Talk by Keegan McCallum (uRun) arguing that generative video's interesting frontier is no longer quality but efficiency and real-time serving. A $10 budget now yields roughly three hours of continuously generated video, and $50 buys fifteen hours; by contrast, a slower, higher-motion clip can cost on the order of 100x more to produce. Helios, the model he serves, is a distillation of a 14-billion-parameter open model, one of at least forty released that year with real-time or long-horizon capability. He highlights new interaction shapes beyond clip quality: webcam-driven applications (e.g., previewing a haircut), visual interfaces for people who don't think in text, and steering generations live in under a second instead of gambling on prompt lottery. The remaining hard problem is serving: globally distributed GPUs, WebRTC with ICE and TURN, and multiple models in one streaming pipeline synchronized frame by frame with user controls. Timestamped sections cover the quality axis, efficiency and long horizons, cost of generation, live-steering applications, and serving infrastructure. # Channel Presentation Observations ## Speaker / Channel Self-Identification - The speaker introduces himself directly: **"I am Keegan. I'm the founder of U Run, um, a new kind of inference provider focused around, uh, interactive media."** This is the only explicit self-description in the material. - The video title identifies the speaker as **Keegan McCallum, uRun**, consistent with the transcript's self-introduction. ## Claimed Scope - **Inference serving for interactive/generative video**: The channel (via this talk) presents uRun as a company that **serves generative video models**, demonstrated by: *"what you're watching here is a demo for a model called Helios that we serve at U Run."* - **Developer tooling**: Claims of building application infrastructure — *"what if there was just a React component that you could drop into your application"* and *"a programmable Python runtime that lets you easily build these complex pipelines."* - **Agent tooling**: *"we've actually built one that will let folks hook into a CLI or an MCP server and build these kinds of applications."* - **Positioning statement**: *"the models are here and the frontier is really in how we serve them."* ## Intended Audience (Observable) - **Developers and builders of AI applications**: The talk references technical specifics the audience is assumed to know — WebRTC, ICE, TURN, MCP servers, React components, GPU deployment. - **An in-person, technically engaged audience**: Audience interaction is present — *"Show of hands, who here has burned 10 or even $50 worth of tokens in an hour with Clockwork? A lot of people."* - **Potential partners and hires**: The talk closes with explicit recruitment — *"we are looking for design partners who want to push the boundaries of human human-computer interaction and we're hiring at YURUN."* ## Demonstrated Presentation Patterns (vs. Claims) - **Educational/demo format**: Walks through a quality timeline of generative video (Will Smith spaghetti 2023 → Sora → Sora 2 → SeeDance) before pivoting to the efficiency axis. - **Quantitative framing**: Uses cost figures as evidence — *"$10 can get you 3 hours worth of generated video continuously"* and *"$50 would give you an entire day."* - **Product pitch embedded in talk**: Transitions from industry overview (*"at least 40 models with real-time capabilities... released this year"*) into uRun's own tooling, a conference-talk structure rather than pure tutorial content. - **Use-case advocacy**: Covers accessibility (*"some people... think visually and learn better that way"*), magic-mirror applications, and real-time content steering as framed benefits. ## Notes on Evidence Limits - No channel-wide claims (subscriber counts, upload schedule, channel description) are present; observations derive solely from this single talk transcript and title. - Technical claims (e.g., Helios being *"a distill of Juan 2.1 14B"*) are creator assertions, not independently verifiable from this material. # Delivery Observation — Keegan McCallum, "Generative Video at the Speed of Light" ## Tone - **Casual, conversational demo-day style.