176 | ChatGPT’s Change Everything (Again) with a New Image Generator, AI Boosts productivity by 37%, Capabilities Doubling Every 7 Months, AI Doctors & Teachers Will Be as Common as Cellphones – and more top AI News Week of March 28, 2025

25:41They're delivering new products and they're in the process of converting from a nonprofit to a for-profit. So it makes sense that it's going to be reshuffling. It will be interesting to watch to see what impact does it have on the performance of the company. Another big piece of news from OpenAI and to the AI world in general is that OpenAI is adopting Anthropic's model context protocol called MCP, which is an open source standard that connects AI models and AI agents to data and tools. The protocol was released by Anthropic a few months ago and has become extremely popular with developers of agents.

26:17And now OpenAI is basically saying, this is a great model. We love it. And we're going to join and use the protocol as well. So as of now, it's already available as part of OpenAI's agent SDK, the desktop app, and the API support coming soon, per Sam Altman's tweet, which said, people love MCP and we're excited to add support across our products. Anthropic chief product officer, Mike Prager, obviously was very happy. And he said, excited to see that MCP love spread to OpenAI, welcome. What does that mean? It means that MCP's approach is successful. It is open source. It is open to everybody and it allows to create a unifying environment and solution for agent development.

26:54I think the more standardization we're going to see across this industry, the better because it will allow for collaboration. It will allow for companies to choose, switch and replace the agents that they're using because the underlying infrastructure and connectors between them are going to be standardized. So overall, I'm myself very excited about this as well. OpenAI are not the first company block Apollo, Replit, Codium, Sourcegraph, and other companies has already adopted MCP. So it seems that it's going to be the infrastructure to allow agents to connect with data, connect with themselves, and connect with tools moving forward.

27:28Now on a different topic from OpenAI, it seems that they are talking to several different providers to purchase billions of dollars worth of data storage, hardware, and software aiming to build its first ever data center for themselves. That's per the information article from March 26. Now, they're planning based on this article to purchase five exabytes of storage. That means absolutely nothing to me and probably to you, but this rivals Apple's iCloud capacity just a few years back. So this is a huge amount of data storage that OpenAI wants to purchase and own. And this will obviously work in tandem with their Stargate project, so their GPU project that they are working together with SoftBank, that is a$500 billion investment over the next few years to build data centers.

28:17So these two will probably work hand in hand. On a different topic, OpenAI, in collaboration with MIT Media Lab, has released a research exploring ChatGPT's effect on emotional well-being of its users based on its huge amount of, as I mentioned, 400 million people using it every single week. What they found is only a small fraction of users were emotionally connected to ChatGPT, despite the very large volume. Now, they've done two studies in parallel that were measuring different things and different emotional connection and impact on people. But what they did find is that users who trusted and bonded with Chachipiti were more prone to loneliness and dependence, hinting that there is a way to develop emotional dependency on these tools, which is something we all need to be aware of, especially those of us who have kids who are already potentially addicted to social media.

29:10this might be a lot worse if it spreads. And it's, again, something that we need to be aware of and at least educate people about. Now, for those of you who are enjoying deep research on ChatUPT, I'm definitely one of those people. As if you are a regular paid users who pay 20 bucks a month, you only have 10 searches in deep research a month. And if you're paying the$200 a month, you're limited to 120 deep research queries, but it was impossible to know where you are in the account that was solved. Right now, it creates a pop-up that tells you how many you have left every time you run it. But something even cooler, if you hover your cursor above the deep research button, it will tell you where you are on your current account and how many you have until the reset date and what the reset date is.

29:53So this is actually very helpful. Just hover your mouse over it and you will know how many deep research queries you have left. On a not so exciting news for Open AI. On March 23rd, Kai-Fu Lee, who's the ex-Google China head and the founder of ZeroOne.ai, which is a really successful AI startup, he's claiming that they're ditching their proprietary models, mostly ChatGPT, to adopt DeepSeek's open source tech. He's claiming that running DeepSeek's free and efficient models is costing him 2 % of what it cost him to run the OpenAI APIs in the backend to perform the same tasks that his company is performing.

