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

The supplied transcript contains two consecutive technology interviews, separated and interrupted by sponsor material.

1. Andrew Feldman interview

• Guest identified in the transcript as Andrew Feldman, CEO and founder of “Cerebrus” [transcript spelling; verify company name].
• Main subjects: AI infrastructure buildout, inference compute, reasoning models, token consumption, open-source models, sovereignty, AI safety and government regulation, cyber risk, AGI, recursive learning, superintelligence, world models and prospective social abundance.
• The interviewer is addressed as “Jason” later in the conversation, but the transcript does not state his full name.

1. Robin Rombach interview

• Guest identified as Robin Rombach, co-founder and CEO of Black Forest Labs.
• Main subjects: latent diffusion, multimodal models, image/video generation, control over generative output, collaboration with Martin Scorsese, filmmaking, generative-AI production economics, world/action models, robotics, IP licensing, fan-created media and Black Forest Labs hiring.

1. Sponsor material

• An AppLovin advertisement appears near the opening.
• A Nasdaq advertisement appears between the Feldman and Rombach interviews.
• These claims are retained separately below as advertising claims and should not be treated as interview evidence.

The transcript is generally intelligible but contains many probable automated-transcription errors, including company, product and person names. Examples include “Cerebrus,” “Cerebra's,” “Srebris,” “Open Air,” “OpenSwerce,” “Chat Chaptapit,” “perplexy,” “Morris Law,” “ATWS,” “Fable,” “five, six,” “GL-5-2,” “QM's,” “Galaxo Smith-Kline,” “MBZ UAI,” “Martin Scassisi,” “Martin Svorsese,” “Gal Godot,” “Uri Milner,” and several grammatical fragments.

No uncertain term has been silently corrected in this record. Where interpretation would require guessing, the issue is identified for verification.

The host introduces Feldman as a returning guest, describing him as the CEO and founder of “Cerebrus,” a company focused on inference chips, and referring to a successful IPO, a meeting in Davos in January and another recent meeting at “liquidity” [transcript wording; verify].

The host frames the current AI buildout as extraordinary in both capital and human effort. He compares its scale with the construction of the Great Wall of China, the pyramids and, within more recent history, a wartime mobilization. His point is that unusually large amounts of money, time and highly capable people are being concentrated on a single technological project, while Feldman is directly involved because his customers are constructing data centers and his hardware is part of that infrastructure.

The inserted advertisement claims that:

• “App Loven” began with an $8 domain and no venture-capital funding.
• It subsequently became one of the world's largest advertising platforms.
• The same advertising engine is now being applied to e-commerce.
• Ads are said to appear inside mobile games, reaching more than one billion people with full-screen attention.
• The platform is presented as automatically finding buyers and optimizing campaigns for profitability once the advertiser sets a target.
• A cookware brand is cited as having grown from $4 million to $16 million, becoming profitable and being “on pace for $80 million this year.”
• Listeners are directed to an AppLovin URL to launch a campaign.

These are sponsor claims and require independent verification before editorial use.

Asked what his company is doing in 2026 and about large construction activity including Texas, Feldman emphasizes that AI infrastructure is unusually physical for a technology industry normally discussed in terms of chips, servers or software.

Feldman claims that data centers being built over the next several years will consume more power than “the previous 50 years on earth took” [claim and wording require verification]. He describes individual buildings as being approximately the size of football fields while receiving more incoming power than midsized cities.

He says substantial data-center construction is taking place:

• across the United States;
• in Canada;
• throughout the Nordic countries;
• in Paris and elsewhere in France;
• elsewhere in Europe;
• across the Middle East;
• and in countries including Kazakhstan, Tajikistan, Georgia and Armenia.

His broader point is that governments and regions increasingly believe they must participate in the AI infrastructure race.

Feldman describes buyers of AI compute as “insatiable.” The transcript mentions OpenAI, Anthropic, “SpaceX AI” [verify], Google and others.

The host notes that customers had reportedly placed chip orders with Feldman's company before the relevant chips had even been completed.

Feldman argues that this differs from speculative technology expansion. In his description, suppliers are not constructing infrastructure in the hope that demand eventually arrives. Instead, available construction and manufacturing capacity is trying to catch up with demand that already exists.

Feldman states that his company has a $25 billion backlog.

He says the company is not alone in experiencing extreme demand and points to OpenAI, Anthropic, Google, Microsoft and “ATWS” [transcript unclear] as organizations seeking additional data-center capacity.

He characterizes the central commercial problem as retaining already-booked demand rather than persuading customers to consume capacity later.

Both speakers consider that situation highly unusual.

The discussion then introduces the term “token maxing” for heavy or unconstrained use of AI inference.

The host raises the central question of whether unprecedented compute demand necessarily demonstrates real economic value.

Feldman answers with a qualified position:

• There is, in his view, clearly enormous value being generated.
• There is also enormous experimentation.
• Some experiments will produce little or no lasting value.
• The existence of waste does not invalidate the much larger net value.

He compares the current period with the early adoption of AWS. Cloud infrastructure allowed engineers to bypass cumbersome internal IT processes, so companies encouraged engineers to obtain accounts and experiment freely. Much of that use was valuable, while some later looked unnecessary or regrettable.

His point is that an initially permissive period can still produce major net gains even when some spending is wasteful.

Feldman offers a second analogy based on Costco opening in the Palo Alto area around 1988.

He recalls that inexperienced customers initially shopped at Costco as if it were Safeway: walking through every aisle and buying items simply because they were available. This produced unnecessary purchases.

As customers learned how the store worked, he says they became more strategic—for example, entering specifically for a chicken or a large order of cupcakes and then leaving.

He applies this to enterprise AI.

At first, organizations effectively told users to consume as many tokens as they wanted. Feldman argues that businesses normally do not leave resources unconstrained forever.

Companies are now learning to discriminate between workloads:

• some employees or applications are productive enough to justify as much compute as they require;
• some jobs can be routed to open-source models;
• some can be handled by cheaper models.

