AI Engineer World's Fair 2026
The Next Medium: Why Real-Time Interactive Video Changes Everything — Ahmed Ahres, Reactor
Read the talk
When Video Becomes a Live, Programmable Medium
Real-time video generation replaces the finished file with a session people can steer, creating new applications and new demands on streaming, memory, latency, and evaluation.
From a talk by Ahmed Ahres
What if you could change the video while watching it?
A prompt produces a video file. You watch it, decide what you wanted differently, and try again. That is the workflow Ahmed Ahres contrasts with real-time interactive video. He uses world models to mean this live, steerable medium, rather than other uses of the term associated with techniques such as Gaussian splatting.
In the prompt-to-file workflow he describes, Veo 3 and Seedance 2 can generate impressive footage, but the returned recording cannot respond to you while it plays. The question is what happens when pixels become programmable like software: generated continuously, with inputs that can change what appears next.
Ahres introduces himself as Reactor’s head of go-to-market, with a background in computer vision and machine learning, experience as a founder, and a hobby of shipping iOS and Android games. He describes Reactor as a Series A company building the infrastructure and developer platform that make these models accessible to application builders. The underlying problem is straightforward: generating a recording does not make it interactive. Real time changes the medium, not just the speed of production.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Immediate feedback creates different applications
A printed map gives you a representation someone prepared earlier. GPS adds a continuously updated position: you can see where you are and track movement. In Ahres’s analogy, the consequence is larger than locating yourself faster. Live location becomes an input to an application such as Uber.
Photography supplies a second feedback loop. With film, you shoot first and see the developed result later. A digital camera or iPhone shows what you are recording as you record it, letting you adjust the shot immediately. Ahres connects that ability to the quality of everyday content creation and, in turn, to Instagram and TikTok, while acknowledging that other factors contributed. The useful mechanism is the shortening of the loop between action, observation, and adjustment.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Steering a continuing video
Programmable video can be addressed, conditioned, and changed. The first model category resembles Veo or Sora, but with a continuing stream instead of a fixed-duration result. Ahres describes three properties together: generation continues, inputs can alter it, and the viewer sees the output as it is generated. His description of an infinite stream concerns continuing generation; it does not establish unlimited reliable memory.
The demonstration starts with an image of a dog. While the video continues, Ahres introduces the prompt “A cat shows up” and describes a cat appearing in the generated scene. The intervention happens during generation, rather than requiring a replacement for an already returned file. He then proposes further directions—a running or jumping dog, a dragon arriving, a trip to the World Cup—as possible continuations of the story. Those are suggestions, not additional demonstrated outcomes.
For creators, this feedback loop addresses a familiar complaint: the model can produce a compelling image, yet still resist the specific change the filmmaker wants. Ahres summarizes the appeal as “Instant feedback is the ultimate level of control.” The creator can observe the developing result and steer again, instead of relying entirely on the initial prompt.
Advertising follows from the same mechanism. Recent search interest could condition a logo inserted into a scene or an advertisement generated during viewing. Ahres is personally unenthusiastic about that application but expects it to grow. He also identifies an adoption barrier: brands care about exact visual identity, down to an incorrectly colored logo pixel. His eventual disappearance of preproduced ads is a forecast, contingent on overcoming that sensitivity.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Navigable worlds and live characters
The second category is a controllable world, exemplified by Google’s Genie 3. An image and, typically, text establish the environment; the user then controls a character within it. Games are the obvious application, but combining navigation with steerable video also creates experiences between games and movies. Ahres reports community projects exploring that combination. Netflix’s Bandersnatch, where viewers select the next scene, supplies a familiar analogy for audience agency.
Controllable simulation also suggests applications beyond entertainment:
- Robotics: Vary an environment and the actions within it to generate training data. Ahres describes substantial activity among robotics labs and the prospect of effectively unlimited simulated examples.
- Education: Place a learner inside a situation, such as a historical setting, rather than only describing it. This reflects his preference for experiential learning over a future centered on textbooks or LLM conversations.
The third category is the live interactive avatar. Ahres acknowledges that existing avatars, including those used for customer support, often feel strange or unconvincing. He expects emerging research-preview models to become more useful when combined with continuing video generation and controllable environments. Proposed applications include customer support, training, sales, gaming, and streaming services.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Applications built around audience input
An interactive livestream makes the control loop visible. Viewers propose what should happen next, vote, and influence the continuing video. Ahres reports Reactor users building this kind of experience, with X, YouTube, and Twitch as possible distribution surfaces. The application coordinates collective input; the model generates the next visible development.
