178 | AI is storming into higher education, Vibe Everything (not just coding), How to Learn AI, AI finds diamonds in Minecraft, and more AI news you need to know for the week ending on April 4, 2025

25:05Now, Musk announced it in a post on X, suggesting that XAI's value is$80 billion and X, formerly Twitter being$33 billion for a combined value of$113 billion. This makes sense from several different perspectives and doesn't make sense from another perspective. So the reason it makes sense is this deal allows Elon himself and his investors in X in the$44 billion acquisition he did that he didn't actually want to do and was forced into doing. It allows them an out and get value in something significantly bigger, piecing everybody and getting rid of potentially a lot of debt and a lot of risk. The other reason is that these companies work very closely together anyway.

25:45XAI is getting fed from X, formerly Twitter, as its main data source, I assume, and it's providing a lot of AI capabilities into the platform for people to be able to search and engage and understand people's tweets. And so that makes sense from that perspective as well. The reason it makes absolutely no sense is the actual valuation, because it values XAI, the creator of Grok, which I absolutely love. I use Grok all the time and I think it's a great tool, but I believe that XAI has really low revenue, if any, right now, and it's valuing them at$80 billion compared to, let's say, Anthropix,$60 billion valuation when they are currently generating$120 million every month.

26:28So what does that mean? It means that Elon was able to clean the cap table, merge two companies that he's managing anyway, get them into one unified environment from both a financial perspective, and it will allow him and these companies to merge their assets as well as human capital to do more with all the assets that they have. So overall, I think this makes sense from multiple aspects, and it will be very interesting to see how they do in the overall race. From XAI to Microsoft on March 27, Microsoft revealed two new reasoning agents for Microsoft 365 Copilot. One is called Researcher and the other is called Analyst.

27:03They're basically a wrapper around the existing ChatGPT tools. So researcher is basically ChatGPT deep research, but with a very interesting direction, it allows you to search through your emails and meetings and web data, as well as other internal data, which I think is brilliant. And it's finally starting to go in the direction that we were all hoping it's going to go in 2024, where AI will allow you to actually see what's happening across your organization. I don't think it's there yet, but I think it's a big step in the right direction. And the analyst tool is basically ChatGPT's O3 mini model, allowing to analyze data across multiple data sets that you're going to give it.

27:39So none of this is new, but the implementation into the Microsoft environment is new. And I think it's needed. And hopefully, Microsoft is not going to screw it up like they have with Copilot so far. I have nothing against Microsoft. I just think Copilot so far has been a poor shadow of the raw tools from ChatGPT that it is based on. And I don't think I've ever saw it outperforming just ChatGPT as is. So it will be interesting to see how this works. As I mentioned, combining it with internal sources is a huge benefit. Another interesting piece of news for Microsoft is that they're stopping or delaying their data center project worldwide, scrapping plans in the UK, Australia, Indonesia, and US states like North Dakota, Wisconsin, and Illinois.

28:19And that's according to TechCrunch and Bloomberg. and they are claiming that this shift is thinking that they're over building infrastructure. We talked about this when we covered an interview with Satya a few months ago where he was stating that there might be overbuilt of data centers and they're thinking that leasing that data center space might be more cost-effective than building it themselves. So it will be interesting to see whether their new bet, their new direction is the right one or not. Going from the US to competition from China. So Alibaba, who recently released their latest QEN model, now released a new subset of that called QEN QVQ Max, which is a visual reasoning model that can see, understand, and think.

29:00So think about a combination of a model that can see and understand images in the world combined with reasoning capabilities, images, and videos that help it solve real world problems. They're claiming that it can solve things such as math and physics and other visuals such as flowcharts and diagrams and so on, which will align it with similar capabilities from Chachupiti and leading models from the West. And it adds a new thinking button that can combine with uploading visual information to allow the model to reason step-by-step on what it's actually seeing. Now, they're also saying this is just the first iteration of bigger upgrades that are ahead, and they're planning handling multi-step tasks and complex problems in more of an agentic approach, as well as allow it to operate on computers and cell phones and play games and generate images, basically trying to align and stay in the competition with the leading models in the world.

