170 | AGI and ASI are coming and we are NOT READY, Is GPT 4.5 better than 4o? And more AI news for the week ending on March 8th, 2025

0:00Hello, and welcome to a weekend news episode of the Leveraging AI podcast, the podcast that shares practical, ethical ways to leverage AI, improve efficiency, grow your business, improve your business, improve,

0:22improve your business. released in the past few days, and it's going to be very interesting to talk about. We're going to talk about how ready or unready our world and our government is to what is coming. We're going to do that with the help of an interview with Ben Buchanan, who was the advisor for artificial intelligence in the Biden administration, as well as a very interesting paper released by Eric Schmidt, the ex-Google CEO, Dan Hendricks, the AI safety expert, and Alexander Wang, the CEO of Scale AI. So two different inputs that will provide us a lot of interesting things to think about and a long list of rapid fire items, including the release of GPT 4.5 to any paid member.

1:03So let's get started. We'll start, as I mentioned, with Ben Buchanan, who's the former special advisor for artificial intelligence to the administration. And he was interviewed on the Ezra Klein show. And I'm going to start with a quote actually from Ezer Klein himself, because I think it's brilliant and it will help us set the stage. For the past couple of months, I've been having a strange experience where person after person, independent of each other, from AI labs, from government, has been coming to me and saying, it's really about to happen. We're really about to get to artificial general intelligence.

1:36And what they mean is what they have believed for a long time, that we are on a path to creating transformational artificial intelligence capable of doing basically anything a human being can do behind a computer, but better than most human beings at doing it. And before they thought maybe it would take 10 or 15 years, but now they believe it's coming inside of two or three years inside the Donald Trump second term. And they believe it because of the products they're releasing right now. they believe because of what they're seeing inside the places they work. I think they're right. If you've been telling yourself this isn't coming, I really think you need to question that.

2:16It's not Web3. It's not vaporware. A lot of what we're talking about is already here right now. And I think we are on the cusp of an era in human history that is unlike any of the eras that we have had before. We're not prepared, part because it's not clear what would it mean to prepare. We don't know what it would look like, what it will feel like. We don't know how labor markets will respond. We don't know which country is going to get there first. We don't know what it will mean for war. We don't know what it will mean for peace. And as much as there is so much else going on in the world to cover, I think there's a good chance that when we look back at this era in human history, This will have been the thing that mattered.

3:03This will have been the event horizon, the thing that the world before it and the world after it were just different worlds. So this is a quote straight out of the beginning of that episode in the interview with Ben Buchanan. And this is, again, Ezra Klein talking about what he's seeing. And I agree with him. I think we are on the verge of a very dramatic transformation in human history, society, government, politics, economy. Everything we know is going to change dramatically. The podcast episode itself is called The Government Knows AGI is Coming. I highly recommend listening to the whole thing.

3:40It's about an hour long and it's really interesting. And they're touching on multiple points. And I will mention just a few. Buchanan stresses the urgency that basically saying that it's very, very clear that the leading forces that drives AI forward are certain that AGI is just around the corner. And all they're saying is very loudly shouting to anybody that they can, that it's coming and nobody is ready. Another big topic that Pekennan talks about is the competition with China, basically stating that achieving AGI or ASI first will have profound economic and military and intelligence capabilities that will be very, very critical to who's going to come out on top in this race.

4:21Now, Now, another interesting point that he mentions that relates to that, he's saying that basically every previous technological revolution that we had was preliminary funded by the U.S. Department of Defense. This has to do with the Internet, various types of communication and almost anything you can think of. The DoD was involved in funding and hence in understanding and also controlling to an extent. And this is the first time that the Department of Defense is not driving the bus and is not involved in the process. And hence, the U.S. government has a lot less control and understanding on what is happening, which increases the risk of this is falling into the wrong hands or going in the wrong direction.

