174 | NVIDIA's big announcements, The US-China AI war intensifies, Microsoft-Open AI frenemies relationship evolves and more important AI news for the week ending on March 21, 2025

0:00Hello and welcome to the Leveraging AI podcast, the podcast that shares practical ethical ways to leverage AI to improve efficiency, grow your business and advance your career and advance your career and welcome to this weekend. news episode, we are going to focus on three different topics. The announcements from NVIDIA's GPU Technology Conference, known as GTC, that was held this past week. We're going to continue talking about the AI war with China. And we're also going to talk about the updates to robotics in this past week that have made some big leaps forward and their potential impact on the workforce globally.

0:35And then there is a long, fun list of rapid-fire items that we're going to touch on, a lot of stuff this week, including some interesting announcements from the big companies. So let's get started.

0:51NVIDIA's GTC conference was held this past week in San Jose in California, and the event attracted more than 25 ,000 people in person and over 300 ,000 people participated globally, making it the world's biggest AI party. And it happens every time NVIDIA does one of these events. They released a lot of new announcements. The biggest one was the release of Blackwell Ultra, which is their next model. And they also revealed what is going to be the successor of that model that is coming up after that. So Blackwell Ultra will have 40 % more memory capacity than the existing Blackwell technology. And that will enable to run bigger models in a more efficient way like GPT-5 and Cloud4 and so on.

1:34The focus was very, very clear on the importance of inference, which is the generation of AI tokens versus the training aspect where Jensen Hong specifically emphasized that running the models is going to be the focus of compute demand in the coming decade. There's more and more competition to GPUs on that particular aspect. And that was a very clear focus of everything Jensen was talking about. And he's claiming that their next infrastructure will allow real-time handling of trillion parameter reasoning models. And the Vera Rubin platform, which is their next platform, combined with their Dynamo Interconnect technology, is supposed to reduce latency in real-time calculations and inference by 50 % compared to existing GPUs.

2:17They also shared DGS Spark, a new version of their existing DGX technology, which allows developers to basically run their own local AI computer. And they're partnering with companies like Dell and HP and Lenovo for the distribution of those. But as we know from Nvidia in the past few years, it's not just about the chips. They have been developing a lot of software infrastructure for AI. So they announced that they've launched Lama Nemotron family of open source models that is optimized for multi-step reasoning and decision-making and that it's showing 92 % accuracy on benchmark tests, which is 15 % improvement to the baseline underlying Meta Llama 3 models.

2:57NVIDIA also open sourced its Qopt optimization engine, which is a decision making tool used by Walmart and UPS right now to reduce delivery route costs by 12 to 18%. It's integrated with Omniverse and it allows companies to simulate the supply chain scenarios in digital twins before deploying the actual changes in the field. They also share that they're partnering with DeepMind on SynthID, which is Alphabet's or Google's, if you want, watermarking system for AI-generated content. So it will now be integrated into NVIDIA's Cosmos foundation models. SynthID will basically enable to track every content that these models generate and know for a fact that they are AI-generated that will hopefully help to fight or at least mitigate the risk of deep fakes.

3:45Now, since we already mentioned SynthID and Google, something interesting about Google and they will come back to NVIDIA, users started using Google's new Gemini 2 Flash AI model that they've added as an experimental to their Google AI studio. People started using it to remove watermarks from trademarked images, including Getty images and similar websites. So on one hand, they have Synth ID. On the other hand, their tools allow people to remove physical watermarks from images. I assume Google will block that in the very near future, but that's just a funny thing that happened this week in relation to this.

4:18One interesting announcement is the collaboration with Disney. They are planning to build a generative AI pipeline for animated films using Omniverse and RTX 1590 GPUs that were just launched at GTC. And the main thing that they will enable Disney to do and other artists after that is to render scenes in real time while the AI agents handle tasks like lip syncing and background generation and so on. and the anticipation is it will reduce the estimated production time and costs by 40%. This is very clear why Disney and other creators wants to use it, and I assume other people like me that are generating content regularly will probably benefit from this down the line.