30:30This is a huge push towards the really advanced capabilities of open source AI. And I think this will continue pushing the pricing of AI down, going back to what Bill Gates dubbed the age of free intelligence. Now we're going to shift gears and talk about contradicting signs of where the AI copyright battle is going. So the first one will stay with open AI. So on March 27, 2025, a federal judge in New York ruled that the New York Times and other newspapers can move forward with their copyright lawsuit against OpenAI and Microsoft. Those of you remember, there's multiple lawsuits against OpenAI and other AI-leading labs on using copyrighted material to train their chatbots.

31:11They claim that this is fair use, and obviously the owners of the copyrighted material think otherwise. And so the judge allows you to move forward, which may go to a jury trial. Now, as you know, a jury trial can go either way because the jury will decide what's going to be the fate of this lawsuit. That being said, that will end up most likely the Supreme Court either way. So whatever the jury decides will stand for a very short while, but there will be questioned in the Supreme Court and whatever they decide will be the way this moves forward. In the current political environment, it's pushing for AI innovation regardless of the costs and the consequences.

31:46And so I think in the current political system and with the way the Supreme Court is leaning since the previous Trump administration. I think the outcome will be in favor of AI innovation, but I might be wrong. But to tell you how confusing this is, Anthropic on March 25th actually got a really big win. And in a case in California, a federal judge rejected a bid by the music giants like Universal Music Group and others to stop the AI firm from using copyrighted song lyrics to train its chatbot Claude. So the judges were calling the claims vague and noting that they couldn't prove irreparable harm to their business in the lawsuit.

32:24Now, the interesting thing here is you have two states that are on the same side of the map, both are hardcore democratic states, and yet we found two opposite rulings by a court, which aligns with the call from OpenAI last week to the federal government to step in and define rules and regulations that will be on the federal level instead of allowing a state-level patchwork that is already in the making across multiple states and multiple regulations. How that evolves, I don't know, but it will be very interesting to track and I will keep you posted. Now, staying on Anthropic, they released a, I don't know if final, but a new update on their ability to look into the brain of Claude and see how it works, introducing new tools and methodologies on how to, if you want, read the mind of how AI works.

33:10So historically, AI was a black box. We know how it's getting trained and we know the outputs, but it was very, very hard to track how and why it's actually doing what it's doing. And Anthropic has been at work at trying to understand how the AI brain works. They now came up with a tool called Microscope that allows them to trace Claude internal steps and how it thinks and solves different problems. It's worth reading their release. It's written in non-technical terms, and it actually reveals that Claude thinks in very universal concepts before translating it into specific language. So it may think and analyze a question in several different languages like English, French, and Chinese to a simple question like what is the opposite of small, and only then it comes up with the language to describe it to the user.

33:55They also found that when given a simple math question, it actually computes and goes step by step, but when it gets a very complex mathematical question, it tries to wing it and if you want, bullshit the user with an answer without actually trying to solve the problem. The importance of this research, it goes way beyond that because the goal is the ability to see what AI is doing as it's doing it, to be able to prevent it from doing things we don't want it to do, either in the current scenario, but way more important when we hit AGI or ASI and beyond, where AI will be significantly smarter than us.

34:25Now, is that really doable at that point? I don't know, but I think it's a very, very important step by Anthropic in the right direction. And I really hope that these tools will also be able to run on other AI models and that other AI companies, such as OpenAI and open source companies, will adopt these tools in a similar way to the way they adopted MCP to allow us to see how these models work and stop them before they do something that we do not want them to do. Still, Anthropic joins forces with Databricks. Those of you who don't know, Databricks is one of the more successful enterprise data management platforms.

34:58and this collaboration will allow the 10 ,000 corporate customers of Databricks to run Claude's models, including the latest Claude 3.7, to get insights from their data. This is obviously the big promise of AI when it comes to corporations and then eventually smaller companies as well, is to be able to look across the entire company data and get insights that so far were either very hard or on the verge of impossible to get. And that will become obvious to anyone just by asking the questions or even just the AI by itself, suggesting insights based on what it's seeing. Anthropic also released version two of their economic index, which is their view on how people are using Claude for which tasks we shared with you when they shared the first one a few months ago.