The vote-selection step can be kept separate from model integration. This JavaScript function accepts a map from viewer IDs to their current choices, so each viewer contributes one active vote. It returns a proposed next prompt; sending that prompt to a live session is a separate operation.
javascript
function selectNextPrompt(votesByViewer) {
const counts = new Map();
for (const choice of votesByViewer.values()) {
const prompt = choice.trim();
if (!prompt) continue;
counts.set(prompt, (counts.get(prompt) ?? 0) + 1);
}
const ranked = [...counts.entries()].sort(
([promptA, countA], [promptB, countB]) =>
countB - countA || promptA.localeCompare(promptB)
);
return ranked[0]?.[0] ?? null;
}
Here, equal vote counts use alphabetical order as a simple application policy. The result remains a selection until the application submits it to generation.
Other reported projects apply the same input-and-response pattern to different activities:
- Medical simulation: Add or remove a medicine and explore what happens next as a training playground for aspiring doctors. These are reported application concepts, not evidence of clinical validity.
- Cooking simulation: Add an ingredient and inspect the generated result.
- Video editing: Use video-to-video generation to add visual effects while editing through prompts, speech, or clicks.
Ahres says the editing platforms are not yet very good because model quality remains limited. Their significance is the interaction pattern: a user directs changes while seeing the developing output.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Four models behind the API
Reactor’s platform presentation turns those interaction patterns into four model choices:
| Model | Intended workflow |
|---|---|
| Helios | Interactive video generation |
| LingBot-World | Navigable generated worlds |
| LongLive 2 | Multiple shots forming a continuing film |
| SANA-Streaming | Live video-to-video editing |
Ahres attributes Helios to ByteDance and describes LingBot as Alibaba’s Genie-like model; the LingBot repository credits the Robbyant Team. He identifies LongLive 2 and SANA-Streaming as NVIDIA models. With LongLive 2, prompts can be prepared in advance to guide a consistent story across shots.
SANA-Streaming starts with existing footage rather than generating every element from scratch. A creator can upload recorded video or a Seedance 2 result, then add effects, remove people, or change the background. Ahres points to Hollywood previsualization as a useful setting for that workflow: explore a scene’s visual treatment before committing to its final production.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
A live session is not a batch job
Batch video generation can be simplified to a request, a cloud job, and a returned file. A live model needs infrastructure that keeps serving an evolving experience.
| Concern | Batch generation | Live generation |
|---|---|---|
| Unit of work | Request or job | Continuing session |
| Delivery | Completed file | Streamed video |
| State | Bounded generation task | Context retained across interaction |
Streaming pixels from a server to a client adds delivery concerns that a file-returning workflow can avoid. Keeping the session running also makes memory a continuing responsibility.
Scene memory is a model limitation as well as a systems concern. Ahres recalls Genie 3 demonstrations where a character looks back and the world fails to preserve what was previously there. Turning left and then right should not casually replace the environment. Maintaining enough context to preserve that continuity remains difficult; a continuing stream alone does not solve it.
Geography adds another constraint. Ahres sets sub-100-millisecond latency worldwide as a requirement for the experience, not a measured Reactor service guarantee. A user in India or Japan should reach a GPU in that country or as close as possible. If compute is too distant, the delay interrupts the feedback loop that makes the medium interactive. Network latency, generated frame rate, and the time until an input visibly changes the scene are separate quantities; a single latency number cannot describe all three.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Connecting an application to the stream
The developer-facing goal is to hide much of that infrastructure behind a small integration. Ahres estimates roughly ten lines of code, while acknowledging that the example is simplified. The basic sequence is to authenticate, load a model, and connect its output and controls to a video application, image workflow, or plugin. The slide shows the intended compact integration rather than the full operational responsibilities of an application.
For onboarding, the talk offers a signup QR code and the promotion code AIE2026 for $75 in credits. That is the offer made during the presentation; its continued availability is not established here. Ahres also characterizes the platform’s models as inexpensive, without supplying a cost comparison.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Frame rate, broadcasting, and simulation rules
An audience member cites a stated 16 FPS for the game generator and asks what it would take to reach 30 FPS. Ahres names three approaches: multiple GPUs, model-weight optimization, and quantization. He frames the work as a matter of priorities, without reporting an achieved 30 FPS result.
