Alejandro Matamala

The through-line

Everything Matamala says traces back to a single origin story and a single measure of success. He and his co-founders met at art school, spent their time at NYU's interactive telecommunications program exploring what the first generative models could do for creative workflows, and ended up building the tool they wanted to use themselves [2]. The papers they were reading around 2017 and 2018 were not written with any creative use case in mind, some of them aimed at self-driving cars or security, and the outputs were 256 by 256 pixels with terrible alignment between text and image [1][3]. That was still the shocking moment: "we couldn't see another scenario where we will not be using this type of models, this type of technology for telling stories" [1]. From that came a north star that has not moved in seven years, allowing anyone anywhere to generate a two hour feature film, with synthetic characters, synthetic actors and synthetic sound [1][2]. He frames every decision against it: does this help us towards that vision [2].

What has changed is the world around the goal. In 2018 the users were R&D departments and creative coders asking what was possible; in 2025 everyone arrives with the same brief, wanting AI in production [1][2]. And the company's own definition of itself has shifted, from an AI company to a media and entertainment company, a reframing he dates to roughly two years ago [1][2].

On being a media company, not an AI company

He is blunt about the label: "defining yourself as an AI company today is like saying I'm an internet company. Everyone is going to use AI in one way or another" [1]. Runway does cutting edge research and publishes state of the art models, and the research team is one of the biggest inside the company, but he insists the motivation was never the technology itself [2][3]. It came from wanting to empower creators, and the yardstick follows from that: "we measure success in how many movies are we going to be able to create how many stories are people going to able to to produce", because if it ends in no new stories and no new movies, he would not count it as success at all [2][3]. The company is fully committed to arts, entertainment, film making and storytelling rather than expanding into other verticals [2]. He also wants Runway to be a user of its own tools, making stories itself and pushing the media landscape toward new formats [1]. This positioning is the same one he has taken into conversations about reimagining media, art and entertainment [8] and about generative AI as a change agent for creative industries [9][10].

On the two hour movie and what still blocks it

He now thinks the fully generated two hour movie is close, "maybe even coming this year", and says he has talked to users already working on long form projects that surprised him [2]. But he is careful about what the remaining bottleneck actually is. It is not simply the length of a single generation. Even with the capacity for longer outputs, user feedback showed that is not how people want to work: there is heavy iteration, you want quick feedback on how a generation looks, and "it's much like directing a scene", with directors running a scene repeatedly until they get the take they want [2]. He rejects the idea of handing someone ten minutes of finished scenes and calling it a workflow [2]. The real gap is scale and coherence, consistency of characters, objects and locations over time, and the known failure modes of the models with complex camera movement, complex body motion and physics [1][2][3]. Progress has come as a series of what he calls pillar points, each step building toward the same destination rather than arriving at it in one leap [1][3].

On world models and simulating more than pixels

The research answer to those failures is the general world models initiative Runway announced around two years ago [2][3]. He describes today's video and image models, now in a fourth generation, as a small representation of the world that is still missing physics, complex body movement and camera motion [1][3]. The ambition is to simulate the world in its entirety and not only its visual aspects, training on a wider mix of inputs including video data, synthetic data and sensor data, so the models hold a better representation of how the world actually works [1][3]. He draws a deliberate parallel to 2018: people thought they were a little bit crazy for looking at 256 pixel outputs and talking about feature films, and he treats the world models bet the same way [2]. On architecture he is undogmatic, saying Runway has doubled down on diffusion for several years and continues to focus there, while running experiments in parallel and not discarding approaches beyond diffusion [3]. Beyond storytelling, he expects a closer representation of the real world to accelerate other things, such as running virtual experiments in a virtual environment instead of the real one, with scientific experimentation as an obvious beneficiary [2].

On carving products out of models

His central design metaphor is subtractive. "we carve our products rather than we build our products", he tells his teams, working backwards from what a model turns out to be able to do rather than forwards from a specification [1]. The reason is empirical: every time the team tried to anticipate what a model would do, they mostly failed, because the latency was different from what was expected and the capabilities were bigger [1]. The traditional process of handing a design off to an engineer has failed almost every time for the same reason, since you have to learn the constraints, opportunities and capabilities of the model firsthand [1]. So the design team includes front end engineers and design engineers alongside product designers, building functional, speculative prototypes as early as possible against the model's early capabilities [1]. By the time a model is ready, the team already knows its latencies and limits, which sets up what he calls the sprinting phase, getting it to users as fast as possible [1]. The same logic governs how the company is arranged: he likes to say everyone is a researcher and an artist, and people with twenty plus years in VFX sit next to AI researchers to identify where research might create value in a creative workflow [1]. That is what he finds interesting about a research company, that you never know when or whether research will be ready, but you can follow it and look for the piece that is valuable [1].

On prompts, knobs and control

He argues the industry over-indexes on the prompt. Text is one way to express an idea and get immediate feedback as an image, audio or video, but he wants to move from prompts to more creative knobs, new controls and new kinds of interaction with the tool [2]. The principle underneath is controllability: fidelity and resolution are worthless on their own, because "if you don't and if you can control and have full Precision of what you want to create or how you want to tweak this generation then it becomes useless" [2]. Precision and controllability are what he says directors and filmmakers obsess over, and he treats them as the reason Runway's work lands with professionals [3]. Product direction comes from a mix of three inputs: intuition, the unchanged vision, and what users are actually asking for [2]. Intuition mattered most early, when nobody was requesting what Runway was making, and it has not been set aside since; he treats it as a skill you nurture by continuing to ask what if and never stopping questioning, and he reads past decisions Runway has abandoned as learning rather than bad calls [2].

