OpenAI’s GPT-6 Astra, Meta Muse & ByteDance Enters the World Model Race | Lightwork

0:00The kind of plain language step change here is that you no longer have to babysit this model. I have been frankly blown away by the work that they can do for us and how much easier and simpler my personal life has become. How much are you feeling comfortable with allowing an agent to do work on your behalf before you really want to take the reins? Like, I'm comfortable with it writing my email, but I kind of want to review it before it sends where we are in the capabilities curve of the models, which I think is fair to say is sort of superhuman in certain areas.

0:38Josh Machiz: We're back with another week, and this time we have Bucky Moore, a partner on Lightspeed's AI and Infer team, joining us as a guest co-host. OpenAI shipped GPT-6 Astra its most capable model, and by its own account, its hardest to monitor. Meta launched Muse, an agent that wants the keys to your inbox, your calendar, and your wallet. And ByteDance is reportedly building a real-time world model, and the race to serve those models is already on. Let's get into it. Bucky Moore, welcome to Lightwork. Great to be here with you both. So our first story today is all about ChachiPT's new model, Astra.

1:17It's its most capable model, but also the one where it's a little harder to understand its chain of thought. Talk to us about this model? What did you guys think? This was obviously a really big release week.

1:30Josh Machiz: Yeah, I think Greg Brockman was like, he personally thinks this is the AGI era. Curious for thoughts on that as well. Yeah. And you had Jensen confirm it also on the timeline, but I don't know. This didn't feel like AGI to me yet. I agree. This is a very, very big moment for the industry. And I think for the first time, the thing that feels different to me is that we're seeing this decoupling of understanding how these models actually work and their capabilities improving. And with that, I think you have some very interesting and perhaps concerning implications. Like the interesting implications, I think, are we have this really difficult choice that we're going to have to continue making, which is it's clear that the tactics that we're using to make these models smarter, improve their capabilities, are coming at the expense of some things that we've regarded as really important, such as understanding how they work, which of course has connotations in how we continue to align them, and more importantly, perhaps connotations in how we manage their adverse impacts, such as in cybersecurity.

2:29And so to me, it feels like there's going to have to be an extremely large thrust of resources and research effort towards how to sort of bend this curve in the right direction again, because it's clearly out of whack for the first time in a way that we haven't seen before. And I suspect that's going to take a lot of compute, and I suspect that's going to take a lot of prioritization from a research standpoint that may come at the expense of other things?

2:50Josh Machiz: I think for me, the kind of plain language step change here is that you no longer have to babysit this model. And I was seeing some demos where it was like operating Adobe Premiere for five hours or, you know, obviously taking these great demos with grains of salt, but someone was running it for multiple days to sort of rebuild an entire SimCity-like game. And so I think for me, the big takeaway was that you are seeing these step changes in computer use in how long you're able to run something like this. I think on like the OS world benchmark, which basically measures how long it's able to drive a full desktop end to end.

3:25Josh Machiz: I think the UK Institute of Safety had it running for 31 minutes over 3.6 minutes. So I think you're sort of seeing that massive change in computer use in how long it's able to autonomously run. But to your point, Bucky, I think that is sort of coming at the cost of how much you're able to monitor this model. I guess to help the audience understand, can you help us kind of ground in like why this is different structurally or architecturally? Yeah. So what you see in the model card and a lot of the discussion around this new model is that they're starting to embrace a new pattern around how inference works, right?

4:02Where previously inference would happen in a single shot. And now what you're starting to see is this notion of looping where outputs are then being fed back into another model and generating new outputs in a loop-like architecture. And that loop-like architecture, I believe, is why this interpretability challenge is arising in a new way.

4:19Josh Machiz: And historically, because you have had that forward path structure, it's given us this sort of clear chain of thought. Now that is sort of compressed a little bit. Is that the right way to think about it? That's correct. In light of this, I know you mentioned a lot of compute going into sort of steering this in the right way. I think a lot of us constantly think about AI safety here. What does that structurally look like in terms of making sure that we put in the right guardrails? Is it new ways of monitoring these models? So I thought Jacob's blog post had some really interesting takes on this.

