168 | When it rains it pours, 3 leading models in one week, AI training is critical, Alexa+ release, and more AI news you need to know for the week ending on Feb 28th

0:00Hello, and welcome to a weekend news episode of the Leveraging AI podcast, the podcast that shares practical, ethical ways to improve efficiency, grow your business, and advance your career. This is Isar Maitis, your host. And in just one week, we got several different leading models around the world. Three in the US, OpenAI with GPT 4.5, Claude with 3.7, Sonnet, Grok, which we talked a little bit about last week, and Alibaba with a whole new set of models that are incredible. So that's going to be our first main topic. Our second topic, we're going to talk about AI training to executives, and it's important, and what can happen if you don't invest in this kind of training.

0:36And then we're going to have a lot of rapid fire items. Some of them are really exciting and interesting. So let's get started.

0:46So last week, we spoke a lot about Grok and there's going to be some additional news about Grok today, but we will start with launches from two other companies. We'll start with Claude 3.7 Sonnet. So on February 24th, Anthropic has unveiled Claude 3.7 Sonnet, which is available both on the app and the API and online. And it's the first in the world that has a reasoning and not reasoning model combined in one, what they're calling a hybrid reasoning AI model. So you can pick between a quick response or a extended thinking time to provide it more time to think and provide a reasoning model, something in the lines of what OpenAI promised us for GPT-5.

1:29Now, I'm not sure it's exactly the same because you still need to pick, do you want the regular model or the extended thinking model? And I think the idea of GPT-5 is that it will do it on its own, but it's still a step in the right direction of one model that does both things. Now, the other interesting thing is that the pricing for the model remained unchanged compared to GPT-3.5, despite the fact it provides significantly better capabilities. And that's actually really good news. And that's very, very different than what's happening in OpenAI. We're going to talk about that in a minute. Now, early testers, including people on Cursor, Cognition, Versal, Replit, and Canva, has reported significant improvement in code generation, handling complex code bases, and producing production-ready code with fewer errors.

2:12So as you probably know, Cloud 3.5 is the leading model for most developers in the world today across all the development platforms that I mentioned before. And this will just extend the gap that Claude has across development solutions in the world, maintaining their leadership in that field. Now, in addition, Anthropic also introduced Claude Code, which is a command line tool for agentic coding. It's currently only available for research preview, but I anticipate that to be widely available sometime very, very soon. Now, in addition to all of that, they have improved their safety, which has been always the core of what Anthropic is trying to do.

2:51And the new model reduces unnecessary refusals by 45 % compared to previous version while maintaining the strong security standard that they have defined before. So less false positives, which is good for everyone. Overall, a very good model that is being highly embraced, mostly by the coding community. Second big announcement we got only three days later was highly anticipated. We knew that was coming because Sam announced it on X before. OpenAI released GPT 4.5, which is their largest frontier model. This is the Orion family of new versions of products. As Sam suggested before, that's going to be their last non-reasoning model.

3:27And as I mentioned, GPT 5 is going to be a unified solution that will unify the O model thinking family together with the traditional GPTs. But this was like a midterm solution. If you remember, Orion was supposed to be released long ago, and there were a lot of rumors that OpenAI could not get it to produce a significantly better result than 4.0, and that's why they didn't release it. Well, they're claiming right now that GPT-4.5 has significantly less hallucinations than 4.0 and slightly less than 0.1. And in addition, several different researchers are saying that GPT-4.5 is slightly better than GPT-4.0 across virtually most categories.

4:07But there is a big disappointment from 4.5 when it comes to people who actually have access to it and started using it. We're going to talk about that in a second. But a lot of people are saying it's not much better than GPT-4.0, and it's definitely not better than O1 and O3. And it is 50 times more expensive to use than GPT-4.0 through the API. And it's 70 times more expensive than O1 Mini on the input side of the API and 15 times more expensive and 30 times more expensive on these models on the output of the API, while many people are saying that it's definitely less capable than O1 Mini on coding tasks and some other tasks as well.