35:42This one now has Claude 3.7 Sonnet in it that wasn't available when they released the previous survey. And it's very, very clear that Claude is the king of coding because it's now taking 37 % of usage of Claude 3.7 Sonnet. If you go to the review, you'll be able to see bar chart that shows the different tasks that people are using Claude for, and the first 10 or so tasks are computer related. Beyond that, the next things are education, science, and healthcare, but these are very far behind in volume compared to coding capabilities. I use Claude 3.7 for coding small things, and it works very well for me.

36:18I'm not a coder, so it's very hard for me to compare to other tools, but I did find the most amount of success building small applications with Claude. Now, from Anthropic to Google, as I mentioned earlier, Google launched Gemini 2.5 Pro Experimental, which is what they're claiming is their most intelligent AI yet, with built-in reasoning, and it jumped to the top of the LM arena, as I mentioned earlier. Now, if you remember a few weeks ago, we talked about humanity's last test, which is trying to come up with the hardest questions on earth to test AI's capabilities. So previously, OpenAI O3 Mini scored 14 % and DeepSeek R1 scored 8.6%.

36:57Well, Gemini 2.5 Pro Experimental now scores almost 20 % with 18.8. But beyond that, as I mentioned, it's on day-to-day tasks that people are evaluating it. It's scoring better than any other tool on the planet as of right now. It also has the same huge context window as the other Gemini tools. So right now, 1 million token context window soon to be switched to a 2 million context window, and it handles text, images, audio, video, and code, making it probably the most versatile tool out there. It is still the only tool that knows how to quote unquote, watch video and understand what's happening in the actual frames, not just listen to the audio track and can understand what people are saying.

37:36So if you need to analyze video, this is a very powerful capability that did not exist before this tool was released. In the demo by Demis Asabis, he showed significant improvement in coding, and they showed a few demos of creating mini web apps and interactive games with just a few prompts. It is currently available at the Google AI Studio, and if you're a paying user of Gemini, of the$20 a month, basically Gemini Advance, you have access to it as well. Staying on the topic of new models, as I mentioned, DeepSeek released a new version of its V3 model. This is dubbed V3-0324 for the date it was released.

38:12It has an MIT open license, so it's an open source model that you can take and use. And it has a significant upgrade compared to its predecessor, mostly in reasoning and code generation and user intent understanding. So it scores very high on all these things. As I mentioned, it's right now on several different benchmarks, the best AI coder out there, period. What does that mean? It means that there's two races that are very active and aggressive. One is the race between China and the US. The other is the race between open source and closed sourced AI. And it is very clear right now that both these races are very close and are going to be very close moving forward with both open source and China pushing models that are as good and in some cases better than the leading labs in the US.

38:57And if that's not enough from China, Alibaba just released QEN 2.5 Omni 7B, which is a 7 billion parameter multimodal AI model that is designed for end-to-end processing of text, images, audio, and video. Again, something that will compete with the US models across the board in a real true multimodal AI solution. Now, despite the fact it is a relatively small model with only 7 billion parameters, it rivals many other larger single modality models and multimodal models as well from the leading labs around the world. And like its previous models, this is also an open source model that everybody can go and use and it's available on both their website as well as other solutions like Hugging Face.

39:39Continuing on this path of this insane week as far as releases, Mistral just launched Mistral 3.1. 3.1 has 24 billion parameters and it is scoring better than some of the other open source models like Google Gamma 3 and even GPT 4.0 mini across several different benchmarks. It supports 21 languages, including European and East Asian languages. It's still lacking on the Middle Eastern side of capabilities when it comes to languages, and it has 128 ,000 tokens context window. And it's an open source under Apache 2 license, which means it's free to tweak, deploy, and allows developers to do basically whatever they want with it.

