forest Labs Hiring

[事实] Rombach says Black Forest Labs has crossed 100 people and is hiring in Germany and San Francisco. [事实] He says the company is looking for researchers with large-scale model training, diffusion, and flow-matching experience. [事实] He says they also want engineers for customer work, customized physical AI solutions, IP-owner partnerships, large-scale compute infrastructure, and getting the technology into users’ hands.

播客点评/总结

[推测] The strongest part of the episode is the bridge between compute infrastructure and model capability. The Cerebras conversation makes inference, reasoning, tokens, latency, data centers, and safety feel like one connected system rather than separate AI talking points.

[推测] The Black Forest Labs interview is valuable because it grounds visual AI in concrete workflows: storyboarding, Scorsese’s ideation process, startup launch videos, production scenery, IP-controlled models, and future robotics. It is especially useful for listeners tracking how image and video generation may evolve into world models.

[推测] The limitation is that both interviews are highly optimistic and founder-facing. Job loss, safety failures, copyright conflicts, and governance are discussed, but mostly at a high level rather than through adversarial detail or counterarguments.

[推测] This episode is best suited for listeners interested in AI infrastructure, open-source model strategy, multimodal generation, robotics, creative tooling, and the business implications of model sovereignty.