4:52And I think to me, the takeaway from reading that and the model card is sort of twofold. The first is that we do have this window in time where we have models that are so capable that it's critical that we apply them to unearthing as many critical vulnerabilities in our infrastructure as possible while we still can. I think it's actually for the first time, I feel like a moment in time where we need to all kind of come together and ask ourselves, like, how do we do that collectively as an industry? And this is sort of a continuation of some of the discourse that spit off of the Mythos announcement as well.

5:24But I think for the first time, it's very clear that the industry is converging around this window in time notion that we really need to take seriously. And then I think the second is, what do we actually do about it that's realistic, right? And I don't think at the moment, And unfortunately, that understanding how to control and interpret the models is an immediate lever to pull. I hope that that changes as the field advances in terms of research and we're allocating more compute to that problem. But today, the best answer that I can come up with is that we need to go back to the drawing board on how we understand the environments that we're exposing these models to and become really, really methodical and rigorous with how we give these agents that we build with these models permissions, access controls, and how we actually govern what they can do.

6:05The act of doing this is nothing new. It's taken a lot of work and there are many great startups trying to make that easier in the cybersecurity field. I think this is a moment in time where those companies are going to take off in sort of a nonlinear way because of just how indispensable it becomes to have those kind of capabilities. And so these are things like zero trust architectures. These are things like just in time access controls that I think are going to increase in popularity and criticality dramatically as a result of what these models can now do.

6:32Josh Machiz: And so if I'm hearing it correctly, I think, at least for me, my sense is historically, there's been a lot of emphasis on alignment at the model level. But what you're saying is it's almost the environment that you're putting this model in, an ecosystem of companies sort of building in this space. I think that these models are becoming so capable that the only way to rein in the negative impacts of those capabilities will be to constrain the environments that they can operate in, in a way that we've never had to before. And that I think will be a goldmine of opportunity for cybersecurity companies, specifically in the access space.

7:06Josh Machiz: Are you tracking any or excited about any or any that we as a team have decided to partner with? We are tracking companies in that space very closely. And I think that there's a set of companies that we've been watching for a long time that we've thought have had great technology and solve real problems whose value propositions just exploded in criticality. And that's going to be a really exciting period of investing for us. I guess it's a little bit less of a scary safety conversation. Josh, have you been playing around with it? Yes, I certainly was playing with Asher, especially after seeing all the demos where people were taking homes, putting them into Blender, and then having Blender render something.

7:46It's crazy the kind of stuff that used to take hours of an individual's human's work just be made in maybe a few minutes by allowing computer use. I'm curious for kind of the both of you, how much are you feeling comfortable with allowing an agent to do work on your behalf before you really want to take the reins. I'm comfortable with it writing my email, but I kind of want to review it before it sends. Is that an age thing? Is that a culture thing? Where's the line? I think everyone is going to have to draw that line themselves. And for me personally, I'm still in the mode of wanting to supervise what they send, at least in a business context and supervise the work that I think has like really a one-way door-like implications, right?

8:31But I think before I turn it to Claire to share her perspective on that, I think the Blender example surfaces another really interesting insight, which is that it's clear now that the models are going to learn new skills such as 3D design and rendering on the basis of the most popular open source ways to do that. And as a result, the value of being open source and the value of being a popular open source project, I believe will dramatically inflect. And I think what you're going to see with something like Blender is that it will become the standard way for models to wield those kind of capabilities at the expense of all other options.

9:08And so I think where this gets really interesting is to think about like, what are some other open source toolkits? What are some other open source projects that are poised to benefit in that way? And I think that's something we're thinking a lot about.

9:17Josh Machiz: It was interesting seeing this sort of emphasis on computer use. I mean, I had a lot of feedback when I had explained this, that, oh, you know, like people have had co-work before, you know, co-work or chatty-peach work already has, or even codex already has access to computer use. So what is the step changer? And it is that ability to run autonomously. I would say, to your question, Josh, I think I'm a little bit, I would probably agree that like, I don't love letting it run for multiple hours and just getting this like one shot output. I've like tested a lot of these tools on even video editing.

9:53Josh Machiz: And I think one of my frustrations is sort of letting it own the full process to even, you know, what Jacob Pachagi mentioned around these models. It is still very much a jagged frontier. And in some ways, it is very good at things. And I love letting it run for certain aspects. However, there are a lot of cases where if I let it run through the entire process of planning, editing, the taste part especially, I feel like it kind of runs astray. So I don't know, I think building SimCity-like games, I can see how just replication is, or replicating websites, you can sort of let it run free. But I think in terms of these multi-stop workflows and projects, my sense is that, at least for me, I would like to also be participatory in the supervising process.