40:19So what does that mean? It means there are more and more capabilities that are available as open source with very powerful tools coming from everywhere around the world for free or almost free. And they're going to keep on pushing the pricing of AI down while the capabilities are going to keep on moving upwards. Staying on the open source topic, on March 19th, Hugging Face submitted a open source AI blueprint to the White House AI action plan. And what they're claiming is that open source models can outperform closed source models. One of the examples they gave is Olympic Coder, which is a 7 billion parameter coding parameter, outcodes Claude 3.7, which was considered the top-of-the-line coder until that point, again, now was taken by the new DeepSeq model.

41:02Now, Hugging Face is definitely a powerhouse when it comes to open source. They're hosting 1.5 million public models that are available for everybody to use and co-develop, and they're claiming that these models can compete with a closed-sourced one at a fraction of the cost. The strategy that they're suggesting to the White House is a combination of three different components, collaborative innovation, resource efficient models for smaller players, and transparent security. And basically what they're saying is they're saying that instead of following OpenAI's suggestion to the White House to be a light touch on regulation, they're claiming the open source world will provide a much stronger and safer foundation because everybody can see what is actually going on with these tools.

41:46Now, I must admit, I'm not sure which side of the fence I'm on. I don't think I know enough to make a fair judgment call. I definitely see the logic in arguments of both sides. One that's claiming that giving it to anyone to change and manipulate is scary. But on the other hand, I'm also thinking that not knowing what OpenAI and Cloud and so on are doing, and definitely similar companies from China and other places around the world, versus allowing everybody to see what's happening in the code and in the systems and in the weights and so on also makes sense to me. I don't think this is going to get resolved.

42:19I think we're going to have both universes running in parallel, just like we had with a lot of other tools like operating systems. Now, how will that impact the Trump administration team that pushes innovation and very little regulation? It will be interesting to follow that as well. Now, a few quick updates from the robotics world. So a new company from California called Dexterity just introduced a new robot called METCH. And it's an interesting combination of old school robotics and humanoid robotics. So it's a moving platform versus a humanoid robot, but it has arms that are mimicking human arms just with more capabilities.

42:53And these arms can lift 65 pounds each. So it's making it the robot that now can lift the most amount of thing and move around a factory. Their first use case is to manage packages on different supply chain lines and put it on trucks or other conveyor beds and so on. And it's very efficient at doing that in the demo that they were showing. It can lift high weights while still doing it very accurately. Now, the interesting thing is they're claiming that one worker can oversee up to 10 Mechs robots at the same time, which means from a labor perspective, it's not just replacing the human workers who are moving the packages around, but also the amount of managers required to do that.

43:34Their initial use case, as I mentioned, is truckloading, but they're definitely looking to scale for solution in the general logistics world. Staying on the US side of robotics, on March 25th, Figure, which is one of the leading humanoid robotics companies in the world, have released a video showing that their robot Figure 2 can walk like humans. So it doesn't have that robotics walk anymore. And the way they achieve that is through a physics simulator that allows AI to get reinforcement learning of years within just a few hours just by running its brain, if you want, in the sim. And now it is working a lot more like a human with heel strides, toe-offs, and arm swinging, which makes it look very human.

44:18The interesting thing here is not the fact that it works more naturally or what seems more naturally to us, but the fact that you can train these robots in a simulated environment to do tasks that previously took months and years to do, and now it can do it in a very short amount of time. And that means that they will be able to mimic literally any activity that we want and learn it very quickly. Moving from the US to China, so we mentioned many times before, China has some of the most advanced robots in the world. So Agibot, which is one of the Chinese successful startups, are planning to produce 3 ,000 to 5 ,000 humanoid robots in 2025.

44:51So this year, this is a jump from only 1 ,000 units in 2024, so a 3 to 5x increase. They are releasing three different models, Adjibot Link C2, Genie Operator 01, and Yangzang A2. The Yangzang A2 is very interesting because it has very delicate capabilities and precision. And one of the things they demoed is the ability to thread a needle by the robot, which requires obviously very gentle dexterity and understanding what's happening in its surrounding. So different robots for different scales, and they're scaling up their production. And the quote that they released is, we aim to deploy new products in industrial scenarios this year, replacing humans in specific tasks to deliver tangible customer value.