5:02Now, I don't necessarily agree that the government is the only way to go, but it is very clear that as an example, from a security perspective, the leading labs are not even nearly as secure when it comes to vulnerability to cyber attacks and people stealing their secrets. We've seen several examples of those in the past because they try to do it, but it's not their top priority. And if this was a DoD project or even a DoD monitor project, it probably would have had significantly higher levels of security. They're talking about the tension between the safety and opportunity on this channel. And as we've seen in many different cases, including the speech that we talked about from our vice president, currently in this administration, the opportunity and maybe the risk that the other side will get there first is taking the driver's seat instead of the safety of what might happen if we get this wrong, which is different than the previous administration.

5:53Now, as I believe, Buchanan also believes that AI will likely have a substantial impact on labor markets, most likely driving significant mass displacement. Now, he's saying that there's no clear path for the government right now to do anything about it. He's talking about how slow governments in general and the US governments are in responding to changes in tech. And in this particular case, because the change is so fast and also so profound in its impact on everything, he's claiming that the US government is not equipped in its current setup to handle these changes and provide any meaningful control on the outcome.

6:29Now, I don't know if I'm excited or terrified with what's happening right now with Doge and Elon Musk and his process to eliminate different government overheads. In general, I'm very much for that. I think when it comes to the push forward with AI, I personally think that they're going beyond the optimum point and there's going to be not enough control on what US companies even can release. That being said, the race with China is very obvious. I don't know if it's a winner takes all, but it's very clear that whoever's going to get to advanced AI first will have significant benefits. And Buchanan is also talking about the potential partnership with China and attempts to collaborate at least on AI safety while still maintaining strong export controls and trying to keep China behind as much as possible.

7:17Bottom line for this particular interview, A, go listen to it. I'm going to put a link in the show notes. B, it is very, very clear that the government is not in control of what's happening in the AI right now and that the current administration is pushing forward as fast as possible to be ahead of China while throwing safety to the back seat. Will that backfire? I guess we are all going to find out. The other thing that is very, very clear is that AGI is just around the corner and that nobody is ready for that. The second really interesting piece that I found this week is called Super Intelligence Strategy.

7:51And it was written, as I mentioned, by three people that even if one of them was involved, it was worth reading. And that's Dan Hendrick, who's an AI safety expert, Eric Schmidt, that is the ex-Google CEO and one of their co-founders, and Alexander Wang, the CEO of Scale.ai. All three are globally known experts on the topic of technology and AI specifically. And they released a very detailed paper that talks about what happens on the road to and once we achieve superintelligence. And what's superintelligence is AI systems that are smarter than humans on all domains, in some cases, significantly smarter than humans.

8:28And they're saying that this could happen within this coming decade. So by 2035, we may have superintelligence available to anybody who can achieve it. They're saying that this is basically the next Cold War and controlling and owning these kinds of capabilities and making sure that the other side knows that you control it and that you can generate damages to their side will be the new nuclear weapons. They're talking about a concept called Mutual Assured AI Malfunction, or MAIM as an acronym, and they define it as the nation's ability to cripple rival AI by cyber attacks or physical strikes on data centers to avoid being outgunned on this race.

9:11Now, they're taking a position of a winner-takes-all scenario where the first country that's going to get to super intelligence could dominate economically, militarily, and in other aspects, leaving everybody else behind, especially if they can use that in order to slow down the progress of others. This makes the race obviously even more fierce than before, at least on a country level, not necessarily on a company level. This is a very long document, and what I suggest to all of you, and again, I will put the link in the show notes, is to download the PDF and upload it to Notebook.lm and then either listen to the summary or create yourself a summary document in Notebook.lm and read that and then go back to the original document to read specific pieces that you're more interested in.

9:56But it's very, very clear that the top minds in the world think we're on a relatively short final to getting AGI and then a slightly longer final to achieving ASI, artificial superintelligence. And again, nobody's ready for the implications and yet everybody wants to run there faster just to make sure they're there first. Now, since we're speaking a lot about the race with China, let's talk a little bit about a quick summary of what's happening in China right now. So we all remember the DeepSeek moment for just a month and a half ago where DeepSeek released V3 and then R1 and claiming that they did it significantly cheaper.

