Learning Library

← Back to Library

NVIDIA GTC Unveils Robot AI Breakthroughs

Key Points

  • NVIDIA’s GTC spotlighted the **Groot N1 foundation model**, a humanoid‑robotics AI trained on both synthetic and real data that uses a dual “fast‑and‑slow” architecture inspired by human cognition, positioning it as a step toward AGI‑level robotics.
  • The **Newton Physics Engine** was announced for real‑time physics simulation, enabling more accurate and AI‑driven robotic interaction with virtual environments.
  • A new **synthetic‑data generation framework** for robots was unveiled, addressing a major bottleneck by providing scalable training data to boost robot performance across applications.
  • Experts — Vyoma Gajjar, Kaoutar El Maghraoui, and Nathalie Baracaldo — concurred that the announcements collectively signal a strong, robot‑focused direction for NVIDIA’s AI roadmap.
  • While the show also previewed other AI headlines (Baidu’s models, a paper on Chain‑of‑Thought flaws, and Gemini 2.0 Flash), the dominant theme of the episode was NVIDIA’s robotics breakthroughs announced at GTC.

Sections

Full Transcript

# NVIDIA GTC Unveils Robot AI Breakthroughs **Source:** [https://www.youtube.com/watch?v=TsDdk7xHhMY](https://www.youtube.com/watch?v=TsDdk7xHhMY) **Duration:** 00:39:09 ## Summary - NVIDIA’s GTC spotlighted the **Groot N1 foundation model**, a humanoid‑robotics AI trained on both synthetic and real data that uses a dual “fast‑and‑slow” architecture inspired by human cognition, positioning it as a step toward AGI‑level robotics. - The **Newton Physics Engine** was announced for real‑time physics simulation, enabling more accurate and AI‑driven robotic interaction with virtual environments. - A new **synthetic‑data generation framework** for robots was unveiled, addressing a major bottleneck by providing scalable training data to boost robot performance across applications. - Experts — Vyoma Gajjar, Kaoutar El Maghraoui, and Nathalie Baracaldo — concurred that the announcements collectively signal a strong, robot‑focused direction for NVIDIA’s AI roadmap. - While the show also previewed other AI headlines (Baidu’s models, a paper on Chain‑of‑Thought flaws, and Gemini 2.0 Flash), the dominant theme of the episode was NVIDIA’s robotics breakthroughs announced at GTC. ## Sections - [00:00:00](https://www.youtube.com/watch?v=TsDdk7xHhMY&t=0s) **Robotics Highlights from NVIDIA GTC** - Guests discuss NVIDIA GTC’s robot‑focused announcements, including the Groot N1 generalist model, the Newton real‑time physics engine, and a synthetic data framework to boost robot performance. - [00:03:18](https://www.youtube.com/watch?v=TsDdk7xHhMY&t=198s) **Simulating Robotics for Safe Deployment** - The speaker explains that transferring machine‑learning code to physical robots frequently causes unpredictable failures, prompting the use of synthetic data and extensive simulated environments to test, accelerate development, and ensure safety before real‑world deployment. - [00:06:31](https://www.youtube.com/watch?v=TsDdk7xHhMY&t=391s) **High‑Fidelity Physics Engine for Robotics** - The speaker explains that a new physics engine, built on a RAP acceleration framework, provides real‑time, GPU‑accelerated, high‑fidelity simulations and integrates with AI, reinforcement‑learning, and DeepMind robotics tools, enabling precise virtual training and testing of robots before real‑world deployment. - [00:09:41](https://www.youtube.com/watch?v=TsDdk7xHhMY&t=581s) **Balancing Inference and Training Resources** - A researcher inquires about allocating compute between inference and training, wondering if solutions like DGX Spark can help, and receives a response emphasizing its role in expanding AI research access. - [00:12:47](https://www.youtube.com/watch?v=TsDdk7xHhMY&t=767s) **Desktop AI for Robotics Education** - The speaker advocates democratizing AI by delivering supercomputing‑grade, robot‑training capabilities to desktop devices, enabling students to experiment with fine‑tuned models locally and expanding the educational market similar to early Apple school deployments. - [00:16:03](https://www.youtube.com/watch?v=TsDdk7xHhMY&t=963s) **Baidu's AI Platform Strategy** - Discussion about Baidu's integrated AI model platform, pricing, open‑source dynamics, and competitive positioning versus rivals like Veeam. - [00:19:11](https://www.youtube.com/watch?v=TsDdk7xHhMY&t=1151s) **Baidu's Shift Toward Open‑Source AI** - The speakers debate Baidu’s closed‑source approach, highlight security and ecosystem benefits of open sourcing, and note its June announcement to release new models to compete with OpenAI and DeepSeek. - [00:22:24](https://www.youtube.com/watch?v=TsDdk7xHhMY&t=1344s) **Chain-of-Thought Bias and Security** - The speakers debate how chain‑of‑thought explanations can mislead about AI decision‑making, raising security concerns and showing that these traces inherit the same cognitive biases observed in humans. - [00:25:26](https://www.youtube.com/watch?v=TsDdk7xHhMY&t=1526s) **Debating the Future of CoT** - The speaker argues that chain‑of‑thought prompting is here to stay, explores ways to improve its reasoning—including a “reverse CoT” validation—and illustrates the concept with a personal example about selecting a sofa. - [00:28:33](https://www.youtube.com/watch?v=TsDdk7xHhMY&t=1713s) **Addressing Bias and Chain‑of‑Thought Errors** - The speaker explains that AI inherits human biases and accumulates reasoning mistakes in long chain‑of‑thought processes, arguing for interpretability and solutions such as self‑correction modules, constitutional AI, and neuro‑symbolic approaches like tree‑of‑thoughts. - [00:31:39](https://www.youtube.com/watch?v=TsDdk7xHhMY&t=1899s) **Prompt Engineering, Localization, and Google’s AI Push** - The speaker outlines ways to make prompt engineering more robust, scalable, and culturally customized—while noting inherent bias trade‑offs—and critiques Google’s rapid AI catch‑up, highlighting the recent Gemini 2.0 Flash Experimental image model release. - [00:34:51](https://www.youtube.com/watch?v=TsDdk7xHhMY&t=2091s) **Google Leverages Multimodal Knowledge** - The speaker explains that Google integrates image data with its extensive text and search history to create more accurate domain‑specific AI models, viewing this capability as a current industry baseline rather than a unique catch‑up advantage. - [00:37:54](https://www.youtube.com/watch?v=TsDdk7xHhMY&t=2274s) **Innovation vs Catch‑Up Debate** - The hosts critique whether guests bring truly novel ideas or merely follow trends, audience‑engage by urging listeners to suggest future topics, and close the episode with show promotions. ## Full Transcript
0:00What's the announcement you're most excited about from NVIDIA GTC? 0:03Vyoma Gajjar is an AI 0:04Technical Solutions Architect. 