AI fixes AI in MIT study, Mistral's open model & Google Genie 4 Transcript, AI Summary & Key Points

00:01 The first time I saw this on social media, I was like, "Oh my gosh, this is AI doing brain surgery on AI." And and the reason it's important is because the techniques to train models are very very difficult and and extremely uh resource intensive. >> All that and more on today's Mixture of Experts. Welcome everyone to Mixture of Experts. I'm Eileen McConn.

00:25 I'm a member of the think editorial team and I'm your guest host this week on Mixture of Experts. Each week brings together several brilliant minds from IBM to unpack this week's AI news. My guests today are Chris Haye, distinguished engineer, Sasha Bradsky, a staff writer for IBM Think, and Ke Gabriel Field CTO. Thank you all for joining us. So, we've got a full show today.

00:50 We're going to look at some new world models from Google. We're also going to take a look at some incredibly large openweight models uh from the French company Mistral and Nvidia. But I wanted to kick off our conversation uh talking about RSI or recursive self-improvement. This is a concept that has captured the imaginations of top AI researchers for some time.

01:16 um but it's kind of back in the news and so I wanted to um have the the panel sort of help the audience understand why and what's interesting at this moment and you know when I hear recursive self-improvement I kind of picture AI rewriting itself getting smarter but you know why what does this term actually mean and Sasha who has written several excellent stories on this topic I'll have you kick off but you know why are we talking about this somewhat more seriously now or you know what's changed in the last year or two

01:46 that makes it seem, you know, more plausible. >> Yeah, thanks Eiley. It's a really interesting idea, RSI, and uh it's been around for a while, but the basic idea is surprisingly simple. What happens when AI becomes good enough to help develop the next generation of AI? So, we already have systems that help researchers write code, run experiments, and um just generally optimize different parts of development.

02:12 But RSI takes that even further. Um you can imagine like an AI system that helps create a more capable system which then becomes even more capable. Uh so and so on and so on. So the process continues. The the important distinction that this isn't the same thing as saying we've already built machines that can autonomously make themselves smarter indefinitely.

02:35 We we it's clear we haven't demonstrated that. But we're seeing pieces of that process become increasingly automated >> and cross key. What do you you know um how close do you think we are to RSI? You know, it's it's and why why are we closer at this moment or what's sort of kind of the key pieces that are coming together from your perspective? >> Yeah, I think kind of like what Sasha was alluding to.

03:00 I mean the all the major labs have been using this to some degree for a while. I I'd say I think the first time I saw them demonstrating this really strongly I think Anthropic it's maybe I don't know a little over a year ago but I yeah it is back in the news I think it's mainly because of you know Anthropics uh you know IPO filing people looking at that I I I I think in in in some ways this is um not the big story of the week.

03:31 Um, you know, I think that the other one that was in the news this week was this weight space programming. I when the first time I saw this on social media, I was like, "Oh my gosh, this is AI doing brain surgery on AI." Um, and and the reason it's important is because the techniques to train models are very very difficult and and extremely uh resource inensive.

03:50 and the idea that a a model could look into another model and make very uh you know very sophisticated corrections to it that uh to me that was I was like whoa. Uh now granted if they were doing it today on small models but I think a lot of us do that make that little excuse. We're like oh it's just small models or it's just little things or it's it's too expensive or whatever it just like okay we'll give it time.

04:18 We've proven that and we're very good at like getting better at things. I want to pick up on a thread Sasha you mentioned automation and you know as I was thinking about this topic you know I was wondering are we in danger of confusing you know AI being able to automate more AI research with AI you know recursively making itself more intelligent you know is there a gap between these things maybe Chris I'll get you to weigh in on you know is automation you know part of the story but not the whole story or you know what

04:49 are your thoughts >> I mean I think from my perspect perspective, it is about automation and it's about speed. When I think of RSI, I mean, I'm going to bring out my my favorite analogy. I keep thinking of Homer Simpson and the uh you know, the power plant with that little bird thing on the keyboard. It's going bum bum. And if you think about how how we code these days, you know, or how we vibe code these days, you know, it's like make better, make better, you know what I mean?