10:30I don't know if they did it significantly cheaper, but they're definitely running it significantly cheaper. So So they're allowing you to use a very powerful AI at significantly less than you can run it with US companies. They're claiming, by the way, a theoretical potential profit of 544 profit margin. If all their users would be paid users, that's obviously very far from the case. But they did a very quick math that says that a daily operation of their current volume is$87 ,000. And if all their members would be paying members, they could make$562 ,000 every single day, which is really interesting.

11:04other than the fact it's completely made up and most of the users are not paying users. But what it gives you the understanding is it gives you the understanding that it's possible today to create very powerful AI systems that actually run profitably versus most of the leading labs right now that lose billions of dollars every single year. Now, while the dust haven't fully settled on DeepSeek, we have some new models coming out. So we talked about Gwen from Alibaba when it was released just a few days after R1. Well, they're now released their thinking reasoning model that is called QWQ32B. The interesting thing about this model is that it's a thinking model that was trained on top of QUIN 2.5, and it has significantly less parameters than many of the other reasoning models, but it's achieving the same level of benchmarks.

11:52So a similar approach to what we've seen with R1, where a smaller, significantly cheaper model can achieve comparable outputs to things like DeepSeek R1 and O1 Mini from OpenAI. This, first of all, puts Alibaba on the global map with its capabilities, and it's driving a significant push forward across entire industries in China because now they have models like R1 and like the new Quen to integrate into new processes and new tools and new products that they're developing without needing to rely on US and foreign AI capabilities. But that's not the final news from China. There's a new force in China that is virally growing in China right now, and it's a company and a product called Manus.

12:36And Manus, spelled M-A-N-U-S, is a new agent tool that, per the reports from China, combines capabilities of deep research plus chat GPT operator plus cloud computer control, all combined into one. Their website is saying that they're focused on real-world complex tasks, and they're giving the example of creating an itinerary and executing to a trip to Japan, providing in-depth analysis of Tesla stock, creating interactive courses for middle school teachers, comparing different insurance policies, and so on. So day-to-day use cases that are aimed either on the business market or on personal market, where an AI reasoning model and a deep research model combined with the ability to take action in a browser are all combined into one.

13:22This is something that hasn't been released in the Western Hemisphere yet, and I assume it is coming in 2025, maybe soon, and it will change a lot of everything we know. And if you are running a business, any business that depends on web traffic, that depends on SEO, online marketing, or web traffic for a significant portion of its revenue, you need to start thinking about it, Because what it means is that more and more of the traffic that comes to the website is going to be agentic versus humans, meaning the way the website needs to be built, it needs a completely different back end to appeal and appease agents versus humans.

13:58And the front end that we invested so much in order to understand our customers and what they think and what are their pain points and how to communicate it and how to attract them will become less and less relevant. And that is a complete shift from everything we've learned since the introduction of the Internet. but it's a shift that's coming. It's coming really fast and it's going to have a profound implication to anybody who is running an online business from giant companies all the way to small startups and even solopreneurs. Another big thing that we're seeing from China right now is Huawei's new chipset called 910C, which is reportedly the first chipset from China that is starting to get close to NVIDIA H100s.

14:36So they're still talking about 40 to 60 % capacity, but they're planning to start churning them at 100 ,000 units by the end of this year, which now makes it a significant option for Chinese companies to power their AI ambitions. Now, this comes in tandem to another piece of news that is saying that Chinese companies are able to put their hands on the latest and greatest NVIDIA chips. So despite the US export control, it's clear that Chinese firms right now are acquiring these top-of-the-line Blackwell AI chips, and they're buying it through buyers in Malaysia and Taiwan and Vietnam that are legally purchasing these and illegally reselling them, or at least portions of them, to Chinese companies.

15:18While Huawei is claiming that they're catching up to NVIDIA, NVIDIA's CEO himself is basically saying that their new GB200 are 60 times faster in token generation than the Chinese competitors. that's obviously very significant. And generating tokens is becoming more and more important part of this race because all these reasoning models compared to the previous generation of models require significant more time compute when generating tokens because the reasoning part actually happens when tokens are generated versus in the training part of the process. And so being able to do that makes a very, very big difference to the output.