29:51Alibaba, very similar to Google, has everything it needs from data distribution, resources, capital, human capital to keep on being a really important participant in this AI race. And now to our weekly robotics section. So agility robotics by ex-Microsoft executive Peggy Johnson just captured$400 million investment at a$1.75 billion pre-money valuation. And that was released by the information on April 1st. Their robot called Digit is already working on some Amazon warehouses doing several different activities. Now, leading this investment was WP Global, but you will also find names there. Not surprising, SoftBank is one of the investors, and that's following$150 million they raised in 2022, and a total of$320 million they raised to date before this latest$400 million raise.

30:39Now, the robot is five feet, nine inches tall, and it's priced around$100 ,000. And there's already about 100 units that is out there performing different tasks. And they're claiming covering for labor shortages, whether that's really all it's doing or actually replacing employees. It's very hard to know, but the trajectory moving forward is very clear. It will replace warehouse employees. Now, they're just recently released a new version that has longer battery life and autonomous charging and better grasping of equipment and so on. And this new raise will allow them to push the limits even further.

31:11Staying in the robotic fields, one of the companies that's been pushing forward very aggressively in this past year is China's Unitree. We talked about the robot G1 several times. It's the one that did the Kung Fu moves and did different flips and the first to do back flips and the first to do side flips and all these really cool things. But despite all that hype, they're delaying their household robot release due to, and I'm quoting, stringent safety requirements. That makes perfect sense to me. I said that in almost every show that we talked about robots. Before I know for 100%, not 99.9, that this thing is safe and that it's not going to harm my kids because it's going to go crazy or think it's a good idea for whatever reason, none of these robots is getting into my house.

31:53Despite the fact it's really tempting because the G1 robot that is about four feet tall is priced between$16 ,000 and$20 ,000. So a very reasonable number compared to its bigger brother, as an example, coming from Unitree called H1 that costs$90 ,000. And we just talked about$100 ,000 price tag from a different company. So for$16 ,000, you can get a small robot that can do a lot of chores inside and around the house. But again, I think safety here has to be at a completely different level than it can be in a factory where it can work in an isolated area without human employees around it. Staying on the integration of robotics into factories, Audi's China facility, Audi FAW, just announced that it's going to be the first that is going to start using robots.

32:34They're using UB-TEX Walker S1 humanoid robot for quality inspections. The first thing that it's doing is actually really interesting, and it's detecting refrigerant leaks in the air cooling system, a task that was previously done by humans, but it's posing respiratory risk because these gases are toxic. So allowing this robot to check with its very high level of accuracy of measuring devices, as well as very quick ability to see and detect things while not risking humans is a great move in the right direction. These are exactly the tasks you do not want humans to do, and there was just no other better choice right now.

33:08So I think starting with hazardous, risky operations in factories are a great way to start using humanoid robots. But as I mentioned, as this thing evolved, it's not going to stop there. UBitech is planning producing 500 to 1 ,000 of these robots by the end of this year. And beyond Audi, they already have started deploying them at BYD and Link & Co. BYD is stating 120 % improvement in sorting efficiency and Link & Co. is claiming 40 % faster warehouse processing and 65 % labor cost cuts using the S model that kind of tells you where this is going. Faster, better, cheaper labor at factories that will eliminate jobs from blue collar workers.

33:53That's not a good thing from a global workforce perspective or from a society perspective or from an economical perspective other than for the companies themselves. But if the companies will not be able to sell anything because people will not have jobs and will not be able to buy anything, then there's no point in saving money in production. I don't know where this future is going, but I think we're looking at some seriously troubling times because of the trajectory and how fast this is moving right now. Switching from robots to video generation and another staggering raise, Runway AI, which is one of the most advanced and one of the longest running AI video and image generation tools on the planet, just raised$308 million in funding, valuing them at over$3 billion in valuation, doubling its worth from the previous round.

34:39General Atlantic was the leading investor, but you will find again, SoftBank and NVIDIA and Fidelity all are investing in this tool. So you see SoftBank are more or less betting on every leading company in the AI race. Now they just recently released Gen 4, which solves maybe the biggest problem that AI video generation tools had, which is the ability to keep consistent characters across scenes and generations. And this is a dramatic change in the ability to generate real movies, videos, and series with that. And if you remember back in September, they signed a deal with Lionsgate to help them create video generations to support their actual work.