10:36Josh Machiz: I know. Sounds like we all want to be on the loop. It's just a little scary that now with the chain of thought being a little more hidden that we don't know exactly what it's doing as it's running without our consent. I guess the good thing about it from at least everything I've read, it seems like it's doing more right than it was doing before. So while you might not see what it's thinking, it does seem to be doing it more correctly and in a better direction. But that's just one model. Bucky, I had a question for you. Where do you see, it sounds like this had a heavy emphasis on agentic work.

11:12Josh Machiz: In terms of the next sort of flagship releases, GPT-7 or whatever's next for Fable, what do you feel like are the vectors that models will continue to advance? I think the main one is, you talked a lot about how in the Astra release, there are these time horizons over which the model is proving it can work in an uninterrupted way. I think what we're going to continue to see is that time horizon just continue to expand. If you think about the work that we do, some of it is over the course of minutes, hours, days, but some of it's also over the course of weeks and months, right? And as you get into more STEM-oriented fields where there are these really long-term, research-oriented, iterative processes, it's becoming increasingly clear to me that we're going to start to see the models be able to take on more and more of that work, which is going to unlock a tremendous amount of economic value in a way that I personally find will be transformative.

12:01Yeah, well, speaking of things that we allow to run without our intervention, I feel like half the internet has been using instinct, or at least the VC-verse is using instinct. Town is getting a lot of buzz. I feel like in all of the mommy chat friend groups, I hear about it all the time. And of course, Meta announced Muse, which is kind of their mainstream answer to these. What are your thoughts on this very exciting arena? You know, this kind of asks the same question, like, are you okay with an agent finding the flight, getting you the best price on the flight, but are you okay with buying the flight?

12:42Like, these are the challenges with this new arena of products that seem to be in vogue. So like many of us, I have been using these products. I have been frankly blown away by the work that they can do for us and how much easier and simpler my personal life has become as a result of entrusting, in this case, instinct with it. I used to just have this morass, like many of us, of personal emails and administrative tasks that were built up, most of which had all of the necessary context somewhere between my inbox and the web. And what I've found is that I can just offload all of that to something like Instinct, which is the one that I've been using primarily.

13:19That's been a pretty amazing unlock for me from a sanity and productivity standpoint. And I personally feel really excited about the future of these products as a result. I do think it's also noteworthy to see that the sort of chat sheet PT moment in the space appears to have come from a startup rather than one of the incumbents. And I think it's interesting to kind of ask why that is and connect the dots to some of the historic analogies that you've seen in terms of like why, for example, OpenAI was able to bring the transformer to market before Google was, who obviously invented it. And what I see is that there's a common parallel between these two events, which is to say that at least as Google tells it, the story of why they didn't release it was they were caught up in their internal bureaucracy of like, are these models going to cause harm?

14:03Are they going to expose us to unwanted risk with consumers and our users and customers? And on the other hand, OpenAI was just saying, let's just give it a try and take the risk, right? And I think what this comes back to is that startups are always going to be more nimble and have a higher risk appetite than large companies. And I think what we're seeing here with Instinct being first to market with this is fundamentally about that. And it just comes back to why we're all sitting here, why we're at a venture capital firm that exists the way it does, which is to say that startups will always have that advantage.

14:32And that is usually why startups are the ones that bring new things to the world before the large companies do. Now, at the same time, what we're about to start to see is whether the advantages that I think are structural to the larger companies, Meta, Google, perhaps OpenAI and Anthropic, are ultimately going to win out over the speed and risk appetite that the startups like Instinct and Town clearly possess. And so that's what I'm watching for most closely. But it's clear that we've reached a point where these models are so good that they can do essentially anything we need them to do on the browser.

15:03And it's clear that once again, we're seeing the power of speed and the power of risk appetite that startups have relative to big companies yield a first mover advantage. Claire, have you been playing with these? Do you have a favorite yet?

15:16Josh Machiz: I was also heavily using Instinct this weekend. I was at a wedding and also editing a video. Instinct was having it had a filter. It was chopping up silences. And so I've been able to unlock the power of Instinct and really like it. I've also been playing around with Muse. So I've been connected with the meta team, and they sort of have been working on this for a while. I think I have a lot of similar questions to Bucky Run, this kind of trade-off between safety and value. I think, you know, I had a lot of people asking, why would you even use Instinct? Like, are you comfortable giving this startup your email?

