AI News: Microsoft Finally Reveals Their Plan!

What does this button do?

Hmm.

I just got back from Microsoft Build out in San Francisco and there was a ton of announcements this week, not only from Microsoft Build, but also out of NVIDIA, out of OpenAI. There's just a lot of stuff. So let's jump straight into it. Let's start with the announcements that came out of Microsoft Build, starting with the fact that they announced 7 new models developed in-house at Microsoft AI. They built a new Thinking model, which is their new flagship reasoning model, and it's a pretty solid model. According to Microsoft, it's preferred over SONNET 4.6, which, as we all know, Anthropic's best model is actually Opus 4.8. So they're not quite comparing it to like the best of the best state-of-the-art models. They also introduced a new coding model, MAI-Code-1 Flash. This one they actually compared to Claude's Haiku 4.6. 4.5 model, showing that it's more accurate and uses quite a bit less tokens than Claude Haiku 4.5. But again, not quite the most state-of-the-art model yet. They mentioned MAI Image 2.5, which we actually played around with in last Friday's video, but they also announced a new ultra-efficient Flash variant. Now, when they announced it during Microsoft Build, it was the third best text-to-image model. It's actually since been bumped down to fourth, and it is still ranked number two in image editing. So when comes to image editing, it's a really, really solid model, only very slightly being beat out by GPT Image 2. Again, if you want to see what that image model is capable of, check out last Friday's news breakdown because we did test it out in that video. They also announced a new transcription model, MAI Transcribe 1.5, and it's currently the best transcription model in the world. It is currently the state of the art in accuracy and 5 times faster than competing models. When it comes to how fast it is, we can see in this chart here that it pretty much stands alone as the fastest model. And they announced MAI Voice 2, which is a speech generation model that works across 15 languages with a Flash version coming soon. So taking a look at MAI Voice 2 here, we can get a little sample of what this is capable of. I was just standing there, right? And then, oh my God, she actually said it to his face.

With everything on the line, the egret makes its move.

Slow through the shallows, watching. Waiting.

And it's a sudden strike.

Got it. You get the idea. Since this was announced, there's been even another voice model, but I'm going to talk about that one and demo that one later on in the rapid-fire section that you might like even better. Now, while I was at Microsoft Build, I had the opportunity to sit down with the CEO of Microsoft AI, Mustafa Suleyman, and ask him some questions about what they've been working on. And when it came to these new models, he was pretty clear with their goal. They want to release reliance on other AI providers, you know, like OpenAI, This is about taking first steps towards true self-sufficiency in AI.

And that means that we have to have the capacity to build our own models from scratch and show that we can achieve the absolute frontier.

Now, another thing that Microsoft seems to be trying to do is to really hammer home that they're a company that can be trusted and that the training data that they got to train these models, you can feel confident that it was ethically obtained.

Part of the challenge is that we don't always know exactly what's in other people's pre-trained sets. So we try to be as transparent as possible, knowing that, you know, our main job is to build trust so that people can come build on our platform. And that also extends to how we buy the data. Like, we have paid a great deal for this data. We've licensed it very carefully. We've been very deliberate about not using open-source datasets, you know, and we don't know what security bugs could be introduced through, you know, some of the more open-source datasets that maybe haven't gone through the same kind of rigor. So we want to give them trust and confidence that the model that we give them is absolutely clean from top to bottom and has, you know, been, you know, acquired in exactly the right way.

Microsoft is also getting into the agent game. They've been seeing the success of things like OpenClaw and Hermes, and well, they're like, yeah, us too, we're gonna give you agents also. So during Microsoft Build, they introduced Microsoft Scout, which is your always-on personal agent. This is part of their new category of agents that they're calling Autopilots. They work autonomously, they have their own identity, and they act on your behalf. Now if you're thinking it sounds like OpenClaw, that's because it is literally OpenClaw. It is powered by OpenClaw's open-source technology, except now Microsoft is giving OpenClaw direct access into like all of the Microsoft products and direct access to literally manage Windows for you. Like it works at the operating system level. It operates across cloud, desktop, and web, connecting to Teams, Outlook, OneDrive, and SharePoint, including chats, email, calendar, and contacts. Now Microsoft Scout isn't available to everybody yet. It's available to some early testers, but it sounds like it's going to be rolling out more widely soon. So if you're someone that's been really scared to set up OpenClaw on your computer, well, it sounds like if you have a Windows machine, Microsoft is going to make it A, a lot easier, and B, a lot less scary to actually operate within the confines of Windows. Once again, this was something I wanted to ask Mustafa about, and here's what he said.