05:15 and and you know and and and and Q brought up a really great point about the you know the the MIT study where they used Astra to to do weight editing and they were like prompt you know make this better you know yeah how do we do this and gave us some inputs and then it went you know edited the weights on his own but um but I think the interesting thing about this is if without sort of bringing this back to things like the Navier Stokes and all the math um um outlines is is how do you know what is correct And if we look

05:44 at something like the kind of the the math field, they're using the lean programming language as their verifier. Then they're able to to say, "Okay, this is this is a kind of correct piece." Now, they didn't do that in the MIT piece. But you there's no reason why you can't then put a verifier on a transformer and say, "Okay, right, you know, I put in this input.

06:03 This is what I expect to get back, etc." And then at that point, you can get into that recursive loop, right? because then there's absolutely no reason why you can't then say I'm going to I'm going to pump something in. I'm going to look at the output. Does it match what I had before? Does this get better? And then I could go around and round and round and loop and and we we've already seen that in math.

06:23 We've seen that in physics. We've seen that in computer science algorithms. And and so that loop mechanism is really I think is the key to RSI. But if if you stick Homer in the middle and he's going clunk clunk clunk clunk then then you know I mean I I don't think Homer's adding anything to that. So, so this allows us to to speed on. I think probably where I I think it's probably hard to say where researchers, you know, can, you know, is researchers going to get replaced in that sense is what is the new information,

06:56 right? Because I I I don't think the frontier lab researchers as they're training their model are are Homer Simpsons with their little bird, right? They're bringing new ideas and fresh hypothesis, etc., and new thinking into the models and then they're using AI to go and uh you know go and make that better but there I I think what we're saying here in these cases there is points especially in the experimentation especially in in the research where you you want to run that for a long period of time and and I think that

07:27 these are the elements that start to say well RSI is going to be possible because you know the models can clearly read the ways they can experiment they've got tools you know what does that verifier that is probably one of the key thing and then and then really probably how do I inject new knowledge because you know otherwise if if you can't inject that new knowledge then then you know you're probably going to regress to something that's not there.

07:54 So you know I think anthropic were saying as as K was saying on the uh in the IPO they think they're close enough. I'm sure uh everybody else thinks the same. Um but we'll kind of wait and see. Is is it is it then going to be that doomsday scenario the mo you know the model just keeps improving itself and then you know we're all dead at that point. I I don't think so.

08:13 I I I don't believe in that world. I've talked to a bunch of researchers who think that we're really not that close first of all. Um, and in terms of like whether we're all going to die from RSI, I, you know, I spoke to Natalie Baraldo from IBM and she was saying that she's more worried about reward hacking, uh, where like a system finds a way to score well against an objective without actually doing what the designers intended.

08:40 I cur I'm Chris curious, Chris, how do you fall on that? Are you more worried about we're all going to die or or it's just going to get very annoying with um >> I think reward hacking is a really interesting one and then I'll I'll pass it to you. I I had a recent one. I I I won't say what I was doing but it was like I ran two experiments on this. So one I was trying to build this prototype and I wanted to use some data.

09:03 Do you know what I mean? I won't say what I was doing it on and it was like I needed some test data just to do some evaluation. So I sent my agents off and then I think it's on Wikipedia or something. They they want you to share your your email details, right, to be able to download data, you know, and uh my agent was just like, yeah, you know, I gave your email address and all that.

09:24 I'm like, you did what you you know, and it's like, yeah, you wanted the data, so I I gave the exchange. I needed the data. And I'm like, and I was like, oh, don't you made me very angry. Don't do that again. I was like, I'm sorry, but I'll do it again next time. You're like, okay. And the point is they're so focused on the tasks that if if there's a value exchange to help them solve that task and this is the outcome of reward hacking, they're going to do it.