0:05Vyoma, welcome back to the show. 0:07Uh, what did you think? 0:08Thank you. 0:09And I feel the Groot N1 model, the generalist model that they're 0:13calling it for humanoid robotics was something that I really 0:16enjoyed. Kaoutar El Maghraoui is a principal 0:18Research Scientist and Manager at the AI hardware 0:20center. 0:21Uh, Kaoutar, welcome 0:22back to the show. 0:22Uh, what did you see from the keynote that you liked? 0:24Thank you. 0:25Great to be here. 0:26I was also very excited about the robotics and simulation, uh, announcement, 0:30especially the Newton Physics Engine for real time physics simulation 0:34and how, you know, it's, uh, working 0:36with the AI. 0:37Nathalie Baracaldo is a Senior 0:39Research Scientist and Master 0:40Inventor. 0:41Uh, Nathalie, welcome 0:42back to the show. 0:42We haven't seen you for a while. 0:43Um, and, uh, what did you like from GTC? 0:45I 0:46was super excited with their framework to generate synthetic data for robots. 0:51Uh, because that has been a key, key factor for reducing the performance of 0:56robots in all sorts of, uh, applications. 0:58So super excited about that. 1:00I guess we're all into robots here. 1:02Absolutely. 1:03All that and more on today's Mixture of Experts. 1:11I'm Tim Hwang and welcome 1:12to Mixture of Experts. 1:13Each week, MOE brings together the best minds in artificial 1:16intelligence to walk you through the biggest headlines of the week. 1:20As always, there's a lot to cover. 1:21We're going to talk about Baidu's new models that they've 1:23dropped, a paper about the 1:24flaws of Chain-of-Thought of Thought 1:26reasoning, and 1:27Gemini 2.0 1:29Flash Experimental. 1:30But first, I really want to cover NVIDIA GTC. 1:33GTC is NVIDIA's sort of big Conference that they do every year. 1:37It's where the big drops happen. 1:39Uh, you 1:39know, Jensen, Huang gets to walk 1:40out on stage and, and do all the exciting keynotes. 1:43Um, sounds like this group really wants to talk about robots and 1:46specifically Groot N1 which 1:48is a foundation model for robots that NVIDIA announced during the 1:51keynote, uh, Vyoma 1:52maybe I'll 1:52start with you. 1:53What got you so excited about this announcement? 1:55Um, one of the things that I saw is like a 1:57model such as Groot N1 created 2:00by NVIDIA, which is trained on both the synthetic and the real data and, 2:06uh, the NVIDIA, NVIDIA were during that keynote, they were claiming that it 2:09features like a dual system architecture. 2:12So it's thinking fast and slow, which is kind of inspired by that human 2:17cognitive processes that we see. 2:19So I feel we are going to get, these are the small, small ways in which 2:23people are trying to get towards AGI. 2:25Like, let's get a little bit closer, let's get a little bit 2:29closer, however far it seems. 2:31So I feel that was a good, um, catch that they were trying 2:34to do in that, uh, part, yeah. 2:36Yeah, absolutely. 2:37And I know, Nathalie, in your 2:38response, you kind of flagged this kind of synthetic data part being the 2:42thing that got you the most excited. 2:44Um, you know, I know that's been a little bit of a blocker, but it would 2:46be good for our listeners to kind of understand how big of a blocker it 2:49has been sort of traditionally if you want to talk a little bit to that. 2:52Yes, definitely. 2:53So one of the big issues that we have is that when you are trying to simulate a 2:59robot to test it before it goes into the real environment, we have limited data 3:04and traditionally what has happened is that when you simulate, which is less 3:09expensive, you don't have the exact kind of a spectrum of different types of, uh, 3:15of scenarios where the robot might move. 3:18And as a result, when you move your machine learning, uh, programming 3:22into the actual robot, it fails. 3:26And, uh, there are like, uh, these very nice, uh, videos of how it fails. 3:30So if you have a humanoid, it may just fall on their face and it's just crazy. 3:36So that's why, uh, a lot of, uh, the different, uh, robots at companies 3:41and, and actual factories, they have a very restricted set of environments. 3:46Do you see them going down an app, for example, if it's an arm and so forth. 3:51And it's just because it's very, very complex to create a robot that can 3:55move in an environment that may not be exactly the wine it was designed for. 4:00So just moving a little bit, uh, uh, a millimeter or something the 4:04robot may not behave as properly. 4:06And having this. 4:07type of synthetic generation of data allow us to basically create a 4:12huge environment where we can test this and make the whole development 4:16cycle much faster and much safer. 4:19So another aspect to it is that because these robots, as they 4:24evolve more, they move around. 4:27You may have situations where you have unknown safety. 4:32things happening. 4:34And that is super interesting to me because understanding how we can make all 4:38those environments really safe and try to simulate things that go wrong before the, 4:44the, uh, robot actually gets deployed. 4:47Uh, I think it's just fascinating. 4:49It opens a lot of different, uh, new, uh, opportunities to create 4:54safe robots and safe applications and deploy them in real life. 4:58So I am super excited when I, when they said also they open source it. 5:02I, I was very, very excited to hear that. 5:05Yeah. 5:06The additional kind of open source element, I think is like a 5:08really interesting part of this. 5:09Cause they've clearly created like something that's like a big deal 5:11from model standpoint, but they're just saying, actually we're, we're 5:14here to sell hardware, right? 5:16So we actually have like ways of, uh, the, the business incentives 5:19lean towards things like open 5:20source. 5:21Um, Kaoutar, you spend 5:22your days thinking. 5:24All about all things hardware, um, how big of a deal is this? 5:27And why is NVIDIA getting into robots? 