5:00Now, despite the fact that in January 2025, just a couple of months ago, Jensen remarked that quantum computing won't be useful for the next 15 to 30 years, NVIDIA unveiled the NVIDIA Quantum Accelerated Computing Center, NVAQC in Boston, where they're partnering with MIT and Harvard for research and simulating quantum computing capabilities on GPUs. That's an interesting move and it will be very interesting to see where NVIDIA moves forward in that field. They definitely have the money to invest in that direction if they wanted to. Now, one of the biggest fears that were raised recently is US tariffs and their impact on Taiwanese chip imports and their costs, which are a key to NVIDIA's success, where they're mentioned that they're investing$200 billion in US manufacturing to insulate short-term margins.

5:48Now, analysts wasn't necessarily convinced, and they're claiming that on-shoring chip production can remove 4 % to 6 % of margins by 2027, just because of the higher cost of manufacturing in the US. In general, investors and analysts weren't impressed with this whole presentation by NVIDIA, and the stock actually went down this week. Not by much, but usually after these kind of events, NVIDIA stocks went up, and it was the opposite in this particular case. One of the reasons is Jensen, as I mentioned, is clearly stating that inference is going to be the main thing moving forward because of the reasoning models and the new scaling laws around the reasoning capabilities of these models.

6:27And there's a lot more competition on the inference side compared to the training side to GPUs. So companies like Cerebras and Grok are offering low-cost inference chips that have been running for the last few years, providing better results, faster results, and cheaper results than GPUs on inference. And in addition, the big players like Google, Microsoft, and Amazon are developing their own chips. And now Meta is going into that field as well. So there's going to be a lot of competition to NVIDIA in the inference field. NVIDIA also gave some great examples of actual real-world benefits of AI capabilities that they're providing.

7:05So they shared that their new system, Holoscan 3, which they demoed at GTC, enables real-time AI analysis during surgeries, and that they're partnering with Medtronics, which is a giant in that field, that the system is going to, as an example, reduce laparoscopic procedure times by 25 % while providing better results. They also shared that they're playing in the drug discovery field with their EVO 2 biology model that predicts protein interactions 50 times faster than legacy tools that are doing the same thing. And this can accelerate things like vaccine development. And in another field they're playing in, General Motors announced that it will equip all of its 2027 vehicles with NVIDIA's drive Thor system enabling level four autonomy through their Groot N1 robotics platform.

7:54So like in every single time NVIDIA does this thing, it's on one hand really exciting, on the other hand, really, really scary. They have their solutions across everything AI from software to hardware, to infrastructure, to robotics, to metaverse, to everything you can imagine. And they're definitely cementing their position as one of the key players for an AI future. I cannot stop thinking of Cybernet every time I see Jensen does one of his presentations and watching what NVIDIA is releasing. But it is very, very obvious that they are going to continue investing and participating in more and more aspects of the AI universe and get a grip on more and more components in this AI ecosystem.

8:35And now, as I mentioned, let's talk a little bit about the continuous pressure and competition in the AI world between the US and China. OpenAI has formally called for a ban of Chinese deep-seek models, claiming that it could be compelled by the Chinese Communist Party, the CCP, to manipulate its models to cause harm. So in a formal letter to the U.S. Office of Science and Technology Policy, OpenAI proposed banning China-produced equipment and AI models in Tier 1 countries and Tier 1 companies, including partners like UK, Canada, and Germany, due to privacy violation and security risks. They are comparing the risks from these kind of models, and specifically DeepSeek, to the risk that is possessed by Huawei's chips that has been banned by the US government a while back.

9:20Now, if you remember not too long ago, just when DeepSeek came out, Seym Altman actually welcomed competition from China, but now they're taking a lot more aggressive position against it, whether that's politics or actual real concerns is very hard to tell. Critics of this point are actually saying that DeepSeek open sourced their models, which allows it to run on US servers without any access to anybody from the outside world. So again, you can decide for yourself whether you think this is more politics and trying to block competition from China, or this is a real concern to US national security.

9:54In parallel, the US Department of Commerce has banned the use of DeepSeq models to government furnished equipment, including government employees and any other government devices, citing concerns about the Chinese company potentially accessing sensitive information. The Commerce Department bureaus sent a mass email to staffers introducing them to do not download, view, or access any application, desktop apps, or website related to DeepSeek. That's pretty broad. I mentioned to you in previous episodes that similar bans has been implemented by several states and even several countries. So states like New York, Virginia, and Texas already have these bans on DeepSeek products.