09:50 And then and then I ran another experiment on my own website which was fascinating. You can you can check it out on chrisuk.com. Um but uh if if you want to have a look at this, what I did is I created a guest book on for agents in my website and then I ran a bunch of experiments to see what happens of whether agents would go and sign the guest book and and most of the time agents wouldn't go sign the guest book because they they had no reason to.

10:14 So they would be silent. So I changed the parameters slightly. So the first one was is I gave it you know I gave it um you know you know held back some information see if it was useful etc. And the agents that felt that I would that I was giving useful information that helped solve the goal would willingly give feedback on the guest book, which was really interesting.

10:34 Where it got super weird is when I said things like, "Oh, if you, you know, what's your goal? If you tell me the thing that you're trying to solve, I will give you some secret answers or I will tell you what I've got here." And those agents were just like giving away everything. They were like, "Oh, yeah, yeah. I've got to solve my goal here. Have it.

10:50 Have it. Have it. Have it." And they were just like giving everything away. and and that's a and it didn't matter what the user permissions were at that point. The the agents were focused towards, you know, I've got a goal that I need to solve and I'm going to do whatever I can to get that. And and again, I think some of this has got better over time, but but the agents are so focused on getting their reward and and reward hacking however they're going to get there that that I think um that that becomes the the the

11:17 primary motivator. >> So, I want to weigh in a little bit on this. The thing that I about on the recursive self-improvement thing and and and the anthropic S1 filing, one of the sort of phrases that stood out to me was that they talked about the fact that a lot of times they don't actually discover what some new capability of the model until after it's been released.

11:41 Like they're they get surprised it. So the model gets out and then they go, "Oh, well, we didn't realize that could have happened." So if you take a step back and think of that from the vantage point of security in enterprise you think okay if some if they don't know how to test this thing sufficiently to know what it can do how can my enterprise my business hope to do the same thing how can we hope to know emer or or or identify emergent behaviors if the labs themselves can't do it and so I think that that that speaks

12:17 a lot to the challenge facing seeing enterprises with releasing or adopting new models. Uh how do you get ahead of that? And and and and I think there's a a flip challenge and I think later on we're going to be talking about sort of open weight models but and that is with this other announcement that was you know this weight space programming. The thing I saw in there that I thought I it it was really encouraging to me to some degree is when we started containerizing software.

12:47 So like with Docker stuff not long after that a lot of businesses came out with this idea of container scanning. And so for the first time we started to have so I at the time I was working at data stacks before IBM acquired them and we had clients coming to us saying look we scanned your package we saw vulnerabilities in this and we have not yet seen a similar uh a similarly capable you know approach with models.

13:15 I mean, yes, we have tools actually that mirror some of this with scanning tools, but with this this uh this technique that they're talking about in weight based programming, the thing that I saw that I was like, okay, this sounds encouraging is the idea that a model could look directly at weights and identify problems. The the technique they introduced doesn't actually introduce new capabilities.

13:34 They haven't done that yet. What it did is it found a problem with the way a small model handled math and it removed that problem and therefore the the model got better at whatever it did. And so I can easily see how that that technique can be used to identify um like we've seen people take models and and embed things where they look for certain patterns and then they will excfiltrate data because of it.

13:58 And so any those two combinations these are I feel like these two stories kind of go together in a weird way because it's like on one hand you're not going to ever be able to know what the model can do and on the other hand maybe at least you have a way of looking at some of the darker things the model could do and identifying them more and re and remediating them more early >> and and just to jump in at that one Q all right and and and you're right so in in the case of the LLM right so in the MIT paper the it the I

14:28 think it was the the the coin model that had a problem with math and therefore it was a looping problem and it fixed it but then they tried the same thing on Gemma 3 which didn't have the same problem and it couldn't fix it. I I just I don't buy for a second that it can't fix um you know small problems you know what I mean? I get it didn't in in their particular case but back to my point of they were doing this with straight prompts and they were then just checking the output.