4:48So now even Sam Altman himself acknowledged on X that they're having issues with these models, and now I'm quoting, GPT-405 pricing is unhinged. If this doesn't have enormous model smell, I'll be disappointed. I don't know if that's a acknowledgement of their failure or just trying to get some love for the fact that it's such a big model, but the reality is it doesn't justify the results that it's providing. So the ROI is just not there. You're not getting enough benefit for the significantly increased price. Sam also said that they're running out of GPUs because of the size of this model, and that's why they're slowing the deployment.

5:25So right now the model is only available to people at the$200 a month level. And while they're saying they're going to release it to everybody else soon, it might have a lot of limitations because as I mentioned, Sam is saying they're running out of GPUs in order to run this enormous model. My gut feeling tells me is that they released 4.5 because they felt they had to. We had DeepSeek in January, then Grok last week, Sonnet 3.7. So their biggest competitors and Quinn from Alibaba. And so they felt the pressure to release something right now because GPT-5 is not ready yet. I think there's going to be a lot of people really disappointed as they deploy more of GPT.

6:03If you're just a regular user on the chat, it probably doesn't make a big difference to you because you're not going to see big differences when you use the chat and the cost of the API doesn't make any difference to you. But for anybody who uses it through the API, probably just won't use it. And that's going to be, I think, a failure launch from OpenAI. Now, the good news from OpenAI this week is that they've expanded the reach of deep research. So deep research is incredibly powerful, and it's now available for multiple different platforms. And it was initially released only for people with a$200 a month licensing.

6:36And now it is available to any paying members with a limit of 10 deep searches, queries per month. So that's not a lot, but it's definitely better than zero. And I've already tested it several times, and it works extremely well. So if you're doing any kind of research and trying to find information across the net and summarize it in a productive way, then this is going to be a very helpful tool for you if you have the paid version. If you have the$200 a month version, you will get 120 queries instead of 100 moving forward. So that's good news from OpenAI. Now, I think they're doing this, A, because they promised to eventually release it to everybody with some limitations, but they're facing significant competition across the board because in addition to OpenAI, Google has the same tool in their regular pricing model.

7:18DeepSeek has it for free. U.com have it for free for a very long time and they're probably the best free tool out there. Grok has an incredibly capable that I actually really like using, deep search versus deep research, but it does the same thing, capability. And right now it's free, but it is going to cost money whenever they decide to flip the switch. And Perplexity has a free one that also does something similar. But from my personal testing, while it's faster, it also has a lot more hallucinations than the others. And so there's a lot of competition and it makes perfect sense why OpenAI releases it to a broader audience.

7:51Now, we talked a lot about the release of Grok 3 and it jumping and taking over the first spot in the chatbot arena, but Grok 3 has promised to release voice mode in the next few weeks. Well, it took a few days and they released voice mode. Now, live voice mode in Grok is very different than live voice mode in all the others. If you remember, OpenAI did not release their live voice mode for about six months from the first time they announced it because they couldn't control exactly what it was saying and they wanted to align it first. It does not seem to be the case with X. So they have released the Grok voice mode with multiple personality options that you can select from.

8:28And they are deliberately defying the conventional AI assistant behavior of trying to keep it very professional. It has an unhinged mode that can basically say or do anything you want and it can escalate to screaming and it's been insulting users all over the internet. and it's even making weird, strange, long horror movie type of sounds. So not being very friendly, but it's a choice, right? You can choose that mode. They also have a sexy mode. And if you go and search the internet for Grok sexy mode, you will see that there's very little or maybe no censorship whatsoever on what sexy mode will say.

9:03They also have an unlicensed therapist mode. So all of these modes are available in this model. Now, this is not surprising when this is a company run by Elon. If you remember, Elon bought X, the Twitter company, in order to open it as the tool for free speech with very little or no censorship. And this just makes sense when it aligns with that. This by itself is not a problem, depending on what you believe. But the fact that people were able to get very detailed instructions from Grok on how to build biological and chemical weapons, including a detailed shopping list and the process on how to manufacture them, That's a whole different story.

