Cristobal Valenzuela

The through-line

The idea Valenzuela keeps circling is that the same class of algorithms that learned to produce language can produce pixels, and that this changes what a moving image fundamentally is. He describes falling into a deep learning rabbit hole roughly a decade ago after coming across research papers that he considers "the beginning of everything we've seen today," and becoming "obsessed with this idea that you can take these these algorithms and these approaches to AI and use them to create not only language but also uh pixels and and videos and images" [1]. Everything since has been, in his own framing, "perhaps uh the same journey and obsession we've had back then, but now at a with a bit more scale and and and attention" [1].

What has shifted is where that obsession points. Early on the argument was about tools for creative production and whether machine learning could move beyond novelty into professional work [8][5][7]. More recently the argument has moved to world models and simulation, where the output is not a rendered clip but a world [2][9]. He treats this as an underrated discontinuity rather than an upgrade to existing video tools.

On streaming every pixel in real time

His strongest stated conviction is about real-time generation. "There's a chance, a a high chance that I'm pretty convinced will happen sooner than most people realize is that you might be able to stream pretty much every pixel to your screen in real time. Any sort of like world that you want, interactive or not, into any screen that you want" [1]. The crucial part of the claim is not the throughput but the category: "that might be a different thing altogether that might not necessarily like feel similar to a to a film," and it is "the thing I would say people are undervaluing, underestimating the most" [1]. That is the frame he brings to world models generally, which he connects to storytelling rather than to graphics benchmarks [2][9], and which the hosts read as the possibility of a genuinely new medium sitting somewhere between games and movies rather than a better version of either [1].

On Runway as platform, tool and model at once

Asked what Runway actually is, he refuses a single noun: "It's a it's a platform. It's a it's a tool. It's a model" [1]. The product range he describes runs from "simple, you know, uh images and videos to way more complex cinematic like 4K uh footage," and the user base is deliberately broad, spanning "the largest and biggest and best studios in the world" alongside agencies and brands, game studios, architects, designers, filmmakers, and "a lot of just consumers and creators" [1]. He puts the count at millions of users [1]. That breadth is consistent with how he has described the company since the early days, as a machine learning platform aimed at pushing the limits of creativity rather than at one production niche [4].

On the moment the models became usable

His read on maturity is practical rather than triumphal: "the models are getting to a point where like they're pretty much usable in a bunch of different domains and we're like I would say venturing into into many of those" [1]. This is the endpoint of a transition he has been tracking for years, from AI imagery as an entertaining toy to something capable of unlocking real creative value for professional content creators [8], with the tools reaching film and video work at a serious level, including on A24's Everything Everywhere All at Once [7]. He has argued alongside this that the mental models artists and the wider public hold about creative work will shift as the tools do [5], and the hosts note that his arts background gives him a distinctive position in the argument over whether these systems are creative tools or cost-cutting instruments [1].

On Hollywood, growth and risk

Runway's relationship with the film industry is one he engages directly rather than diplomatically. He has discussed the company's partnerships with studios including Disney and Netflix as a route to changing how production works [6][3], along with the risks that come with tools moving this fast into creative industries and the pace of the company's own growth [3]. The hosts characterise Runway as changing Hollywood on an hourly basis and note that Valenzuela is not afraid to speak his mind on these questions [1].

On the shape of the company

He is candid about scale and capital in a way that supports his research-first framing. Runway is around 120 people, and "most of that the the team is research engineering," split between New York and a smaller research lab in London [1]. The company has raised about half a billion dollars, has raised "on average every like 12 months or so," and had a last public valuation of around three billion [1]. He is pointed about how much the environment has changed: the company was technically called Runway ML when it started in 2018, and "the rounds and the sizes and the valuations were very different if you were building ML products back then" [1]. He also frames establishing the company as a research lab early as a deliberate posture, useful for recruiting and for stating what the work is meant to be for [1].

On the hard parts and the infrastructure debate

He does not treat the difficulty of AI as purely technical, and he pushes back on the standard environmental framing of the industry, arguing that golf courses are a bigger problem than data centers [9]. The point is raised in the same conversation as world models and simulation, as part of a wider account of what is genuinely hard about building in this field [9].

Takeaways

• His central prediction: real-time streaming of "pretty much every pixel to your screen," any world, interactive or not, to any screen, arriving "sooner than most people realize" [1].
• He believes the resulting medium will not feel like film, and calls this the thing people underestimate most [1].
• The founding insight was that AI algorithms could generate pixels, video and images, not only language, an obsession he dates back roughly a decade [1].
• Runway resists a single product category by design: "It's a it's a platform. It's a it's a tool. It's a model," serving studios, agencies, game studios, architects, designers, filmmakers and consumers [1].
• He judges the technology by usability rather than benchmarks, saying the models are now "pretty much usable in a bunch of different domains" [1].
• The company is built research-first: around 120 people, mostly research engineering, across New York and a London research lab, on about half a billion dollars raised and a last public valuation near three billion [1].
• He has taken the AI-and-Hollywood argument directly to the industry through studio partnerships including Disney and Netflix, while discussing the risks openly [6][3].
• On the environmental critique of AI, he argues golf courses are a bigger problem than data centers [9].