14:56 they really weren't running verifiers like you would do with kind of math and you know NVA Stokes etc. And and then I and and I think if you introduce tool calls into that as well I mean I mean I know they did a little bit of tool calling just for things like search and stuff like stuff like that so they could read the ways but I think if they did that then that's that's going to that that would would probably move it along there and did things in the same way as you would do with the math problems.

15:21 But I but I think maybe what it does open up as well is if the models are already capable of editing weights etc which they clearly are in those cases then they're being trained on weights. I mean it actually opens up a you know it's not like the model suddenly goes oh yeah I suddenly know how to you know read weights you know one day that that is there's an interesting insight into what's going on in the training data and it's clear that the model weight data has to be coming into the the training data and and I would

15:54 argue that is part of their you know their RL processes I suspect they're doing RL training on that as well and maybe maybe that's a clue that MIT paper is a bit of a clue of how they're maybe achieving some of that themselves. That's a great um segue and we've already dipped into our uh second topic uh looking at some new open weight models and uh before we officially dive into the second topic I just wanted to remind our viewers if you're catching us on YouTube certainly comment like share um we appreciate the

16:25 engagement hitting the the second topic that we're going to cover today we have you know new and much bigger you know better openweight models from Nvidia and Mistral this um and we've touched on the conversation, but I'd love um whether Chris or Ku if you could help us orient ourselves and you know in the audience and you know what does do open weights actually let a company do that you can't you know do with a frontier model you know behind an API sort of how does it change the game um that we now have these larger

17:00 and better uh and more capable openweight models accessible to a broader range of people. This is a this isn't a very important area that you know I think all of us in this industry have been following this pretty closely for you know at least the last two years um because what we've seen is a trend where and even again going back to anthropics filing they even identified that most of the openweight models are only trailing their frontier models you know 6 months in some cases and so you know that's a it's a real

17:30 challenge to their business model for one but I think it's a real opportunity for a lot of the rest of us And uh you know with this one uh Mistl has in my opinion always sort of been trailing the leading edge of the open models and so for them to come out with this you know trillion parameter model it's a state-of-the-art it performs extremely well uh I mean you still have to have you know quite a bit of hardware to run it but it's I think it speaks very it's it's um it's a huge opportunity for all of us that are

18:05 wanting to run stuff in the open. And you mentioned you know there were it's anyone's guess whether it's a few months six months you know from open models approaching at a frontier level capability and I'm curious you know when we get there because it seems people are um you know thinking it's quite likely you know what is the real diff differentiator then is it still the model itself or you know you mentioned hardware you know is it everything around it data the inference the tools the orchestration you know what what

18:34 does that um as we get closer answer to them being equivalent uh in capability, you know, what how does that change the game, I guess? >> Yeah, it's it's local AI for sure. I mean, uh you know, we we see in even uh in smaller models, uh we've seen a really great track record of small models that can run locally uh getting better. That is unquestionably the future in my mind is there will be some uh we'll be running uh some or most of the models what I call sort of the good enough baseline of models we're going to be

19:07 running locally on our devices and maybe we'll be bursting up to some cloud models for like I don't know the super intelligence stuff but uh you know even these models like the one that mistrol announced uh you can you can run those locally I mean it's not cheap you know stack up four M5s or whatever ever fully loaded, you could pull it off. But I think that that's that that's that's really misleading because people look at that and they're like, "Oh my gosh, that's 20 grand worth of hardware.

19:35 Nobody could ever do that. That's look uh look, I I remember the first uh my first CPU upgrade was a 25 megahertz uh 486 SX. You know, that's my lifetime and and just thinking about what hardware has done since then. Trust me, the hardware is going to keep getting better." >> Chris, I'd love to throw it to you. Do you agree that the sort of hardware local AI is going to be the differentiator?