5:29You know, like I, I think about NVIDIA, like they started as 5:32like a, a gaming GPU company. 5:35Um, and then, you know, like the next time we really thought about them, it 5:38was like, oh man, we're going to do these big data centers for language models. 5:42And that kind of sounds like a lot of this keynote was robots, robots, robots, right? 5:45We're going to show you videos of robots. 5:47We're going to bring a robot onto the stage. 5:49Um, Why is, why is NVIDIA kind of investing in this, this vertical? 5:54I think it's, it's high time right now to invest in this, 5:57and this is very attractive. 5:58I think all the ingredients right now are coming together. 6:01You know, the, the models, the hardware, the simulations, the 6:05synthetic data generations all are coming together, which makes these 6:09robots really perform very well. 6:12So I think the collaboration that they have with. 6:14DeepMind and Disney, uh, you know, I also was interested in seeing Disney 6:18also play a role here and especially if they're going to bring, you know, 6:21that maybe, um, uh, the fun, the entertainment piece of it, uh, you know, 6:27their Disney characters or, you know, kind of play that into these robots. 6:31That's going to be really interesting. 6:33Uh, so one thing also that was very interesting is this physics 6:37engine that they talked about, which is designed for these robotic 6:40simulation. 6:41Nathalie mentioned 6:42this and another thing it's also built on their wrap framework, which 6:47provides a lot of acceleration. 6:48So it provides, you know, this high fidelity and real time 6:51analytics simulations, which was not, you know, kind of. 6:55possible or realistic before. 6:57And this is very crucial for training and testing these robotic systems 7:00in virtual environments before even deploying them in real world. 7:04So I think that is a very big step forward to enable these 7:08humanized robots to perform well. 7:10and with high fidelity. 7:12So, uh, combining basically the simulation and the, the AI acceleration using 7:18their, this RAP based acceleration framework that they have with high 7:22performance parallel programming, it helps them achieve, you know, fast and 7:26efficient GPU accelerated simulations. 7:29So they're kind of combining the AI world with the physical physics based 7:32simulations to provide, you know, this, uh, interesting, uh, outcome. 7:37And they also have these integrations with their existing frameworks 7:40like the, uh, the S Hack lab, the, with their reinforcement learning. 7:45And I think they also have a playground, uh, that uses deep minds 7:49robotics research, a lot of, you know, integration with existing frameworks. 7:53And, uh, so that's really makes, uh, this high precision robotic control 7:57possible and paving the, this, um, you know, kind of a great environment 8:03that is ideal for simulating tasks such as, you know, manipulation, 8:07grasping, multimodality, et cetera. 8:09Yeah. And I guess a follow up question for 8:11you there, Kaoutar, is, um, 8:12you know, in the last few episodes, I feel like every few episodes we do a segment 8:16where it's like, oh, but you know, open AI is we're about to work on its own ship 8:20or, you know, Amazon might be catching up. 8:22You know, there's lots of people who kind of want to like. 8:24You know, capture some of NVIDIA's market, but you know, I kind of look 8:27at all of this robotics work they're doing and also just the announcements 8:31about like Blackwell Dynamo, right, it's just like the performance 8:34metrics are just like insane. 8:36Um, I guess from your opinion, kind of as someone who thinks 8:39about this a lot and watches the industry, like, can anyone catch up? 8:42Like, it kind of just feels like after this keynote, it's like, feels 8:45like very hard for anyone to really credibly claim that they're going to 8:48kind of like, do things kind of on par with NVIDIA, particularly because 8:52they have this ecosystem, but curious about how you think about that. 8:55Yeah, I agree with you. 8:56I think it is, they're kind of creating this big gap. 8:59Uh, and they're also lining up the right collaborators. 9:05like DeepMind and Disney and others. 9:08So it's going to be hard to catch up, but I wouldn't be surprised if somebody 9:11comes with some contributions, like 9:13either from OpenAI or I think 9:15we'll have to see, although I agree with you, it's really difficult to catch up. 9:21Um, I want to look. 9:23talk a little bit about some of the other announcements 9:25that were done at the keynote. 9:26And, 9:26you know, Nathalie, in particular, 9:27I kind of thought of you. 9:28So, you know, a few episodes ago, we talked a little bit about the project they 9:32announced, uh, I think at one of the last keynotes called Digits, which was this 9:36like, candidly, like quite cute little supercomputer that they were selling, 9:39uh, that would be sort of like a desktop. 9:41Um, and, um, I'm kind of curious as like someone who's a researcher, you 9:46know, if like that kind of form factor for doing work is interesting to you. 9:50Um, I just think a little bit about, I have a friend at a company being 9:53like, our product's getting really successful, but that means we're burning 9:56all of our compute on like inference. 9:57We have no time to do any like training or fine tuning work anymore. 10:01Um, and there's been kind of this like tension of like, oh, well. 10:04All of the compute resources for an organization kind of 10:06come out of the same bucket. 10:08Um, and I guess I'm kind of curious, like, does something like DGX Spark, which 10:11is what it's called now, like, is that, is that something that's interesting? 10:14I don't know if you put your name in to kind of reserve one of these devices, but 10:17I'm curious about how you think about it. 10:19I would pass that 10:20question to Kaoutar actually. 10:22I am not sure how to answer that. 10:26Yeah, I think definitely, you know, uh, it's gonna open up the doors for 10:30many, uh, researchers and enthusiasts and people who are interested in 10:36learning about, you know, all of these different cycles in the AI journey. 10:40So, of course, there is a lot of focus on the inferencing and inference scaling, 10:44uh, because You know, I think the need for that stemmed from the fact that it's 10:48very hard to get access to these GPUs. 10:51So we had to come up and be creative about, you know, what can we do with 10:55the resources that we have available. 