A subsequent question suggests showcasing the technology at IBC, the International Broadcasting Convention in Amsterdam. Ahres says he was unfamiliar with the event and had not planned a showcase, but would look into it. The exchange connects the proposed medium to an existing industry concerned with producing and distributing video.
The next question asks about deterministic grounding: checking generated simulations against rule sets so that their behavior stays constrained. Ahres says Reactor does not provide those checks itself. He reports developers building such layers on the platform, open-sourcing them, and Reactor reusing community work. The division of responsibility is explicit: Reactor concentrates on the substantial infrastructure workload, while application builders can add rules that constrain a particular simulation.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
How do you know the world stayed consistent?
The final substantive question concerns measurement: how should visual consistency be evaluated? Ahres distinguishes fidelity, which he considers comparatively easy to assess, from the broader challenge of evaluating an interactive world over time. He acknowledges evaluations for real time, consistency, and fidelity, but describes overall evaluation as unsolved and current assessment as heavily dependent on looking at the result and applying human judgment.
That should not be read as an absence of quantitative evaluation. The LingBot-World paper reports VBench evaluation, including temporal and consistency measures, while acknowledging that world-model evaluation protocols remain immature. Existing video metrics and a solved general test for interactive worlds are different things. Ahres’s claim that nobody, including DeepMind, has solved the broader problem is his assessment of that research gap.
Asked what Reactor is doing about it, Ahres says its research team is working on the problem, without describing a method. The presentation therefore ends with an open engineering question: once users can act inside a generated stream, the system must be judged not only by how its frames look, but by whether the experience remains coherent as those actions accumulate.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
The autoregressive video-generation model's design, long-video training approach, and reported single-H100 performance.
Action-conditioned world generation, streaming distillation, memory demonstrations, and quantitative video evaluation.
NVIDIA's long-video generation infrastructure, including multi-shot support and quantized inference.
Streaming video editing through a hybrid diffusion transformer, consistency training, and GPU optimizations.
The original overview of navigable generated worlds, promptable events, agent experiments, and acknowledged limitations.
Updates since the talk
Current session setup and commands for steering a continuing video stream with prompts and reference images.
Current API examples for scheduling shots, making hard cuts, and managing a continuous multi-shot session.
Read the complete timestamped transcript
- 0:00
[upbeat music] Hi, everyone. Uh, welcome to the talk. First of all, thank you all for making the time.
- 0:16
I know it's the last talk of the day probably, or I think the last one is at, uh, three forty-five. But, uh, yeah, thank you all for your time.
- 0:22
I, I know you're all probably very busy. Um, today I'm gonna be talking about something that is a little bit slightly futuristic, though not for San Francisco, and that's World models.
- 0:31
Um, I know here it's written real-time interactive video, but the way we think about World models is really in the real-time interactive video, and I'll explain why. Um, and in today's world, I think World models is a little bit of a marketing term that people think about it from a Gaussian splatting standpoint or others from video.
- 0:47
Uh, but the way we define World models is really real-time interactive video, and I have strong evidence or, like, beliefs that this will actually change everything in how we produce and consume content.
- 0:59
Oops. What happened? Sorry. [clears throat] Sorry about that. Cool. [clears throat] So if you think about video, video has always been something passive. Before, people used to produce videos and movies, and then we would watch it.
- 1:19
And in today's world, we have these models, the Veo 3, the CDense 2, and what they do is you prompt them, you get back a file, you watch, and good luck.
- 1:28
It's a slot machine. You cannot change it. You cannot do anything about it. So there's a question that, um, we like to think about in our company is what happens when video becomes programmable like software?
- 1:39
And what happens when pixels can be generated in real time? Once this happens, it actually changes completely how we think about consuming content and even producing content, and I'll be talking about how in history we've seen that real time has always been the future and the co- kind of applications that it unfolded and how you can get
- 1:56
started today. Quick, quick background here. Um, so my name is Ahmed. I'm the head of go-to-market at Reactor. My background is in computer vision and machine learning. Um, and I state machine learning because this is the time where we used to actually train our models.
- 2:12
Um, I built and shipped games on iOS, uh, on iOS and Android for fun. Um, I was a founder, um, and today, I'm the head of go-to-market at Reactor.