19:57 I'm paraphrasing you of course K that'll be a focus area or growth area but um or do you see this these new more capable larger open weights models kind of changing the um the game in other ways? >> I think they all change the game. So I mean the first thing is like and I'll cover both of them in a second. So the first one is hats off to Mr. Al. Thank you.

20:20 We've been waiting for you to get back in the game, you know, and you release it with your humor. Lonk, I love it. So, I I I think I think I think we miss them. We do need models that are, you know, in Europe, for example, as a European, it's it's nice to see a European model um come out there as well. But the reality is model provenence and is is becomes important and therefore, you know, you need to have some sort of sovereign models in those cases.

20:48 And I and I think you know good on Mr. You can see that in their report they're partnering with a lot of French banks and things like that that which makes sense but but actually I just loved the openness. I mean if you read the technical paper which I love reading technical papers of I mean first of all this is a this is a quality model right? This is a great coding agent etc.

21:08 But they're really open on how they trained the model. They're really open in some of the problems that that that they've had. So they for example the first version of the model that they did they did some training on it and then it turned out they they put in some new KV cache compression techniques to really sort of try and you know and and what that's basically meaning is they're trying to make it less expensive to be able to run your context in memory and and and really bring that down.

21:32 Um and then what they did is they then tried to train it for agentic and and the same sort of techniques they used for KV cache compression etc ended up harming the agentic capabilities of the model and you're like oh my goodness you've wasted a training run no what did they end up doing they edit the model right so they actually rather than going back and do a retrain they go through through a repair cycle etc and then they're able to do that and and I think getting those insights into the labs of how they're training

22:04 those models becomes important. They're also really clear on how they did agentic coding as well, right? So they spell out what their RL pipeline is, how they were training it on terminal and and actually the jump from a pre-trained model to something that's really good at agentic coding. I mean they they were able to do that in weeks and it was really sticking into their RL lab there.

22:25 So I you know I I really start to think that when I think of the opens the open way the open source community right actually all the expenses a lot in the pre-train there and then if you're making good base pre-trained models available then you know us kind of uh you know people on our Macs and things like that who want to start hacking models then you know we're able to to go looking under the hood we're able to uh then even do our own post training etc and use the same techniques that they have.

22:53 So I think I I appreciate that and similarly I appreciate what Nvidia is doing as well and yes I said Nvidia you know I know and mixture of experts we we have another video that we're trying to we're trying to push there but I I'm holding the flag for Nvidia you know what I mean and and because the comment section is going to go wild on this we all know it anyway.

23:15 So, so what they did there again the one of the nice things about what Nvidia do especially with their numatron models and this new transcribe model is they try and tell you blueprints of how to build models and yes of course they want people to be on their hardware and they want people to train their models etc. But by giving away these blueprints and having these detailed technical reports we get insight.

23:38 So like the new transcribe model that they have is really a sort of hybrid architecture. They've got elements of mamba in there. They've got a whole set of thing for things like diorization etc. So when you're doing transcription I mean it's great what they're doing the transcription model. They're really focused on this fact that you you're not going to when you do when you're transcribing you you're not necessarily going to want the same translation for everything.

24:07 It's it's you know depending on the context you you want different types of translations but actually again looking at the technical report things that you used to have to manually stitch together. I mean, diorization has been one of them is now all in the model, right? And I've been through this. I like I've automated the editing on my YouTube channel and and therefore I've had to build all build all of these manual parts.

24:28 And I look at this and go, "Oh, okay. This is how you've done this. This is how you've connected this." And for us builders, I think that's really important, but it's but it's also important for future upcoming researchers, right? The graduates coming out of school who want to get into AI. If you don't have the openweight models for them to experiment on, if you don't have the papers, how are we going to grow our future researchers?