15:56And NVIDIA chips are still much superior to their Chinese counterparts on that particular aspect. And so the combination of China being able to make chips that achieve 50 % capacity in training and maybe complement that by illegally getting their hands on NVIDIA chips for inference will allow China to develop advanced AI capabilities. And the US export control needs to find other ways to monitor how to control this in order to sustain the gap that is very critical based on the first two topics that we talked about in this episode. Now, if you think that AGI and ASI are the only two things we need to worry about when it comes to how our world is going to look like.

16:32Let's switch gears and talk about humanoid robots. We've been talking about this a lot in previous episodes, but Figure AI, which is one of the leading companies in robotics in the world right now, just raised$1.5 billion in funding, pushing their valuation to$39.5 billion. This is a huge leap from their$2.6 billion valuation just in 2024, so a year ago. Now, if you remember, we discussed they already have several robots being tested in BMW factories working in assembly lines, and they're planning, based on CEO Brett Abcock, to mass-produce 100 ,000 robots in the near future or built on their in-house Helix AI platform.

17:11Staying on the humanoid robots topic, many of you probably have seen, and if not, go check it out, and I'll put a link in the show notes. Unitries G1, which is their smallest model, has several Kung Fu videos released online that are absolutely insane. I literally watched this video, I don't know how many times to see if it's real or maybe it's AI generated or it's computerized, but it's actually a robot doing Kung Fu, even doing a roundhouse kick, kicking a stick that a person is holding. This level of dexterity and control is incredible and it just shows you how capable these robots are getting right now.

17:47Now to put things in perspective, this G1 robot is already being sold and its price tag is$16 ,000. So while it's not a cheap toy, it's something that many people can afford to put in their house. Now, I wouldn't put one in my house right now because I don't know what it might do to my kids or other things in my house. But once there is clear controls on the safety side of things, paying$16 ,000 to have a robotic assistant that can do any chore you want to do in the house or in the yard is not a crazy amount of money. And when you think about businesses, there's no reason for a barista. There's no reason for people packaging packages in different logistics centers.

18:26There's no reason for cleaning services. All of these things can be replaced by robots at a price point that even right now is reasonable. If you take this two, three, five, 10 years out, the prices are going to be significantly lower. The safety is going to be significantly higher. And these robots will be able to take blue collar jobs left and right on almost every task we know. Now, to take that to the next level, UB Tech, which is a Chinese company that generates humanoid robots, just showed an amazing demo of their S1 robots working together, collaborating on complex factory tasks. So dozens of these humanoid robots are working together on a single assembly line, doing assembly inspections, sorting, all of that powered by a new system that they're calling BrainNet Framework, which allows a super brain of all these robots to talk to one another and allow them to collaborate in a very effective way.

19:19If that sounds really scary to you, I'm on the same page with you because if you take that to the Terminator scenario, we're in very big trouble. If you take that to the factory scenario, we're only in trouble from the fact that I mentioned before that what are we gonna do with all factory workers around the world once you have millions of these robots that can do the work significantly more effectively than any person and they do not require breaks and they do not take vacations and they do not have medical issues or medical insurance and all they need is to be charged and maintained every now and then, but for significantly less money than any human employee at a significantly higher output.

19:54Where does that lead us to and how quickly will we get there is two questions I don't think anybody knows, but I think we'll need to start finding answers within the next three to five years, probably when these robots start to mature and are manufactured at a high scale. And now let's switch to rapid fire items. As I mentioned, the first item is that GPT 4.5 has landed in any paying member's chat GPT account. So if you have the$20 a month or the Teams account, you can now use GPT 4.5. We talked about it last week. This is the largest model ever by chat GPT, maybe the largest model ever, period.

20:31And in theory, its strength is in EQ. It feels more human, and it's a better conversational partner than the previous models. I must admit that in my original test, I couldn't see a significant difference between GPT-4.0 and GPT-4.5. And I've actually stumbled upon a very interesting post on X by Andres Karpathy. Andres Karpathy was a part of the OpenAI team. He was the lead AI in Tesla, and now he's an independent sharing very interesting things about everything AI because he has very in-depth experience. And he actually did a survey on X, basically a mini chatbot arena where he gave the same five prompts to ChatGPT 4.0 and GPT 4.5.