15:55Right, all of these things. It's all of these things are all trade-offs. And, you know, for better or worse, I feel like Meta already has enough of our data, but is that overall after years of social media that people, you know, are they going to be willing to give Meta their email inbox in addition to everything that it already knows about you from WhatsApp to Instagram. Yeah. It almost gives, in some instances, I think Instinct, even though they're a startup, they're a U.S.-focused startup, I feel like there is a little bit more trust, at least in the startup community with Instinct.

16:29Josh Machiz: And I also think you can bake in your own guardrails. I haven't connected it to my personal email. I mean, they just set this up yesterday, I think. Right, they launched email addresses. So I think there are steps being made in the right direction. And, you know, I only give it access to things that I feel comfortable with. So and have already derived value from it. So I think like I don't know. I think it's actually an example where as consumers, we underestimate how much people are willing to trade safety sometimes or, you know, trust sometimes for true value. And I think in this case, like I've derived a ton of it.

17:04Josh Machiz: But yeah, Bucky, I'm curious, like, have you thought about this sort of question around safety, trust? Do you think Meta, it sounds like they're leaning heavily into that. Yeah, I definitely have in my, I've just been going off of my natural instinct, I guess, pun intended in this case. But what I did was I connected it to my personal email right away. And I have been fairly comfortable giving it credentials to different accounts I use. Not to say my bank account, but I have given it credit card information. I have given it access to a number of different, you know, user accounts that I have on websites from travel to restaurant bookings to, you know, things perhaps even more high stakes than that.

17:41And what I found is that so far that the feeling I get in response to doing that and the way that they go and carry out these tasks for me makes me want to trust it more. And I'm not sure where that breaking point will be, but it's going to be really interesting to find out. I definitely haven't had it like moving money around for me, but I have certainly had it buy flights and book hotels and do things like that that I previously would have been very skeptical of. And the reason I was willing to do that was just based on how rigorous, proactive and comprehensive I've noticed it being. And so another interesting example is that I saw it actually go through my email, review an invoice that I had and spot a discrepancy that I wouldn't have caught myself.

18:19And so this just comes back to where we are in the capabilities curve of the models, which I think is fair to say is sort of superhuman in certain areas, certainly relative to my attention to detail on administrative matters like that. And I think that's just like speaks to the amazing time we're in. And it's finally really starting to hit our personal lives, I think, in the way it has for many people who write software for a living, for example. And that feeling of singularity, I think, is nearer for more consumers than ever before as a result of these products. And I'm really excited to see where they go.

18:49And then I think coming back to Meta's efforts here, what do they have going for it, right? Right. Well, given meta owns WhatsApp, there's probably some very interesting things they can start to do in terms of like responsiveness, interactivity, probably even like creating unique ways to interact with with their assistant in WhatsApp that they may not make available to other parties. And then you have all these interesting implications downstream of like how the government's going to think about that. And will that be a fair playing field? And will we start to talk about antitrust and things like that that I don't know the answers to, but I'm very much excited to see play out.

19:21And then you also have the fact that these are very inference-intensive products, right? And they have to operate at very, very large scale because they're targeting all consumers. And as the three of us have seen clearly, when you start using them, you use them a lot more over time. And so what I think will be also very interesting is to see is like, one, I think instinct is going to have to, and these companies that are startups are going to have to come up with solutions to very new infrastructure problems around how they can do really efficient inference, how they can do really fast inference and very, very large scale and with great amounts of uptime.

19:51And I think then it's going to be a question of whether the obvious structural advantages that a company like Meta has in this area, not just from a distribution standpoint, but also from a compute and from an infrastructure availability standpoint, will those be deciding factors in the success of who wins this market?

20:08Josh Machiz: So when you're looking at these, it sounds like one, compute and infrastructure is a key factor here. I've seen a lot of discussion around multiplayer. Is that another? I'm curious what variables or spaces you think there are ways that these agents and agentic assistants can improve. So I think both in our working and personal lives, most of the things we do, we're doing it in concert with other people. We're emailing other people. We're coordinating tasks with other people. We're relying on other people to do something before we can take the next step. And as a result, I do think multiplayer is going to be critical.

