AI is on Record Pace to BOOM! o3 mini, Grok 3, Operator & More!

Hello everybody. Welcome back to the Matvid Pro AI YouTube channel. This past week has been insane for the AI space and I think it speaks volumes for what we're going to see in 2025 in regards to AI progress. I'm going to be giving you guys as much information as I can today. I can't cover everything because this video would just be far too long, but there is plenty to talk about. First up here folks, I want to talk about O3 mini. It's an upcoming model, as I'm sure you guys know from OpenAI. It's in their thinking realm of models. O3 mini is obviously smaller, cheaper, faster than the large O3 model while not being as performative as that O3 model. But still, O3 mini is supposed to be better than OpenAI's O1 model. This post right here from Smoke Away, who typically is known for having pretty good Insider information into OpenAI, states that we are going to be seeing O3 mini released on January 28th at 10am PT, which sounds pretty close. This could get pushed back back further, we don't really know, but I think that this is a reasonable expectation. As you can see though, this post is just a little outdated from January 20th where people in the comments were speculating about when Operator was going to release. Obviously now we know Operator already is available for pro level chat GPT subscribers and we will be talking more about OpenAI's operator in today's video. Next up folks, we also know that this O3 mini model is going to have some level of free tier usage and really like to see that these highly intelligent models should not just be limited to people who are paying monthly for a chat GPT subscription. But as Andrew Caron points out right here, free access to O3 mini is a direct response to the astonishing capabilities of the deep seq R1 thinking models. Basically, the point that Andrew Koran is trying to make here is that the fact that O3 mini is going to have some free tier usage is probably just due to competitive reasons in the space. DeepSeek R1 released fully open source weights and all while pretty much being as good as OpenAI's current flagship model. Next up, we have even more confirmation about O3 mini and some potential timeline updates. OpenAI's Kevin Wheel says that O3 mini is very soon, so that lines up pretty well with what we saw earlier from Smoke Away's insider take of the 28th. We can expect full O3 in February or March, which also aligns pretty well with what I've seen OpenAI do in the past. AI agents or computer use agents are expected to arrive in Q1 of 2025. As we know now, OpenAI already released operator in Q1 of 2025 to the Pro ChatGPT subscribers. So we kind of already have these computer usage agents, but they are still very much in the early stages. So I think we could see a broader rollout of potentially better, more refined computer usage agents throughout the rest of 2025. And finally, OpenAI is already training the model that and obviously we expect that. Oh, and as always folks, all of these posts are going to be linked down below if you want to dive a little bit deeper for yourself. All right folks, continuing on with O3 mini, Aidan Clark posts a quick little demo. Aiden Clark here is also a researcher at OpenAI. O3 Mini was able to generate this demo of a rotating square and a little yellow ball bouncing around inside of it in just 20 seconds, obviously through coding, which is really cool. And this was in response to a post where Deepseek actually beat out 01 Pro for the same exact prompt and you can see both of them don't do as good of a job as O3 mini, which appears to be very telling at first. But in the comments here we can see our friend Sean Ralston shows off what GPT4 omni was able to do, and it was honestly a pretty good demo. Maybe Sean Ralston just got lucky, or maybe there's more to the story. Either way, this isn't the last you're going to see of this yellow ball bouncing in a square demo today, so keep this one in mind. Next up folks, we have a confirmation from OpenAI's Sebastian Bubek on how 01 reasoning came to be. There was no specific tactic given to the model. The O1 thinking model. Everything is emergent. Everything is learned through reinforcement learning. Now this comes out in light of the Deep seq R1 Thinking Models research paper, the open source model that we talked about earlier. And essentially this is pretty good confirmation based on the Deep Seq R1 paper that Deep Seq was able to essentially replicate exactly what OpenAI did with O1. And to be honest with you guys, this does kind of confirm for me that OpenAI is a little bit on its heels regarding those deep seq R1 models. They were able to replicate what OpenAI was able to do with the O1 thinking models. And that is the reason that the further thinking models such as O3 mini are going to have free access. But don't wait folks. There is even more Google is also on OpenAI's rear end here rolling out a Gemini 2.0 flash thinking update. So it also uses the in context reasoning, but this one supports a 1 million token context window native code execution which is nice to see. OpenAI also has that now inside of Canvas. Longer outputs and less frequent model contradictions. Google has been releasing a lot of these more experimental, quick snappy updates to their large language models and I think it's benefiting them greatly. They also post post some quick benchmarks here to show better performance on math, science and multimodal reasoning benchmarks. Logan points out that scaling is alive and well which insinuates that yeah, these thinking models that OpenAI is into Google's now into Deep SEQ replicated with open source that test time compute is super valuable and it scales well. Model parameter count isn't the only factor anymore. It's also how long do you spend inferencing and processing the information or the task? In other news related to Google, Imagen 3 is finally on LMSIS now. Guys, I'm not quite sure if this necessarily is the final release of Imagen 3, but you can see it is right now a top charter. Much better than Recraft V3, much better than Dall E3, and this model is also