** Filler words appear throughout: "um," "uh," "you know" — e.g., "a new kind of inference provider focused around, uh, interactive media." - **Enthusiastic and optimistic**, expressed through superlatives rather than dramatic phrasing: "SeeDance this year, um, absolutely incredible," "it's pretty incredible to see how fast these are improving." ## Vocabulary & Framing - **Heavy use of concrete numbers and cost anchors:** - "$10 can get you 3 hours worth of generated video" - "$50 would give you an entire day" - "generated for about a 100th of the cost" - **Relatable, informal metaphors:** - "nightmare fuel" (Will Smith spaghetti clip) - "a slot machine type approach" (current content creation) - "magic mirror" (repeated twice for video-to-video use cases) - **Narrative arc via named model progression:** Will Smith spaghetti → Sora → Sora 2 → SeeDance, building a timeline of quality improvement. ## Recurring Linguistic Habits - **Repeated "So," as a transition/launch word:** "So, we hear a lot about generative video…", "So, what you're watching here is a demo", "So, one such use case would be", "So, you're going to need GPUs". - **Repeated "And so," for summarizing turns:** "And so we're at a place right now…", "And so, what does it look like to actually build an application like this?", "And so the models are here…" - **Doubled/echoed words from self-correction mid-sentence:** "but it it's getting there", "it's it's no wonder", "can can transform", "we we don't just need platforms" (transcribed as "don't just need platforms"), "experiment and and build". ## Teaching / Presentation Habits - **Audience interaction:** a literal show of hands — "Show of hands, who here has burned 10 or even $50 worth of tokens in an hour with Clockwork?" — followed by audience response ("A lot of people."). - **Live demo commentary**, narrating on-screen visuals: "the generation in the bottom right corner, you'll see, is a long continuous generation." - **Self-deprecating asides about the demo:** "I kind of ruined it with the last club [clip], but you can guess which one is real time." - **List-structure organization** for use cases and requirements (magic mirror, accessibility, content creation, world models; GPUs, WebRTC/ICE/TURN, streaming workflows) — delivered as rapid parallel bullets in speech. - **Explicit reframing move** to steer the talk's thesis: "I'm here to talk about another axis which models are improving along, which is efficiency." ## Closing Pattern - **Pitch-style close:** time self-awareness ("I went way over I went way under time"), a call for design partners, a hiring mention, and an invitation: "come see me after the talk." ## Uncertainty Notes - Audience reactions ([laughter], [gasps and sighs]) and speaker sounds ([snorts]) appear in the transcript; the snorts seem to accompany specific lines but the cause is not observable from text alone. - Only one talk is available, so all patterns above are **single-sample observations** — "So,"-launching and enthusiasm markers are frequent within this talk but repetition across talks cannot be verified. # Position Extraction: "Generative Video at the Speed of Light" — Keegan McCallum, uRun ## Speaker Identification - **Keegan McCallum**, founder of **uRun**, an inference provider focused on interactive media. All substantive claims below are his. ## Central Thesis - Most discussion of generative video focuses on **quality** improvements at the frontier; McCallum argues the equally important axis is **efficiency and long-horizon generation**, and that "the models are here and the frontier is really in how we serve them." ## Claims and Reasoning ### Quality progression (framing, not his main point) - 2023 "Will Smith eating spaghetti" was "nightmare fuel"; Sora (2024) improved but had "an AI feel"; Sora 2 better; SeeDance (this year) "absolutely incredible, so photorealistic." - Reasoning pattern: quality is improving so fast that last year's frontier quality is now cheap and fast. ### Efficiency claims - **Helios**, served at uRun, is a **distill of Wan 2.1 14B**, released in March. - Demo: continuous long-horizon generation plus clips generated "faster than you can consume them," roughly at last year's frontier quality. - One generated video was "arguably a bit better" than the comparison (better motion) at "about a 100th of the cost." - **"At least 40 models"** with real-time and long-horizon capabilities released this year. - Cost benchmarks (his figures): **$10 ≈ 3 hours** of continuous generated video; **$50 ≈ 15 hours / "an entire day"** of visual AI interaction. (Audience show-of-hands: many had burned $10–$50 of tokens in an hour with Clockwork — his implied contrast.) ### Model taxonomy he cites - **World models**: consistency over long horizons, fine-grained camera/viewport control. - **Avatar models** (mentions "lemon slice" as a recent example). - **Video-to-video models**: real-time transformation, "almost like a magic mirror." ### Envisioned use cases - **Magic mirror**: webcam-based, try outfits, cars, haircuts in real time. - **Accessibility**: text-heavy AI interaction disadvantages people who "think visually"; visual companions could broaden access to what coding models provide. - **Content creation**: shift from "slot machine" prompting (~$10/minute per shot attempt) to **real-time steering under a second** while