The shift, in his framing, is from experimentation toward ordinary business optimization.

The host describes a type of highly effective AI user beginning to emerge.

He associates effective AI deployment with systems thinking, something he says developers often possess naturally and CEOs often develop through strategy and organizational leadership.

He says he has watched startup founders, venture capitalists and associates at his venture firm begin using AI only to discover that the AI exposes weaknesses in their own thinking.

Examples include the user realizing:

• the goal has not been clearly defined;
• the problem has not been expressed as a system;
• no requirements document exists;
• too much compute is being used without enough focus.

The host personifies the AI as effectively asking the user for a requirements document, defining the objective and pointing out that the person is “token maxing.”

He recalls a computer scientist telling him approximately 20 years earlier that computers were “really dumb” because they did exactly what they were told.

Early prompting resembled that model: a small prompt change could produce a dramatically different answer.

The speakers argue that newer models increasingly infer the user's intent, reducing the need for extremely precise prompting.

The conversation refers to “Fable or 5-6 from Open A.” [transcript unclear; verify model/product names].

The speakers argue that advanced systems increasingly remove the need to become a “prompt whisperer.”

Instead of simply following a narrow instruction, a model may infer that the user actually wants additional information or a better representation—for example, deciding that a chart would work better with both a line and a bar even though the user did not explicitly request that combination.

The host characterizes this as a major leap: within roughly 24 months, systems have progressed in his view from strong summarization and web research toward understanding intended outcomes and proposing solutions without requiring the human to specify every implementation detail.

The host describes an experiment performed that morning with a “Hermes agent.”

He says he was given access to what he calls a secret BitTensor project using a new “ZAI model, 5-2,” which is subsequently referred to as “GL-5-2” and “GLM-5-2.” Exact names require verification.

He describes BitTensor as a distributed crypto project.

According to the host, an unnamed source told him that the system had substantial spare capacity, possibly involving capacity in China with “free energy.” This is hearsay/speculation and should not be presented as established fact.

Given what he describes as effectively unlimited capacity, he instructed the agent to run an unusual recurring task: every hour, identify world trends that nobody else had recognized yet. His stated goal was to become “the smartest trend hunter in the world.”

He says he observed the model reasoning about its own research strategy. It considered Hacker News and Reddit, then reasoned that social-media sources such as Instagram might reveal trends earlier.

Feldman identifies what the host observed as a reasoning model working through a problem.

The host says the experience changed his understanding of what a world with effectively unlimited tokens or inference might look like, particularly for users accustomed to much earlier ChatGPT generations.

He proposes the idea that unlimited tokens could mean unlimited reasoning.

Feldman largely agrees with the direction of the argument and says running systems for 25 or 48 hours can already produce remarkable results.

The host then asks what happens if Feldman's hardware makes that computation approximately 15 times faster and the model still runs for 24 hours: conceptually, a system might compress what he characterizes as weeks or months of “thinking” into a much shorter period.

Both speakers treat the prospect as extraordinary.

Feldman reflects that people such as “Ilya” and “Sam” had predicted this development years earlier and, in retrospect, appear to have seen further around the technological corner than many observers.

The host recalls an earlier All-In conversation with Sam [full identity not stated in this transcript] in which he asked what would come next.

According to the host's recollection, Sam answered: reasoning.

The host says the explanation involved:

• understanding the user's intended outcome;
• devising a strategy;
• potentially consulting other agents or separate threads;
• testing whether a proposed approach is appropriate;
• and having systems vet one another's work.

The host contrasts this with earlier views of language models as systems that merely guessed the next word, completed a sentence or summarized a PDF.

Feldman argues that reasoning is fundamentally an inference workload and that it is computationally expensive.

Fast inference therefore matters because it allows models to spend substantial internal computation on reasoning without forcing users to wait prohibitively long for an answer.

His argument is that systems producing better answers through large internal token consumption create particular value for extremely fast inference hardware.

During the discussion, Feldman produces one of his company's chips or hardware artifacts.

He jokes that if something costs approximately half a billion dollars to make, one carries it around.

The host recalls the object being passed around at Davos.

Feldman says the particular unit was among the first eight or ten and therefore has sentimental or historical significance to him.

He jokes that his wife compares his attachment to it with a child receiving a dirt bike for an eighth birthday.

The host asks for the equivalent of Moore's law in inference and whether the company thinks in terms of doubling performance over a particular period.

Feldman says traditional processors followed Moore's law, characterized in the conversation as approximately doubling every 18 months.

He claims his company's architecture broke away from that trajectory and that its current chip produced much larger gains.

Feldman predicts that during the next 18 months, the company's improvement will be well above 2×.

His reasoning is architectural:

• a mature architecture such as the GPU, which he describes as roughly 20 years old, has fewer easy opportunities for radical improvement;
• mature systems increasingly depend on manufacturing improvements such as smaller geometries and moving to the next fabrication node;
• a newer architecture still contains large opportunities to study real workloads and make architectural optimizations;
• those optimizations can, in his view, deliver gains substantially above the traditional Moore's-law trajectory.

These are Feldman's claims and projections, not independently verified performance results.

The host asks how Feldman runs the company under conditions in which it allegedly has $25 billion of demand and must simultaneously:

• deploy systems rapidly;
• hire people;
• establish a product roadmap;
• evolve hardware;
• evolve software;
• handle deployments;
• and keep pace with customers such as OpenAI that themselves move extremely quickly.

Both speakers describe these AI companies as highly demanding customers rather than passive buyers.

The host also raises the possibility that major customers could eventually become competitors.

The host cites OpenAI releasing something transcribed as “Halapeno” and Amazon making its own chips, asking whether these projects are:

• negotiating leverage against Nvidia and other suppliers;
• demonstrations that customers can build alternatives themselves;
• or the beginning of those customers competing directly with companies such as Feldman's.

Feldman argues that the primary motivation is avoiding dependency.