Well, increasingly you're going to be step out of the CLI and just use your agent as an app. You can use those to organize your work on a day-to-day basis, track your health, run background tasks for you, find you the best price of things that you care about, you know. I mean, anything that you could— you would think to give to an assistant, or things that you would previously write code for, obviously an agent is now going to help you do.

During Microsoft Build, they also showed off a new GitHub Copilot app, which when they showed it off, I was actually pretty excited about it. If we take a look at this sort of demonstration video here, it pretty much looks identical to Codex, which right now is my favorite sort of IDE to do vibe coding in, right? However, if I pause this video here and you look at the bottom, there's one major difference about this new GitHub app that you don't get in Codex, and that's the option to choose whatever model you want from any model provider that's out there. So obviously when using OpenAI's Codex, you're gonna use one of the OpenAI models. When using this new GitHub Copilot app, you get seemingly like the same experience the experience that you get with Codex, but the ability to choose any of the models. So you can pick whatever the current state-of-the-art best model is, whatever the fastest, cheapest model is. You know, you take your pick based on what's available. So I got home and this was like the first thing I wanted to play with, cuz this excited me the most. I went and downloaded the Copilot app. You can see I have it downloaded here, except there's one problem. It says you need an active GitHub Copilot subscription to use the app. Okay, cool. I'll sign up for that cuz I want the Codex experience, but with all of the models available. I click over to sign up and well, for some reason, new plan signups are temporarily paused as we ensure a high quality experience. So I can't even sign up for GitHub Copilot right now. The thing that I was the most excited about from the keynote, the one thing I wanted to play with the most that I thought we would have access to is for some reason not available. Who knows though, by the time this video is live and you're seeing this, it might be available to sign up and upgrade your plans so you can use this new GitHub Copilot experience. Microsoft also showed off something called Project Solara. It's basically a platform for putting AI agents inside of physical devices, and you can kind of see some silhouettes of the types of devices they're talking about, but they really showed off two main devices during the keynote, one of them being this desktop device that kind of looks like an Amazon Alexa something like that. But it's a device that sits on your desk that you can, you know, communicate with your agents from and, you know, see what it's doing and see your calendar and things like that. But probably the more interesting one they showed off was this badge, which is like a digital key card with a camera and a microphone on it, meant to kind of, you know, connect to your belt from a lanyard. And it just seemed like a very, you know, different idea. Like, here's the little desktop device that they showed off that you know, kind of looks like that Alexa. But then here's that badge keycard looking thing, and they're showing it scanning packages and, you know, giving you access to devices. And they actually brought one out on stage and showed it off. But one of the more interesting use cases was like in a doctor's office where the doctor was using it to keep track of patient details and scan arm badges. So I don't really know who this is totally for yet. We've obviously seen a lot of like device launches with like the Rabbit and the Humane Pin, and now they're trying to do this like badge thing, but this almost seems more like it's for workplaces to use to scan packages and scan patients and things like that. It looked interesting, it looked different, but I'm a little confused on who it's for still. So I went to one of the sources and asked Mustafa about this as well.

The badge has been around for like 30 years or more and basically hasn't evolved at all. And yet, you know, most of us who go to work every day in a big company have to carry out a badge to open the doors and so on. So I think it's really exciting that there could be an ever-present, you know, aware kind of sensor system on your badge that basically makes it easier for you to ask any question of your ever-present AI with obviously camera and vision. So it's pretty cool. And I like the kind of bold experimental nature of it. I'm very excited to try and use it. And it's a very flexible platform. So anybody can like, you know, ship their own models, you know, ship their own harnesses, integrate it in their own products. So it's really supposed to be a general purpose platform.

What can we do with that that we can't do with our phone?