45:34A, I agree. It will provide more value because the products that they handle will be able to be cheaper, at least from a supply chain and operation perspective. But the question again is what will happen to the people who held these positions before? And if they don't get paid and don't have money, who will buy the products that the robots will pack and ship and do everything with. I don't think anybody has answers to that, but we're going in that direction faster and faster. And another interesting robotics thing that happened in China this week, Chinese influencer Zhang Zhenyuan rented a Unitree G1, which is their smaller robot that we talked about a lot in the past few weeks because it learned how to do ninja moves and kicks and do different cool things.

46:12Well, you can now rent them in China for about$1 ,400 a day. And he rented it and documented the whole thing and showed the world how it was doing kicking, cleaning, and even attempting dancing. It was cooking eggs, sweeping floors, and even running alongside Zeng in the park. It couldn't do the dance move, but it still was very funny to watch. So you can go and watch that video online. But the whole concept that you can now rent these robots because people bought them because their price is now$14 ,000. You can buy a robot for$40 ,000 and it created a whole new market of robot rentals. So think about you want to throw a party or you need to clean your garage or you need to do any other big project.

46:49And And instead of buying the robot, you can rent it for the day to help you with that particular task. Now, going back to what I said before, until I know for sure that these robots won't destroy my house or more importantly, hurt my kids, there is no way any one of these is coming into my house. But as soon as these safety measures are in place, for a$14 ,000 amount, being able to do all the chores in the house and around the house, plus some other things that we can't even think of, it's a no-brainer, even at today's cost, which will dramatically drop as these things scale up their production.

47:20Switching gears to a different topic, Rev, I'm not sure how to pronounce it, I must admit, is a new company in the field of image generation. They just released their first image generator, codenamed Halfmoon, and it is running in a different business model than most of the other models. You're just paying for tokens and it's one cent per image. So you're paying$5 to generate 500 images, which is cheaper than most of the tools out there today.

47:49So think about a monthly plan in Me Journey starts at$8 a month. Here you can pay $5 and until you run out of credits, you can keep on using it. Most people probably are not going to generate 500 images per month, which is actually really good. In head-to-head tests by independent evaluators, it actually was as good and better than Me Journey Flux and Ideogram across multiple types and styles of image generation. So going back to what we started with ChatGPT and Gemini generating amazing images as part of their current payment plan, and now with other quote-unquote professional tools that do it significantly cheaper, this will also drive the cost down of image generation across the board between open source and closed source models.

48:31Now the cool thing is you can try it for free. So you get 100 initial free credits and then 20 daily freebies. That should be enough for most users unless you're a professional image generator. I said that about a year and a half ago. I do not see a future in which image databases survive. There's absolutely no reason for them to exist anymore because you can generate any image you want in any style that fits exactly what you need in seconds. Why would you ever spend time searching an image database for an output? So the whole mess that came out last week with Google Gemini that can eliminate watermarks from existing images on platforms like Getty Images is completely irrelevant from my perspective because I haven't downloaded or searched an image on the internet for the last year and a half.

49:16And that was before the tools were as good as they are right now. That's another industry that is going to be probably wiped off the face of the earth as more and more people understand how to use AI image generators. And the fact that they're now built into the actual chatbots is going to make that a lot more widely available. Staying on the creator's universe on March 18th, the US Federal Appeal Court declared AI-generated music uncopyrightable, reinforcing its initial stating of that situation and aligning it with the Copyright Act that covers music from 1976. And it aligns with everything else that the Copyright Office has said.

49:50If humans are not heavily involved in creating something, it's not copyrightable. I think that's actually a very good thing as a musician myself and somebody who loves to play music, these tools can generate hundreds of thousands of new songs every single day. Even if very few of them become successful, it will wipe out any human generation of music very, very quickly as we will drown with music generated by AI and then musicians will not be able to make any living. Now you're asking, what does one have to do with the other? People can still create music and people will love it. Yes, but there's no way to monetize it.

50:22So if there's no way to monetize it, there's no reason for a lot of people to play it. And then there's no reason for this industry to exist, which means the human side of things will still be able to make money, maybe challenged a little bit, but not wiped out by AI-generated music. Staying on the topic of creation of things with AI, two very interesting releases when it comes to voice models. The first one is Maya, which is a model that was released by a company called Sesame. Maya is a very punchy, human-like, edgy, if you want, voice assistant, and it has blown up in the internet with really cool, crazy examples of how human-like and cheerful it sounds compared to the relatively dull other voice models that sound too robotic.