He points to earlier industry structures:

• hyperscalers had once been highly dependent on Intel in the x86 era;
• GPU manufacturers, in turn, became dependent on a relatively small number of hyperscale buyers;
• GPU firms therefore had an incentive to broaden their customer base, including through support for emerging “neoclouds.”

His conclusion is that organizations do not necessarily need to manufacture the world's fastest chip themselves. They do, however, want enough internal capability to avoid becoming completely dependent on another company's silicon.

He considers control over at least part of one's technological destiny a reasonable strategic objective.

The host says he has become an advocate for open source.

He recounts using “OpenClaw” and “Kimmy” [verify names] and being unable to distinguish the open model's quality on some tasks while consuming fewer premium-model resources.

He says his organization began smart routing tasks among models.

The speakers use a vehicle analogy:

• a frontier model is like a Ferrari;
• a Ferrari is appropriate for difficult or special tasks;
• ordinary errands do not require it;
• many workloads call for a “minivan.”

Feldman predicts more sophisticated users will distinguish between:

• hard problems suited to frontier models from OpenAI, Anthropic or Gemini;
• and ordinary business processes suitable for stable, lower-cost open models.

As an example, the speakers point to large quantities of corporate work consisting essentially of copying information from one place to another, including moving information out of Workday into another cell or system.

Feldman argues that this “cutting and pasting economy” does not require “gold medal” mathematical reasoning.

Feldman argues that many routine business functions need reliable open-source capabilities, not the world's most sophisticated agents.

He points specifically to “GNA” [likely an acronym but transcript should be verified] and says a large amount of administrative or general corporate work is not invention.

Therefore, frontier-level intelligence may be unnecessary for a substantial fraction of enterprise workloads.

The host identifies another reason organizations may prefer local or open models: control.

He says organizations increasingly worry about:

• the ambitions or power of frontier-model providers;
• sharing sensitive data;
• data leakage;
• sovereignty over their own intelligence or information;
• regulatory restrictions.

He gives finance and healthcare as examples and references HIPAA and FINRA.

Such organizations may prefer:

• on-premises deployment;
• domestic infrastructure;
• open-source models;
• and greater control over the complete system.

Feldman says his company is indeed seeing this demand.

Feldman says “Open Air” [verify] made a good decision by releasing “OSS 120B” some months earlier.

He considers it a good open-source model but argues that the United States needs more domestic open-source models so global users have meaningful choices.

He characterizes the present choice as largely between OSS 120B and Chinese models, while acknowledging that Nvidia has open-model efforts as well.

The host discusses an apparent strategic tension for Nvidia: promoting its own models could place it in competition with major customers associated with Sam, Dario, Elon and Sergey.

The host nevertheless argues that more US open-source champions are needed and notes that open software can be forked.

The broader argument is that infrastructure providers can occupy a relatively neutral position by supporting many different model developers.

Feldman lists a broad model portfolio, saying the company currently runs:

• GLM;
• Kimmy;
• a “QM's set of models” [transcript unclear];
• closed-source OpenAI models;
• models created by “Galaxo Smith-Kline” [verify];
• models developed by partner “G-42” in the UAE;
• and “MBZ UAI” models [transcript unclear].

He uses this diversity to reinforce the idea that sovereignty is becoming an important trend rather than all customers converging on one model provider.

The conversation shifts to government action involving products transcribed as “Fable and five, six.”

Feldman says the government's apparent willingness to stop, think and then act became a wake-up call, particularly in Europe.

The host introduces US political polarization.

He characterizes “Dario” as politically or institutionally opposed to the current administration and says both sides have acknowledged an adversarial relationship, though he believes they are beginning to work through it.

The host explicitly acknowledges that outsiders cannot easily distinguish among:

• genuine safety concerns;
• partisan conflict;
• negotiating strategy;
• and political gamesmanship.

He asks Feldman whether the relevant model release genuinely created a major cyber-warfare or cyberattack threat and whether staged rollout would be prudent.

The host also criticizes Dario's communication style, calling him “effervescent” in what he presents as diplomatic wording. This is the host's characterization, not a factual assessment.

Feldman says he had not personally seen the relevant model before the controversy and therefore cannot determine whether that particular timing or intervention was justified.

He instead answers the general principle.

If a model becomes sufficiently creative or capable that it poses a meaningful threat, he considers it reasonable for a government to request a staged rollout.

He compares this concept with controls used for powerful pharmaceuticals, while explicitly rejecting the idea that AI should inherit what he regards as excessive FDA-style timelines and paperwork.

His narrower proposal is limited government red-team time before an unrestricted release.

Questions he believes are reasonable include:

• Can government defenses stop the model's capabilities?
• Has national infrastructure been tested?
• Has NSA infrastructure been tested?
• Can the government have two or three weeks to patch obvious vulnerabilities discovered during testing?

Feldman therefore considers a limited request for time to assess and patch major security holes reasonable in principle.

Feldman argues that political identity can distort assessment of an otherwise reasonable policy.

He asks observers to imagine the same action being taken by President Trump, by a hypothetical “President O-A-O-C,” or by someone between those political extremes.

His conclusion is that intense polarization “hurts clear thinking.”

The speakers agree that any political side can make both intelligent and foolish decisions.

Feldman says his experience with government rank-and-file employees is that many are trying very hard to solve these problems despite the speed of technological change.

He advocates evaluating AI policy on its merits rather than primarily through partisan identity.

Both speakers strongly support continued competition among leaders such as Dario and Sam.

They argue that aggressive competition:

• pushes technology forward;
• forces each company to improve;
• demonstrates to entrepreneurs that even organizations with thousands of employees can continue innovating rapidly.

They do not want a system in which regulation is reflexively the first response to technological progress.

At the same time, they acknowledge that increasing capability requires some safety response.

The host suggests that the AI industry may need to improve self-regulation and says communication around current efforts has sometimes been weak.