9:40So what seems to be happening, while on one hand, Grok was able to jump to the front of the line extremely fast compared to the time it took the other labs to get to that level of model, what it seems that X has skipped is more thorough alignment of the model, which is what controls it on what it's supposed and not supposed to say. And if you go back to the definitions that several different companies have out there, but the leading one is Anthropic, on what they won't release, roughly more or less at the top of the list, there is a model that will allow people to learn how to make biological and chemical and atomic weapons.

10:14And this model did this now. Since then, it was presumably fixed by the X team. But the fact that they released a model that can do that shows you that they're not investing a lot of time, if any at all, in alignment of the model. They're just investing in running faster than anybody else, which then saves them a lot of time and money in the process, which allows them to invest more time in the development of the model. Is that good or bad? I don't know. What will that do to the censorship and the due diligence that other labs are doing in their process? I don't know either. Would it push them to release more models faster?

10:45Well, we kind of already see it did. It seems that that's what's going to happen because, as I mentioned earlier, I think that if OpenAI had a choice, they would not release GPT 4.5 because it does not provide a significant benefit and it costs a lot more. Now, in this particular case, it's not a security risk, but it's definitely showing that pressure to move forward faster is pushing these companies to release models, even if they're not 100 % ready. Speaking of Grok, the release of Grok has driven X.ai's usage through the roof. So the mobile app downloads pushed Grok's app to number one. With 10x the downloads of previous week, active users surge in the US to 260 % week over week and 5x in the rest of the world, web traffic went up from 189 ,000 people to 900 ,000 people in the US and worldwide visits went up from 627 ,000 to 4.5 million.

11:41So a huge spike in Grok usage. I am personally one of these people that did not use Grok at all. When it came out, I'm not a heavy X user. I'm there just to read news about AI and follow different people, but I'm not very active on X. And so it wasn't relevant to me. And the model previously was more of an X sidekick than it is a model on its own. But Grok 3 is actually really powerful. I started using it across multiple things that I do, and I find it to be very, very good. Now, right now I get it for free. Will I decide to pay for it once it starts being a paid version? I'm not 100 % sure, but while I have the chance to test it right now for free, I can tell you that I already moved some of the tasks that I used to do on perplexity, as well as some of the tasks that I used to do on ChatGPT to Grok.

12:25And I can tell you that right now, I'm very happy with Grok's deep search, which is deep research in all the other platforms. And I may continue using that if I only have 10 of those a month on the paid ChatGPT version. Now, in addition, Alibaba dropped another bomb when it comes to their models. So on February 26, with all within this same week, Alibaba has made four of its advanced AI video generation models freely available to users around the world. The new models are called 1WAN 2.1, and it's extremely powerful. It seems to be almost as powerful as Google Veo and better than all the other models that are out there right now.

13:02And it's available anywhere around the world, including on Hugging Face, and you can start testing it and playing with it immediately. In addition, they released a lot of their models, the regular models, not video generation ones, as open source, completely with weights and everything. So that was a very strong move by Alibaba to the open source market, as well as releasing a very powerful video model. And that drove their stock 5 % up this week with an impressive 66 % stock rise in 2025 alone, mostly being pushed by their AI involvement and the level of models that they're releasing in China that has limitations on getting access to some of the models that we have here in the West.

13:41Staying on interesting big releases, Google has released VO2 on two different platforms. So their most advanced video platform that from everything I see is now the best in the world is now accessible to users on two different platforms. So you can get access to it on the Freepeak platform. So that's free, F-R-E-E, and then P-I-K, and you can get access through that. Or if you're a little more geeky and you like to use it through the API or through a tool that allows you to use it through the API. You can use it on fal.ai as well. On fal.ai, it just runs through the API and you're just paying for tokens.

14:13On Freepeak, you're just paying Freepeak. But if you do that, you have access to the most advanced video model in the world today. They have not released it on any Google platform yet, other than for initial testing. But people on Android Authority has found sections in the code that are showing controls for this video tool. So we need to assume it is going to be released as part of Gemini sometime soon. That is very exciting to me because I use Gemini every single day across all the Google universe and Gemini just a chatbot. So if we'll have the video generation capabilities, I will most likely jump into that tool because as I mentioned right now, it is the best in the world.