Well, I, I think it's ease of access, fast, stable, reliable. Obviously it's secure. I mean, technically speaking, you could also use your phone to unlock a door anyway, right? So I mean, I, I, I think that many of these deployments are gonna be somewhat overlapping because clearly all your devices, you know, do some portion of all the capabilities of other devices. I think that the, the main motivation is it would just be great to add a few more features and, and create an open platform, um, you know, with this project that people can experiment with.

So again, it sort of feels like an experiment to see what people think of it, but even more than that, it kind of feels like they're just trying to see how people are going to feel about putting agents in all sorts of different devices and seeing what kinds of devices agents work inside of. Because if we look at this image here, it looks like they're gonna be putting it in all sorts of devices like headphones and watches and little scanner guns and glasses. And is that Rabbit R1 or something. I don't know what that's supposed to be. Another announcement that I know the Microsoft team was pretty excited about was the fact that the Mayo Clinic and Microsoft are collaborating to develop a frontier AI model for healthcare. And when I asked Mustafa, like, what he was most excited about out of everything that's coming, this is what he had to say.

I think that people are going to be very excited that they're going to have a perfect health assistant in their pocket. Yeah, I really think that we're close to medical superintelligence.

Awesome.

I think 2 to 3 years and it'll be possible to get access to the absolute best healthcare in the world, which currently you can only get if you go to Mayo. Right. And that's just going to be, you know, abundant and available to everybody.

And there were a ton more announcements made during Microsoft Build, but I'm trying to narrow it down to the ones that I found the most interesting that I think most people would find cool or exciting or useful. But while I had Mustafa, I had to ask him what is his favorite current use case for AI. And this was his response to that.

Every morning I get a pretty long and detailed briefing of all the activity that's been happening inside of Teams, on email, with all the docs that have been changed. And instead of me having to go and read all of those things, I just get this clean synthesized summary with links to everything I need to do. Recommended actions, a prioritization for my day, briefings before each one of my meet— I mean, it's kind of remarkable, right? Right. I mean, this is something that it would take 2 or 3 people to do pretty much full time. And that, that is kind of a little superpower that I just feel like prepared for the day, clear-headed and, you know, just armed with information, which I like.

