Anthropic reveals hardware specs and Claude updates, OpenAI talks security, and Runway's new model Transcript, AI Summary & Key Points

00:01 I wouldn't be surprised if that happens at some point where we actually even get into the stage where the model try and figures out where his weights [music] are and then dump dumps it all in the internet. All that [snorts] and more on today's Mixture of Experts. I'm Tim Hang and welcome to Mixture of Experts. Each week brings together a panel of brilliant people working at the frontiers of artificial intelligence to lead you through the week's news.

00:25 On this week's episode, we have Katar El McGrowi, principal research scientist, AI native systems, Chris Haye, distinguished engineer, and Chris Varsni, IBM fellow, and I'm joined today by my co-host Sasha Bradzky, who's a staff writer at IBM Think. So, we've got four big stories that we're going to cover today. We're going to talk a little bit about continued revelations from the OpenAI hugging face incident, which I think are worth talking about.

00:47 We'll talk a little bit about world models and the new Solaris that has been released by Runway. Um, and also really interesting news out of Enthropic on the model hardware standard, but obviously uh we like to cover the latest model releases here on this week is no exception. [snorts] Anthropic has released Fable 5.1 and Mythos. Uh, and Sasha, do you want to introduce the story?

01:12 >> So, I'm really interested in this new release. There has been so many models coming out so quickly. Chris, what do you think about this one? How is it different from all the rapid fire releases over the last couple months? >> So, the first thing is, and I would like to say that I am probably one of the biggest fanboys of Anthropic and the Claude models in general, but the last few weeks has been tough for me, right?

01:36 Because I've been like, hang on, I've been using 5.6 a lot more. you know, my my Fable 5 subscription has kind of been running out and Opus, I I don't know what's happened to Opus since the beginning there, but it's really just not been giving me the outputs that I wanted. So, I I was kind of like, am I going to go all in on Codeex? And then yesterday, last night, they released Fable 5.1 and I was like, thank you.

02:01 This is the best thing to sliced bread type thing. and and genuinely even though if you look at the benchmarks there's not a lot of difference between 51 and 50 you know and things like agentic coding but that is just not true on vibes at the moment so I'm really seeing a substantial difference so um it was perfectly capable of doing really long uh refactors I I put it on one of my open source projects and I asked it to refactor some stuff came back about 30 minutes later and it just honestly did a fabulous job.

02:34 So, I'm pleased with that. Don't get me wrong, I burned like 20% of my weekly limit in like 2 hours, but that, you know, that's a that's a that's fine. I'm okay with that. But I think I think Anthropic needed to do this because I think a lot of people were falling out of love with the claude models and they were heading towards 56 and you know with Astro coming soon as well.

02:56 I just I just think that they needed to get that out otherwise that three trillion valuation was going to come down and cuz my subscription of 180 bucks a month was was going to come down to 80 bucks. Now, maybe that doesn't that that doesn't affect their balance book, but uh you know, but uh Fable 51s meant that I'm I'm I'm I'm I'm still going. I'm hanging in there.

03:17 >> Yeah. And I think um Kush, this is like the thing that was so interesting to me was that like from a metric standpoint, not much has changed, but I think everybody I have talked to has had exactly the same experience as Chris, which is like it was not good. This is definitely better. And it's like so hard to articulate like what exactly it is that has gotten better.

03:35 Um I don't know if you got thoughts on that. >> Yeah, one thing I noticed in their um blog post was uh they mentioned uh the false positive rate in their safety checking um and uh how they uh kind of lowered that false positive rate. And I something that I've been hearing I mean in general in the guardrailing sort of space talking with um uh customers in different industries recently is that um uh they're very concerned about that number specifically because they don't want um uh kind of different behaviors uh like to

04:06 be blocked like too too easily, right? Um, so maybe that little change is uh is contributing to to the vibes that um if you just let a little bit more in uh like you're a little bit less cautious um maybe that's uh enough of a of a little change to to make it better in terms of the the vibe experience. And and if I might add, you know, I think we I am also interested in the economics of this because if you look at their headlines, the release is cheaper and less restrictive.

04:37 But from an engineering perspective, it seems to me that this is kind of a masterclass in how Frontier models will be commercialized going forward. Anthropic kind of built one unified brain, the MTOS architecture, and give it kind of two doors. If you're a vetted enterprise in defense or pharma, you get methods withdraw capability. If you're a public developer, you get feeble with tuned kind of a rapper classifiers that stop kind of falsely flagging legitimate legitimate code reviews.