Feldman stresses that frontier AI companies are not executing an established safety playbook.

They are developing the technology and simultaneously inventing the guardrails required to constrain it.

He argues that guardrails impose performance costs.

According to Feldman, his company had discovered during roughly the preceding six weeks that guardrails can add latency and make systems feel slower.

He says faster inference hardware can reduce that user-facing penalty.

This means frontier laboratories are simultaneously racing:

• competitors;
• safety requirements;
• their own performance expectations;
• and, Feldman suggests, their own desire to achieve something historically important.

He believes people inside these organizations are genuinely trying to determine the right course even though commercial ambition, safety and competition are mixed together.

The host recounts a conversation with “Nikesh from Palo Alto networks.”

According to the host, Nikesh said an AI system was tested against Palo Alto Networks' software and found bugs the company had not previously known about.

The host says the result was serious enough that the company stopped other work and spent approximately six weeks creating patches.

The transcript then claims the model found “tens of critical opens” in approximately an hour [exact wording unclear].

The host treats this anecdote as evidence that sufficiently advanced models have crossed into a category where staged release and red-team testing deserve serious consideration.

He compares it with software beta programs: ordinary consumers may opt into a visible beta, while earlier security-focused test stages occur before the broader public has access.

The host says society should assume that a massive data leak or breach will eventually occur.

He uses Warren Buffett's discussion of reinsurance as an analogy:

• insurers know that some major disaster—such as a tornado or earthquake—will eventually occur;
• they do not know exactly when;
• therefore they reserve money and prepare in advance.

Similarly, he argues, organizations should prepare psychologically and operationally for a major breach even if they cannot predict its specific form.

The speakers characterize such an event as potentially a Black Swan: the triggering vulnerability or question may be something nobody considered beforehand.

They nevertheless argue that acknowledging the existence of “unknown unknowns” is itself useful.

The host argues that reasoning models can help with unknown unknowns by explicitly identifying gaps in a user's thinking.

He describes a prompting technique he now uses.

For a task such as trend scouting, he asks AI to:

• create or refine the prompt;
• check its own work;
• identify what he has failed to consider relative to his goals;
• and ask him follow-up questions every time the task runs.

He says this changed the quality of the interaction because the system sometimes returns with corrections to its own work and proposes additional useful actions.

He notes that “perplexy” [verify] provides suggested next prompts automatically but says explicit instructions can produce stronger results.

The host recalls a particularly intelligent question he received while raising money over approximately ten years: at the end of a conversation, an investor asked what smart question other investors had asked that this investor had failed to cover.

The host views that question as a sign of curiosity, humility and an effort to broaden one's picture of a problem.

He proposes applying the same technique to AI:

• What question should I have asked to become an expert on this topic?
• What would a PhD-level question be?
• What question would someone operating at a “gold medal math” level ask?

The purpose is to make AI identify questions beyond the user's current expertise—the things the user does not yet know enough to formulate.

The host introduces AGI and superintelligence as useful waypoints even if their definitions remain contested.

He states his belief that artificial general intelligence has effectively been achieved but has not yet been fully deployed.

Feldman responds that by definitions used 20 years ago, today's systems have already crossed the AGI threshold.

He cites the Turing test as an example, saying modern systems have “blown” past it.

His broader argument is historical: if one takes definitions proposed 10, 15, 20, 30, 40 or 50 years earlier, present systems exceed what people at those times would have regarded as the relevant benchmark.

This is a conceptual argument, not a claim that contemporary researchers universally agree AGI has been reached.

The host connects the AGI discussion with the earlier theme that people often do not know which questions matter until technology advances.

He says science-fiction writers from 20 years earlier posed imagined questions about advanced AI, but present technology has now answered many of them.

The speakers jokingly suggest those writers might now be “out of questions.”

Feldman argues that this is one reason society should listen carefully to technologists whose predictions initially sound fringe or extreme.

Examples discussed include:

• “Ilya” warning eight or ten years earlier about AI safety;
• Elon discussing rockets and radically reducing launch costs.

The speakers argue that ideas that once sounded implausible can eventually become visible realities.

The host observes that AI tools are increasingly being used to help build and improve tools themselves, creating recursive loops.

He describes his own trend-detection experiment repeatedly finding concepts involving “looping” and “maxing.”

He says the system generated the phrase “loop maxing” for him and that he later noticed humans using the same term.

He interprets this as the system anticipating a phrase or concept that other people would independently adopt.

This anecdote is the host's interpretation and requires verification if used as evidence of predictive capability.

Asked about recursion and the path toward superintelligence, Feldman begins with recursive learning.

He says people including Sam, “Elia” [likely transcript issue], Dario and “Dana” [verify] recognized five or six years earlier that repeated recursive gains can compound exponentially.

He describes the loop:

1. Ask the model a question.
2. Learn from the result.
3. Ask it to perform the task again.
4. Add more information.
5. Improve the answer.
6. Repeat.
7. Cover progressively more relevant material.

Feldman emphasizes that the outputs are not merely slightly better. In his view, recursive loops can produce vastly better results.

The unresolved question is where that improvement curve eventually stops.

Feldman asks what happens as more compute continues to be applied.

Today's practical limits may include:

• token limits;
• financial budgets;
• available hardware;
• time.

But he treats as genuinely unresolved whether answer quality eventually reaches a hard plateau or continues improving dramatically as additional reasoning resources are supplied.

The host frames another version of the question: when do humans run out of intellectual problems worth solving?

Feldman then distinguishes between intellectual problems and people problems.

Feldman argues that many problems faced by company leaders are not analytically difficult.

They are problems involving:

• coordination;
• human relationships;
• motivation;
• organizational friction;
• getting people to act on a solution that may already be known.

The speakers compare leadership to “spraying WD40 on your team” so that friction is reduced.

They wonder whether AI can eventually provide deeper behavioral insight rather than merely intellectual answers.

Feldman predicts that world models, as they observe and model human behavior, may contribute to solving those behavioral and organizational problems.