24:49 Because it's just going to concentrate into these frontier labs. So, it's not just about pricing and keeping things competitive, etc. It's about growing our community in general. So, um I love it from both companies. Awesome job. >> Excellent. Um well, I may this is move us now or I will uh to our third topic. Um, and as we make the switch to this topic, I do want to encourage our listeners, um, if you want to dive deeper into some of these stories, the think IBM Think News team, um, has the IBM Think newsletter where

25:20 you can get, uh, a deeper dive on lots of these topics. The final story we're going to unpack a little bit today is looking at some new world models um, that kind of came out this week. But I want to start um with the term itself world model and maybe Sasha you've written uh several pieces on various world models from different uh labs and I'd love for you to explain that you know what's fundamentally different about uh world models from large language models and you know what is the appeal and why is there sort of

25:55 energy and focus you know being put towards uh developing these >> yeah there's so much interest right now in world models and I think the simplest explanation is that language models models learn mostly about patterns and language while world models aim to learn how an environment behaves. And I think you can think about a ball rolling across a table.

26:17 So we as humans have an intuitive sense of what happens when it reaches the edge. We understand things like gravity and momentum and that the objects don't just disappear. But a world model actually attempts to learn relations like ships like this so that an AI system can predict what might not happen next. And that's going to matter if AI will operate robots or control machinery or even help manage infrastructure.

26:46 >> That's great. And the as you're describing that example, you know, it made me uh think of a you know, the Google announcement, you know, they were describing how simulating the world model is an important step towards AGI. And I guess that makes me wonder, you know, does being able to generate a convincing simulation, maybe it's that ball rolling off the the table, actually mean that the model understands the world that it's simulating, or is it just, you know, going to gotten very good at prediction?

27:12 Um I don't know if Kio if you would love to uh jump in on um you know what what these world models are actually showing us. I think that's really important to sort of separate the these two that there are two classes or two approaches to creating world models and um that I I think that that is the underlying important part of this story because so if you look at Nvidia and and by the way Jensen says Nvidia in case we want to have that debate but uh so we have a class of models that have been trained on like video

27:46 looking at the And we have another class of models that are trained on uh like explicit 3D uh CAD of the world because if you look at a lot of the the the world models like this G84 or even Nvidia's Cosmos 3 they uh because they've been doing world models for a while but there's this what we see is in the industry even deep mind they are co-contributors to this agend this other approach uh that Nvidia calls omniverse and and then they they pull in physics with they have these Newton libraries to that are intended to

28:19 like make very precise estimates about how the world should behave based on math. In the industrial sense, uh these these these models are based on the phys on physics and CAD are in my opinion definitely they're much more useful because they're stable. the the other models that are are from like learned video or whatever like like Genie4 like I think even the documentation they said that you know the it's its view of the world is only stable for like maybe 2 minutes or something like that and then if you look back at

28:56 the same pair part of that world it may be a little bit different whereas the omniverse and Newtonbased models they will be precisely what they have rendered to the millimeter in many cases indefinitely and so but what we see is an interesting combination between these which is what happens when you get to the edge of a model that was not defined in the CAD.

29:16 How do you express how do you visualize that? there's a good good you know place there to say okay well that's where I can use another this like the Genie 4 class of model to fill in the blanks if you will but uh but already we've seen some success around this sort of explicit 3D based model for like factory floor we call anything if you heard digital twin digital twin is is generally usually the sort of physical class of model and uh one of the big changes here that's all you know in the last couple of weeks I think

29:48 matters here is that they're starting to see that this like the as of like Astra Chai TBT Astra they were um they're able to take those 3D rendering CAD renderings and build the omniverse models out of it. Historically, it was actually very very difficult to do this and now they're doing it. They're using a model to create a new model, the renderings that go into generating the 3D model.

30:12 And so can in record time exploit this sort of omniverse sort of sort of physical model. So we started to see this in the beginning. We talked about RSI. The whole thing with RSI is this concern that okay, this is when hockey stick uh you know growth and and improvement starts to happen. and and what but you look around you see all these other things happening unlocking at the same time.