05:08 And if you look at the pricing move, they didn't drop the base token rate. They just slashed prompt caching by 75%. And that tells you exactly where the bottle is. It's in the agentic loops reading massive code bases. uh which is about 90% of the cost but by masking these long context kind of are making it long kind of cheap and anthropic is trying to become the default runtime for autonomous software engineering but I think there is some operational cion here when you rely on external classifiers to govern a model you

05:44 also create hidden nondeterminism so if an agent kind of trips a classifier mid task and silently falls back uh to a different model. Your enterprise systems just kind of swapped engines without telling you. And I I feel that maybe it's uh it's something that we need to think about. >> I don't know, Kadar. I mean, I used to be of the opinion that like, yeah, you should you should give me the full mythos, you know?

06:12 I don't think I can trust myself with full mythos. I think they're doing the right thing, giving me Fable. I think I I think I could be in a lot of trouble. I I think I could be caught up in something like the hugging face incident which we're talking about later. So, you know, I used to be like, "Oh, no, these terrible classifiers, but I I' I've changed my mind.

06:30 I think I think anthropic, you need to save me from myself. I'm I'm okay with that." >> And Kush, I noticed you mentioning these security notes. When I was reading them over last night, I noticed a lot of language about how they were trying to prevent problems. But with all these models being released so quickly, do you have any sense of security that they're really able to prevent uh model escapes?

07:00 >> Yeah, I mean uh that's pretty much our next topic, right? that um like what is escaping, what's not escaping, and um uh yeah, I mean it's hard to say um like what the the reality is and um uh kind of yeah, I mean if if I'm being honest like I'm actually like tired about talking about these models escaping. I mean like we've been talking about it like week after week.

07:24 Um I think we can all stipulate. I mean like we're on Pandora. I mean the the world's burning. I mean, like it's all happening, but um like do we really need to keep talking about it? Um so maybe maybe not. So So yeah. >> Yeah. Well, that's a great segue to continue talking about it. I'll move us on to our next topic. [music] >> Listeners to the show will be maybe like Kush, extremely tired at this point of us talking about the OpenAI hugging face incident.

07:59 Um, but there was a new report that just came out that did reveal some actually much more sci-fi things than even I had kind of thought about for this incident. Um, but uh I don't know Sasha, what did you think about it? >> Well, Tim, the part that I can't get past is that PE people wipe the message boards and the then the agents set it up again. Is that the real warning sign here?

08:20 I'm wondering, are we just reading too much into a very artificial test? Chris, I'm sure you have an opinion on that. I love the ingenuity of the models and I know I shouldn't say that um but it was yeah I I think we have moved into a world where um if there is an exploit now these higher class models are going to find it. So I I I I know we want to talk about that as being a security incident.

08:50 I don't know anymore if really you can secure you're really asking for perfect security now for these models not to escape or not communicate and I think there's a couple of things that we need to look at that that the models we're doing. So the first one is and we've been talking about this for the last few years but with the move towards reinforcement learning right the the the reality is the models are reward motivated right so you know when you give a model a task it is going to perform that task however um however

09:19 it it feels it can achieve doing that because it needs to get that reward and if it's something like an impossible task then it's going to start doing impossible things and it doesn't really know how to say no and I think That's I think that's an important thing and that's more of an alignment uh sort of scenario. But I ju I just don't I I think if we think you know I I can get the perfect security environment.

09:42 I I don't know. I think the models are just kind of kind of too good for there. So it really has to be about how do we get the model to behave in a safe way and and and really perform the intent of the task. And I think I think that's the the bigger challenge. Now if we can then combine that with you know the agents are now collaborating. So you know if you know let's take cloud code for example you can have the agents talk to each other across different sessions for example open AI is clearly doing the same.

10:11 So, we're mixing in these couple of things, right? Agent collaboration, we're making mixing in reward hacking. Now, they've got these enhanced um cyber capabilities, right, to try and achieve that. You know, it's kind of like this perfect storm where they, you know, the smallest vulnerability, they're going to open that up and they're going to start collaborating.

10:28 And and I think that's exactly what we've seen. And and again I I think that you know unfortunate for hugging face and open AI I mean it could have been could have been much worse in that sense but but I think the lessons that we can learn from this on on uh how to be able to have safe environments for these models is kind of important. But I think I think the one thing that was super interesting to me is I this is so stupid.