51:04They have released their model as open source for anybody to use. So you can go to Hugging Face or GitHub and get the code for Maya and run it as your own and get a very human-like voice used in any application that you want. In addition, on March 26, Grok with a Q, so Grok, the AI infrastructure inference company, has released that they are teaming up with Play AI to release what they call Dialog, which is a text-to-speech model that runs 10 times faster than real-time speech, meaning its ability to synthesize and create speech is significantly faster than the speech itself. They're claiming that it's very, very flexible, and it uses full conversational history to nail rhythm, tone, emotion, and make voices sound very natural.

51:46So it's currently available on Grok's high-speed inference platform, making very little mistakes, sounding very real, And it has English and Arabic as the languages that run on it. So far, the Middle Eastern languages, Arabic and Hebrew, were usually not available on any of these voice platforms. And Arabic is the fourth most spoken language on the planet. So it's obviously a big deal. So if you are an Arabic user or you have Arabic clients and you want to use an AI voice agent, you can now do this on this new platform. What does that mean? It means that we will see more and more voice agents working across multiple aspects of industries.

52:20I assume the customer service world is going to be the first one taken by a storm, but later on, it might be salespeople and a lot of other customer or internal facing capabilities. You just have conversations on any topic. Think about everything that we talked about before with agents becoming a part of the workforce. You'll be able to talk in natural language with an AI agent that is not human and not know the difference between that and a human employee on a conference call. Does that make sense from a productivity perspective? Yes. Does that connect very well to what we started with as the research from Harvard and Ethan Mollick?

52:52Yes. Is this going to be a very weird future? Absolutely. Now, we're talking a lot about moving forward to AGI and where we are in that process with all the different advanced capabilities that we're seeing almost on a daily basis right now. We talked about the ARK Foundation previously, where ARK AGI was the benchmark that was blasted by the leading models right now. Well, they just came out with Arc AGI 2, and they came out with it on March 24th. And the goal is to push the limits of what AI can do from a reasoning perspective. So similar to Arc AGI 1, it is a visual pattern puzzles that are geared to challenge AI in solving them.

53:31And it's very interesting to see the results with over 400 humans averaging 60 % success rate on this test, while DeepSix R1 varies between 1 % to 1.3 % accuracy. And GPT 4.5 and CLOD 3.7 Sonnet are around 1 % as well. So significantly far behind humans, showing that throwing more compute and even reasoning capabilities has its limits, at least right now, compared to the human brain's agility and understanding of specific problem solving. The ARC is also offering a prize called the ARC Prize 2025, and the prize is going to go to the first company or organization that's going to hit 85 % accuracy on this test, but only for 42 cents per task.

54:13So what they're claiming is that it's not just about being able to solve it, but also being able to solve it efficiently, which will actually make sense from a financial investment perspective. The last piece of news that we are going to talk about is actually an interesting trivia. So on March 20th, the Computer History Museum, together with Google, has dropped the original 2012 AlexNet source code on GitHub. This was the first neural network that actually worked because it took the error rates from 25 % to about 15%. And it was basically what started the current AI era that we know. So this is very, very cool.

54:53If you want to run the very first neural network that worked properly, that is actually really small and can run on a single computer with the right GPU, such as GTX 580, and you can run it in your house. If you're really geeky and you really want to play with it, it is now available. That's it for today. We'll be back on Tuesday showing you exactly how to create amazing videos, including all the different steps between research of what you need in the video to scripting the video, to creating the video, to creating the audio of the people in the video, to actually rendering the video and editing the video, all with AI, all within less than an hour.

55:29So if you need to create videos, and most businesses do, this is going to be a fascinating episode. We're also opening a spring cohort of our AI Business Transformation course. So if you have not had AI efficient and effective business-related training and you're looking for one, check out the course in the link in the show notes, share this podcast with other people who can benefit from it, and have an amazing rest of the day.