free to generate with. You can try Google's Image in three on Image Effects. It's also going to be available on their AI studio. I will say though, I do have some gripes in regards to the arena, the text to Image Arena. Yes, Imagen 3 is a top charter right now with the highest arena score. But you'll also see that ideogram V2 down here has a slightly lower score than Recraft V3, and I've used both of those models. In my opinion, Ideogram is a lot better. So when it comes to more subjective things sometimes like image generation, a lot of these top models are going to compete with each other, and depending on what you're trying to specifically do, you might prefer one model over another. Just thought I'd point that out. I also wanted to mention that if you haven't joined my Discord server yet, you definitely should. The people in my Discord server, and you know who you are, have been on the Imagen 3 hype for quite some time now, and you guys have known how good this model is. The Discord server has a fantastic community and if you want to stay up to date, if you want to see the best in AI, I recommend going to that server next. Up next up. Yes, we're moving back into large language models. I know, but we now have some actual posts regarding the next series of the Grok models by Xai. So here's a researcher actively testing Grok 3. The Grock 3 will be the first model trained on a 100,000 H100 cluster, which is massive. And Elon Musk has also boasted that this will apparently be the most powerful AI model out there. But we'll just have to see if that's the case. And I told you guys this particular test Python script with the bouncing yellow ball in the square was going to show up again. This is a side by side performance comparison. Grok3 apparently nails it the best out of everybody. Perfect bounces with the yellow ball there. 01 Pro is struggling and Deepseek R1 has some issues with with the ball not being confound to the box perfectly. But again, if we move down to the comments here, we can see 01 Pro also nailing it. And also Grock 2 apparently doing a pretty good job with the same task. So take this with a grain of salt. But you know what I might not be able to take with a grain of salt? This next test with Grock3 creating a ball in a rotating tesseract. This is a little bit more difficult. Obviously it's going to be 3D instead. So I'm not sure exactly how is able to produce this or if this is even correct. But yeah, there you go. That's. That's another thing that Grok3 did. I have yet to see something this complicated looking be recreated with 01 Pro. Either way, I think Grok 3 is going to be a very interesting model. I don't know if they're doing the test time compute. I don't know if they're doing the thinking. Maybe it'll just be a big bad model with a high parameter count. We'll have to see when everything releases. All right, let's finally talk a little bit about OpenAI's operator. As you can see, it is locked down to the 200 doll month. OpenAI Pro plan gives you the research preview of Operator. I'm going to be real with you guys. I don't have the $200 OpenAI Pro plan. I'm not paying for that right now. I got to say though, Operator and seeing people use Operator has pushed me closer to dropping that $200. But it's still kind of a ton of money. Let's be real. I also believe that this is only available in the us so don't go out and get a pro plan if you don't live in the United States just for operator. OpenAI also did little live stream showcasing Operator and how it works. You've got your messages on the side and then essentially it opens a virtual browser and clicks around. In this example they're able to look up availability for a specific reservation at dinner for a specific time it's able to do so. They also sent a simple grocery list through a image to the Operator model. They also just said I like Gus, which could really mean anything, but the model was smart enough to know that means Gus is market. The model in this case was actually able to use Instacart to go search for all of the items, add everything to cart and prepare the order, which is pretty impressive. But I think there are some other AI agents that would be able to do this right now. The UI also allows you to take control at any point you want to make sure that Operator doesn't do anything you don't want it to do. Would be pretty annoying if it bought 10,000 eggs and you didn't need 10,000 eggs. They also have Operator picking out seats for a concert, which is pretty specifying everything within a specific budget as well as you'd expect an LLM to do. I know I said I don't have access to Operator right now and I'm not shelling out the 200 bucks, but there are some people who have, so let's hear what they have to say. First up, Matthew Berman, fellow AI YouTuber. Here are some things that he noticed after using Operator Number one. It's kind of annoying to use it without your own browser because you're not logged into any of the sites that you use frequently, which is absolutely valid. That would also annoy me and is probably one of the main things that is preventing me from shelling out that 200 bucks. Logging into everything, especially with two factor authentication, is super annoying. When taking over for Operator, the interactions are laggy because it is running remotely. It's streaming all that information over to you from some other machine that's running in the cloud. So that's to be expected for the most part. Hopefully though, in the future OpenAI is going to allow us to have Operator control our own browser entirely without needing to use a remote machine. Operator is also mesmerizing to watch, according to Berman here. And the potential is insane. I think we all know that the potential is insane. I really hope that Operator vastly improves over the course of this year. I would love to see something insane in terms of capable AI agents come out of OpenAI this year. A reply down here from another AI YouTuber Matt Wolf. He says that he finds it still constantly gets stuck in loops. He tried having it find a specific person on thumbtack and it kept opening the same person over and