generating; piloting/overseeing agents; rendering full-fidelity clips with "Google Gemini Omni." - Wants to expand horizons beyond "basic world models." ### Infrastructure argument and uRun's pitch - Building such apps requires: global GPU distribution, user-to-GPU routing, **WebRTC/ICE/TURN**, multi-model streaming pipelines, frame-synchronized controls. - uRun's solution: a **drop-in React component** plus a **programmable Python runtime** for asynchronous pipeline building (avatars, video-to-video, experimentation). - Position claim: "in 2026, [we] don't just need platforms, we need **software factories**" for agents — offered via CLI/MCP server hooks. ## Qualifications and Tentativeness - "I kind of ruined it with the last [clip]" — acknowledges demo imperfection; asks audience to "guess which one is real time." - "arguably a bit better" — hedged comparison. - Use cases framed speculatively ("what if," "I also think about," "I'm excited about"). - Notes he "went way under time" — talk ended early, claims may be underdeveloped. ## Interests Disclosed - Explicit commercial interest: uRun serves Helios, is **seeking design partners and hiring** — the talk itself functions partly as a pitch. # Speaker Attribution Record ## Video: *Generative Video at the Speed of Light — Keegan McCallum, uRun* ## Speakers - **Keegan McCallum** — **creator / presenter** (self-identified: "I am Keegan. I'm the founder of U Run"). Sole substantive speaker throughout. - **Unidentified audience members** — audible **[laughter]**, **[gasps and sighs]**, and a show-of-hands response; no spoken lines attributed. ## Roles & Affiliation - Keegan speaks as **founder of uRun**, an **inference provider focused on interactive media**. ## Promotional Segments (Flagged) - **uRun / Helios demo**: pitches a model ("Helios") served at uRun; a distill of "Juan 2.1 14B" (as transcribed). - **Product pitch**: a "React component" and "programmable Python runtime" for real-time video pipelines. - **Closing call-to-action**: seeking **design partners** and **hiring at uRun** — explicit promotional close. ## Referenced Third-Party Sources & Claims - **Will Smith spaghetti clip (2023)** — early generative video example. - **Sora (2024)**, **Sora 2** — OpenAI models cited for quality progression. - **SeeDance (this year)** — cited as photorealistic frontier. - **Lemon Slice** — avatar model reference. - **Google Gemini Omni** — for rendering full-fidelity clips. - **Clockwork** — referenced in audience question about token spend. - Claim: **~40 models** with real-time / long-horizon capabilities released this year. - Cost claims: **$10 ≈ 3 hours** continuous video; **$50 ≈ 15 hours/day** of interaction. ## Provenance & Availability - **Comments availability**: not stated in source material. - **Curated metadata**: none supplied beyond title; no timestamps available in transcript. \>\> I am Keegan. I'm the founder of U Run, um, a new kind of inference provider focused around, uh, interactive media. And I'm here to talk about generative video. So, we hear a lot about generative video improving along the quality axis at the frontier. We have the classic Will Smith eating spaghetti from 2023. It is nightmare fuel and not something you would ever mistake for reality. In 2024, we got Sora and it gets a little better. It still has a bit of, you know, an AI feel to it, but it it's getting there. And Sora 2, you know, even better. But SeeDance this year, um, absolutely incredible. So photorealistic. And it's it's no wonder that we talk a lot about quality, but I'm here to talk about another axis which models are improving along, which is efficiency and, uh, the long horizon generations. So, what you're watching here is a demo for a model called Helios that we serve at U Run. Um, the generation in the bottom right corner, you'll see, is a long continuous generation. And the other video, um, is a bunch of clips, um, that have been generated faster than you can consume them. Uh, and they're about at the same quality as the frontier models were last year. They're Helios is a distill of Juan 2.1 14B. Um, and I'll talk a bit about the techniques that are used in the various models that are hitting the scene right now, but there's been an explosion in just the last year uh, terms of efficiency and capabilities. Um so like looking at this, I kind of ruined it with the last club, but you can guess which one is real time and which one was generated in a number of minutes. And the one on the right is and arguably a bit better. It's got better motion and it was generated for about a 100th of the cost. And these are just some of the charts showing the quality bar for both long and short video generation. Helios came out in March and it's it's pretty incredible to see how fast these are