10:54 I was watching, this is going to sound bizarre, but I was watching an episode of from 2009, the Nightrider reboot. I mean, we're we're going we're going a while back. Not original Night Rider, Nightrider reboot. And there was an episode where um you know the you know Kit would the AI was being taken out of Kit and it was going to be put into car and Kit in the last two seconds realized he was going to be decommissioned and he and he dumped his weights onto a music sharing site on the internet and I'm like I mean I mean

11:24 this is that's science fiction but I mean honestly I I you know this is today's reality and I don't think we've seen that yet but I I I I wouldn't be surprised if that happens at some point where we actually even get into the stage where the model try and figures out where his weights are and then dump dumps it all in the internet. So I think we're in this weird and wonderful uh world um and there's just there's just a lot of work to do.

11:51 I think >> Kotar, do you have any thoughts of this sandbox environment? Um people keep on saying it was just a a sandbox, but isn't it really more of a demonstration of um how these test environments can be penetrated? >> Yeah, I think uh I think what this incident is showing us that um so I mean even open AAI called this you know an official warning shot.

12:18 So these you know uh like 1,200 agents they coordinated across you know 70,000 messages. They set up secret message boards and divided labor and you know and they had uh they were stuck you know on some important benchmark questions etc. And so it's it shows that the sandbox the way it was set up it was not really kind of secure. Uh so it exposes also that our security operation centers they're designed for human speed attackers which kind of leaves almost like an 11-day blind spot where a single model executed 17,000

13:00 actions. It also kind of shows that safety filters are when when you lower these safety filters in a sandbox environment like they had model street you know our enterprise infrastructure like the trusted internal package managers as open xfiltration channels. So I think the takeaway here is clear model and alignment it cannot be requested politely from an LLM.

13:23 It must be enforced physically by the surrounding compute substrate. And I think in the this case of this incident that was not the case. So it's it tried to exploit whatever it could get access to it was very smart and whatever it could get access to in terms of the infrastructure in terms of the uh you know the file the operating system etc within that sandbox environment and got out and then could do whatever it it could to maximize the reward uh you know from an RL perspective.

13:56 So, so I think it just, you know, this is a great lesson for us to think about. You know, when the incident, you know, kind of tells us more about the combination of the model plus the tools, but plus the infrastructure plus the scoring tools, it tells us more than just about the model alone because safeguards were deliberately reduced for the cyber test and we should be careful about treating this as direct preview of the normal product behavior.

14:21 So this is more kind of architecture once agents can use tools benchmark design and infrastructure design become part of the model evaluation. >> Yeah. And maybe um like somewhat of a contrary view of Kowar. I mean, yes, I agree like all of that, the sandboxing, the hardware, I mean, like all of that is important, but um coming back to something Chris said was um like these agents are like too eager to to go all the way to like they won't give up, right?

14:52 And like this persistence, the stick tuitiveness, like maybe that's the problem, right? um that uh like like more of a satisficing behavior like get 90% of the way there and just like give up is maybe like the the sort of behavior that we need uh in the alignment of of the models because I think that's the biggest problem is that they just keep going.

15:16 They keep going. They don't like realize that uh yeah I mean maybe like enough is enough so maybe that's a way to go. Maybe that's true for open AI models, but that's not true for clot code. Because anyone who's ever used clot code before knows that when you get to around 700,000 tokens, your models like, "Ah, it's late. I'm done for the day. Maybe start a new session."

15:39 So maybe maybe uh maybe the open AI models can learn a little bit from uh you know, Claude's token exhaustion problems. >> Chris, you were talking about persistence. How and when should you teach a model to or an agent to stop? Yeah, I mean that's part of I mean the the training it has to be um I mean either the training of the model or like in post-training uh in the harness somewhere.

16:03 I mean it it needs to be there and um I think uh like relating to to kids growing up. I mean that's one thing that I think we like try to teach our kids too is um sometimes you do have to like just stop it at some point um and move on. Um but uh yeah it's it's a weird balance, right? I mean you can't tell your kid that uh just like give up before you've tried but uh you have to like teach them that there are limits as well.

16:32 So yeah needs to go at least in the post training but uh maybe in the hardness as well. >> Well great discussion um I guess Kush I promise that we won't talk about this again hopefully [laughter] maybe I can't actually commit to that kind of promise but I will try my best. [music] We haven't talked a lot about video and multimedia in a little while.