over again in new tabs for some reason. But I guess that's why they call it a research preview, right? Stuff like this is expected to happen and will be hopefully ironed out before a more official release. Next up, we have another post from I Rule the World Mo. A controversial figure here on the X AI spaces, but at least in this post here, our Strawberry friend is providing some good first impressions of Operator. He has apparently already hit usage limits. First impressions being that Operator is incredible and by the end of the year it's hard to say where we will end up. But as of right now, the average user won't get much value out of $200 to be able to access Operator. Just being honest. Appreciate that says it's slow, misses enough clicks to be annoying, and doesn't provide a ton of upfront value. McKay Wrigley responds, saying it's not reliable enough to be useful for work yet, but he's glad that they shipped it early. He thinks that it will certainly be there in a year, probably in six months even. And finally we've got another little take or another little note from Smoke Away. In this circumstance, Operator encounters a captcha and it actually asks the user, could you take over for me and complete the captcha? Even though it probably has the capabilities as an LLM to just complete the captcha alone, this meme right here is becoming all too real. Next up. This is pretty darn cool. This is Pig Dev and it's Windows desktops for AI agents. A lot of you might not be surprised to learn that AI agents similar to what we have now from OpenAI's operator, have existed for some time. This, though, is an API to launch and automate Windows apps for use with AI agents. And in this case, it's not just control over a web browser, it's actually control over the entire Windows computer. So as expected, it can still do things like open up a web browser, search Amazon, scroll around. But it can also do things like open up Excel and start a new Excel spreadsheet and begin to actually fill it in correctly. Which is pretty crazy. So while OpenAI still is making consumer level strides towards agents, the rest of the AI community is making more open and accessible strides towards more powerful tools, even though they might not have the same level of support or ease of use. I thought I would share this with you guys in case any of you devs are interested. Pretty darn cool. Okay guys, let's talk about the Stargate project coming directly from OpenAI. Stargate Project is a new company, separate from OpenAI in some way here, that intends to invest $500 billion. 500 billion, guys. Yeah. That's a lot of freaking money, dude. Over the next four years, just four freaking years, man, dedicated to building new AI infrastructure for OpenAI in the United States. Their first goal here is to deploy $100 billion immediately. This infrastructure will secure American leadership in AI, create hundreds of thousands of American jobs, and generate a massive economic benefit for the entire world. According to OpenAI. They're stating that the project is re industri centralizing the United States for this next. I mean what a lot of people would consider industrial revolution of AI, they go through some of the initial equity founders. We've got SoftBank, obviously OpenAI, Oracle is in there and MGX, they announce a chairman as well. They announced some key technology partners like arm, Microsoft, Nvidia and of course Oracle. The build out for this, by the way, is already underway starting in Texas and they're also evaluating apparently other potential sites across the country. Basically what Stargate really seems to be at the end of the day is a computing system, lots of GPUs, lots of power to train new AI models. And you guys might have also seen this, some drama going on over here. Elon Musk saying they don't actually have the money. We move on here. He says SoftBank has under $10 billion secure. He has that on good authority. Sam Altman says he's wrong, invites him to come visit the site. He says it's great for the country. Marco here, AI and design makes a really good point though. Replying Sam, hey, if you're really doing this all for the country, all for the United States of America, why not bundle forces with all major AI players? Shouldn't it involve xai, Google, Anthropic, etc, And obviously no, he's not going to do that because OpenAI loses a huge competitive edge. But it does make Sam sound a little bit more hypocritical in regards to putting necessarily America first. It does seem like at the end of the day it's sort of OpenAI first. I don't know, I don't like to get, get too deep into this kind of a thing. So let's turn this on its head. Let's bring a new perspective here. I'M going to read this from David Shapiro. This is his take on Project Stargate. He says we are about to witness something that happens maybe once in a generation. The full might of American innovation, capital and willpower focused like a laser on a single world changing goal. And if the project is successful, that exactly is what this seems to be. He references this to the Manhattan Project or what NASA achieved when brought down the challenge of taking us to the moon. When America decides something matters and backs it with this kind of money, it just happens. Period. And this time we're not just shooting for the moon, we're reaching for something that will fundamentally reshape human civilization as we know it. And I actually absolutely agree with David Shapiro on this. The target couldn't be clearer. Build and deploy AGI. Simple, elegant and absolutely transformative. Shapiro goes on though to bring a point that everyone seems to be missing. It's not just about building the most powerful AI system in history. It's also about ensuring American technological dominance for the next half century. As he predicts. 