- 2:21
And who we are, quickly, is we're a Series A company building the platform for these real-time World models. So, so far, main-- most of them have been still models, but we are building the infrastructure and the developer platform to make them usable so that anyone can integrate these real-time interactive video, and we can democratize access to this
- 2:40
technology. I'll start with a problem. We can today generate pretty much anything, but we just can't change it, right? A generated video, as I mentioned earlier, is still a recording.
- 2:52
You get it back. You can't do anything about it. And real-time changes what the medium is. It doesn't just make it faster.
- 2:59
And I'm gonna talk about two examples that actually show us what real time has unlocked in the past. Before, in the nineteen fifties and even before, we used to look at a map to know where we are.
- 3:10
Someone produces a map, you look at where you are, and that's it. You cannot do anything about it. Then GPS came. GPS made it real time. Suddenly, I can know where, where I am at the instant, and I can, you know, track it.
- 3:21
Now, you think GPS has just made it a bit faster to know where I am, but actually, Uber would not exist if we did not have the GPS. Another example, which is even a bit more powerful, I'd say.
- 3:34
Before we used to d- to use the film to produce content, right? We had a film. Someone shoots something. They can't see what they're shooting. They go somewhere. They produce that film, and then you can see it.
- 3:45
And then it became digital. You can start seeing what you're shooting. If today you pick up your iPhone and you start recording a video, then you can see what's going on, and that's why we can produce high-quality content.
- 3:56
It's because you are able to see what's going on in the screen and adapt accordingly. That gave rise to In- Instagram and TikTok. Instagram and TikTok would not exist if we could not produce high-quality content, and the only reason why we're able to produce high-quality content, among many reasons, is because we can see in real time what's
- 4:14
happening. It's not a slot machine. We can actually just edit and see, and that's what unlocks all of these new use cases.
- 4:24
So when video becomes programmable, you can address it, you can condition it, you can change it. I can show you on the screen whatever I wanna show you on the screen, and it becomes programmable a little bit like software or like anything that is programmable in the world.
- 4:39
And in, in the market today, we are seeing three kinds of models that do this. Some of them you will be familiar with, others you will not be familiar with.
- 4:48
The first one is these are Veo-- think about Veo or Sora, but real-time and interactive, meaning that, first of all, they're infinite. So they don't stop after five, ten, or thirty seconds.
- 5:00
They actually continue forever. They're interactive, meaning you can change what's happening on the screen,
- 5:06
and they're in real time, so you don't need to wait to see what's going on. And assuming this works. So this is an example of a, of a video that I passed an image with a dog, and this was all generated in real time.
- 5:17
And at some point, I'm gonna prompt, "A cat shows up," and you will see that a cat showed up in the, in the video. This would not be possible in the existing batch regular video generation models because you would get back the video, and you cannot do anything about it.
- 5:30
I could have added anything. I could have gone on to create an entire story with it. I could have said the ca-- the dog starts running, starts jumping. A ca-- a dragon shows up.
- 5:38
It goes to, I don't know, to the World Cup. All of this would have happened in front of you.
- 5:43
These type, these types of models pro-- like, unlock a few things. First of all, control. So if you think about generative media today, the big problem that content creators all have is I don't have the control I need.
- 5:55
Like, yes, it's great to use CDense 2 or Veo 3 to generate videos, but I just don't have the control. And this is always the thing that any filmmaker or movie producer or any content creator will tell you.
- 6:07
And so real-time actually ends the slot machine type of mentality and actually gives you the cr- the control that you need. And a big saying I like to say is, "Instant feedback is the ultimate level of control," and you-- we will never be able to have this level of control if we don't have real-time.
- 6:24
The second thing it unlocks which, you know, among other things, which is a field I'm not particularly fond of, but I think is gonna be big, is advertising. If I can know what you looked for a minute ago, why can't I insert the logo of whatever you've been looking for?
- 6:39
Why can't I produce an ad in real time in front of you? We don't need to produce, pre-produce anything. Now granted, this is gonna take some time because, you know, brands afraid of AI, afraid of, uh, you know, if their logo has one pixel that is white instead of dark.
- 6:55
But it will happen eventually. And so-- And I think at the moment that happens, we will not be to-- need to produce any ads anymore. Everything will be happening in front of you in real-time.
- 7:06
The second type of model is the one that probably you're most familiar with. It's the Genie 3 like from Google. These are models where you can pass an image and a text typically, and you can control a character.