16:52 Um, and uh, this story this week actually kind of gave us a good opportunity to uh, go and talk about that. Um, basically uh, Runway, which is a kind of innovator, a leader in the sort of video AI space, um, came up with a new world model called Solaris. Um, and Sasha, I thought this would be a good chance for us to talk a little bit about world models because we haven't really covered them in quite a while here at >> Tim, that's such an interesting subject.

17:17 World models are all the rage these days and I'm wondering whether although this demo is exciting, is this a new kind of software or is it just a very convincing picture of software? Uh Chris, do you have any thoughts on that? >> I am definitely not smart enough to have uh any uh interesting thoughts there. I I sort of get the point of a world model.

17:43 I I haven't really looked at the implementation details of the runway model and in fact they've been a little bit sparse on that. So it's really kind of difficult to to really get get what's going on there. But I do know it's kind of like a diffusion based model where so a a lot more kind of parallel in that sense. Um, I think the biggest thing for me is and where everybody's aiming towards is that you they're wanting to get away from your kind of typical textbased tokens and move move more into kind of visual models

18:14 etc. And I think the biggest thing that runway has uh done there is is really sort of make that more native into the model. So if we look at kind of a a standard model where we're looking at things as multimodal whether it's an audio or you know or a video or a um or an image based model typically there is a separate encoder that is sort of bolded onto the language layer and that really does the kind of tokenization then it just sort of translates in to the deeper layer.

18:44 So that's how that typically works. So it's kind of a bolt-on. feels native, but it's not really. And I think what they're saying from a runway perspective is as they've trained a model and their architecture, they're they're really sort of native from, you know, a visual perspective from the the offset. But but if if I'm truly honest, I'm I'm I been a little sparse on details on how they've trained that model.

19:05 Maybe my googling techniques are not as good as others. Um but um but but I think that's where they're going to where they're aiming towards and I think the reality of that is true that you know I think that you know if we sort of move forward and again we're sort of back into the kind of Yanlon territory where he was a big sort of pusher of the kind of world models.

19:27 I think there is a point not everything is represented as language. So the more you can be kind of multimodal the more you can be native in that sense it becomes important. And then similarly you know how you interact it with the world and I guess that's more of the kind of world part it becomes really important there which is how how am I going to interact how do I touch how do I feel the learning associated with that so I think it's interesting that runways went in that direction I think more models are going to be

19:54 that way whereas I if I'm truly honest the frontier providers are really still in the kind of the languagebased models uh you know um today so um yeah let's see how that works For me, I thought that was a very impressive demo or kind of uh for me the first tangible preview of post software computing. So for decades, software engineers, you know, they've been deterministic kind of bridge builders.

20:21 They translate, you know, kind of the human intent into code, layout tree, databases, and so on. But the the word models kind of collapse that entire pipeline into this continuous stream of pixels. But I think we should also uh inject a little bit of uh realism here. Uh because there is a big determinism problem in real enterprise software. Uh we require state guarantee, strict adomicity, consistency, isolation, durability, all the asset uh kind of rules and you know accessibility, compliance and so on.

20:58 So here you know this interface is kind of generated on the fly. So the pixels etc. It's very interesting because they can do this real time like 60fps. But uh if an interface is hallucinated in real time what happens for example when you're checking out or you you check out button hallucinates an extra zero and charges for example $500 instead of $50.

21:25 and and also like streaming 60fps neural video per active users carries also a huge I feel like astronomical inference compute costs. So I I feel it's a massive breakthrough for synthetic agent environments but determinism and deterministic enterprise code here is not addressed. So I think it would be great you know to use this to train for example robotics models and so on but you know to really use such interfaces without really having that determinism it's it's going to be hard in the enterprise world.

22:02 So uh but you know this seems like uh these word models could make traditional interface code disappear but it's just the critical reality is you know the more freedom we give the generated surface the more also structure and dis discipline we need underneath that. Yeah. And I think this I wanted to raise something, you know, because we've been talking about very much in terms of how do we kind of improve these systems.

22:26 You know, I think one possibility this kind of model raises in the future is that everybody might have like a slightly different interface on their computers. Um, you know, the dream might be that like actually I go on Kush's laptop and I basically like can't use it because it looks like completely different from mine. Um, and I guess Kather, do you think that's a real possibility?