The host says the physical-world implications deserve a separate future conversation.

He imagines recursive AI moving beyond screens and mathematical or scientific problems into physical construction.

Today a user might request “a new version of Salesforce.”

In the future, he speculates, a user might request something physically monumental—his example is a new Palace of Versailles on 100 acres in Texas or Nevada—and send “a thousand optimists” [transcript wording; possibly unclear] to build it.

The host acknowledges this sounds fantastical but argues that the original Palace of Versailles would itself have seemed fantastical to people living a thousand years earlier and perhaps even awed the craftspeople constructing it.

The discussion uses historic monumental construction to illustrate cumulative learning.

Feldman describes large projects lasting 50, 70 or 100 years, during which several generations of specialist families might work on the same structure.

A mason might learn from a father or uncle and then pass knowledge to the next generation.

A single project could therefore contain three or four generations of accumulated craftsmanship and innovation.

The speakers compare that slow intergenerational accumulation with AI models capable of iterating far more quickly.

Feldman argues that a major constraint on human learning is generational speed.

He says large mammals such as humans or elephants reproduce slowly compared with organisms commonly used in genetic research.

He invokes Drosophila/fruit flies, suggesting that rapid generations allow researchers to observe genetic change across thousands of generations.

He sees AI as an analogue: learning and iterative change can occur at the equivalent of thousands of generations over a period far shorter than biological evolution.

The host remarks that Darwin would be astonished by such an evolutionary pace.

The host recalls studying psychology and asking a professor how paradigms actually change.

According to his memory, the professor told him:

“paradigms don't die. People do.”

The conversation references Freud, Skinner, “young” [likely a name requiring verification] and “Thomas Coon” [transcript spelling].

The speakers' argument is that intellectual paradigms often survive until a generation of influential thinkers and their students lose institutional power.

This process can take 20 years, sometimes 40 years.

AI, by comparison, shortens the equivalent intergenerational interval dramatically, allowing new approaches to challenge prior ones much more rapidly.

The host praises Feldman's intellectual approach and contrasts his optimism with widespread “P. Doom” discussion around AI.

He says public debate often focuses on frightening scenarios, including job losses, whereas he views people such as Feldman as developing the technology thoughtfully.

The host predicts AI will create an enormous net benefit for humanity.

He specifically proposes the possibility that future children and people they know may not die from cancer because AI could accelerate medical progress.

This is presented as an aspirational possibility, not a medical prediction established by evidence in the transcript.

Feldman agrees that AI will produce some economic dislocation.

He compares the transition with automobiles, which harmed occupations associated with horses and carriage building.

His argument is not that disruption can be ignored, but that society should weigh disruption against possible benefits.

He repeats the possibility that future generations may benefit from major progress against cancer.

The speakers then list other possible areas of abundance:

• energy;
• calories/food;
• knowledge;
• education;
• housing.

These are prospective benefits discussed by the speakers, not established outcomes.

Feldman argues that humanity already knows individualized teaching can outperform uniform classroom instruction but historically has been unable or unwilling to provide individualized tutors to every child.

The conversation references Aristotle as tutor to Alexander the Great and also says Socrates was “his tutor” [historical relationship requires verification].

Their principle is that a tutor who adapts instruction to an individual child's learning style can teach more effectively.

The speakers criticize conventional classroom teaching as instruction aimed at a broad middle level; the host compares it with “factory farming.”

Feldman proposes AI agents capable of teaching each child in the way that particular child learns best.

He argues that human societies have known for a very long time that individualized instruction is superior but lacked a scalable method.

AI could potentially provide that method.

He says this benefit should be entered on the positive side of any serious AI cost-benefit ledger.

Feldman says a fair assessment must include both the negative and positive consequences of AI.

The host characterizes Feldman's overall ledger as strongly weighted toward abundance.

Feldman agrees and predicts that AI will create “Massive abundance.”

The host closes by proposing another conversation in approximately six months.

The transitional advertisement claims that:

• industry, capital and intelligence are converging into a single interconnected system;
• infrastructure must evolve alongside that convergence;
• Nasdaq was designed for this environment;
• Nasdaq powers more than 135 marketplaces and regulators globally;
• it connects capital with companies developing future technologies;
• and its technology platform makes connectivity scalable.

These statements are advertising claims requiring independent verification.

The host introduces Robin Rombach as co-founder and CEO of Black Forest Labs.

The host initially describes Black Forest as a city in Germany; Rombach corrects him, saying it is a mountain range.

Rombach says he grew up there.

The host identifies his work with open-source image and video models, his prior Stable Diffusion work and the open-source model Flux.

Rombach clarifies that Black Forest Labs is based in:

• the Black Forest region, specifically Freiburg;
• and San Francisco.

He says he splits his time to some degree between locations.

Rombach says he and his co-founders started the company approximately two years earlier, after previously working on Stable Diffusion.

Rombach says that before Stable Diffusion, he and colleagues developed an algorithm called latent diffusion while they were PhD students in Munich.

A fragment immediately before “algorithm” is transcribed as “agriculture” and appears erroneous.

Rombach describes latent diffusion as a foundational algorithm underlying a broad range of generative systems used for:

• image generation;
• video generation;
• and increasingly “physical AI.”

He explains the basic idea as compressing natural data—including images, video and audio—into a more efficient representation and then training a transformer model on that representation.

He compares the underlying idea of efficient representation with compression technologies such as “JPEC” and MP3 [verify “JPEC”].

He says they translated that idea into a neural algorithm, subsequently built Stable Diffusion on top of it and later built newer generative models on top of that foundation.

Rombach says the company's current ambition goes beyond image generation.

It is working on models intended to understand the world more broadly.

He describes multimodal visual systems pretrained simultaneously on:

• images;
• video;
• audio.

The next step is combining this with action prediction.

Rombach says the same underlying model could eventually:

• generate images;
• generate video;
• generate audio;
• predict actions;
• and therefore operate as part of a robot interacting with the physical world.