- 7:18
They're fun. The first thing that you think about when you think about this is games, right? It's a character, it's a world, you can generate anything. But actually--
- 7:28
It actually goes way far, way beyond games.
- 7:32
It creates entire new interactive experiences where, combined with the first types of models that I talked about, we've already been seeing people in our community building a mix of games and, and movies.
- 7:42
If you've ever watched Bandersnatch from Netflix, which is the ga-- the movie that you can pick your next scene, this is one of those things that gets, uh, that, that, that becomes possible, that you can control a character, you can control what's happening and create entire new types of interactive experiences that were not possible.
- 8:00
The second thing is robotics. So because you can simulate and you can control, you can actually create as much training data as, as you want. And robotics-- World models in robotics is actually a giganormous market today.
- 8:12
Um, I cannot tell you the number of robotics labs that, um, are training and building these models. But because you can control whatever you wanna control in any environment, this creates a new opportunity to generate infinite amount of data for robotics.
- 8:27
And finally, something I like to think about, this is more maybe a passionate thing that I have, is education. Um,
- 8:33
because you can step into anything, who-- With today's world in AI, I don't actually believe that edu- in the future of education is LLM-based or textbook-based. If you can put any, any kid in the situation, for example, in a history lesson, that enables entire new types of edu-- experiences that, uh, can be educational.
- 8:56
The third type of model is probably the type of model that, you know, is more, let's say, um, co- something that we've been seeing before, which is avatars, but live and interactive.
- 9:06
The thing with avatars though is it hasn't actually been cracked. They're still all kinda weird. If you speak to an avatar in any customer support or anything, it's still kinda off, right?
- 9:16
Um, and these types of models, and we're seeing a rise of these live and interactive avatar models in research preview, that combined with Model one and Model two is actually gonna be, I believe, a, a big change in what we've been seeing so far.
- 9:30
And this will be applied to things like customer support, training, sales, gaming, streaming services, et cetera.
- 9:39
And so ju-just to give you a glimpse of what our users are building today at Reactor with these types of models, some of them are building interactive livestream, right?
- 9:47
A livestream where people are watching, and then the users can type what happens next, and then they vote. Why? Because pixels can be generated in real-time. So there's no reason why I cannot put a livestream on X, YouTube, or Twitch and enable users to pick what happens next.
- 10:02
Something that was surprising to me is a little bit on, on the medical simulation. So we've seen users create applications where you generate a world, and then you simulate what happens next.
- 10:12
What if I put this medicine? What if I remove this medicine, right? And this creates-- can be a training playground for people wanting to become doctors.
- 10:21
The third one, which is also kinda surprising, is cooking simulation. So people are building applications where you can simulate cooking and what happens if you put this ingredient.
- 10:32
And finally, video editing. With video-to-video models, video editing becomes very interesting because I'm able to just add visual effects in real-time, and we've seen people build entire video edit-editing platforms.
- 10:46
Now granted, they're not very good yet, just because of the quality of the models, but it's a new paradigm once you're able to v-- edit videos just via prompting or via talking to it or via clicking.
- 10:59
And now for the final part, how do you actually do all of this? And this is why I like to say the world behind an API.
- 11:07
At, at Reactor, we have four types of models today. First one is called Helios, which is the interactive video model that I talked about. This one is from ByteDance.
- 11:15
LingBot, which is a world model like Genie 3, trained by Alibaba.
- 11:20
LongLive 2 from NVIDIA, which is multi-shot film. You can prompt things in advance and create a consistent story over time.
- 11:30
And SANA-Streaming, also from NVIDIA, which is a model that does video-to-video editing. So people are already using this, for example, by shooting something or creating something on Seedens2, uploading it, and then adding visual effects, removing people, adding background.
- 11:44
And this be-- gets very interesting in previsualization, for example, for Hollywood movies.
- 11:51
And under the hood, when we talk about infrastructure,
- 11:54
the thing that I think I like to drive home is building infrastructure for regular video generation models is very different from real-time. Because in regular video generation models, you're talking about requests.
- 12:06
You just send a request, a job gets run in the cloud, and I'm oversimplifying here, but you know, a job runs in the cloud, and it gives you back a file.
- 12:14
With real-time, it's a different ballgame. Uh, you cannot just take what works for batch inference and apply to real-time inference. For example, you need to r-think about streaming, right?