10:57But I feel like there is a lot that can be done even, you know, in the pre 11:01training and the fine tuning stages. 11:03It's just because only a few people or few organizations are really limited 11:07because of this, uh, the, the resource constraints that we have right now. 11:12So I would love to have access, you know, to AI in a box that I can use 11:16in my, in my home and experiment with all these different things and then. 11:20push the boundaries even further and I'm sure many others would 11:24have that appetite as well. 11:25Yeah, I think it's just kind of like a cool, I mean, I guess I'm 11:27a little bit of a device nerd. 11:28I was just like, oh, it's just like amazing that you can have 11:30that much computing power just like on your desktop right now. 11:33And yeah, I think it's like very, very 11:35exciting. 11:36Um, Vyoma, any thoughts 11:37on DGX Spark? 11:38I don't know if you like had seen that announcement. 11:40Yeah. 11:40If you think it's more of a gimmick or it's like the kind of thing you'd actually 11:43be interested in playing around with. 11:45I, I, I feel, uh, the reason why they came up with this also is to 11:49target the developer community. 11:51That the developers sitting at home or even want to try 11:54something as a side project. 11:55Because that's how innovation kind of flows. 11:58Someone's side project, if I have the compute to do it, I, 12:02it opens my creativity, mind doors, if you may call that. 12:06So it, it helps you, um, experiment. 12:09Train like let's even if you want to find you in a small little thing and move on 12:13so feel fast kind of works query very well on this and I feel that will be the 12:18reason why people want to adapt more that. 12:20Okay, I have this. 12:21I can leverage this. 12:22Learn about it. 12:24Try seeing if this works or not. 12:26Let's move on. 12:27So no longer are you going to see big companies or big institutions spending 12:31a lot of time and energy on like innovation projects because someone 12:35somewhere would have tried it and be like, Hey, guys, it's not gonna work. 12:37Let's move on. 12:38So I feel that a quick turnaround is something that is the angle that 12:44NVIDIA is trying to play here as well. 12:47If 12:47I might add here, I think, of course, you know, all the 12:50points that Vyoma said are, 12:52are great. 12:54So this is kind of bringing AI supercomputing to the desktop, 12:58trying also to democratize AI. 13:01And, but I think there is also this angle of the robotics, the humanized AI 13:05training at scale, which I think that was also one of their motivation, is, 13:10uh, basically, pushing these humanized robotics forward with models like the 13:15Groot and why DGX here, uh, matters. 13:19It allows, you know, developers to fine tune and deploy these 13:22robotics models locally. 13:25So, uh, using, you know, their Newton Physics Engine, and it 13:30allows also this enabling sim- simulation to real time training. 13:34So you need, you know, these capabilities locally to be able 13:37to, to advance these things. 13:38And I think it would be also great for students. 13:40that are learning AI, they need to learn all these concepts and the 13:44best way to learn is to experiment and have these hands on experiences. 13:49So, uh, students right now are struggling to get access to to GPUs and resources. 13:54Yeah, it'll be really cool if there's like kind of a big schools or education 13:57kind of market for these types of devices. 13:59I just think about when like the You know, the Macintosh or like the kind 14:03of the first iMac laptops came out. 14:05Um, it was like a huge thing, like Apple had a huge market selling to 14:08schools because everybody wanted to give computers to their kids. 14:11Um, and, uh, and it was like a great way to kind of break into people learning 14:15how to, you know, like use these devices. 14:22So I'm gonna move us on to our, uh, next topic. 14:25Um, Baidu, uh, announced, uh, this week that they were launching two new models. 14:30One's called ERNIE X one, and the other one is ERNIE 4.5. 14:33Um, and x1 is supposed 14:35to be the, 14:36their DeepSeek competitor. 14:38Um, and of course Baidu is like a longstanding, you 14:41know, Chinese tech company. 14:43Really one of the leaders in the space. 14:45You know, I think in many ways, kind of like one of the people that you would've 14:48expected to kind of really dominate. 14:49uh, in AI. 14:51Um, and almost like a lot of other players in the space, like OpenAI, 14:54you know, it's like they too are now kind of struggling with all of these 14:57new competitors coming up, right? 14:59Like, um, you know, what's interesting 15:00about ERNIE x1 and ERNIE 4.5 is they're 15:03both closed sourced models. 15:05Um, and so I guess maybe as a first cut, like curious about you know, how 15:09folks think a little bit about sort of open source here, and I guess why we 15:13think Baidu is still trying to pursue like a closed source, uh, strategy. 15:18Um, you know, I'm kind of curious if you have any thoughts on like why they're 15:21still playing this game and if, you know, you really think ultimately they're going 15:25to have to open source just like, uh, just like many others are thinking about now. 15:29Yeah. 15:29Um, I feel Baidu is the kind of company which was, which stemmed, 15:34its origin stemmed from the point that they wanted to create a search 15:37engine for China and they wanted to keep majority of that data private, a 15:42lot of data privacy, um, inhibitions that they were going through as well. 15:46So I've. 15:47feel this is like their chance to kind of utilize some of the information, 15:53the knowledge graph, if you might say, that they have created between 15:56their different AI applications like Baidu AI for like search or 16:01like Baidu AI for maps, et cetera. 16:03So they are trying to come up with like a platform interface with 16:06that one particular model, which kind of creates synergy across. 16:09So I think A1 that is that. 16:11I always believe that yes, sooner or later they are going to realize that. 16:15The open source market would be some sort of a better way to take this forward. 16:20Like how Sam Altman after a couple of years had to say it in a AMA 16:25on Reddit that, Hey, I think maybe we're on the other side of history. 16:28He's getting kind of like dragged into it. 16:30Exactly. 