500 billion might sound like a lot of money, but what it's really buying us is a billion quote unquote Einsteins that are working around the clock on every problem humanity has ever faced. One Einstein gave us relativity and fundamentally changed our understanding of the universe. But he says what happens when you multiply that billion by a billion? Every single doctor gets unlimited cognitive resources for diagnosis and treatment. Every scientist gets an Einstein level partner working on their research. Every economic challenge gets solved by super intelligent analysis. We're not going to just remove the intelligence bottleneck, we're going to shatter it, giving us plenty of overhead. And he makes a really good point. The first world wars were won with steel and manufacturing. The Cold war was all about nukes and economics. But this war is already over and America won it with silicon and algorithms. China's entire IT infrastructure is about to look like a pocket calculator compared to what America is going to build in Texas. I don't know. Tell me what you guys think down in the comments below. I honestly agree pretty closely with David Shapiro's take on this. If the money can be secured, and some part of me does feel confident about it, specifically because it is tied to this brand new administration we just got here in the United States. It could put a serious dominance prominence in the United States for leading the entire world in AI. If they are successful though, and they do put super intelligence into the hand of every doctor, every scientist, even the average Joe like me or you. What happens next? That is the big question. Anyways, for the sake of time I'm going to move on to my next take. Matt Wolf comes in with some random thoughts regarding Stargate project. Mr. Wolf has a keen eye for red flags. He points out that that Mark went on Joe Rogan and said that he was in closed door meetings and the government basically said don't build an AI. We've got our companies we're working with and there won't be room for others. He's paraphrasing, but that was essentially the message at the end of the day and I have seen that specific clip and I do remember that and can vouch for it. Matt Wolf is now wondering if this is specifically what he was referring to. He's wondering if there's been plans for months for OpenAI to be the one true king and to have the serious compet processing power resources to be that AI king. Matt Wolf also wonders if this is the reason that a ton of people left OpenAI last year. Did they see this? People ask why would so many leave on the brink of AGI? I can't blame Wolf here. It does sound like a logical reason to step away from a company. If they're cozying up to all the three letter agencies and you don't want to be associated with that, it seems like a pretty good reason to get out and move on to one of the many other people people or businesses that will pay you half the amounts of money to focus on AI research. Wolff also points out that Oracle was founded to service CIA contracts. He says to this day Oracle still works closely with government agencies and it's interesting that they are attached to this project. Wolf also makes the great point that OpenAI now has the backing of the US government and now will be able to better steer regulations in whatever direction they that fits them best. He's saying that we might no longer wonder whether or not Google or Anthropic or X will ever catch up to OpenAI. It's sounding like the answer might be from here on forward, no. Wolf swears up and down I'm not an AI doomer. We know, we know Mr. Wolfe. He's just putting some pieces together and thinking out loud. I thought that this perspective, overlapping the previous perspective from Mr. Shapiro would bring a good two sided argument to this. I really want to hear your thoughts down in the comments below. This is a big deal. I know it's a little bit more political, it's a little bit more economic, some of you might not be as interested, but it has huge Implications for the future. Here's my thoughts though. Competition is going to breed innovation for everybody, right? That's why I'm a huge fan of the deep seek R1 model. I don't care if it comes from China. It's going to push American companies like OpenAI or Anthropic to do better. And when they do better, they're going to provide more value for us, the consumer. Now if we are to believe OpenAI in regards to the Stargate project, $500 billion in four years to build AI infrastructure is insane. And if it is going to be allocated to just OpenAI, well that gives them, you know, kind of a monopoly, which is not good for the consumer. But I still believe in the world, I still believe in other American companies and I believe that OpenAI does not have to be the sole winner here. This Stargate announcement should be a wake up call, a serious wake up call to China to dig their heels in and actually go all in on AI man. Because if you want to win, it's looking like the price tag is around $500 billion. Monopolies are also illegal in the United States of America, but obviously that is its own can of worms. I want to see AI built at an acceleratory rate, but I want to see it done with competition. Not just one company ruling over everybody else and having way better AI than everybody else. Especially considering OpenAI doesn't release anything open source anymore. Alright guys, I am going to cut off this particular news video here. I do have other things that I want to talk about however, I think I'm gonna save them for early next week. So be on the lookout for that. We're gonna have topics like a brand new AI music generator. We're going to have topics like 11 labs coming to desktop, Netflix actually doing some research and building an AI animator that looks really good. Cling elements and perplexity updates. So. So probably that and more coming early next week also. Yes, now the embargo is lifted in regards to the RTX 5090. So I'm going to put some sort of a video together. I don't know if it's going to be a super benchmarky scientific review, but I've already been running models locally on the RTX 5090 like the deep seek R1 models and it is blistering, man, it's insane. So yeah, expect a video on the RTX 5090 and what it can do locally soon. I think that's going to cover it for today. Like and subscribe. Subscribe. I'M doing my best. Things are super busy right now. I'll see you in the next video. Goodbye, guys.