- 12:25
Once you need to think about streaming, um, pixels, uh, from a server to the client, um, it adds entire new-- entire complexities that batch generation does not have to think about.
- 12:37
The second one is that everything is a live session, so everything runs constantly, and there's memory to be kept into account. Now, granted, one of the things that live vi-- live real-time models struggle with is memory.
- 12:49
If you've seen demos from Genie-3, for example, we've all seen that the character can look back and then not remember what-what's going on. So there is a lot of work that needs to be going into maintaining that context window so that you can remember what happened if you turned your character left and right.
- 13:06
And finally, global scale. If you th-if you think about real time, it needs to be sub one hundred millisecond latency anywhere you are. And if you're deploying applications in the world, then someone based in India or someone based in Japan should be routed to a GPU that is based in India or Japan, or as close as possible
- 13:25
to it. If not, if you don't have the compute worldwide, then the experiences are not real time anymore, and it breaks completely the medium.
- 13:35
And with Reactor, this is as easy as it gets to integrate these real-time models. Okay, I kind of maybe oversimplified it a little bit here, but it really is maybe ten lines of code.
- 13:44
Um, and we have docs to do all of this. But essentially, you can just load the model with an API key and just start integrating it in whatever video, image, plugin, anything that you're building.
- 13:57
And if you wanna get started, there's a QR code there, and I've added a promo code, AIE2026, which will give you seventy-five dollars off. Well, seventy-five dollars worth of credits, which is, um, significant amount of compute, uh, in our case because we make the models extremely cheap.
- 14:15
Thank you very much. [audience clapping] I'm happy to take questions if anybody has a question.
- 14:25
Oh. Hi.
- 14:28
So I saw the sixteen FPS was the stated, uh, frame per second on the video game generator. What would it take to get that to something more like doable thirty FPS?
- 14:39
Uh, well, one thing that actually we do is multi-GPUs. So we use multiple GPUs, um, optim-optimizing the model weights, applying quantization techniques. So there are ways, it's just a matter of priorities, but it is, uh, there are ways around it.
- 14:56
Yeah. Good question, though. Hi. Yeah.
- 15:08
Uh, I'm, I'm, I'm wondering if you are gonna be showcasing something at IBC in Amsterdam?
- 15:15
Sorry?
- 15:16
Are you aware of what the IBC is?
- 15:18
Uh, no.
- 15:19
The International Broadcasting Convention.
- 15:21
Okay.
- 15:21
There is, uh, I was wondering if you're gonna be presenting some showcasing regards of the, uh, next generation.
- 15:27
I wasn't, but now I'm gonna look at it. Yeah. [laughing]
- 15:31
Yeah. Yeah.
- 15:35
Yeah. How do you feel like about comparing, uh, some deterministic engines too inside-
- 15:41
Yeah
- 15:42
... of the traditional 3D Russell? Can we get a lack of feedback maybe?
- 15:45
Yeah.
- 15:46
Do, do you guys experiment with that? Can you share something?
- 15:49
When you say deterministic engines, what do you mean exactly?
- 15:53
Oh, any, any kind of, like, deterministic rule sets that you check against maybe, so the simulation stays, uh, grounded. I would, that's how I would describe it.
- 16:03
Hmm. So we don't do any of that today. Um, the reason why also we build a developer platform, and but we've seen people build that on top of us.
- 16:11
Um, building the infra for this is already a lot of work. Um, and I think we, we were seeing already developers building it and then open sourcing it, and then we reuse it.
- 16:20
Um, but so the community is doing it for us, which is even better. Yeah.
- 16:25
Is that a question? Yeah.
- 16:28
How do you measure and evaluate the visual
- 16:32
consistency?
- 16:33
Uh, you're asking a question that the entire research community in world models has not answered. There are evals for real time and consistencies, uh, and fidelity. Well, fidelity is easy.
- 16:43
It's just like pixels, right? But, um, evaluation for these real-time models is an unsolved problem. So today, it's literally just look at it and
- 16:54
human judgment. That's what it is today. And this is including, by the way, G-DeepMind and everything. Nobody has solved this problem yet.
- 17:03
So what are you guys working on?
- 17:04
Uh, we're working on it. We have a research team. Yes. [laughing]
- 17:11
Awesome. Cool. Well, thank you everyone. Thanks for your time. [outro jingle]