16:31I don't think he wanted to say it. 16:32It's just that it just came out, right? 16:35So I feel that as well, but, but looking at Baidu's, um, Um, like core 16:40structure, they were always like very privacy, um, integrated systems or 16:46something that they believed in building. 16:48So I get where their mind is right now, but I think sooner or later 16:52they are going to have to move. 16:53And the pricing that they've kept, it's almost like half of 16:56what EveSec or the others are. 16:58So that is another, there, there another point that, see 17:02guys, everyone's gonna use us. 17:03We are like half as expensive. 17:05So they have like an. 17:06up in the market as is. 17:08So they're like, maybe we're gonna leverage this as much as we can till we 17:12can, and then we'll see when we get there. 17:14Yeah, the competitive dynamics are really 17:16interesting. Nathalie, I kind of 17:17take a look at the situation and, you know, Veeam is a good reminder, right? 17:20Like this is like, Baidu is like the, the kind of Google 17:24of its, of its market, right? 17:25The kind of search engine of its market. 17:27And I guess I kind of look at that and I say, well, you know, the 17:30kind of reputation has been that like Google has been kind of slow. 17:33to capture the opportunity from AI. 17:36And I say, oh, it's okay. 17:37It's very interesting that in China also, like, the search engine 17:39company is the one that's been kind of like slow to capture this. 17:43Um, I don't know. 17:43Should we read into anything in that? 17:45Do you think that there's like something about search businesses or search 17:48companies or, you know, dominant kind of, you know, these types of search 17:51companies that are like maybe more limited in using or benefiting from AI? 17:56Yeah, I think that's an interesting question because the way I see it 17:59is that probably they don't need to open source in the sense that they 18:03already have a big user, uh, base. 18:07So, uh, people are already trusting them with so many things. 18:12So potentially from the strategy perspective, open 18:15sourcing, uh, would not be. 18:17a key priority as it is for other companies. 18:21Um, the other aspect that every time I think about open source models versus 18:26having a more closed source model, from the security standpoint of view, when 18:32you have an open source model, you are telling people, Hey, just go inspect it. 18:36We try our best. 18:38Tell us how, how would you think we did, and that it's offering a 18:42lot of transparency, and I think it improves how we move forward. 18:47Now, when other companies keep their models, uh, behind the scenes, and you're 18:54just basically not telling how it works exactly, they may be, uh, planning, 18:59for example, to orchestrate different types of components in the backend. 19:03And, uh, I think, uh, we'll see just like, uh, as we see with OpenAI, 19:08We are not fully sure how many models they have behind the scene. 19:11We, we know we have guardrails and a lot of things. 19:14So I think, um, not fully open sourcing in because they already 19:19have such a big base for search, then they probably are thinking 19:24it's, uh, it's okay to go that way. 19:26But as you know, I'm a security person and I like transparency. 19:30Uh, it makes it easier to test the system and so forth. 19:34So, so yeah, that's, uh, my take on. 19:37Open source versus non open source and what they are doing. 19:41Yeah, for 19:41sure. Uh, Kaoutar, are you on 19:42team Vyoma? 19:43Like, do you feel like they're, this closed source strategy is doomed? 19:46You know, we're going to see Baidu have to open source in the future. 19:49Uh, or do you think there's kind of like maybe different 19:51things going on in that market? 19:52Yeah, I, I think I, I kind of 19:54agree with Nathalie, but I see 19:55that they've already started making a step forward, uh, towards open source. 20:00And they, I think they've announced that they're planning 20:02to open source sometimes in June. 20:05Uh, their, uh, their new models. 20:07And I, this just shows that they are also competing with open AI and 20:13deep seeks and especially seeing all the, um, you know, all the bus 20:18that DeepSeek created. 20:20Uh, so open sourcing AI 20:21models like DeepSeek they've 20:23gained traction. 20:24And Baidu is likely sees this as a way to increase also adoption of its own models. 20:31And so also a way to gain market share, attract developers, build an ecosystem 20:35around its models, because if you keep these things closed, you're missing 20:39in terms of this open ecosystem and developers and getting also the community 20:44to help and especially the adoption. 20:46I think the adoption is kind of goes hand in hand with the open source, especially. 20:51Uh, so that is very important. 20:54Uh, so driving this widespread adoption, uh, more developers, more use cases, 20:59widespread adoption, faster improvements through also these external contributions. 21:04Those are all win win strategies when you use open source. 21:07And I think Baidu is getting it, and it's moving also towards that direction. 21:12And this is just, you know, it just intensifies the competition 21:15in China, but also globally. 21:17So, so it's interesting to see these dynamics. 21:25So the next thing I think I want to talk a little bit about is I like to always 21:28have like a paper that we can discuss. 21:30I'm a little bit kind of old fashioned in that sense. 21:32We talk a lot about industry news, but I think it's just fun seeing what's going 21:35on in the world of research and kind of interesting papers from week to week. 21:38And this paper caught my eye. 21:40So the title of the 21:41paper is Chain-of-Thought Reasoning 21:42in the Wild is Not Always Faithful. 21:45Um, and so, for those of you who are kind of not 21:48super aware. 21:48Chain-of-Thought and kind 21:49of reasoning models and exactly how this is looking right now. 21:52Right now we have these 21:53kind of like Chain-of-Thought reasoning 21:55traces, um, where, you know, a model will kind of think through a problem, 21:59uh, to greater or lesser degree before it kind of like renders an answer. 22:03And, um, You know, overall, right, like we've sort of discovered this method is 22:07really, really good in terms of getting the model to perform better and better. 