21:14And then he showed the results to multiple users on X, asking them to vote which option they prefer. And 80 % of responders, basically four out of the five different prompts, prefer GPT 4 over the newly GPT 4.5, which basically comes to show you that to see the benefits of 4.5, you need something that is very nuanced and that aligns with its benefits. I remind you, and we talked about this in depth last week, that on the token side, if you're using it through the API, using 4.5 is 15x more expensive than using GPT-4 without any very clear benefits. So I think nobody's probably using it through the API right now.

21:55And on a personal note, I must admit, like I said, I couldn't find any big differences until yesterday when I was trying to develop a new custom GPT. When I developed custom GPTs, I usually start with a regular chat with ChatGPT. And then I try to finesse the process. And then I sometimes ask ChatGPT to help me write the instructions for the custom GPT. But in this case, I was testing 4.5 just to see how we were doing that particular task. And the task has to do with analyzing the core values and the why behind companies based on their websites. And then after I was able to do it successfully in the chat, I took the same exact instructions and put them into a custom GPT.

22:33Now, it's unclear what is running behind custom GPTs. Some people saying it's GPT-4 Turbo. Some people are saying it's GPT-4.0. I couldn't find an answer I'm 100 % certain of, but I am certain it's not running GPT-4.5 because the differences in that particular task were very extreme. GPT 4.5 was very good at actually reading between the lines and understanding the why of a company and the core values of a company based on what they write on their website. And GPT 4, or basically the custom GPT, whatever runs the custom GPT, did an okay job, but not even close to the same level. So there is something there.

23:08And again, it's the first time I see that. And I guess it really has to do with these kind of tasks. If you have tasks that are more emotional and conceptual than very practical and factual, then GPT 4.5 will do a better work in doing that. And for everything else, you probably won't be able to tell the difference, and hence it's not worth switching. Staying on the topic of OpenAI and pricing, in an article on the information on March 6th, there has been news from OpenAI that they're reportedly planning to charge up to $20 ,000 per month for specialized AI agents designed to support PhD level research.

23:44They're obviously not going to release this to the public. They haven't even mentioned exactly who is it going to be released to based on what criteria and when, but it's very obvious that they're seeing in their labs stuff that is going to be worth more than$20 ,000 a month to the right entities. I definitely see big research corporations either on academia or on different large corporations who are doing research across different things, as well as obviously Department of Defense paying these kind of amounts without even blinking to get the benefits that they can get on the other end. So just showing you that beyond what we're using on the day-to-day and people that are considering, should they pay the$20 a month on specific tools?

24:23And the answer is yes, pay the$20 a month, probably on multiple tools, because if you know what you're doing and if you have the right use cases, it will pay back very, very quickly. But putting that aside, there is a huge market for significantly more expensive, more specialized tools out there, and they're coming. Another cool piece of news from OpenAI has to do with different news about Sora. Sora is their video generation model that everybody was waiting for for about a year, and they finally released in December of 2024, and now they are planning to integrate some of it into Chachapiti. So Sora right now is running on a Sora website, not within Chachapiti itself, but if you have a paid account, you can still use it.

25:00Well, the plans are based on a Discord office hour session that happened at the end of February, they're planning to bring Sora into ChatGPT, where you'll be able to have a chat and generate videos on the fly right in there. You will not have some of the fancy capabilities that the Sora website has. It will be less sophisticated, but for most users, it's probably going to be enough. The other big benefit of having it within ChatGPT versus just on its standalone website is obviously the fact that it understands the context of what you're working on. This is one of the main reasons I like generating images in Gemini now more than I generate images in mid-journey.

25:35I still think mid-journey as a tool provides you more control and more capabilities, but it lacks context. It does not understand what you're working on and you cannot iterate in the same way that you can within a chat. So having the ability to do the same thing with a video, that the video understands what you're working on, your customers, the project you're in, the previous steps that you took, the images and why they're there, and the thing that you're trying to achieve will be huge for video creation as well. So I'm personally really excited. They also mentioned that they're bringing the ability to generate images with the Sora engine, which will replace DALI 3, which I must admit is a big disappointment compared to everything else that is out there now.