And Microsoft Build wasn't even the only big event to happen this week. Out in Taiwan, there was the Computex event where Nvidia made a series of announcements themselves. The biggest announcement, in my opinion, is the RTX Spark. This is a new chip that Nvidia developed that's both essentially a GPU and a CPU in one. And the current version they're showing off has up to 128GB of unified compute, meaning you could run some pretty large, pretty smart LLMs locally on your device. It's also going to be a beast for gaming. I actually got to demo some computers that had these chips already installed on them, and the graphics on those computers were insane. And they were running some really, really large, smart local AI models directly on like Microsoft Surface laptops. Now these aren't available to the public yet. Just sort of announced during Computex. And the whole idea here, they're trying to shift more and more of the inference, the stuff that happens after you give AI a prompt, directly onto your computer. You know, this has privacy implications. If you're running local models, you don't have to send them to the cloud where they could be possibly training on your models or getting access to private details that you send over to those cloud LLMs. You can also run them completely offline. So if you're like on a plane without internet access and you want to use a solid LLM, well, if you bring the inference onto the computer. You don't even need to be connected to the internet to run AI models anymore. I mean, a lot of what people use large language models for these days is fairly simple things that, for lack of a better word, the dumber models, the smaller models, can actually handle for you. Like, you don't really need Opus 4.8 or GPT-5.5 high max level models to summarize documents for you, or to help you rewrite an email, or to take a list of things and reorganize them for you. These are all fairly simple things for these smaller AI models. So these bigger, more compute-intensive models that you're sending stuff to the cloud for, well, we can start saving those for the things that actually need that level of intelligence and let most general things run locally on your computer. That's what these companies are starting to push more into. And like I mentioned, jumping back to Microsoft for just a second here, they did introduce a new Surface Laptop Ultra, which is being built with this new RTX Spark chip. These are what I actually got a demo of while I was out at Microsoft Build and was actually really blown away by the graphics and what they were able to do from like an LLM running locally standpoint. There is one caveat about these. They're probably going to be expensive AF. Like the DGX Spark, the little box that has a similar sort of level of compute power to run models locally, start at about $4,000. So I would say that these laptops are probably going to be starting in that range. They haven't announced prices. We don't have an exact date of when these are coming out yet. The RAM shortages that we're seeing are also going to impact the price. But when these first versions come out, expect them to be pretty damn expensive. Now, outside of the models we got from Microsoft, there was a handful of other new large language models that got released this week. I'm going to run through them quickly, but I never really want to skip over new large language models that got released to the public. I just don't think I need to go too deep into them because for most people they're gonna feel like fairly marginal improvements. The first one I want to mention came out of Nvidia this week in their NemoTron 3 Ultra. Now they showed off a bunch of NemoTron 3 models during GTC earlier this year, but we weren't getting access to the Ultra model yet. This new NemoTron 3 Ultra is an open model, so I mean theoretically if you had a strong enough computer you could run it locally. Now most people are going to be running it in the cloud because it is a 550 billion parameter model and no consumer-grade GPU is going to run that. But it is a new state-of-the-art model among open-weight models designed for agent productivity. It's also much more cost efficient than other models within the same sort of class. Google also released a new model in their Gemma series of models, which is their open-weight models. They introduced Gemma 4.12B and this one, well, you will be able to run it on your laptop. If we look at the benchmarks here,. This model is pretty much as smart as the Gemma 426B model. So a little more than twice as many parameters, but pretty much just as good. I mean, just slightly underperforming the much larger version in most of the benchmarks. And we also got a Minimax model in Minimax M3, and this one is a coding model with a 1 million token context window. When we take a look at the benchmark comparisons for this one, it actually claims to beat out GPT-5.5 and Gemini 3.1 on SuiBench Pro. Pro, and it's a pretty inexpensive model to use as well. Okay? So I covered probably the most important announcements out of Microsoft Build, the coolest stuff out of the NVIDIA event, the handful of large language models that came out this week, and I feel like I've barely scratched the surface of all the announcements. So instead of going deep on a whole bunch more stories, let's jump into a rapid fire and break down everything else. Pew pew pew pew pew! Let's start with some updates out of OpenAI. If you're on Windows and you'd like to use the Codex app app, you got a new upgrade this week. Computer use now works on Windows, so Codex can operate desktop apps by seeing, clicking, and typing in the foreground. And remote control now supports window devices from ChatGPT on iOS or Android, so you can actually tell your Codex app on your computer to do things from your phone. A really cool feature that everybody who uses Codex is so happy they rolled out. Speaking of Codex, it's slowly turning into that That super app from OpenAI that just kind of does everything and is not just for coders anymore. This week they rolled out some new features for people that may not specifically need to use it for coding. They rolled out plugins that adapt Codex to your role and tools, annotations that help you refine the result in place, and a preview of the ability to create interactive websites and apps you can share with your workspace using a URL. Some of the plugins they rolled out, the data analytics Analytics plugin, the Creative Production plugin, the Sales plugin, the Product Design plugin, the Public Equity Investing plugin, the Investment Banking plugin. And they also added this new Sites feature, which allows you to actually get a URL directly from Codex that you can then share with other people. So if I pop open the Codex app here on my computer, come up to the top left for plugins, you can already see some of the brand new plugins here: Sales, Product Design, Data Analytics, Investment Banking, etc. So you install the plugin that's most helpful for your own use case and your responses in Codex will be more geared towards that specific use case. OpenAI also rolled out a better memory for your ChatGPT. And well, I'm not going to try to pretend I understand what's going on here. However, if we scroll down here, they sort of show the evolution of memory. So in the beginning you just had saved memories, essentially stuff you told it to remember. It would save in like a memory bank. Then in April of 2025, they updated it with the first version of dreaming. Which allowed ChatGPT to actually look at past chats to understand context of current chats. And this new dreaming feature appears to be like an evolution of that. It's better at understanding and saving past conversations into context. So when you chat with it, it does a better job at remembering things from past conversations. Again, I don't quite totally understand myself, like how this was updated exactly, but you should notice that your memory in ChatGPT is better than it used to be. I thought this was pretty cool as well. Noose Research, the company behind the Hermes agent, just rolled out a desktop app. If we take a peek at what the desktop app looks like, it has some similarities to Codex as well. I feel like I'm saying that a lot lately, but a very similar design style to Codex, but it's to keep track of all of your agents that you've built in Hermes. Now, I haven't actually played with this desktop app yet. I have played with Hermes a little bit where it chats with me in Slack, but I'm looking forward to testing out this desktop app. Let me know if it's something you want me to go deeper on and maybe we'll do a whole video just playing with this Hermes agent. If that's something you're interested in, you know, let me know in the comments. We got a whole bunch of new image generation models this week, starting with Ideogram. They introduced Ideogram 4.0 and it's an open image model. You can download the weights, you can fine-tune it on your own data and it will run on your own hardware. And you can also use the API if you don't want to run it on your own hardware. Jumping back to Arena.ai, a site that does sort of a blind taste test of image models, we can see the new Ideogram 4.0 ranks at number 9. But this is an open model. Every single model above it is a proprietary model. None of these you can download the weights of and run on your own computer, fine-tune them, do what you want with them. But Ideogram, well, you can't. And it does a lot of stuff that the more state-of-the-art models do. It's really good at getting text into your images. It's got native background transparency, which is actually something that GPT image model still struggles with. And this is also really interesting— they trained Ideogram 4.0 with bounding boxes tied to region descriptions. So it teaches the models where every object, text region, and layout element belongs. So the way they trained it basically like trained in composition to make sure the design always looks like really solid with the rest of the design. And it makes, you know, decently realistic looking images. Maybe not the most realistic we've seen, but not not bad either. Again, the key to this one: open weights. We also got a new image model out of Reave, a company that I'm not even that familiar with. I can't remember if I've talked about them in the past or not. But jumping back to our Arena page here, this Reave model is now in the number 2 spot. The MAI Image 2.5 model was in the number 3 spot for like 3 days, and now it's bumped down to number 4 because of this new Reave 2.0 model. And it even beats out Nano Banana in text-to-image. According to their own blog post here, it's now number 2 in text render in photorealism. It's number 3 in portraits, number 3 in art, number 2 in cartoons, and number 2 in commercial design. The only model that really beats it out in almost every single category is GPT Image 2. So if you're looking for another text-to-image model to play around with, Reave is another one to give a look. The company Krea introduced a new model this week called Krea 2 Turbo. And well, the big deal with this one is that it does what Krea 2 does, just a lot faster. It generates images in just 2 seconds. So fairly similar quality to what you'd get out of Krea 2, you know, just fast. Grok rolled out a new model this week in Grok Imagine 1.5. Here's an example from JS Films with both video and dialogue. So let's take a peek slash listen.