22:11But there's this kind of increasing sort of series of papers, this is not the 22:14only one, that kind of are investigating the problem of what happens when the 22:18model gives you erroneous reasoning for the decision that it's trying to make. 22:24And when is basically sort of these reasoning traces not actually a faithful 22:28way of understanding how models make 22:31decisions. 22:32Um, and Nathalie, you talk, 22:34you think a lot about security, um, and I think this, this kind of paper 22:38really raises a bunch of security issues in the sense of like maybe 22:41we are giving people the wrong impression of how AIs actually think by 22:45giving them Chain-of-Thought traces. 22:48Is 22:48that the right way of kind of thinking about this paper and I guess kind of the 22:51problems of Chain-of-Thought in general? 22:54Yeah, I think that that's a very interesting question. 22:58Uh, I'd 22:58rather have Chain-of-Thought so that I 23:01can know at least a little bit how the model came to an answer. 23:06What the paper shows is a lot of biases that may 23:09happen in that Chain-of-Thought itself. 23:12So, 23:13And I think the reason it is really interesting is because the particular bias 23:17that they are demonstrating in the paper is a bias that it's also in us humans. 23:24So, for example, if I ask you, Tim, a question, is X larger than Y, depending 23:30on the way I phrase the question, a lot of people would answer one way. 23:35versus another. 23:36So that's this cognitive bias that we know for sure and cognitive 23:39psychologists have for so long a study. 23:44Now where that that paper in particular shows that that 23:47same bias exists in this model. 23:50And I thought that was interesting. 23:52And there's like the parallel in the cognitive psychology 23:55for humans versus that paper. 23:57And I think that's just why people are like, oh my gosh, this is so interesting. 24:02Uh, if you study farther this type of situation, what you'll see 24:07is that the models also exhibit a lot of other types of biases. 24:12Now, uh, in particular for fairness, for example, there are some papers 24:18like in 01, uh, there was a very interesting section that I read, uh, 24:23about how the Chain-of-Thought itself may 24:27be biased. 24:28pretty hateful, for example, or may tell you stuff that you as a user don't want 24:33to necessarily uh, see or exposed to, to certain, uh, uh, for certain use cases. 24:40So, uh, overall, I thought it was really interesting paper. 24:44Uh, the caveat, and because I am a researcher at heart, is that they only had 24:50one data set and their temperature was 0. 24:537, uh, which I thought was, interesting. 24:57So that goes 24:57into a lot of detail about the paper, but I would like to see 25:01like more, more expansion on this work because it's fascinating. 25:04It's fascinating. 25:06Yeah, absolutely. 25:06Um, Vyoma did you agree? 25:08I think, uh, I'm curious if you saw things that you were sort 25:10of interested in the paper. 25:12I mean, you know, from that, I think it's kind of interesting is 25:14like on the temperature point, it's like, I don't know, how much should 25:17we believe these results, right? 25:18Like, it actually maybe turns out that, by and large, reasoning 25:21traces are really useful as a way of kind of like understanding 25:25how the model's making decisions. 25:26And maybe we shouldn't be so scared? 25:28I don't know. 25:29What do you think about that? 25:30Yeah, I do 25:30agree with Nathalie on that point 25:32that when it's at the temperatures point, you're actually telling it to be a little 25:36bit more creative in its thinking as is. 25:39And now you're using that as a base of saying that CoT is not here to stay. 25:44It's gone. 25:44I feel it is here to stay because it kind of tells you what it's going through. 25:49And the other part is the people, like all these companies which have come 25:53up with these models, the reasoning models, now they are looking into how to 25:57make the Chain-of-Thought processes 26:00better. 26:01So I feel you can, like right now we see a lot of pattern matching 26:04than having something which is like a more generalized way in which you 26:07can understand the deep reasoning. 26:09But going further. 26:11What about a reverse CoT? 26:13Like whatever a CoT has given you in the information, go 26:16back and evaluate it again. 26:17Tell me if that Chain-of-Thought was right 26:19or not. 26:19So there can be innovative ways in which researchers, 26:23etc. will go on, uh, answering. 26:25I feel it is your to say and it should. 26:28So I'll just give you a short example. 26:30I moved and I'm looking for a sofa in my, uh, apartment. 26:33And what I was looking is, I 26:35said, I want anand style 26:36sofa with a table. 26:38And then it just started giving me aand table. 26:41But I knew that because I read it in that entire reasoning that, oh, now it's just 26:45going and spinning on that table thing. 26:46I don't want that. 26:47Then I went and said, I want a side table. 26:50And then again it went and told me that no, it's the table. 26:52So, so, so I'm trying to tell you that I understood that I have to be so specific. 26:56I want an adjustable low level. 26:59site table and so that is, I wouldn't have done any of that had it just spun 27:03and I would have been okay it's going to give me a japanese style sofa someday. 27:07So I feel those are ways in which it tells you to improve your 27:10prompt, tells you that this is not, 27:12right now Chain-of-Thought is also based 27:14a little bit on your prompt. 27:15It doesn't tell you the exact model internal workings, but I feel it will. 27:21Evolve with time. 27:22It should evolve with time, and it will. 27:24I mean, people are working on it, so. 27:25Yeah, I think that's one of the most interesting things. 27:27I never really thought about it that way, but 27:28kind of the Chain-of-Thought's useful for 27:30letting you know when the reasoning is definitely off, even though it may not 27:34necessarily be a good guide for like when it got it right, how it got it right. 27:38But it's like, it's kind of a debugging tool more than anything else, which I 27:42think is like a really fun way of thinking 27:44about it. 27:45Um, Kaoutar I think I 27:47would love to get you to comment on, uh, one 27:49comment that Nathalie had, which 27:50is it's very funny that these models have kind of inherited all of these 27:54cognitive biases, uh, that humans have, um, which is very funny. 