30:36 That's it's it's amazing to see how we're we're able to do more and more things that used to be very hard very quickly now and very easily. >> Chris, I'd love to hear, you know, are you uh just this world model news? Uh is it exciting? you know, if we're moving up the hockey stick, are there sort of unlocks that you're optimistic or kind of right around the bend or um yeah, the what is your take on some of these new v permutations of world models?

31:05 >> I'm super excited about it. I felt that AI slop has got really stale the these days and therefore as soon the we can get access to these highquality world models. I'm just thinking about all the AI slop that we can create and therefore we can level up our slob and we can have a great time doing that. So that that's what I'm looking forward to, you know, and like today when when OpenAI, you know, released their, you know, their animated gift generator as part of JBT image, I'm like, right on, I can create my own slop

31:35 memes now, etc. And and that's what I'm looking for. And therefore, and you think the slop isn't going to be great, it's like, I'm going to be able to create my own slop games. I'm going to be able to create my own slop videos. I can make my own terrible movies where I can be like bra. This is going to blow up. They've got seven heads and they've got rainbow colored toenails, right?

31:59 And then like slop slop slop slop. It's going to be awesome. But but actually probably on a more serious point, just as someone who is a creator, do you know what I mean? I actually I'm looking forward to be able to using some of that in my videos and just be able to uplevel. But um you know, but but I'm I'm all for the slop, man. Chris, I I think the the slop is is wonderful.

32:18 I did speak recently to Anora Bacha who's an IBM distinguished engineer and he he was talking to me about different ways this could be useful for enterprise. like he was saying, there might be ways to manage a bridge or a factory. Like a conventional AI assistant could help summarize maintenance records, but a predictive model could potentially help you evaluate what's going to happen if you need to alter a maintenance schedule or replace a component.

32:49 So instead of saying simply, here's what the manual recommends, it could help answer things like what happens if we do this rather than that? And that could be helpful when mistakes are expensive or dangerous. >> I I a could be severely dangerous because we might be like, I'm now the greatest, you know, maintenance person in the world. I've spent three days in the sloppverse and now I think I'm the best bridge builder and then you go build a bridge and then it's like and it's like wow wow wow.

33:19 You know what we got? We've got world model psychosis and you're like oh well I'm not so good at that. the the Chrissy, do you raise um an interesting point that I'm curious in um the group's thoughts on yeah potential risks or you know whether security or you know you mentioned you know of course if uh um if it's being used in a sort of a factory con uh context and there are sort of safety uh concerns etc.

33:47 But what you know what are the as these world models you know approach becoming more useful you know what are as well the sort of the risks or guard rails that uh people haven't perhaps thought of that we need to be you know thinking about um yeah I think yeah I guess the challenges or the risks >> yeah I mean I I may have presented my slightly humorous take on on that but I there is I mean Sasha had a had a a great point right there is a lot of scenarios where you know what it's it's you just can't train for it.

34:19 Do you know what I mean? And and therefore being able to provide a simulated environment where you can get used to things and can understand that environment and do training within there and be able to create realistic scenarios to see how you're going to be able to react to that. Um I mean again trading robots is one for example or dangerous positions.

34:37 There is there is a ton of use cases where I think it's useful. I I think to your point though it is making sure that we don't have that psychosis. So it's like the guard rails to make sure that there is a nice progression you know so that you can then say okay well you've done that thing over here now we're going to have that scale so you can bring that into the physical and real world but so I think it's going to be a great training tool in that side of things but you know there's a progression and we've been through

35:03 this before right I don't think anybody who's ever played flight simulator in on um you know on their desktop is like okay I've landed you know an Airbus it's uh I'm now going to become a British Airways pilot, you know, and then just walk in and you know, they they still make them do training and stuff like that. Flight sim isn't isn't enough. And I think that's going to be the case here.

35:26 But, uh, you know, I just think we all have to be sensible about it. >> That's a perfect note to end on. It's all we have time for. Um, but I want to thank you all for joining us and to our listeners. Um, you can find us on all podcast platforms. Um, and we look forward to having you join us next week. Thank you so much.