35:15So with this combination of things, the ability to generate consistent characters, as well as partnerships with actual studios, shows you that the direction in this field is clear as well. We will need less actors, less camera people, sound people, and so on, and more and more work will be able to be generated by AI. I assume in the beginning will just complement the traditional way of doing things and over time we'll just replace it and I mentioned that in 2024 I think in 2025 we will see the first series or the first movies that will be released done completely by AI that will be provided to an audience that will be wanting to see it and will not care or will actually be excited by the fact it's 100 % AI generated I still think this is coming and I'm just waiting for the creator to actually make it happen.

35:55Notebook.lm, which is one of my favorite tools, just released another really cool capability, which is called Discover Sources. So far, Notebook.lm, you had to give it the sources that you wanted it to summarize or create into podcasts, etc. This could have been files that you have or links to different websites. And now you can ask it to find relevant sources for you. It's called Discover Sources. There's a button. You tell it what you want to know about and you click the button and it gives you up to 10 relevant sources. And you can pick with checkboxes which ones you want to include in your summaries and which one you don't.

36:25This is already available for me. I don't know if for everybody, but they said that they're running it out for everybody in April. Now, they also added another functionality, which I don't really get. It's called I'm feeling curious. And it basically allows you to find information on random topics that you don't really know anything about. And this is more of if you're just bored and instead of playing a game or watching a series, you want to learn something new. I guess you can click that button and it will find you several different sources on a random topic. And then you can pick and learn about it through the platform.

36:52So I found this functionality not very helpful, but the first one is actually really great. Another really interesting feature is actually coming from Amazon. Amazon is now testing a new AI shopping agent called Buy For Me, which allows users of the Amazon app to search third-party websites for stuff that is not available on the Amazon platform. So think about you're going to Amazon, you're searching for something, and instead of showing you other options, it's showing you the actual thing, but not coming from Amazon. The other cool functionality is in addition to finding it in several different sources and showing you the pricing, since it has your credit card information saved on Amazon, it will actually do the checkout for you once you pick the one you want to purchase.

37:29Now, they're encrypting your billing information and obviously your credit card information, so it's perfectly safe to use per them. The only difference between shopping through the Amazon platform and using this tool is that if you need to return the thing that you're buying, you have to return it through the third-party platform and not through Amazon. But I see that as a brilliant move from Amazon to get even a bigger market share than their complete dominance that they have right now. But B, I think it's a very smart use of AI agents that allows you to shop through multiple websites from a platform you're doing most of your shopping.

38:00Anyway, this is a very clear move by Amazon to block or at least compete with some of the new functionality of shopping on ChatGPT that was released earlier this year. In addition, Amazon released a very interesting new capability called Amazon Nova Act SDK, which is a toolkit that lets developers build AI agents capable of tackling over the web browser's tasks. So basically take over your browser and do multiple actions. And they release that to the public, not just on Bedrock, which is how they released our previous models. So it's an SDK that's available to anybody, whether you're using Bedrock on the AWS platform or not.

38:36This is, again, I'm sure comes to keep them relevant in this new frontier of computer control and browser control that is intensifying seams every single day. You can get access to it through nova.amazon.com if you want to test it out. And it's just available with text images and video generation capabilities, all available, not just, as I mentioned, on the Bedrock platform. Staying on the topic of agents, Augment's Code just released an AI agent solution that wins 70 % of the time over GitHub Copilot Enterprise, basically allowing it to generate better code per developers 70 % of the time compared to one of the leading code generation capabilities.

39:17This new tool integrates Anthropic Cloud 3.7 together with OpenAI 01 reasoning model, hitting very high results on various programming benchmarks. And this is just another tool in the AI coding wars that is now available to people to use. The other interesting thing in this tool is that it has a$200 ,000 context window, which allows it to handle very large pieces of code. Now, it also is SOC compliant from a safety perspective, and it integrates with tools like GitHub and Jira, allowing people to just work in their current environment and not taking any extra risks while using it. And maybe the most interesting and the most exciting and the most scary announcement in the agent world is on April 1st, Emergence AI unveiled a new platform that auto-generates agents in real time based on user-defined goals, basically eliminating the need to code agents.