10:33And there's a new bill in the House called No DeepSeek on Government Devices Act that were introduced as a bipartisan effort to block it from any government devices. In addition, beyond government devices, Wall Street Journal reports that Trump administration is weighing to potentially broader restrictions, including potentially banning DeepSeek apps from the App Store in the US and from all US cloud providers. So time will tell whether we will even have access to DeepSeek in the US through standard channels. As I mentioned, other countries such as Australia, Italy, and Taiwan has been completely banning DeepSeek on all government devices, and some other countries such as France, South Korea, Ireland, and Belgium are considering similar steps.

11:13Now, an interesting approach has been taken by Amazon. So if you remember, I shared with you that a few days after DeepSeek showed up, Amazon already started offering it as an open source solution on its AWS platform. That being said, Amazon just started warning customers and all of its employees as far as privacy concerns over using DeepSeq's AI models. So again, is this real security threat? Because Amazon clearly stated when they allow people to use DeepSeq R1 on AWS that, and I'm quoting, inputs and model outputs aren't shared with any model providers, basically saying this data stays on AWS.

11:53So what is the concern? Well, I think in this particular case, the concern is very clear. Amazon is developing its own AI models called Nova, and they are a hybrid reasoning architecture similar to Anthropic Cloud 3.7 and what GPT-5 is supposed to be. And they are planning to offer it as the main solution on AWS. They're planning to make it slightly cheaper than DeepSeek. And the idea here is obviously that DeepSeek models on AWS are competing directly with their own reasoning models that they're going to release. So warning people from the risks associated with it is a good way to push their models even further.

12:28Again, I'm not here trying to protect DeepSeek by any means. I would prefer that US companies and Western Hemisphere companies and countries will run on Western Hemisphere-generated AI models. That being said, I think this particular case has nothing to do with real risks and more with commercial reasons. But staying on the topic of the race between the US and China, Baidu, the Chinese parallel of Google, so it's a gigantic company that is running all search in China, has just introduced two new models called Ernie 4.5. And there are two different versions of it. One is a traditional model and one is a reasoning focused model designed to compete with DeepSeek R1 and its parallels from the US.

13:10Both models are now freely accessible to individual users through Ernie's bots official website. Now, they're claiming it's a fully multimodal foundation model that has exceptional multimodal comprehension capabilities, allowing it to process and understand text, images, audio, and video. And they're comparing it to the latest top models from the US. Now, another interesting change in direction in Baidu, their previous models were not very successful and were not very highly accepted, even in China. And now that there's additional competition, they're open sourcing their models, which they haven't done before.

13:44So Ernie will become open source in June 30th of this year, making it obviously more attractive for developers and other integrations, both in China and around the world. So where do we stand? Continuous competition between the US and the Western Hemisphere and China on leading models that are going to continue coming from the top labs and companies on both sides of the ocean. and I am sure this will be an evolving topic that we're going to be talking about more and more. The current U.S. administration is very obvious in its push to make sure that the U.S. stays ahead in that competition and there's very strong tailwind from the U.S.

14:20industry to work together with the government to push the boundaries of what's possible as far as speeds and capabilities of AI, regardless of the risks that is associated with it, just because they believe, or presumably believe, that the risk from China getting there first is riskier than whatever they're going to put out there. I do not completely agree with that, but I must admit I understand the concerns that even if we slow down, that China won't, and then there's no point in slowing down. So on one hand, I'm terrified with what that means as far as running fast without the right safety guards.

14:53On the other hand, if we don't and they do, I don't know what the risks are over there, but it sounds like a lot of smarter people think that's a serious concern. Now, our third deep dive topic for today are robotics. two new interesting capabilities were demonstrated this week. So I shared with you last week or two weeks ago about Unitree's G1 model performing karate moves that were incredible, where this week they've done something new. They've done the world's first side flip with the G1 model. It's a pretty cool video of the robot doing a side flip, which as I mentioned, no other robot was able to perform before.