Excuse me, is this seat taken? No, go ahead. Thanks.

Not bad. Still has that like AI lip sync quality to it where you just instantly know like this is an AI syncing. But again, another model for you to go and play with. Runway showed off their Runway Aleph 2.0. This to me feels very similar to what we are getting out of Google's Omni model where you give it a video, you prompt it to change something about that video, and it makes the desired changes in the video. I'm gonna jump into Runway here. I'll use this video that I shot at Microsoft Build Mustafa up on stage. I can select a specific frame that I wanted to reference. Let's do change the man on stage to be a humanoid robot. All right, so it gave me a couple variations here. We could see there's our humanoid robot. This one was generated with ChatGPT image. Here's our Nano Banana version, and here's our Runway Gen 4 version, which just totally altered the image way too much. I think I think the GPT version did it best, so let's go ahead and generate a video based on that. And well, here's what the final result looks like.

Computation in just 15 years. It's now clear that a consistent exponential increase in computation leads to predictable—

I mean, it's okay, but I'm probably gonna stick with Gemini Omni for a little bit. And finally, I teased another AI voice model earlier in this video. There's this new model called MISO-1, and they claim it's the most emotive voice model in the world. And not only that, but it's also open source. Here's some examples.

I'm honestly in love with him, and I'm so tired of pretending that I'm not. You know, I just don't really get it why people don't approve of the age gap thing. You know, NBA teams that won a championship with no superstars. Back in 2004, nobody thought that the Detroit Pistons were title contenders. Over the past few years, they had been building up a roster of NBA outcasts. This is everything I ate in a day as a startup founder in San Francisco.