40:08Users can put in their tasks and their goals in natural language and the system will spin up the relevant specialized agents to do that task. Now the CEO Satya Nida explained it this way. When a new task comes in, the orchestrator figures out if it can solve it by checking the registry of existing agents, creating new ones if needed, and refining them over time. What does that mean? It means complete democratization of agent generation. So far, everybody hears about agents, but to create them, you needed to be a developer or have an army of developers to build them at scale. And now this is the first platform, and I'm sure others will follow.

40:43We just talked about Claude working on stuff like that that will allow you to just prompt it and tell it what you're trying to do and explain your goals and explain your limitations. And it will create the agents in real time to do the tasks that it needs to do. This is a very significant step in the acceleration, in the creation and adoption of agents. And I think this will dramatically reduce the time to when we're going to start seeing more and more agents in the actual workforce and in everything that we do. Now, something very smart that they've done, it's built with specific guardrails and human oversight verification for the things that the agents are doing.

41:14And their CEO specifically said, you need to verify that the new agents spawned are doing the tasks that you want. So think about it as new employees that you hire, you still need to check what they're doing. You can't trust it completely 100%, at least yet. And I think adding these checking steps in everything that you do with AI, and that's true from the basic things of having to write an email for you, but definitely once you start getting into a genetic world where these things could do more complex tasks with multi-steps and taking action, we need supervision. And I think what's going to happen over time, we'll become more and more supervisors of tasks than performers of tasks as humans.

41:48And from practical agents to an interesting study from Carnegie Mellon and Stanford, heard, they have found that actually overtraining large language models is reducing their performance. So beyond a certain size, overtraining an AI will actually reduce its performance. The study that was released on March 28th was showing that they increased the number of tokens from 2.3 trillion to 3 trillion tokens in the training and huge amounts of data. They actually got a reduction of 2 % in performance after fine-tuning the model, and it was harder to fine-tune. So they're claiming that this push to just get more and more data and more and more and more training to get better outcomes is not the right way forward.

42:29We heard rumors about this from a conceptual perspective at Q4 of last year. And this seems to be the first research, at least that I'm seeing, that's proving this scientifically. If you think about how long they've been training GPT-5 and the fact they haven't released it, even though they run at least two complete training runs to get it out there, each training run costing them conceptually half a billion dollars, and yet they wasn't released potentially because this exact fact that the bigger the models get once you get to these numbers of tokens it actually becomes harder to make them actually work as efficient as the previous model forget about getting better but that is pushing the new kinds of innovation with better and new types of algorithms and the reasoning models that we now so like in the stuff that are getting integrated into everything that is doing and distillation which is the concept that was used by deep seek to create their models as well as we don't know what's coming next, but I'm sure there's going to be new innovations that will be able to drive the scaling laws beyond just brute force.

43:23Regardless, we're more or less running out of new data because we use most of it training the models so far. Now, staying on the research side, on April 2nd, MIT Research revealed a new framework called LLMFP, which allows to boost large language models' ability to solve long, intricate planning problems, achieving 85 % success rate on benchmarks, which broke those benchmarks and achieved significantly higher results than anything before. What it basically does is it guides the LLM to mimic human problem solving by breaking the tasks into smaller steps and using optimizing algorithm to call specific steps and break them down further and further, and then combine them together into the final output.

44:02And it allows turning vague prompts into actionable detailed plans that can be executed. Now, to show you how powerful this is with the framework, it slashed the error rate by 50 % compared to unguided large language models. And on some benchmarks such as the blocks world, this new framework nailed 88 % of the task versus 33 % of the underlying large language models that was used without the framework. So again, talking about innovation and new things that are being developed regularly, this is a great promise for people who are not great prompters that basically shows that there's a path forward for just explaining in regular English what you're trying to do and having the AI understand it much better and be able to define for itself what it needs to do and achieve significantly better results.

44:45Now, speaking on interesting milestones, on April 3rd, Futurism reported that OpenAI ChatGPT has passed the Turing test. And the way they've done this is they had 300 participants run three-way chats with humans and AI. And at the end of each conversation, they had to guess which of the participants were AI and which were real. And ChatGPT 4.5 was picked as human 73 % of the time. This is really interesting because the humans were picked 66 % of the time as humans. Now, ChatGPT 4.5 dramatically outpaced other models in mimicking people, with Mera's Lama coming second at 56%, ChatGPT 4.0 at 21%, and they also used Eliza chatbot from the 1960s with 23%.