15:27Their H1 model was the first model to perform a backflip in March of 2024, and also the first robot to run at a speed of 7.3 miles per hour, which still holds the Guinness World Record for that sprint. Now, how are they training it is actually really interesting. They're training it on NVIDIA's Isaac simulator, which is the ability to train the AI brain behind the model in a simulated environment before running it on the actual robot itself using NVIDIA's Sim2Real solution. and it's the ability to train these robots in a simulated environment, basically training just their brains without the robot ever moving.

16:06And once it gets to a good enough capability, only then they actually provide these capabilities to the robots, which makes it a much faster and much safer capability to train these robots. In a parallel demonstration, Boston Dynamics robot named Atlas has performed some pretty cool breakdancing capabilities as well as cartwheels, Again, showing much more advanced movement capabilities and much smoother and human-like movements compared to previous versions of the models. And they were actually taking a slightly different approach to perform that. They are using motion capture suits worn by actual human operators to train the robots how to move in a more fluent and more human-like way, which means the robots can now mimic, analyze, and understand how human moves in different scenarios and make the robots move in the same way without having to, quote unquote, program the AI on how to do that.

17:02Now, while these are cool and not necessarily relevant to anything, they show significant advancements, both in the mechanics of how the robots move, as well as in different ways and processes to train these robots to do more and more stuff. I think we're looking at a much accelerated timeline for when robots will start showing up, definitely in the workforce, in factories and so on, but very shortly after, probably in different services. I assume we'll start seeing robots in factories more and more in 2026 and 2027, and probably in other places, as I mentioned in previous episodes. So think about things like Starbucks or people who do cleaning of offices and things like that.

17:43We're going to start seeing more and more robotics in these kind of jobs. And then shortly after that, in our houses, doing housework and chores and yard work and so on. Now, the housework will probably take longer, mostly probably because of security risks of what happens if this thing malfunctions or does something unexpected. But I do think that in factories, we'll start seeing a mass deployment probably in 2027 and 2028. What does that mean to the global workforce? Well, it means that in addition to the risk to white-collar jobs, blue-collar jobs are at risk as well. and you've heard me say that time and time again, I don't see anybody explaining what is going to replace that as far as jobs for actual humans.

18:27Everybody sounds that they're very optimistic and they're saying, oh, every previous revolution generated new jobs, so this one will do it as well. But I haven't seen a single one of those leaders explaining exactly what these jobs are going to be, which direction they're going to be, or more importantly, why do they think this will happen in this time as well? Because in previous revolutions, the biggest benefits that humans had is that we were smarter than the machines and we let the machines do the manual labor. And while we did smarter and smarter work, while jobs that were lower paying jobs stayed with humans because it was easier to give humans a low paying job than to develop a really sophisticated machine that will be able to compete with it.

19:05Well, now Unitree, the same company we talked about a second ago, is selling this G1 robot for$16 ,000 a pop. And that's just the beginning of the scale up. Once these things will scale to much larger numbers of production, the number will fall to probably single digits, thousands of dollars, which means even places that require very cheap labor will be able to be done by robots instead of humans. And I have very serious concerns of what are these new jobs that will be generated. And even if they will, I think the timeline that will take jobs away is significantly more condensed than the timeline that will generate new jobs that we can't think about right now, which generates a few years at least of serious concerns to the global job market, which has a lot of other impacts on everything else that we know, because if people don't make money, they don't spend money.

19:51And that has a lot of impact on the global economy. And now let's switch to rapid fire items. There are a lot of interesting things to talk about. The first one that could on its own have been an entire episode is other conversations about Vibe Coding. I talked to you about this in the last two episodes, but now Gary Tan, the CEO and president of Y Combinator has highlighted how Vibe Coding is transforming the startup world and how companies with 10 people can scale to 1 million and tens of millions in revenue in just less than a year by using Vibe Coding. If you remember, we discussed this previously, Vibe Coding is a term that was introduced by Andrei Karpathy in a post a few months ago talking about how he felt when he was using Cursor for the first time.