I started the day by going to a beautiful bakery. I mean, this model is really good. Like, it would fool me if I was just skimming TikTok or something and came across one of these videos. They sound like they were recorded by a real human. I don't know why this last one was like a woman narrating a dude's day, but it's a pretty good model. And for being an open-weight model, even more impressive. If you want me to do a walkthrough in a future video of how to install this and run it locally yourself, let me know in the comments, and maybe this is something I'll make a demo of in a future video. But that's what I got for you this week. I mean, it's been a very, very busy week, especially with Microsoft Build and Computex. GPT-X all happening in the same week. And of course OpenAI and Google and Minimax and all these other companies throwing models out this week. It made for a fairly overwhelming week. But you know, there is so much happening every single week. I mean, these videos themselves, I'm usually starting with about 70 news stories and then narrowing it down to just a handful that I think most people are going to be interested in. My goal with these videos is to try to cut through the hype and cut through the noise, and I'll drink from the fire hose all week. I'll read all the news, I'll try to keep up with it, and then turn around and just bring you what I think you need to know, so ideally it's not so overwhelming for you. If that's something you're interested in, consider liking and subscribing. I do make these videos on Thursdays, publish them on Fridays, so if there's any big news that happens on Friday that I missed in this video, I'll do my best to get it into a video next week. Also, if there's any tools or news that you think I didn't go deep enough on and you want to hear more about, let me know. These videos are getting longer and longer every week with the pace of the news that's coming out, and well, I don't want to turn these into 2-hour videos, so I do have to cut things down to keep the videos moving a little bit and to squeeze in all of the info I think you should know. But that doesn't mean I can't make some side videos here and there and go deeper on some of these topics. So I do want to hear what you think and what topics you want me to go deeper on, what tools you want me to test. I'm happy to pay for some of the tools and put them through their motions so that I can demo for you before you have to go pay for them. That's something you want, again, let me know in the comments. And, uh, yeah, thanks again for hanging out with me, nerding out with me. I'm having a blast making these videos. Hopefully you're having a blast keeping up with everything with me. And, uh, yeah, I'm done rambling. Thanks for nerding out, and hopefully I'll see you in the next one.

Bye-bye. Our hope and belief and core motivation is that we can create a humanist superintelligence. There are lots of people out there in the industry who are publicly saying that their motivation for building superintelligence is to create something that exceeds the intelligence and capability of all humans combined. Mind, right? And then can keep improving, you know, for many, many years to come, way beyond our intelligence. And I think what scares people is when you say that, it's hard to imagine how we would control something like that. So not only might it compete with us for resources, but if it were autonomous, it might be pretty hard for us to kind of keep it contained. Yeah. And this is something I have personally been thinking about for a very long time, since I started DeepMind and have worked for many years on AI safety and ethics. And I think that the only test that we can put technology to is a very simple one, like, does it make our collective lives healthier and happier? Right. Um, and that's why we invent things. That's the point of science and innovation. That's why we're excited to, you know, have new drugs that, you know, cure diseases, that, you know, relieve our pain, that help us to live longer. Right. And I think that that is just a very simple test of technology that we put, you know, superintelligence to next. Does it enhance human progress? Does it make everybody healthier and happier? Mhm. And that I think is quite, quite clear. Like, we've, we've got to subject it to that test.

Yeah. So, so somebody that's, um, you know, maybe about to go into college, do, do you have any advice of like, what, what would you tell them to study right now moving into the world that we're moving into?

I think you still have to study software engineering, just as, you know, people have been focused on for the last 20 is. I think that more than ever you have to study science and maths. But, you know, we are also going to have to think hard about the philosophy and the ethics behind this because we're creating something that is incredibly powerful. It will feel very human-like. It's going to be very dynamic. It's going to be integrated into our lives in ways that we've not really seen from any other technology before. And so thinking about the governance of these things and how they play into our politics and our civil service and our public duty to create the world that we want to live in, I think is going to be really, really important. And maybe it's not so much the primary muscle that we've exercised in the last 30 or 40 years, but it's going to become really the main focus, I think, of people in the next 5 to 10 years.