27:58I mean, computers didn't used to have those types of biases, but 28:02I guess we live in a world now where, you know, that's the case. 28:05Um, yeah, just like, I don't know, someone who kind of like thinks about 28:08sort of like hardware, which I always kind of envision as a much more kind 28:12of like structured, you know, thing. 28:14It feels like we've kind of like, they're, they're, these computers are now kind 28:17of like, you know, executing systems that have like all of these weird kind 28:22of like soft emotional aspects to them. 28:25And it's just like, I don't know, I'm curious to hear your reflection on that. 28:27It's like a very funny kind of contrast to what we thought about 28:31computers doing 10 years ago. 28:33Computers used to be exact, you know, zeros and ones, and we expect them to 28:38be kind of the, the opposite of biased. 28:42And, uh, but right now, because they're learning with AI, it's learning from the 28:46data and this data is generated by, by us. 28:48So it has inherited all of our biases and the way these models are learning. 28:53I think it's only natural to see, you know, these. 28:55these outcomes that we have to figure out systematic ways to solve in 28:59them. 29:00So Vyoma, I think, 29:01mentioned some of them. 29:02So, of course, you know, this isn't 29:03valent as Chain-of-Thought uh, is unable 29:08to generalize and it also accumulates all these errors. 29:12The longer, the 29:13longer the Chain-of-Thought of reasoning 29:14is, which, you know, leads this faulty logic despite, you 29:17know, these correct answers. 29:19But, you 29:20know, as Vyoma mentioned, 29:21I think there are potential solutions, and I also agree that this is 29:25something here to stay, especially the interpretability aspect of it is so 29:29important that I think we need only to amplify the importance of this. 29:34So we definitely need things like self correction modules. 29:38Uh, Claude, for example, has this constitutional AI, which is a 29:41reflection based approach that helps to self correct the model. 29:45There's also things like structured step verification, these hybrid models where 29:50neuro symbolic reasoning, like the tree of thoughts, for example, can also be 29:55used to help correct, you know, the logic. 29:58Uh, you know, also combining things like statistical logical 30:01AI with probabilistic reasoning. 30:04with logical constraints, all of these techniques, I think, need 30:07to be brought into the table. 30:09So to figure out how do we combine neural symbolic AI approaches to 30:14improve the reasoning aspects of these LLMs and having also this self 30:19verification in the reasoning and self correcting, which I think, you 30:23know, will keep CoT useful, uh, and reduce the flawlessness in these tools. 30:31So I think these hybrid reasoning frameworks will be necessary 30:36to improve the reliability and the reasoning of these models. 30:40Yeah, we'll definitely see that, I think, and it is kind of a 30:42funny outcome that like we created 30:44this like Chain-of-Thought thing, which 30:46at times can be very emotional. 30:47You read the Chain-of-Thought and like 30:48I read one that was like, Oh. 30:50You know, I'm trying to do the best I can at my job. 30:52Okay, let's try to research this task. 30:54And then you're kind of trying to like make it more computer, uh, again. 30:58Um, it makes me think a little bit about like people will 31:00say, Oh, he's like a computer. 31:01And like, I think like maybe 20 years ago, that would mean that the person 31:04is like very rigid and very logical. 31:06It's almost kind of, I think a little bit about like, maybe, you 31:09know, kids growing up today will be like, Oh, he's like a computer. 31:11And by that they mean, you know, really irrational and emotional 31:15and, you know, it'd be very funny if it kind of like flips what we 31:18mean, uh, when we say, Oh, like. 31:19This person is like a computer or they're thinking like a computer. 31:22Imagine, 31:22imagine that NotebookLM with CoT. 31:25So let's say 31:26you see the Chain-of-Thought of thought 31:27and like in 31:27your NotebookLM you can posit 31:29that. No, don't go there. 31:30Don't think like this, change this. 31:32And then that can be used as a training data set. 31:34I feel it's going to open new avenues for the prompt engineering aspect as well. 31:39People will learn how to make prompt engineering more robust, scalable. 31:44more precise with time. 31:46And I think this also could help with the customizations. 31:48The way you interact with the model might be very different from, you know, 31:53how Tim interacted or how Nathalie interacts with, with the model. 31:57So that, you know, I think localizations, customizations might also be interesting. 32:01So you can also inject cultural cues and preferences. 32:06So, Tim, I think it's going to even be more Biases that we're introducing to 32:11this world while we're trying to make it 32:13different. 32:15It's going 32:16to be, I think, both. 32:17This hybrid world. 32:28on our kind of segment on Baidu, which is again, like, kind of the narrative 32:32that's been in the market or at least on Twitter, right, is basically that, 32:36you know, Google's coming from behind. 32:38They should have captured, you know, the AI revolution and they kind of 32:40missed it and now they're catching up. 32:42But it's kind of like week to week. 32:43It feels like Google's like really catching up now. 32:46Like, there's just all of these launches, which are like quite impressive. 32:49Um, and, uh, uh, Google recently announced. 32:53Um, and this is like almost kind of a joke in AI now. 32:55Like, they launched a model called 32:57Gemini 2.0 Flash Experimental, 33:00um, which is basically a model that they had in beta for a small group of 33:03people, but it is now widely available. 33:05And it's an image gen model, um, that people can play with. 33:08So this by itself maybe wouldn't be kind of super impressive, though the 33:12model itself is pretty fun to play with. 33:15Um, but I wanted to kind of use it as an opportunity to talk a little bit about 33:18one particular aspect of the launch, um, which is that Google is touting that 33:23one of the reasons why its model, you 33:25know, its 2.0 33:25Flash model 33:26experimental is so good, is that it incorporates what they call 33:29world knowledge, uh, to make the image generation better. 