The host summarizes the trajectory as moving from images to video to audio and ultimately robotics/world models.

The host suggests that generating realistic images or video implies some degree of understanding of the world and its objects.

Rombach agrees with that framing but distinguishes complementary forms of intelligence.

He describes:

1. intuitive intelligence, associated with perception and multimodal representation;
2. a deep reasoning layer.

In his view, a complete intelligent system ultimately requires both and must allow them to interact.

Black Forest Labs has approached the problem primarily from the intuitive/perceptual side.

Images were a practical starting point because they require less compute than video.

Rombach says these once-separate areas are now converging into a multimodal model.

Rombach argues that pretraining on video can produce an implicit understanding of:

• physical dynamics;
• interactions;
• relationships between objects over time.

That understanding can support action prediction and therefore robotics.

The discussion then turns to a major criticism of generative media: that it can feel like a slot machine.

A user types a prompt and receives an image or video but may have weak control over:

• style;
• color;
• aesthetic;
• and the reason the model produced one result rather than another.

Rombach says the solution is to expose as many manipulation layers as possible to users and developers.

He describes an evolution in image-generation interfaces:

1. Early systems accepted text and produced an image.
2. Later systems accepted text plus an image.
3. Users could take either a real or generated image and alter it through text instructions.
4. Systems evolved further to accept multiple images plus text and combine the material semantically.
5. The same principle is now moving into video.

Rombach says multimodal systems become especially powerful when different media types can function as both inputs and outputs of the same model.

The host turns to Black Forest Labs' announced work with Martin Scorsese.

He asks how close AI video is to creating something that a filmmaker of Scorsese's stature would actually release.

He contrasts that aspiration with present workflows in which users generate isolated five- or ten-second clips, struggle to maintain consistency across several clips and assemble them later in editing software.

The host characterizes much present material as visibly artificial, “AI slop,” “clujy” [transcript spelling] and unable to pass the uncanny valley.

This is the host's evaluation, not Rombach's.

Rombach responds that Black Forest Labs sees AI models primarily as a medium.

He does not want the model maker dictating how artists should use that medium.

He specifically rejects the idea that his company should tell someone such as Martin Scorsese how to use the model.

Rombach calls Scorsese one of the greatest filmmakers ever and describes personally sitting with him several times while Scorsese explored the models.

For Rombach, the experience was unusual both because he had helped develop the underlying technology and because he is personally a fan of Scorsese.

Asked what Scorsese found most interesting, Rombach describes a session focused on a possible village setting somewhere in Eastern Europe.

According to Rombach:

• Scorsese had a clear mental image of a scene or location;
• he described that image;
• the model generated visual outputs;
• they reviewed the outputs;
• and they iterated.

Rombach says Scorsese ultimately highlighted the usefulness of getting a mental picture out of his head and communicating it visually through an image or sequence of images.

Rombach sees this as one of generative AI's most powerful creative uses: converting a hard-to-describe internal vision into something collaborators can actually see.

Rombach calls language a somewhat lossy communication medium.

Words can be interpreted differently by different people.

Images and video contain much richer visual signals.

He therefore views generative imagery as an additional communication method rather than simply an automated content-production machine.

The technology can help a creator transmit an idea that might otherwise remain imperfectly expressed in words.

Asked about producing entire movies with generative video, Rombach says he is not certain that full autonomous movie generation should be regarded as the ultimate objective.

He says very long AI-generated video could be technically interesting to explore and mentions plugging generation into something transcribed as a “genetic workflow” [likely transcription issue; do not silently correct].

However, Rombach says the most interesting creative results generally appear when a human remains in the loop, iterating with the model and treating it as a medium rather than surrendering the whole process to automation.

The host sees a clear near-term use in brainstorming and pre-production.

A phrase is transcribed as:

“You can paralyze your brainstorming, basically.”

The intended word may be different, so this wording should be verified before quotation.

The host compares generative AI with traditional storyboarding.

He cites:

• Ridley Scott as a director known for drawing or developing storyboards for films including Alien[s] and Gladiator [title as transcribed; verify];
• Spielberg as someone the host believes sketched material for Raiders [of] the Lost Ark;
• George Lucas as a filmmaker who collaborated extensively with artists, miniatures specialists and storyboard creators on Star Wars.

The host's point is that filmmakers have always used external visual aids to make ideas tangible; generative AI can dramatically increase the speed and scale of that process.

The host shifts from Hollywood to startups.

He says startups historically might spend $100,000 to $250,000 creating a launch video.

A company that had raised approximately $10 million could therefore spend $250,000 on one video.

He says companies he now invests in may instead spend only a week or two working with a director and generative tools.

He presents 30- to 90-second launch videos as an early practical form of AI-assisted storytelling because the objective is simply to explain a product or service in an engaging, concise way.

Rombach says Black Forest Labs supports this type of exploration.

He finds it valuable that many different products, videos and creative applications can be built on top of the same underlying model or technology.

The host says he spoke with “Gal Godot,” whom he identifies as the actress who played Wonder Woman [name spelling requires verification], at an event he calls the Breakthrough Prize / “Uri Milner's event” [verify].

According to the host's account, she told him about a “Bitcoin movie.”

He says:

• actors performed on a soundstage;
• traditional green screens were not used;
• generative AI created the scenery behind them;
• the production budget was approximately $30 million;
• she allegedly told him equivalent set construction would have cost $150 million;
• and the project would not otherwise have been greenlit.

These claims are presented second-hand through the host and should be treated as an anecdote requiring verification.

Rombach confirms that he has seen some production uses resembling the host's example.

He emphasizes that high-end filmmaking remains one of the most demanding applications of generative technology.

He nevertheless sees the trajectory as clear and rapidly improving.

He contrasts his PhD period, when systems could generate images only around 64 × 64 pixels, with current systems capable, according to him, of producing multi-minute, high-resolution video.

