Runway's Chris Valenzuela on Why Taste Beats Tools in Generative Media

Chris Valenzuela tells Jacob Effron that Runway's field is early: most of what generative media can do today is just the beginning of what will appear over the next couple of months. He frames the moment as one where scale has unlocked something useful, but the work is far from finished. The company's product and research trajectory are built around the expectation that capabilities will keep moving.

Gen-3 Alpha was a moment where realism, control, and fidelity improved substantially, Chris says, yet he still lists better control tools, more fidelity, longer generations, and greater consistency as open needs. He is excited about real-time generation, more customization for specific styles and art directions, and multi-model controls that let creators build sequences of media with audio or other inputs beyond text and images.

The model is flexible enough to be fine-tuned with particular data, art directions, or mood boards, Chris says, and that lets creators steer it toward a consistent style. The best steerers, in his view, have good taste: they know how to pick ideas, evolve them, and fit the right data to the model. That is true across creative forms, not just AI media.

Chris says the people who steer generative models best have good taste: they can pick promising ideas, evolve them, and feed the right data into the model. He describes his own approach as coming with concepts—characters, camera angles, or themes—rather than fully specified outcomes, then letting the model surface combinations he did not anticipate. Speed matters because a few seconds per generation lets creators visualize and explore ideas faster than before.

The best results come from an exploration mentality, Chris says: set yourself up to be surprised rather than demanding a perfect first output. That iterative loop is central to how Runway users discover what a model can do. The model is not a vending machine for a finished idea; it is a system for exercising creative judgment.

One example is the bee-cam, a first-person camera attached to a bee flying through grass and landscapes. Chris started with the idea of visualizing insects in different locations, and the model produced an interesting camera angle he had not thought possible; he then iterated roughly 50 times across locations. He sees it as a summary of what these models can do: create versions of ideas that were extremely hard to make before or had never been seen.

Chris describes internal testing with artists and creatives who spend hours in flow, exercising parts of the brain they were not used to using; he likens it to going to the gym for the mind. He argues that artistic expression should be detached from craft and tools, because creativity is a state of mind that can appear in soccer, science, or any discipline. The point is not to become an award-winning filmmaker on the first session but to end invigorated and want to do it again.

Chris describes Runway's users as a wide spectrum: professionals in studios, production teams, filmmakers, art directors, and editors on one end, and casual creators making content from their bedrooms on the other. Runway has not settled on serving only one segment because flexible models can support many use cases. The needs are not as different as they might seem: both ends want to tell a story, though media formats, output quality, and subject consistency vary.

Chris expects a new type of professional to emerge, just as visual-effects and CGI roles that seemed unthinkable 40 years ago became established professions. He is less interested in fitting new work into existing user personas than in building for people who want to express something. That means serving a wider spectrum rather than narrowing the product to a single category.

The blank-canvas problem is real: a first-time user who expects one prompt to produce a finished result will probably be disappointed. Chris says ideas and taste matter most, and the common misconception comes from chat interfaces, where people expect a correct answer or summary from one query; video generation is different. Users have to iterate, see what comes back, and repeat until they unlock the tool's potential.

For people learning the tool, Chris recommends starting with work they have done in the past—if they are traditional post editors or work in 3D systems, they can try to recreate that in Runway. Constraints help with the blank page, and Runway Academy and other resources teach new users how to work iteratively. The comparison to filmmaking is direct: a camera alone does not make a filmmaker; intention, timing, and editing do.

Chris's provocative claim that UI does not matter is really about model velocity: when models are improving quickly, over-engineering a fixed interface can be a mistake. He says Runway has made that mistake before. A beautiful UI on top of a weaker model can be steamrolled by a better model whose output is good enough that users do not care about the interface.

He also imagines a more radical design shift: creative software today is prescriptive, with fixed sliders and primitives that require weeks or months of planning per update. In the future, he hopes the model will generate or adjust the interface based on what the user is trying to do—2D animation, 3D short film, or hyperrealism may need different controls. The model would prompt and create those interfaces for the user.

For AI product builders, Chris acknowledges the hard problem of deciding how much workflow to build around today's capabilities when they may be irrelevant in six to nine months. Runway tries to think long term, identify truths likely to hold in a couple of years, and build toward those rather than optimizing for a single point. The world will be very different in 12 months and extremely different in 18, so the right system matters more than a transient feature.

Those truths include quality and temporal consistency of video elements, plus real-time generation as inference times fall. Gen-3 is Runway's best public model and significantly faster than earlier generations while also being better, Chris says. He describes technology moving in waves of expansion—discovering what is possible—and optimization—making what works cheap and efficient—and art benefits from a different error regime than chatbots, where hallucinations are bad but weirdness and unique interpretation can be good.

Chris wants interaction with media models to resemble directing on a film set, where people point to references, cite previous films, use gestures, and speak about intent. He wants models to understand those references and the intent behind them, then suggest something in response. That is a different interface from text prompting alone: it is closer to collaboration with a director or editor.

He also thinks music and other sensory inputs should be able to drive generation, because music triggers ideas and states of mind for him. Jacob Effron connects this to a previous conversation about music generation, where inspiration need not start with text. For Chris, the broader point is that creative work draws on many kinds of input, so models should accept them.

Runway's advantage in combining cutting-edge research and product comes back to its vision: Chris sees this as eventually a new form of art, built for people who want to tell something to the world. Doing that requires deep empathy for creators and fundamental research at the lowest level. The sweet spot is people who can speak both the language of art and the language of science.

Chris says artists and researchers need their own processes and a shared language, and researchers must remove preconceptions about how things should work. Research often limits variables to what can be measured, but art should not be over-measured or over-prescribed. Generating a video is not the same as making a film or art; the hard part is turning a sequence of pixels into something interesting, so teams need room to find that.