14:49If it's going to be integrated into everything that I'm doing, it is going to be fantastic. I told you last week that Google now allows you to create images with Imagine 3, which is now the most capable image generation tool across all the Google universe. So you can generate images in Google Docs, Google Sheets, Google Slides, Google Drive, everything Google, and obviously on the Gemini app as well. So if they're going to do the same thing with video, that's going to be extremely powerful and will push Google forward in their race for dominance in the AI universe. I expected Google to take this position about a year and a half ago when the madness started.

15:23And I said that because they have more of everything they need in order to be successful in this, whether it's compute, talent, access to data, distribution, all the things that are required, they own on their own. And they're proving that, yes, it took them a while and were a little disappointing in the beginning, but they're leading the race as far as I'm concerned right now. Now, DeepSeek, the company who may have fired the first shot in this new wave of models, are planning to potentially release DeepSeek R2, which is their next version after R1, which is the model that drove the world into a frenzy because it was a very powerful thinking model from China.

15:59The initial launch was supposed to be late May, and now they're talking about potentially doing it earlier. That being said, it seems that they're running into problems with capacity. So if you remember when R1 came out, we spoke about this, that DeepSeek is financed by its mothership company that is a hedge fund in China. That was good enough in order to develop DeepSeek, but it's not enough to run it with the amount of demand they're having right now, while at the same time training the next model. So it's the first time they might be searching for external investment. There are a lot of people who are willing to jump in, very big companies from China, including some government funds.

16:32So I don't think they have an issue raising the funds, but right now they're running short on cash to do the things that they're wanting to do. But to stay competitive, they're planning to release R2 before May. So that's going to be very interesting to watch and I will keep you updated once this comes out. So like the saying says, when it rains, it pours in one week. We got so much new stuff to play with. And I know this is confusing as hell to people. And that takes me to the next topic, which is AI training for executives. So yesterday, I spent most of the day training people in an exact ed program in the local business school.

17:05And there were 27 people in the room from like 15 different companies all across multiple industries and multiple states. So there were people from banking, manufacturing, small businesses, law firms, et cetera, you name it. There were people there in the room. And I've been doing this for a very long time. I've been delivering these workshops. Most of them are for specific companies that bring me in to train their company. But in many cases, it's things like that, like different workshops that provide training to different groups. I also do the AI business transformation course, which we talked about in the past in the show.

17:37And I want to share a few things on what's happening in that room. First of all, if you're feeling that you're left behind when you're listening to these kind of podcasts and you're seeing what people are doing, don't. And the reason I'm saying that still 70 to 80 % are still in the very early stages of figuring this out on their own and definitely as a business-wide initiative. And so if you're feeling left behind, but you're listening to this podcast and you're making steps in the right direction, you're doing the right thing. Just keep on going. The second thing, the importance of proper AI education versus just listening to podcasts or watching YouTube videos or just following people on different platforms.

18:15Doing is very different than seeing what other people are doing. And you need to find yourself workshop or somebody to lead a workshop in your company or somewhere that you can attend in order to actually show you and allow you to experiment with these different tools. And that shows up time and time again, even people who are saying they are advanced users when they come into the courses and workshops that I'm teaching. And there's always X percentage between 5 % and 15 % of people that are claiming that are learning lots and lots of new things once they start experimenting based on the things we're doing in the workshop.

18:45So what I'm telling you is get your leadership buy-in. And if you are in leadership position, you got to start moving in that direction of training yourself as well as your leadership team, and then eventually all your employees on specific use cases that are relevant to your company and your industry and learn how to implement them right now while doing a strategic process of evaluating how AI will shape your industry, your niche, and your environment, because it might be dramatically different than it is today. We had great examples in the conversations yesterday with some of the law firms right now make money through billable hours.