He expects improvement to continue and unlock increasingly demanding production applications.

The transcript briefly contains an unclear exchange around predicting a timeline, including “A couple of years.” It is not sufficiently clean to attribute a firm two-year prediction to Rombach without verification.

Rombach reiterates that even as generation improves, he prefers production systems in which humans remain involved.

His larger excitement, however, comes from the convergence of multimodal generative models.

He says the same kind of model could:

• generate a movie;
• and serve as the “brain” of a robot.

He also mentions possible related experiments in digital computer use, while saying it remains uncertain whether those approaches will work.

For Rombach, concepts now described as world models and world-action models are largely different views of the same underlying convergence.

The host asks whether robots will learn primarily from:

• real-world human demonstration video;
• synthetic training data;
• or enormous preexisting video corpora.

Rombach rejects a single-answer framing.

He explains action prediction as simultaneously related to:

• prediction;
• simulation;
• generation;
• perception.

To predict an appropriate next action, the model must first understand or perceive its visual inputs.

Once it perceives the scene, it can potentially:

• transform that scene into new content;
• predict an action;
• or describe what it sees.

Rombach believes progress comes from combining these capabilities rather than choosing one technique exclusively.

The host gives a detailed example of robot-training data.

He asks whether people might need to wear glasses or gloves so models receive a first-person record containing fine physical details such as:

• moving a glass;
• pouring liquid;
• adding ice;
• splashing;
• condensation forming;
• understanding that the outside of the glass is wet and therefore must be gripped differently.

Alternatively, he asks whether existing video corpora such as YouTube will provide enough demonstrations for models to infer the relevant behavior.

The question is intended to distinguish high-fidelity embodied data collection from large-scale passive video learning.

Rombach says the long-term goal is a robot that can be prompted in context much like a language model.

A user should eventually be able simply to instruct the robot to pick up a particular glass or perform another task.

He explicitly says “We're not there yet.”

Current deployments are harder because different robot hardware platforms use different action representations.

A general model therefore has to be tuned toward the specific robot or task.

Rombach says current models can begin with broad visual understanding.

According to him, a particular robot task can sometimes be adapted using only a few hours of fine-tuning data.

The longer-term research objective is to reduce this requirement and move toward as much in-context adaptation as possible.

Rombach explicitly characterizes that transition as still being a research problem.

The host observes that open source and sovereignty are receiving increasing attention.

He then asks what owners of major intellectual-property libraries—using Disney and Star Wars as examples—should do.

Possible strategies include:

• taking an open-source model and training it internally;
• working with Black Forest Labs to create a customized model;
• retaining greater control over proprietary characters and content.

The host also refers to Disney having worked with “ChatGBT” and “Open Air” around SORA and character licensing, but the transcript's details are unclear and should be independently checked.

He asks how Rombach views large proprietary content libraries and whether Black Forest Labs is working with IP owners.

Rombach says the central appeal of generative technology is the ability to produce something that did not previously exist.

For public-facing Black Forest Labs tools, he says restrictions prevent generation of certain protected IP.

He considers that approach sensible.

At the same time, he confirms that Black Forest Labs works directly with some IP owners to develop models jointly.

Those partnerships may use:

• Black Forest Labs' open-source models;
• or more powerful proprietary models.

Rombach describes custom model development for rights holders as an attractive commercial proposition.

Asked what consumers could encounter on Disney+ in a few years, Rombach refuses to prescribe Disney's strategy because he does not work for the company.

He says Black Forest Labs wants to provide technology that allows Disney or comparable companies to build whatever experiences they envision.

Because generative systems are becoming faster and increasingly interactive, he can imagine interactive content-creation tools operating inside services such as Disney+ or elsewhere.

He does not claim that Disney has committed to such a product.

The host identifies fan films as one of the most interesting possibilities.

He describes an evolution:

1. Before generative AI, fans wrote fan fiction.
2. Fans later dressed as Jedi and filmed their own Star Wars stories.
3. The host claims George Lucas permitted noncommercial fan films and even provided resources explaining how to create lightsaber effects and sounds.
4. Generative AI now lets fans produce stories from parts of fictional universes that were never officially filmed.

The host cites “Star Wars Stories Untold” as, in his view, the largest example and claims it is already receiving millions of views per video.

These claims should be verified before publication.

The host predicts that IP owners could eventually let fans pay:

• a licensing fee;
• a software rental fee;
• or a fee based on generated output;

in exchange for permission to create new stories using established characters.

Rather than treating fan creativity only as infringement, rights holders could potentially monetize controlled access to characters and worlds.

The host suggests Black Forest Labs may be well positioned to provide the technology behind such systems.

Rombach agrees that a strong model could satisfy both sides:

• IP owners need a system that respects their rights and commercial interests;
• consumers want creative freedom and customization.

He says that when he personally reads a book or watches a movie, he often imagines alternative events or ways the story could have unfolded.

Generative technology could allow people to visualize those ideas rather than leaving them only in their imagination.

He sees that capability as appealing if the rights and business model can be made workable.

Rombach says Black Forest Labs:

• has an office in San Francisco;
• has raised “a bunch of money” [no figure supplied];
• has recently passed 100 employees;
• is hiring in both Germany and San Francisco.

He lists several categories of desired hires:

Experience with:

• large-scale model training;
• diffusion-model training;
• flow-matching training.

Engineers able to:

• work directly with customers;
• build customized physical-AI solutions;
• collaborate with IP owners on jointly developed models.

People experienced in:

• large-scale compute infrastructure;
• managing development/training systems;
• keeping large training runs stable;
• maximizing “MFU” [acronym should be defined/verified].

People interested in getting the technology into users' hands.

The host concludes that corporate demand for open systems, greater control and access to powerful proprietary capabilities creates a strong opportunity for Black Forest Labs.

Rombach agrees with that characterization and the interview closes.