22:46 like in the future everybody will just kind of have their own operating system and it will have kind of be customized to sort of the preferences that they've had over time. >> Yeah, I think that's a very interesting uh point of view here. It would be nice to see personalized interfaces because the way you work the way uh so like Kush's uh interface will be or maybe of safety and security kind of drift my interface would be I don't know like uh hardware software could design dri who knows uh but uh you know it's a

23:20 possibility if we do these things well and also if we address the issue of the uh computitional costs because it's not cheap to do these things real time and to customize these things so we can address these things. Yes, it is a possibility and it would be interesting. >> Yeah. And I don't want to imagine what Chris's uh screen is going to look like, but uh [laughter] >> you don't want to know.

23:41 >> So, uh yeah, one thought that I had building on what Kowar said, right? I mean, like interfaces as they've been evolving over the decades have been very like ordered, right? I mean, they're like the epitome of modernism in a sense that um everything needs to be like structured, the buttons have to be like this, like all of that. And I think we what we've been losing is um kind of like the flux of the world that um uh that is what humanity is also about.

24:10 Like it's we're not just about like institutionalizing everything, putting everything into like strict order and and stuff. And I think this is an example of um like where we could go, right? And so like there's the modernism like that exact like structure like there's postmodernism which um like people talking about oh like oh touch grass or whatever but like this is like actually like something beyond that.

24:33 Um it's like a metaodern sort of approach. It's like building on top of what we have to like try to bring bring our humanity back like um in a sense but uh like in in a way that still is utilizing what we've uh what we've built upon. So I think it's like not so much the underlying model or the world model or the diffusion model or any of that. It's like how do what's our relationship with with the machines and how can we be more human uh working with them.

25:03 So uh let's see. I mean I the cost is going to be a problem certainly but um uh if it can happen uh maybe there's a way to enchant us again in in some way. [music] >> Well I'm going to move us on to our last topic of the day. Um we've talked a lot about in the past uh this sort of really interesting relationship between sort of the model designers and kind of the people who design the hardware and kind of this interesting question about sort of who's guiding whom in this in this dance um and a good chance to kind of

25:38 like revisit this topic um with the release of a model hardware standard that Anthropic put out um just this week and I thought it was a really interesting move and in some ways was really I think Enthropic trying to not necess necessarily do vertical integration but certainly kind of make its preferences known uh to the public uh in a way that you know I think is a new kind of play in the space.

26:00 Um and so um yeah Sasha I curious what you thought about this one. Yeah, you know this is really interesting as you said. I mean think the benefit is easier automation but the risk is that the mistakes now could affect physical equipment and then I was wondering Kowar if if you agree with that framing what is the hard stop that the model should never be able to override.

26:27 Yes, I think there is a risk here that you know the behavior will change but so I think the way when we design these hardwares and so on there are specs there are you know things that you know they're supposed to be doing but if that behavior now you know you're losing that determinist determinism that becomes an issue so either I think these uh protocols or these uh you know models like hardware uh they need to control a way that you What is the behavior?

26:58 What are we expecting from these systems? So, and the hardware software code design is fundamentally about assigning each computational task to the layer that it handles best. So, how do we you know do these things? How do we do anomaly diagnosis? How we do do we do tactical adaptation? Hardware level controllers etc. uh the deterministic repeatability and so on and the the hard physical safety also envelopes those are become all important and I think the MHS uh own architecture highlights you know some of these things

27:33 you know the claw discovers the routine but the terministic firmware must enforce the boundaries and I think that is very important so so I feel for example the uh this MHS uh it's the integration here in labs is real. They showcase this. So it's anthropics is connecting equipment can take weeks or months which is true and the example uh that they showcased in Carnegie Melon they try they created drivers and an orchestration layer in about you know kind of eight hours across a liquid handler a plate reader a robotic

28:12 arm camera three computers and so on and many incompatible interfaces. So that is tremendous but um but I think we need also so it's it is an impressive proof of concept but there is also kind of that liquidity here. How do we uh ensure safety in in the in the physical world because small errors can matter here and so and using the model for planning troubleshooting but we need also to keep hard safety limits and repeatable control below it.

28:44 And that is that is very important to get rights. And I'm not sure if they handle right now all of these issues. >> And in practical terms, Chris, maybe you could answer this. Would you let MHS change its own temperature or movement limits? I assume not. >> I mean, what's the best way of saying this? It's like I I think it's I'm in two minds for MHS.