20:40And it's clear that these products were largely designed for single player to start. They can email on your behalf and communicate very effectively and rapidly with other parties. But it will be very interesting to see once, you know, for example, my assistant is talking to your assistant, like how they can kind of create synergies and efficiencies there that I think is going to be interesting. But, yeah, it's going to be really exciting to see what the, like, vectors of innovation are in this space. And right now, I think multiplayer is definitely one of them. And I think speed and accuracy are others.

21:10And I think it was just yesterday that there was sort of a release of like a personal assistant eval, which is to say I think we're also going to start seeing a more like methodical way of tracking like who's actually good at what tasks. So, you know, there might be some assistant that's particularly good at travel. There might be some assistant that's particularly good at restaurant bookings or replying to emails. And I think just like we've seen in like frontier models as it relates to coding and reasoning and, you know, other sort of dimensions of capability, we might see sort of a fragmentation of which assistants are good at which things.

21:42But I do think it will be also very interesting to see on the basis of that which assistant gets the most adoption.

21:47Josh Machiz: I'm so curious, like, is the move to go vertical and really focus on the travel use case or do you want to be super horizontal like a meta? I think it's really interesting thinking about even just seeing some of the tasks that they had these agents run on. But yeah, definitely, I think a lot happening in the ecosystem. Zooming out, even thinking about a lot of this has happened for the consumer. Obviously, enterprise is a major market where this sort of proactive agent hasn't been unlocked yet. But do you see that being a different product? Do you see them sort of as a consumer? Ideally, at least for me, I would like something that traverses my work and personal context.

22:26Josh Machiz: How do you see that playing out? My instinct is that this will be a projection of how each of us handles that today, which is to say that we live in this sort of seamless transition between our working and personal lives, right? You may be cooking dinner and responding to a work email. You may be in a meeting thinking about some tasks that you have to do at home for your child or for a loved one. And so it's hard for me to see a world in which there will be a perfect bifurcation between working and personal assistance because that's just not how we live. Part of that is like a risk appetite question.

23:00And I think another part of that is a privacy concern of once I connect something that knows everything about my personal life, how do I know what's flowing into sort of the space at which it understands my work? And so I think there's some, again, some very interesting sort of like permissioning and segmentation problems that you can probably think about solving. And whether there will be like an enterprise focused personal assistant and a personal focused assistant product, I think is like a really interesting question that I don't know the answer to yet. But all I can go off of is like what I've been doing so far and I've been keeping it bounded to my my personal life and personal set of to do's and tasks.

23:36And at the same time, I think it's hard to imagine that doesn't change given just like that's how we live. Right. We live in a more seamless way between our work and professional lives.

23:44Josh Machiz: Yeah. I guess the last question is I've been curious. I've been seeing a lot of discourse around the I think Resi had blocked instinct. I was going to say, there's a lot of people who don't have Resi accounts today because they maybe abused their use of instinct. Yeah. But I think even broadly, this question around like our internet is, yes, you have like Resi that has like these anti-bot postures. But, you know, when there are so many ways our internet is structured where it's like Ticketmaster or Resi. There's so many things that are so clunky to operate on the internet from buying tickets to booking flights.

24:18Yeah.

24:19Josh Machiz: But even the idea that like, so the way that we buy certain things on the internet, there's like scarce inventory and this idea now that you have swarms of agents that you can throw at it. I'm just curious, like what, how that changes. Yeah, I think there's been an ongoing conversation since we started to realize that these models would allow us to do interesting things in the browser and on the web around how you sort of refactor the web to be agent friendly. right? And most of the work that has been done in this area has been around bots for like ad related purposes, meaning go to a website and make sure that it understands how to serve ads, right?

24:55What I think we're going to start to see is that a lot of the work in this area is going to go towards like actually enabling agents that do tasks for us to operate effectively on these websites for one. And for two, I think, you know, some of the properties you mentioned, whether it's resi or ticket master, these properties will be disrupted by alternatives that have refactored the experience to be more agent friendly if they don't do it themselves. And so I think there will be a rush from the incumbent web to refactor itself because of how fast this is happening. And I also think that we're going to see the rise of new platforms that maybe take like similar inventory from the real world and bring it online in a way that makes it much, much more seamless for agents to interact with.