45:32So it's showing you that these models are getting better and better at mimicking humans, and now, as I said, sounding more human than actual humans. Now, another interesting fact here that is very important for all of you who are thinking or already building chatbots for different things, without defining a persona, without telling ChatGPT which role it's taking, its rate of success dropped to 36%. But when building a backstory like who you are, what's your background, what you've done before, what your knowledge is, which industry you're in, that jumped to the 73 % success rate. that tells you that when you're defining personas for large language models, you're actually going to get significantly better results and definitely better human-like results.

46:13Now, we didn't talk about Manus, I think, at all last week. So those of you who don't remember, a few weeks ago, Manus, which is a Chinese company, rolled out a viral agent platform that really took China by a storm that allows you to run multiple operations on your browser autonomously. And now they introduced two paid subscription plans that will allow you to do a lot more with this tool and will allow them to obviously pay for some of the cost of running these tools. There's two tiers, one$39 that gives you 3 ,900 credits to run two tasks simultaneously and also a$200, or if you want to be more specific,$199 per month tier that provides you with 19 ,900 credits and allows you to run five tasks simultaneously.

46:54They're also now introduced a mobile app that you can run this on your phone. This is just another interesting step in the direction of agents will do everything for us in our browsers and beyond in the near future. Now, there's been many tests, including by Tech Crunch recently, that showing it's not perfect and that it's making a lot of mistakes. But I think the direction is clear from even just what we talked about in the last 10, 15 minutes. And I love ending with a fun fact. Well, DeepMind's Dreamer, which is a new platform that they released, was able to find diamonds in Minecraft without being taught anything about the game.

47:28So the concept behind Dreamer is it's an AI that imagines future outcomes of what it is going to do and then make decisions based on that. So this way it was able to learn everything in Minecraft and be able to fly around and dig and do everything it needs to do and was able to find diamonds within just a few days from starting to play the game again without any prior training or instructions. What this shows is it shows AI's capability to tackle really complex open-ended tasks with zero hand holding and or training on that particular task versus pre-programming it or explaining it the rules of the game or in this particular case it's a very complex game right it's an entire universe with rules now i must admit i'm on the fence of where this is going on one hand this is a really cool as somebody who plays minecraft with his kids and seeing my kids play a lot more minecraft so seeing being able to see an ai basically learns it on its own and does very well in it very quickly is exciting.

48:24B, it shows its ability to do things way beyond playing Go or chess, right? So this is how it beat humans in previous games, but this is a much more complex game. So it's very interesting. That being said, despite the fact it's a very complex game, it's still a game with very limited rules compared to the real world. Does that translate directly to the AI's ability to understand the real world? I don't really know. I think it's a step in that direction, but it is showing the ability of the AI to learn on its own and develop tasks for its own. So combine that with everything we talked about this episode about the new agent capabilities, and you're walking into a very different world than we know today where the AI will not need us to give its prompt.

49:06It will understand on its own what it needs to do, defining its own tasks and executing on them. This could be on one hand really exciting, but also really scary. If you want to learn how to navigate this new future with AI and you want to be able to apply it for your career and your business, don't forget to check out the AI Business Transformation course. There's a link in the show notes. You can click on it right now and go and learn about our course. As I mentioned, it transformed hundreds and maybe thousands of businesses based on these eight hours of training. You can also join our community on our Friday AI Hangouts.

49:36There's a link for that in the show notes as well. We meet every Friday, 1 p.m. It's completely free, completely open to everybody, and we discuss real-world AI implications and implementation. So if you want to join us and learn more about AI, come and join us. And please recommend this podcast to other people who can learn from it. If there's one thing that's very clear that this thing is not stopping and that it's going to profoundly impact everything that we know. And the more people know about this, the better the outcome can be. So the role that you can take is share this podcast, click the share button right now and share it with several different people who can benefit from it, write them a short message.

50:08And while you're at it, it will be really nice of you if you also rank this podcast, give us whatever star rating you think we deserve and write a nice message or connect with me on LinkedIn and give me any feedback you want about this podcast. I regularly chat on LinkedIn with people who are podcast listeners who approach me. So thank you for all of you who are doing this. And if you want to chat with me, please do that as well and have an amazing rest of your weekend. I'll see you again.