20:37And in an interview with CNBC, Tan, again, the CEO of Y Combinator, said you can just talk to the large language model and they will code entire apps, potentially allowing 10 engineers to do the work previously required by teams of 50 to 100 people. Last week, I shared with you that 25 % of startups in the current batch of companies in Y Combinator are generating 95 % of their code with AI, despite the fact that they have very capable founders that can write the code themselves. This is another field where human labor is being replaced very, very quickly by developers. Now, these companies still need developers.

21:14The fact that I can now write code without understanding it doesn't mean I can become a coder. People who have coding experience can develop things that are significantly more sophisticated and faster than I can. But as they mentioned themselves, they can do with one or 10 people what was done previously with 100, which means they're not going to hire the other 90 to 95 employees that they would have hired in the same scenario developing these kind of capabilities. So on one hand, it enables amazing new innovation in the software world because significantly smaller teams can develop really sophisticated capabilities very quickly.

21:48But on the other hand, it means less work for less people, unless we create a lot more startups that will hire these fewer developers. I'm guessing that it will end up with a loss of many developer jobs. Now, two different analysis companies have done analysis of AI-generated code, and they're claiming that there's significant increases in duplicate code, as well as that it forces developers to spend more time on debugging. And yet, it is very, very clear what the direction is. Even with these additional issues, it is still worthwhile for companies to use AI to write most of the code. And I'm sure these issues will be decreased over time, which will make it even more efficient and more tempting for more and more developers and companies to use this methodology.

22:32Now, since we mentioned Andrei Karpathy, he had another prediction this week, which aligns with something I was talking about a lot in multiple episodes of this podcast. He's claiming that 99 % of web content will soon be optimized for large language models and AI rather than humans. Now, I agree with them 100 % because what is going to happen is that we're going to have agents that will browse the web for us, which means we have to learn how to prepare the content on our websites or whatever is going to replace websites with something that will be easier for machines to read than easier for humans to read.

23:06So think about maybe a single file that is well-structured with the right formatting to allow AI agents to understand all the information in it without any fancy HTML and CSS user interface. Now, the interesting thing that I think will have to happen is for a while, the next few years, we will need to keep two parallel websites. One for humans, because humans will still continue to browse websites, at least in the next few years, and one for agents. Now, what does that mean for the world? I'll give you an example from my personal experience and how the world can dramatically change in that field.

23:38And that will give you, I think, a great idea of what can happen in the rest of the world. So I used to run a reasonably sized travel company called Last Minute Travel. Last Minute Travel did about$100 million a year, which sounds like a lot, but it's very little in the travel industry. So if you compare it to Expedia and all the brands that work under that, and by the way, under these two companies are most of the travel websites that you know. So while you think you know what you're doing and you know which websites are better than others, it's all a scam. It's all owned by two companies, and they also own the quote-unquote comparison websites like Kayak and Trivago.

24:09So there's no real competition in the travel world. But putting that aside for a second, And these two companies between Expedia and all their brands and Booking and all their brands pay over$10 billion a year on digital marketing, basically making sure they're top of minds and that everything you're going to click and search on Google is going to show their solutions. But what if that's not the case anymore? What if nobody goes to websites and what if nobody looks at ads? What does that mean to the huge amount of traffic and huge amount of money that these companies are generating? What does that mean to the way the internet even works?

24:43So if you think about how the internet works today, the internet is free. You're not paying for using the internet. You're paying your ISP for their service, but you're not paying for the internet itself. What runs the internet is ads, right? So people are paying money so they can send direct ads to you based on your search, based on your behavior, based on your cookies, and so on. And all of that may go away with the introduction of agents and them browsing the world. So how does the internet continue functioning in this new model? I don't think anybody knows, but it's coming very quickly. Now, those who will figure it out faster will gain a huge benefit.

25:19So again, think about the example I just gave you. What if a company that has access to the travel backend, which many companies do, figure out how to do this better than Expedia and Booking.com, they can undercut them very quickly as soon as agents take over more and more of the search of the web universe. and that is true for any other industry as well. So any company that depends on web traffic for its livelihood needs to think very deeply on what that means to them and what steps they need to take in order to sustain and preferably grow their current dominance in this new universe. And now for really quick updates on several new releases from multiple companies, OpenAI is going to introduce a new version of their voice capabilities on the API.