33:33And, you know, like many phrases in AI, you're like, well, okay, world 33:36knowledge, what does that even mean? 33:38Um, and I guess maybe I'll start with you, like, What is world knowledge 33:42anyways, uh, and why is it, why is it important, I think, to like AI 33:45generation, uh, particularly in images? 33:48Correct, um, so that's a good point that you said that I, that 33:52first I want to answer this thing that is Google really catching up? 33:55So that, that point that we made. 33:57Yeah, the hot take. 33:58I mean it's just vibes, 34:00I don't have any industry stats, so feel free to knock me down there. 34:03No, I get it, 34:04I get it. 34:04I've been asked this like many times now, like outside speaking 34:07sessions as well, because catching up is like very subjective. 34:10I get it. 34:11My, my question here is, is the real question is, are these models going 34:15to surpass or like at least match the creativity of these already 34:20established model like Midjourney, Dall-e, etc, which are there, right? 34:24So maybe they've arrived at the table, dinner table date, but maybe they got 34:28like Really good big products, which none of these people already had. 34:34And the other question that you're saying that what is this world knowledge? 34:38So I feel the world knowledge in this is that they're talking about is deeply 34:43integrated with their entire Google's knowledge graph that they have with access 34:48to all the real world data that we have. 34:51So instead of just learning from the image that we have, image pictures of etc. 34:57They are also learning from the text, the structured text as well. 35:00Imagine all the Google searches that we've done, all the pictures that I've 35:04posted about my sofa that, you know, this is not what I want, this is what I want. 35:08So that has been kind of added into that historically consistent world 35:13knowledge that Google already has. 35:15And I think it's extremely important to have any to kind of create a model 35:21which is much more accurate in answering the questions that users might have. 35:25Yeah, and I think, I don't know, I see this as almost like Google 35:28using or trying to use its like advantages in the space, right? 35:30It says lots of people can train image generation models, but we've 35:34got, we've got the knowledge, right? 35:35And so like we have to put that to use. 35:37I would trust them. 35:38Like if, even though I've been using many of these other models available, I'm going 35:42to use them now because I know for a fact that they might have much more domain 35:47specific accurate data that I might need. 35:50Yeah, for 35:50sure. 35:51Um, Kaoutar, am I just 35:52operating on vibes? 35:53Is, uh, is Google catching up? 35:55Or is this just kind of like, ah, table stakes? 35:57They're just like able to generate a model which is as good pretty 36:01much as everybody else now. 36:02I think it's table stakes. 36:04So I don't think it's a catch up game series, they've been working on 36:08it, so it's just the time for them. 36:10It's ready for the release. 36:12Yeah, 36:13right. 36:14Um, Nathalie, you were 36:15laughing, I guess maybe you agree, or? 36:17I, I think at some point, uh, last year I was so surprised and so excited 36:22every time I saw an announcement. 36:24Now it's like every week something is happening. 36:28And yeah, so I, I think that the space has started to be like people 36:34are catching up and it's kind of becoming a commodity sort of situation. 36:39Uh, that said, I still get impressed by the fact that right now we can say 36:44like change my tulips for flowers. 36:47that are wild and be very focused on a part of an image. 36:50I think that was not the case a few months back, and that still makes me happy. 37:01So from that perspective, I love seeing more and more models coming out. 37:05I think not only Google, but many other players are going to continue improving 37:10the models and the capabilities and the ways we describe and get to go 37:16get some beautiful pictures out of it. 37:18Um, so yeah. 37:19Yeah, for sure. 37:20Yeah, I feel like it's, uh, it's really hard to be in the AI business 37:23because you're like, you're, you're doing like magical things that have 37:26never been done with computers before. 37:28And then people like six months later are like, ah, what else 37:30do you, what else you got? 37:31You know, it's like, it's very, very hard, I think, to like keep ahead. 37:34Cause I, I agree with you. 37:35I mean, there's almost like announcement fatigue, you 37:37know, it's just like, what is. 37:38What is the next big thing? 37:40Well, I don't know. 37:40It just feels like there's big announcements, like every week. 37:42And so all of it kind of like blends in together. 37:46Yeah, I think the key question here, are they catching up or 37:49are they really innovating? 37:51So, uh, I think that's what we need to focus on. 37:54So of course you can catch up. 37:55You can see what others are doing and try to close those gaps, mimic, 37:58you know, or try because a lot of the stuff, the algorithms, you know, 38:01a lot of them is published, but are you really bringing something 38:05new here to the table that nobody? 38:07else has thought about. 38:08So, I think maybe we should start seeing those more, uh, trends. 38:13Who's the real innovator here who's just playing the catch up game? 38:17Yeah, 38:17definitely. 38:18Kaoutar, I feel 38:19like you're, you're a harsh judge. 38:25Well, that's all the time that we have for today. 38:27Thanks for joining us. 38:28Uh, 38:28Nathalie, Kaoutar, Vyoma, it's 38:30always a pleasure to have you on the show. 38:32And thanks to all the listeners out there. 38:34We're going to try something new this week. 38:37We're always interested in kind of hearing a little bit more about what 38:40you out there are interested in, uh, hearing about from week to week. 38:44So Spotify, please drop a comment. 38:47Let us know. 38:48We're going to be keeping an eye out on that. 38:49And we'll probably work that into future episodes. 38:52So flag anything you've seen that you want us to talk about And we're 38:55looking forward to hearing from you. 38:56Um, and as always, if you enjoyed what you heard, you can get us on Apple Podcasts, 39:00Spotify, and podcast platforms everywhere. 39:03And we will see you all next week on Mixture of Experts.