25:36Very interesting take. I feel like there are now a whole generation of new agentic companies that could be born to take some of these seats in the traditional web. So Bloomberg reported that ByteDance is developing its own world model. And for context, ByteDance, as we know, owns TikTok. Bucky, you've got a lot of knowledge on world models. What's your take on this? Like, is this a big moment for the world models industry that ByteDance is taking this so seriously? So I think it certainly is interesting in that it suggests that they regard this modality of model to be important enough to allocate tremendous amounts of capital towards and give away for free, right?

26:18That is clearly an interesting takeaway from this. And I say that because I think currently the conversation around these real-time video models and world models has been very interesting, very captivating when you see what's possible with them. But the set of use cases that they'll ultimately serve, I think, has been a little bit more nebulous than, say, LLMs, where it's very obvious what they do, right? So to see that the Chinese ecosystem is starting to kind of repeat history again by saying, hey, we actually want to propagate the ecosystem with these models in a way that anyone can use them and basically give frontier capabilities away for free would suggest that they think it's important to do that, right?

26:55And they think it's important to do that because in an effort to compete with the West, they think there's going to be a lot of economic value created with these models. And they want to make sure that that economic value is diffused to as many people as possible rather than just Western companies. That's at least my takeaway. It's also obvious that ByteDance has so much relevant data given the consumer properties that they own that you mentioned that it sort of gives them a leg up to do great things here in the same way that Google had with its Genie effort. So I find that to be another interesting takeaway.

27:22way. And then the third thing I'd say is, I think we're starting to near this point where, you know, some of the performance characteristics that are being touted about this model suggests that we're entering this point where these models are now able to operate at a level of latency and fidelity that makes them genuinely useful, right? So a model that can generate continuous video frames is useful if those continuous video frames look realistic and are happening at a frames per second rate and a latency rate that makes the content actually enjoyable by the consumer or relevant, say, in a marketing context or robotic simulation context or any one of the other places where it's clear that these models are going to start to be applied.

28:03And I think we're getting to that point now where these models are becoming so advanced that people are actually going to start to use them in these domains.

28:09Josh Machiz: How do you see this manifesting on, you know, Doion or TikTok? Is it the idea that we have like these personalized, endless world models that we can explore, I guess, to make it tangible for an everyday consumer, how would this world model be useful for their surface area products? Yeah. So I could imagine a world in which we go from giving people the ability to use AI to create short video clips and maybe edit existing videos to giving consumers the ability to actually create endless streams of interactive content. And so I could imagine creators beginning to start to deliver worlds and identify their brand or on, hey, this is a world that you can step into and interact with, and you can really do anything with it.

28:53And you can experience me as the creator of this world in any way that you like, in a way that's personalized to you. And I think that's going to be a very profound shift in the way content is generated and consumed that I think we're probably not fully appreciating yet. And I say that because it's just, it's so different than what we've been used to that I think it's hard to imagine the implications. We have things like Roblox that are these sort of continuous, you know, interactive worlds, but to give anyone the ability to sort of create their own Roblox, so to speak, or, you know, create their own, you know, video interactive world is a very new capability and one that I think is going to be really exciting.

29:28I also think that there's a world in which the content that we interact with from like an advertising and marketing perspective starts to get hyper, hyper personalized, right? Like why, for example, should I create five different versions of the same video for the five countries that I want to interact with when I can render video on the fly based on the firmographics of the person in the country, in the city, in the area of that city that they're in. And I think this is going to have a lot of economic impact in a good way where we're going to be able to deliver, basically reach people in a more effective and personalized way and give them the products and services that they need in a higher leverage way.

30:03And I think that's both good for the consumer and good for the companies behind those products and services. Yeah, it's like marketing tailored to the stream that you're watching. It feels very like Fifth Element or one of those things where that, Sure. Where this screen is giving you what you want, not just what's being served to everybody. But at the same time, I think these videos and these interactive worlds are still very like inference hungry. They cost a lot of money to generate. And so I think there's going to be a really big opportunity in the infrastructure layer coming back to Reactor and some of the other companies that are doing great work in that space around how you actually, now that we have models that can do things that are useful in this area, how you actually ensure that you can do things that are useful at a cost and economic profile that makes sense.