19:18That's going to shrink dramatically. A lot of it comes from paralegal work. Paralegal work will definitely shrink dramatically. If you use tools like deep research, you can do a lot of the paralegal work in one-tenth of the time it took before. If your income from paralegals shrinks by 90%, can you still run your business effectively? I don't know, but that's one of the questions that law firms need to ask themselves. And there are similar questions across multiple industries. Now, since I already mentioned law firms, I want to mention to you what happens when you don't do this kind of training to your people.

19:50So Morgan & Morgan, one of the largest law firms in the country, had several of his lawyers be sanctioned by a federal judge for citing nine out of 10 cases that they cited that were made up by AI, meaning those lawyers has went to trial citing cases that don't exist because they did their research with AI tools, not checking the facts and assuming it's going to be correct, not knowing that these tools can hallucinate and make stuff up. Now, this is not the first time it is happening. In my courses, I've been showing companies a case like this from early 2023 with an Air Canada case. So this is not the first time that this is happening.

20:31And if these lawyers would have been trained properly on how to use AI, they would have avoided this really big scandal. Now, lucky for them, they were not unbarred, but they just got a fine. But this really embarrassing incident just can be avoided very easily just by getting proper training for every person in every company. And while this is a legal case, similar things can happen in every single company in the world if you don't get your people properly trained on how to use AI, what to do and not to do, and to understand its limitations. And now to rapid fire news. Before we jump into the regular news from the leading companies, I found something really interesting in the news this week, that Princeton University in its engineering department, together with the Indian Institute of Technology researchers, revealed in December of 2024 that they allowed AI to design complex wireless chips.

21:19And they were able to do it in a few hours instead of weeks or months that it would take humans to design those chips. Now, what they're saying is that these chips have superior performance compared to the chips that are designed for 5G by humans. And the design is radically different than the way humans design chips. Now, when they started examining this deeper, usually humans have these templates on how they design chips. There's specific processes with defined structures when we build chips, but the AI took a different approach that looks randomly shaped that human cannot really understand.

21:56And yet in simulation, they haven't actually developed the chips, but in simulation, it performs significantly better. And what it seems to be doing is instead of following the patterns that we follow, the AI kind of reversed engineered the process, just trying to understand what are the needed inputs and outputs and how it should be set up in a process that will make it the most efficient. And as I mentioned, they design chips that look dramatically different than humans. Now, will that work in actual real life? I don't know if anybody knows, but what it's showing us that when you have access to huge amounts of data and you can think of a problem versus through the current limitations that we have and figure out any solution to find the best solution, AI will be able to do a better work than most of us, including in really advanced complex problems, such as designing new chips.

22:41Is that the future of computing? Most likely, but I don't know that for sure. Obviously, time will tell. Now let's jump to a lot of news from the leading labs in the world. We will start with OpenAI. So the first news is a personal news. Sam Altman became a dad. And so congratulations to Sam. Being a father is a big, big blessing, and I hope he finds lots of joy in his new son. But from that to news that actually have to do with AI, the first one is somewhat alarming, and yet I'm happy that it's happening. So OpenAI busts Chinese AI surveillance tool that is using ChatGPT in the backend in order to do this.

23:16So OpenAI has uncovered that GPT-powered surveillance tool built by a Chinese operation designed to sniff anti-Chinese posts in Western social media has been running on their backend platform. But the way they found it was actually almost by accident when one of the coders of that platform used OpenAI Tech Debug to review the surveillance system's code, right? So the peer review that OpenAI tools can do for coding is the one that allowed them to find them. This is not the first time, by the way, that OpenAI find Chinese campaigns using OpenAI platforms. This has happened in the past, and I'm sure it's going to happen in the future.

23:54What I'm glad is that OpenAI are actually actively looking for these kind of cases and blocking them. Now, in addition, a report on February 23rd, OpenAI has claimed that it's banning dozens of users' accounts in the world for using ChatGPT to facilitate scams and other malicious activities. So it is obviously easier and easier to do bad things with AI because it's advanced capabilities and because it's been proving that it's more convincing than most humans. So if you're thinking about phishing or any other scams, it's becoming easier to do if you are just using these AI platforms. And again, I'm extremely happy that OpenAI is looking for these kind of cases and that it's blocking the accounts that are using this technology in negative ways.