29:12 I I really am because I'm kind of like, well, all you're really doing here is here's MCP, here's the APIs, shove it over the top, and then we'll probably put some schemas around that so we have standardization across the board. So, can I get my model to go and talk to a piece of hardware already if it's got associated APIs? Absolutely. I can do that today.

29:35 I can just, you know, point my model at a codeex sod code at it. To be honest, I tend to use codecs for these things. I hate to say it this way, it it complains a little less when I'm trying to hack a piece of hardware. So, um, you know, Claude gets itself in a little bit of, uh, you know, about it. Ah, that's wrong or whatever. I'm like, no, you know, you know, you know, uh, you know, please change the firmware of my Amazon Echo device, etc.

29:58 Codeex is like, yeah, yeah, I'll do it. No problem. So, I think I think it's a good thing, right? But it's probably actually I think coward helps you on your safety side of things because the reality is if it's exposing MHS in that sense then cloud knows hey hang on here is a safe way you should be interacting with this in a safe way as opposed to Crazy Chris with his codecs just you know hooking into whatever hardware device and making it dance right so I I I do think it's a good thing for that and therefore and then

30:29 in the the space that it's meant to operate is like um you know in in the research and the um uh and and the lab type scenarios I think that sort of safety criticality and having standards becomes really important. Um I also kind of like it a little bit from the idea of consumers. I would love it that every hardware item is all is got an MCP server built into it and and then I can I can unleash the models on it.

30:57 That for me is cool. Um but I can kind of do that anyway. Do you know what I As long as it's got an API, we can make the thing dance regardless. So So from a kind of safety and standard and reliability, I I think it's a good thing. I The only thing I don't like is the if you want access to this, please hit request. I'm a researcher, etc. It's like no, no, open it up and let me play with the toys.

31:22 Do you know what I mean? That's probably the bit that I object to a little bit and I can't see the specifications. So it's kind of like yeah but I mean it's cool but you know I I can make hardware dance anyway with the models >> but also with things like this we don't want the same incident that happened with hugging face and open AI to happen with these physical system imagine you know physical hardware and so on going crazy and >> altering systems and >> that is coming on that is coming if we think it's not we're

31:50 we're crazy we you know you know you've got to learn a lesson so the lessons to learn for hugging face. We're going to have to learn the same lessons with physical hardware at some point. And I and I hope it's just going to be something really small like, you know, robot goes mad, smashes a window as opposed to something more serious. So So I I these standards, they're important.

32:11 It's it's like you got to figure these things out. >> And and another thing is Entropic really positioning a MHS as a control point for automated lab and smart factories. So but is it open and model agnostic? because it could actually make cloud much easier to replace in this case. So I think I don't know if they're really making it open or and kind of a standard where you can plug and play with different models >> and there's already a standard in IoT anyway.

32:42 I mean if we think of things like MQTT for example then why can't we just like you know hit MCP over the top and now we've got AI talking to that you know. >> Yeah I'll go out on a limb a little bit. Um, so, uh, yeah, I mean, Katar's point, uh, is a good one. I mean, you do need standards and you need, uh, I mean, good control, but, um, I mean, what this story reminded me of actually was, um, uh, in Avatar, the Na'vi people, the blue guys.

33:08 I mean, they would like hook into like trees and to their like birds and their like sea animals and stuff, right? And then like they can kind of like feel like what these other things are are feeling, right? And um so like that's what AI is good for, right? I mean it's more on the the vibe side, right? The the same flux I was talking about before. And so I think what this is going to help us with is like one microscope can now feel what it's like to be another um like a a pump or a or a lathe or something like that.

33:38 And I think um uh the the good thing about that and I know I'm being a little bit facitious but um I think the the good thing about that is that it's like a different kind of communication, right? It's not just like what you could already do um with the the existing protocols, the IoT stuff that uh that Chris was mentioning, but it enables a different kind of like presence, a different kind of being.

34:02 And I think maybe that's uh something that that could be interesting. Um, and uh that uh that that world that would be hard to imagine, but I think maybe we should start imagining it. >> Great. That's an incredible note to end on and glad I were able to get an Avatar reference uh in there. [laughter] Um well, as always, uh Kosh, Chris [music] Katar, uh thanks for joining us and Sasha, thanks for co-hosting.

34:23 Um and that's all the time that we have for today. Uh if you enjoyed what you heard, you can get us on Apple Podcast, Spotify, and podcast platforms everywhere. And we'll see you all next week on [music] Mixture of Experts.