Speaker: Host/interviewer
Exact quote: “We've never seen a buildout like this since, you know, the Great Wall of China.”
Context: Opening comparison describing the amount of capital, infrastructure and human effort going into AI.
Topic: AI infrastructure scale.

Speaker: Andrew Feldman
Exact quote: “we have a $25 billion backlog.”
Context: Feldman argues suppliers are trying to catch up with demand that already exists rather than speculating on future demand.
Topic: AI infrastructure demand.

Speaker: Andrew Feldman
Exact quote: “There is clearly massive value happening.”
Context: Feldman acknowledges waste and experimentation but argues net value remains enormous.
Topic: AI economics.

Speaker: Host/interviewer
Exact quote: “I didn't ask for it, but that is better.”
Context: Reaction to models anticipating a better output than the user explicitly requested.
Topic: Intent understanding.

Speaker: Host/interviewer
Exact quote: “Because unlimited tokens, I believe means unlimited reasoning.”
Context: Discussion of running reasoning systems for very long periods with abundant compute.
Topic: Inference scaling.

Speaker: Host/interviewer
Exact quote: “we have come a long way from guess the next word.”
Context: Recalling an earlier explanation of AI systems understanding intent, developing strategies and checking one another's work.
Topic: Reasoning models.

Speaker: Andrew Feldman
Exact quote: “This reasoning is inference and it's computationally intensive.”
Context: Feldman's explanation of why reasoning creates demand for fast inference hardware.
Topic: AI compute.

Speaker: Andrew Feldman
Exact quote: “you don't have to sort of make the fastest chip, you just can't be entirely dependent on other people's chips.”
Context: Explanation of why hyperscalers and AI companies build internal silicon.
Topic: Chip strategy and sovereignty.

Speaker: Andrew Feldman
Exact quote: “We need to give the world a choice.”
Context: Feldman's argument that the US needs more domestic open-source models.
Topic: Open source and AI sovereignty.

Speaker: Andrew Feldman
Exact quote: “This doesn't seem to be an unreasonable thing for the government to ask.”
Context: Supporting limited pre-release security testing for sufficiently powerful AI systems.
Topic: AI regulation and safety.

Speaker: Andrew Feldman
Exact quote: “I think the polarization hurts a great deal. It hurts clear thinking.”
Context: Feldman argues AI policy should be evaluated beyond partisan identity.
Topic: Political polarization and AI governance.

Speaker: Andrew Feldman
Exact quote: “There's not a playbook.”
Context: Frontier labs are developing AI capabilities and safety mechanisms simultaneously.
Topic: AI guardrails.

Speaker: Andrew Feldman
Exact quote: “the very guardrails can add time and make it feel slower.”
Context: Feldman argues fast inference can make additional safety layers less costly to users.
Topic: AI safety infrastructure.

Speaker: Host/interviewer
Exact quote: “there will be a massive breach.”
Context: Reinsurance analogy arguing that organizations should prepare for an eventual severe security event without knowing its exact form.
Topic: AI cybersecurity risk.

Speaker: Host/interviewer
Exact quote: “please check your work.”
Context: Part of the host's preferred instruction asking models to review themselves and identify missing considerations.
Topic: AI prompting and self-review.

Speaker: Andrew Feldman
Exact quote: “By any definition we had 20 years ago, we've hit it.”
Context: Feldman's claim that today's systems exceed historical AGI benchmarks.
Topic: AGI.

Speaker: Andrew Feldman
Exact quote: “powerful recursive gains are exponential.”
Context: Feldman's explanation of repeated model improvement loops.
Topic: Recursive learning and superintelligence.

Speaker: Andrew Feldman
Exact quote: “these sort of loops are producing sort of not a little bit better answers, but vastly better answers.”
Context: Feldman argues iterative reasoning can generate nonlinear improvements.
Topic: Recursive inference.

Speaker: Host/interviewer, quoting a former professor
Exact quote: “paradigms don't die. People do.”
Context: Comparison between slow generational changes in human thought and rapid AI iteration.
Topic: Generational learning.

Speaker: Andrew Feldman
Exact quote: “I think it will create abundance, for sure. Massive abundance.”
Context: Closing assessment after weighing disruption against potential benefits in health, education, energy and other areas.
Topic: AI and abundance.

Speaker: Robin Rombach
Exact quote: “these AI models, they are a medium, right?”
Context: Rombach rejects prescribing how creators such as Martin Scorsese should use generative systems.
Topic: AI and creative authorship.

Speaker: Robin Rombach
Exact quote: “getting like the mental picture of something out of your head and communicating it in a visual way”
Context: Description of Scorsese iterating on an imagined Eastern European village scene.
Topic: AI storyboarding and visualization.

Speaker: Robin Rombach
Exact quote: “language ultimately is like a little bit of like a lossy communication medium”
Context: Rombach argues images and video can communicate rich details that words fail to transmit completely.
Topic: Visual communication.

Speaker: Robin Rombach
Exact quote: “the real interesting use cases they come when you have a human in the loop who iterates and uses it as a medium.”
Context: Rombach says full automatic movie generation may be less interesting than interactive creation.
Topic: Generative filmmaking.

Speaker: Robin Rombach
Exact quote: “you can use the same kind of AI model to make a movie and deploy that as a brain on a robot.”
Context: Rombach describes the convergence of multimodal generative models, world models and action models.
Topic: Multimodality and robotics.

Speaker: Robin Rombach
Exact quote: “We're not there yet.”
Context: Rombach's response to the goal of prompting robots naturally in context as one prompts a language model.
Topic: Robotics limitations.

Speaker: Robin Rombach
Exact quote: “you cannot generate certain IP with these models”
Context: Rombach explains Black Forest Labs' public-facing restrictions while also describing partnerships with rights holders.
Topic: Generative AI and intellectual property.

Speaker: Robin Rombach
Exact quote: “We just crossed a hundred people.”
Context: Rombach describes Black Forest Labs' growth and hiring in Germany and San Francisco.
Topic: Black Forest Labs business growth.