Chris says Runway uses benchmarks and human evaluations to check preferences, but he considers that approach somewhat flawed. The company relies heavily on taste: people with strong judgment review outputs and ask whether something is interesting, even if it does not win a particular benchmark. Those tasteful evaluators then communicate back to research how to push the frontier.

Jacob asks about the tension between shipping individual features—avatars, dubbing, and similar tools—and building an end-to-end product. Chris says Runway has been through stages where it optimized for the wrong product at the wrong time, and it has learned from six years of building. The question is whether a feature is a durable product or a temporary point on a moving line.

He gives rotoscoping as an example: Runway trained a model to remove objects from foregrounds or backgrounds, a task that is still manual, frame-by-frame, and expensive in traditional editing. Their system could do about 95% of the job in minutes or seconds on 4K footage, a huge saving, and Runway built APIs and interfaces around it. But Gen-3 Alpha can do rotoscoping and green screen zero-shot, more effectively and cheaply than the narrow task model, so the point solution was subsumed.

Chris does not regret the rotoscoping work: it made sense at the time, taught the team a lot, and helped build streaming capacity and other capabilities. The lesson is to focus on the line over time—Gen-3, Gen-4, Gen-5—rather than any single point, even if a point can be justified in the short term. Runway saw its tools as algorithmic improvements that would unlock new creative tooling, with stepping stones along the way.

Chris names better control as a critical unsolved problem: traditional computer graphics offers pixel-level control over almost everything, but at high cost in time and money, while Runway offers speed at the expense of fine control. He expects Runway to move toward pixel-level control that unlocks fidelity and fine-grained aspects comparable to traditional CG, and estimates that is one to two years away. That control is part of the path to a new media format.

Gen-3 Alpha is a smaller early-stage model, magnitudes smaller than Gen-3, and one of the main enablers was an entirely new infrastructure built for training at this level. Chris says people underestimate how much developer tooling, infrastructure, and the right research environment matter; the cohesive foundation lets Runway do quick fine-tunings and iterations. The unit economics also matter because running the model must be accessible: if generating 10 seconds is too expensive, users are not incentivized to experiment.

When Jacob asks where future improvements come from, Chris points to scale, data, and the team's oral knowledge of what works. Scale still matters and Runway is far from a tipping point, he says, but data and taste matter just as much: what data is chosen, how it is captured, and how it is worked with. The combination of research and infrastructure, not one ingredient, drives progress.

Runway organizes research into teams for pre-training, controllability, quality and safety, and fine-tuning with studios, and it embeds artists alongside researchers. Chris favors ensembles and less prescriptive goals so teams can wander toward true invention rather than only marginal improvements. Motion Brush, a tool for brushing over a subject and defining its motion, emerged from a researcher and an editor tinkering, not from a top-down plan; as the team grows, Chris watches for coordination costs and hires people comfortable with uncertainty.

On pricing, Chris says Runway is pricing for a world still being explored rather than optimized. Prices will come down and the cost of creating media will approach the cost of inference, but optimizing for that future too early could make the company miss the discovery phase. Runway therefore aims for unit economics that let it and its users wonder, then optimizes once the use cases really work. When OpenAI released Sora, Chris tweeted 'game on'; he welcomes competition because it validates the market, incentivizes innovation, and comes from a much larger company, while Runway has managed to build better models and products early. He expects most markets to condense into a few winners, and he says media models rather than video models because video is a transitory stage toward combinations of pixels and audio.

Runway is building models in the audio domain because its world-model approach is about translating between modalities and understanding the objects and content being represented. Chris frames the direction as sequence-to-sequence, less about text-to-video or image-to-video and more about anything-to-anything. On IP, Runway is working with studios, IP holders, and media companies on custom models that may never be public: a studio can use the model internally, and audiences may never know which scenes were made with it. He wants to stop distinguishing AI-generated from non-generated content, because audiences care whether a movie is good, not which tools made it.

Chris places the moment in the history of storytelling and technology, from pigment and fire to the invention of filmmaking in the early 1900s. He notes that many early camera pioneers did not understand what was happening, and the Lumière brothers thought cinema had no future, while critics said photography was not painting. With time and experimentation, a new art form emerged, creating industries, professions, and shared experiences; Chris sees generative media at a similar early stage, where the first impulse is to compare it to previous art forms but the more interesting question is what the new form will be. He points to unusual camera angles, perspectives, customization, and real-time generation as early glimmers of that new format.

On the business, Chris says it is hard to predict whether Runway will be mostly enterprise or consumer in five to ten years, because storytelling appears in every persona, industry, and market; professionals using 20-year-old tools may soon do with one or two people what once took teams of hundreds, but they will still be professionals working differently. Runway raises enough to train for at least the next 24 months and stays efficient with burn and team size; Chris says the goal is not AGI or a 'god' system but tools for people to express themselves. In the quick-fire round he calls text-to-video prompting overhyped and prompting the wrong interface, while underhyped is what Gen-3 Alpha can do with simulation, fluids, and dynamics. He says state-of-the-art models do not require tens of billions of dollars, just a focused team with clear goals; he expects a ChatGPT-like moment when hundreds of millions of people make things, motivated partly by his background in Chile, where capital and resources constrain stories. He advises artists to explore weird stuff, focus on ideas over tools, and be original, authentic, and weird; he cites Waymo as an AI experience that resembles the fear-then-shrug reaction people will have to many new tools, and says his favorite story is still Runway's own—building the company without Silicon Valley, fundraising, or research experience and finding that what seemed impossible kept working.