improving. But these are techniques that are being applied all over the place, not just the one model. There's world models which can keep consistency over long horizons and you can control in a fine-grained way, the camera and the viewport. There's avatar models like we just talked about with lemon slice and there's video-to-video models that can can transform what you're seeing in in real time, almost like a magic mirror. There's actually been an explosion of innovation. There's been at least 40 models with real-time capabilities and long horizon generation capabilities released this year. Show of hands, who here has burned 10 or even $50 worth of tokens in an hour with Clockwork? A lot of people. And so we're at a place right now where $10 can get you 3 hours worth of generated video continuously with most of these models and $50 would give you an entire day interacting with an AI in a visual medium. 15 hours. And so I want to talk a little bit about the different things this enables in terms of the way that we interact with computers. Um and I'll talk a little bit about what we're doing at You Run to try and make it easier for folks to experiment and build out applications like this. So, one such use case would be a magic mirror. You \[snorts\] could have your webcam and you could ask to see yourself in any outfit. You could ask to see yourself in a car you like or with a haircut you're considering. Um a lot of different possibilities because these are open-ended models that can transform what they're seeing on a webcam in real time. I also think about accessibility a lot with these models. Um you know, working with AI involves a lot of reading and a lot of text. For some people that's more difficult. Um for some people they just don't think in in text. They think visually and learn better that way. Uh so, there's more opportunities to have companions or visual mediums that are going to allow more people to experience the things a lot of us have with coding models. And I'm excited about content creation. Um so far we've very much had a slot machine type approach where you're setting up a prompt and maybe some key frames and spending about $10 a minute to try and get the shot that you want. But with these models you can actually steer them in real time in under a second while they're generating and get the actual shots that you want. Maybe you're piloting an agent that you're able to look over its shoulder and see what it's generating in real time. Um but you're able to more granularly control the content you're generating and with modern models like Google Gemini Omni, you can actually render these out as a more full fidelity clip. And of course, we all are thinking about world models, but I want to take the the focus off of just kind of the the the basic world models that we talk a lot about and just try to expand the horizons of what we can do with this technology. And so, what does it look like to actually build an application like this? So, you're going to need GPUs all over the world potentially if you've got a global audience that's going to be using these. You're going to need to think \[snorts\] about where you're connecting the users to, what GPUs you're going to use to serve them. You're going to need to set up probably WebRTC and ICE and TURN. And for the most interesting use cases, you're going to want a model wire multiple models together in continuous streaming workflows. Building those real-time harnesses, and you're going to want things synchronized with your controls that you're providing to your end users with every frame and continually providing a smooth streaming experience. And so, our idea is what if there was just a React component that you could drop into your application to make it easy to provide video interactively inside your applications with any model. And behind the scenes, there's a programmable Python runtime that lets you easily build these complex pipelines generating asynchronously so that you can build avatar models, you can build these video-to-video transformation models, you can experiment and and build whatever you can really imagine on top of these. And I argue that in 2026, don't just need platforms, we need software factories and ways for agents interact with these. And so we've actually built one that will let folks hook into a CLI or an MCP server and build these kinds of applications. \>\> \[gasps and sighs\] \>\> And so the models are here and the frontier is really in how we serve them. Uh I went way over I went way under time. Um \>\> \[laughter\] \>\> but we are looking for design partners who want to push the boundaries of human human-computer interaction and we're hiring at YURUN. Um so come see me after the talk if uh if you're interested in chatting more.