# The Next Medium: Why Real-Time Interactive Video Changes Everything Ahmed Ahres, Reactor | AI Engineer World's Fair 2026 | 17:30 Source: https://www.youtube.com/watch?v=5dCAmSDOAjI Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks. Page: https://aietalks.com/talks/the-next-medium-why-real-time-interactive-video-changes-everything Published: 2026-08-18 Tags: gpus, inference, latency, video ## TL;DR - Real-time video changes the medium because viewers can alter what appears instead of receiving a finished file. - Interactive world models could support controllable stories, robotics training data, education, live streams, medical simulations, and video editing. - Real-time video infrastructure needs streaming, persistent session memory, sub-100-millisecond latency, and GPUs close to users. ## Summary Ahmed Ahres argues that real-time video is a new medium rather than a faster version of generated video. Current systems return a file after a prompt, so the user cannot steer the result. He compares this with the shift from maps to GPS and from film cameras to digital viewfinders. In both cases, immediate feedback enabled new products and new ways to create. Ahres defines world models as video that is interactive, effectively infinite, and generated fast enough to control while it is running. He describes three model types: real-time interactive video, controllable character worlds, and live avatars, which he says still feel wrong. Possible applications include interactive entertainment, robotics data, education, medical simulation, cooking simulation, and prompt-based editing. The engineering requirements differ from batch generation because pixels stream continuously, sessions retain state, and compute must be distributed near users. He also says evaluation remains an unsolved problem that currently depends on human judgment. ## Key ideas ### Generated video is still a passive file [01:06](https://www.youtube.com/watch?v=5dCAmSDOAjI&t=66s) Ahres says traditional video generation works like a slot machine. You prompt a model, receive a file, watch it, and have no way to change what happens inside it. He asks what happens when video becomes programmable like software and pixels can be generated in real time. That would change both consumption and production because people could affect the scene while it is running. His company, Reactor, is building infrastructure and a developer platform for real-time world models so applications can integrate interactive video rather than simply download generated clips. ### Real time changes what a medium can do [02:43](https://www.youtube.com/watch?v=5dCAmSDOAjI&t=163s) Ahres uses GPS to explain why real time is more than speed. Before GPS, people consulted a map made by someone else to work out where they were. GPS made position available continuously, which enabled services such as Uber. He makes a similar comparison with film and digital viewfinders. With film, a camera operator could not see the shot while recording. With a digital camera or phone, the operator can adjust immediately. Ahres connects that feedback loop to the ability to make high-quality content for platforms such as Instagram and TikTok. ### The first world-model type generates infinite video that users can steer [04:48](https://www.youtube.com/watch?v=5dCAmSDOAjI&t=288s) The first category is similar to models such as Veo or Sora, except that it is real time and interactive. Ahres says these models continue rather than stopping after a short clip, respond to changes on screen, and do not make the user wait for a finished file. In his demonstration, a scene begins with a dog and a cat is prompted into it while the video is still generating. He says the same mechanism could add a dragon, make the dog run, or build a longer story in front of the viewer. ### Immediate feedback gives creators control [05:40](https://www.youtube.com/watch?v=5dCAmSDOAjI&t=340s) Ahres says content creators struggle with the lack of control in current generative video tools. Real-time generation lets them see the result and adjust it instead of repeatedly accepting or rejecting a completed clip. He calls instant feedback the ultimate level of control. He also describes real-time advertising, where a system could insert a logo or generate an ad based on what a person had just looked for. He expects brands to resist at first because they care about exact logo rendering, but believes real-time ads will eventually be generated in front of viewers. ### Controllable character worlds extend beyond games [06:57](https://www.youtube.com/watch?v=5dCAmSDOAjI&t=417s) The second model type lets users control a character and its environment from an image and text. Ahres connects this to games, then extends it to interactive experiences that combine movie-like scenes with user choices. He cites Netflix's Bandersnatch as an example of choosing what happens next. The same ability could produce training data for robotics because developers can control environments and generate many situations. Ahres also sees education moving beyond textbooks or language-model conversations, with students placed inside simulated historical or other learning situations. ### Live avatars have useful applications but still feel wrong [09:13](https://www.youtube.com/watch?v=5dCAmSDOAjI&t=553s) The third category is live, interactive avatars. Ahres is direct about their current limits: when people talk to avatars in customer support and similar settings, the result still feels strange. He expects better avatar models to combine with controllable worlds and real-time video for customer support, training, sales, gaming, and streaming services. His point is that the category exists and is appearing in research previews, but it has not reached a convincing level yet. ### Early applications let audiences and users change simulations [09:49](https://www.youtube.com/watch?v=5dCAmSDOAjI&t=589s) Ahres describes several things Reactor users are building. Interactive live streams could let viewers type what should happen next and vote on the outcome. Medical simulations could let trainees change a treatment and see what happens. Cooking simulations could show the effect of adding an ingredient. Video-to-video models could add visual effects, remove people, or change backgrounds while editing. He says these editing platforms are not very good yet because of model quality, but prompting, talking, or clicking to edit video creates a different workflow. ### Real-time video needs a different infrastructure stack [11:47](https://www.youtube.com/watch?v=5dCAmSDOAjI&t=707s) Batch video generation sends a request to a cloud job and returns a file. Ahres says that approach cannot simply be reused for real-time inference. A real-time system must stream pixels from a server to a client and keep a live session with memory. Memory is still a problem because a character may look away and forget what happened. The system also needs sub-100-millisecond latency around the world. Users in India or Japan should reach GPUs in or near those locations, or the delay breaks the real-time medium. Ahres says Reactor exposes this through an API and describes integration as roughly ten lines of code. ### Evaluation for real-time world models remains unsolved [16:37](https://www.youtube.com/watch?v=5dCAmSDOAjI&t=997s) In the question period, Ahres says developers can use multiple GPUs, optimize model weights, and apply quantization. He says deterministic rules can be built on top of the platform, although Reactor does not do that today. He is less confident about evaluation. For real-time models, researchers still judge results by looking at them and relying on human judgment. He says consistency and fidelity are part of the unresolved research problem, including for major research groups, and that Reactor is working on it. ## Notable quotes - "What happens when video becomes programmable like software and what happens when pixels can be generated in real time?" (01:28) - "Real time changes what the medium is. It doesn't just make it faster." (02:43) - "Instant feedback is the ultimate level of control." (05:58) - "You cannot just take what works for batch inference and apply to real-time inference." (12:07) - "Evaluation for these real-time models is an unsolved problem." (16:37) ## Tools & references mentioned - Reactor - Uber - GPS - Instagram - TikTok - Veo - Sora - Bandersnatch - Netflix - Helios - ByteDance - Lingbot - Alibaba - LongLive 2 - NVIDIA ## Who should watch - You are building generated-video products and need users to steer scenes while they are being produced. - Your application depends on simulated worlds, interactive characters, robotics data, or immersive education. - You are designing inference infrastructure and need to account for streaming, session memory, geographic GPU placement, and latency.