30:46Josh Machiz: And for those less familiar with Reactor, what role does it play? Say, I'm not by dance, but I still want to run a world model like Infinite Content Loop. Yeah. So today in the real-time video space, you sort of have a similar market structure forming as what we've seen in LLMs, where you have frontier models that are delivered via API. Google Genie would be sort of the crowning example of this. And then you have an increasing increasingly clear open source ecosystem that's emerging. And obviously, we've been talking about what ByteDance is doing here. And I think that's sort of the bellwether of that effort.

31:18A lot of those models similar to the LLM world are also coming from China. And the moment you have open models, the question of like, how do you actually serve them becomes topical. And today, what we're seeing is that a similar set of inference platforms are emerging, as did in the world of LLMs for these models. And Reactors are very focused on taking these open real-time video models and world models and making them performant and making them efficient to serve and then exposing that experience via API that any developer can consume. One thing you should play with, I went to the live.reactor.com site and there are actually all these people like prompting the live model.

31:55And it's really fun to see what people are coming up with. Like it's almost like the audience is both the viewer and the director of the content and its evolution. And it's really neat to see. I do think that, you know, particularly for us as creators, we've been making video and just being so focused on what the audience wants from you as a creator. It's interesting to think about this whole new modality of like being able to world build and bring people into this new conversation. I mean, I could see people wanting to go to Claire's world. And it's important to note that these models aren't perfect yet.

32:31So, for example, you might be interacting with a world and you see a mountain with snow on the top and then you veer away from it and try to come back to that mountain and it may not have snow on it. This is for deeply technical reasons around the fact that the amount of memory that is available for these models to remember what they've generated and our ability to use that memory efficiently is limited today. So there's sort of an open research problem around how do we actually more efficiently use the KV cache in these models so that that problem that I just cited in the example of the mountain with snow is addressed.

33:03At the same time, you're getting a lot of people who are critical of this sort of real-time video model that isn't necessarily grounded in some kind of like real-world physics the way video game worlds are because of issues like this. And it reminds me a little bit of in the early days of, say, LLMs, where they would do some fairly foolish, misguided things, and we would be critical of those. And I think those that took a fixed view of the models being incapable of moving past those limitations, I think were proven wrong in a pretty flagrant way. The question I ask myself is like, will the same kind of capabilities improvement and scaling occur in real-time video?

33:39And if the bitter lesson holds here, then I think it will. And I think that those critics will probably be wrong. And so I think having a mindset of just remembering that these products will continue to improve and that these limitations that you see today may not be limitations to tomorrow is something I've been thinking a lot about and trying to remain optimistic about because history has shown us that taking the more optimistic side of that question has been the right place to be.

34:03Josh Machiz: You know, if we follow this path, is the idea that game studios, media companies, startups with Reactor, they can basically have the infrastructure sort of a layer for anyone who isn't ByteDance to sort of run all these world models. Is that the vision? That is absolutely the vision. And I think there are sort of two worlds you can envision here. One is where Google gets to generate all this content or whoever becomes the leading sort of closed source model. And then the other, and again, this is sort of what we're starting to see in LLMs right now, is that there are a lot of companies that want to own their own intelligence.

34:35And the way they're going about owning that intelligence is they're reaching for one of these open weight models. They're choosing a platform upon which they can serve that model. And they're building their own AI applications that they own end to end that way. And I think the question that we're in the process of answering in video is whether that plays out as well. And Reactor, I think, takes a view on that that's more similar to the latter than the former.

34:54Josh Machiz: It's just really interesting to think about. Like, you know, we interact with these world models or video models, and we sort of take for granted some of the infrastructural challenges. And even just like frame-by-frame generation or the latency that you mentioned are having millions of concurrent sessions. There are all these little technical challenges. And I think from the outside for an everyday mainstream consumer, we often underrate how important an infrastructure like Reactor is to being served a final output at the end. So we covered a lot this week. Certainly we saw Astra and how we're now allowing agents to kind of run somewhat unsupervised.

35:34And then on the consumer front, the same thing with instinct and town and muse and all of these ways we're allowing agents to go and act on our behalf. And then finally, with world models, this whole new step change is coming to light with potentially like new ways for people to interact.

35:52Josh Machiz: Yeah, it's just a really exciting time to be a builder, a consumer even. I mean, I personally have seen it with instinct. But thank you, Bucky, for joining us. It's just been so helpful to have you break down a lot of these changes and capabilities, James, and also share a little bit more about the spaces that you're spending time on. Funny to be here with you both. Thanks. Thanks.

36:24Thank you.