On editing and transformation as the next capability

The most recent video model moves beyond generation from scratch, which is what the Gen 1 to Gen 4 family did from a prompt or a reference image [1]. It edits and transforms existing footage: changing style, going from summer to winter or day to night, adding a prop that was missing on set or removing cables from a shot [1]. What excites him more are capabilities that had no prior equivalent, in particular generating new angles of a scene that was shot with a single camera, so that footage suddenly supports a drone view or any other vantage point [1]. He describes it as a new series of models aimed not only at generating content but at transforming it in ways that were previously unimaginable [1].

On games, interactivity and formats that do not exist yet

Having judged the feature film goal nearly within reach, he has been looking for the next one, and concludes the models are capable of far more than traditional storytelling [1]. Games are the most interesting candidate, and Runway has released a platform for generating video with its models to create stories on the fly [1]. He is watching the gaming industry alongside interactive AI video and real time video generation, and expects those to overlap as both mature [1]. He also expects generative AI to be the thing that unlocks VR and augmented reality, immersive formats that have traditionally been very hard to produce [1]. The surprises are already arriving from users: people building escape room and trivia games as expected, but also learning games, including one where the player acts as a graphics engineer to learn how to work with GPUs and train models, a use the platform was never designed for [1]. Looking three to five years out, he thinks the current phase is people using new tools to make the things they already know how to make, and that the interesting outputs come later, through interactive video, real time generation, personalised content, stories that never end and follow you across mediums, and formats where "you are part of the story" [1]. He is clear that the push will come from artists, with Runway providing the tools [1][3]. His historical analogy is the camera: it was first pointed at the landscapes people had been painting, and film cameras at recorded plays, with cinema arriving as a consequence [2][3]. Impressionism, likewise, followed the invention of the paint tube that let artists work outdoors [3].

On jobs, budgets and who gets to make things

Asked repeatedly whether AI destroys creative jobs, he reports seeing the opposite [1][3]. The headline effect is more projects getting green lit, because films and shows have become so expensive that many simply never got made, and thinner budgets and different kinds of talent now make them viable [1]. He sees creators working on several projects at once, which in his account requires more teams rather than fewer, and notes that while one person can make something alone, most of what he sees is made by teams with mixed skills [3]. The geographic effect matters as much to him: filmmakers in Chile and teams from everywhere he travels are finally able to make projects they had been waiting on for budget, team or timing reasons, and he expects visual effects heavy and post heavy work to start coming from places it has never come from [1][3]. He also frames the tools as a way for artists to triage: artists are people of ideas with many ideas at once, and generation lets them pitch or visualise the smaller ones to see whether they are worth pursuing [1][2]. To sceptics he offers the industry's own record, that Hollywood is "a story of technological breakthroughs and progressions", from the camera to black and white to colour to sound to CGI, and this is one more tool it is fine to use or not use [1][3]. He predicts the distinction dissolves entirely: "I don't think we will have AI films. We will just have films" [1]. He also notes how far there is to go, estimating that probably less than one per cent of what audiences watch today is generated, with future sets mixing virtual production, generated environments and fully synthetic characters and objects [3].

On partnerships, data and working with studios

On copyright and training data he is candid that it is unresolved, saying Runway is still at day zero and still figuring many of these things out, while pushing to build alliances with the right companies, including announced partnerships with Getty and Lionsgate [3]. He describes those relationships as more than data deals: they are a way to understand how studios want these models inside their workflows, how to build tools around them, and possibly to discover new economies that emerge from the collaboration [3]. With Lionsgate the work involves models designed for their specific use cases and content, and he attributes the deal to Runway's north star matching theirs, plus the obsession with controllability and precision and the fact that the solution is targeted rather than general purpose [3]. Usage runs across the pipeline: ideation, the writer's room, pitching, getting a short film made in a day to test whether an idea holds, storyboarding and pre production planning, post production, location scouting, animation and dubbing, and outside film into advertising, commercials, and unexpected places like architecture and e-commerce [3]. Asked whether studios will build their own tools instead, he is sceptical that it is the right division of labour, noting it took Runway five years to reach its first video model and that the better outcome is people focusing on the technology while others focus on telling and distributing stories [3].

Takeaways

• Runway defines itself as a media and entertainment company, on the grounds that calling yourself an AI company today is like calling yourself an internet company, and measures success by how many movies and stories its users manage to make [1][2][3].
• The founding north star, unchanged since 2018, is letting anyone generate a two hour feature film with synthetic characters, actors and sound; he now thinks that is close, possibly within the year [1][2].
• Product is carved, not built: attempts to anticipate a model's latency and capabilities have nearly always failed, so designers and design engineers prototype against early model capabilities and work backwards [1].
• Fidelity without control is useless; the priority is moving from prompts to granular creative knobs that give artists precision over outputs [2][3].
• Longer single generations are not what users want, because creating is iterative and closer to directing a scene repeatedly; the real bottlenecks are consistency over time, physics, complex camera movement and body motion [1][2][3].
• General world models trained on video, synthetic and sensor data are the bet to fix those failures, with virtual experimentation as a side benefit beyond storytelling [1][2][3].
• The observed effect on the industry is more projects green lit on smaller budgets and from outside traditional production centres, not fewer jobs, and eventually no category of "AI films", just films [1][3].
• Games, real time and interactive video, VR and AR are where he expects genuinely new story formats, with artists rather than tool makers driving what those become [1][3].