AI News: This Video Model Has Everyone Freaked Out!

This is Will Smith eating Spaghetti back in 2023 and this is him in 2026 with the new model that just came out. That's the good stuff and that's a wrap. It's been a heck of a week in the world of AI. I don't want to waste your time, let's just jump straight into the news. The video model we were just looking at was Seed Dance 2.0 from ByteDance, the same company that makes TikTok. This new model actually supports four different input modalities. You can give it text, you can give it an im, can give it audio, and you can give it video. I don't believe there's any other video model available on the market right now that lets you generate with all of those modalities. The model supports 15 second high quality multi shot audio video output. It features dual channel audio and ultra realistic audio visual experiences. It's some of the best lip syncing and most realistic lip syncing I've seen from any of these video models yet. Now this model is rolled out in a lot of the world already and people have been able to use it, but it's not available in the US yet. Supposedly we get access on February 24th, but people have found workarounds and loopholes to be able to use it, but most of those have been closed and well, I haven't been able to use it myself. Now this model does generate both video and audio. Good luck. The realism on these videos is to a level I have absolutely never seen before and the consistency of the characters is just mind blowing. I'll link up to the official article down below this video so you can go and check out a lot of the demos. But instead of showing the demos that they sort of cherry picked and put on their website, let me show you some of the best demos I've come across on places like X. This account on X called Chat Cut claims that this was made with Seed Dance 2.0. They said that they built an open claw agent to get this. The agent crawled the page, extracted product info and photos, then fed the right assets into seed dance 2.0 to generate this UGC product video. So check this out.

I live in an apartment with no tap outside. I just unfold this stupid frog washer, fill the 12 liter tank, snap on the four in one nozzle and spray. It's cordless, compact and powerful enough for a full car wash.

Here's another clip shared by Charles Curran here. If this is your first shift at Waffle House, you have to. Apparently you can even do motion graphics as well. Check out this video. And if you're wondering if it can do, like, real celebrities and stuff, well, yeah, it could do that too. Check this out. You know, that was completely uncalled for. And it even got the voice pretty spot on. Here's a clip from Ilker about what the Lord of the Rings could have been if it was condensed down into 15 seconds. Mr. Froto, why don't we just take the Eagle straight to Mount Duel? That actually makes sense. Here's somebody named Dork Sense who managed to make a video that looks straight out of Dune. Trademarked ip. No problem. Matvet Pro managed to make a clip straight out of SpongeBob. It's working. He's really feeling it now. And then you have Butoshi here, who made a clip of Luffy from One Piece, throwing an Apple laptop overboard. So, yeah, this model is probably one of the best models we've ever seen to make video. Which means expect to see more videos online where you go, holy crap, I can't tell if that was real or not. And I think one of the big differences here is that, well, this is from Chinese company ByteDance, who apparently is less concerned with copyrights and trademarks than some of the American companies. Sora and Google and some of these other companies based in the US are not going to be able to get away with allowing people to generate copyright work like this. Where I feel like the Chinese companies, they just kind of do it anyway, and it's going to be so hard for the US IP holders to like, actually go and battle with these Chinese companies. So they managed to make models that are kind of better than what the US companies are able to make. But saying that there's going to be a new VO model pretty soon, and I have a feeling that will also blow people. So we'll just have to wait and see. Now, while we're talking about AI Video, last week I actually tried to generate a video in Korea with this prompt. I mentioned that it was taking over an hour and it never generated. Well, I actually checked in. It did finally generate. I have no idea how long it took to generate because I just left the site and came back a few days later. And this is what it managed to generate using the CLING 3.0 model, which was released last week. But I wasn't able to use last week because. Because it just was taking forever to generate. So go ahead and pause the video if you want to see my actual prompt. But here's the video that that generated. I think it's capable of generating with audio, but for whatever reason it didn't. I must have had it turned off or something. But this is the video you can see the character stays very, very consistent. He walks into another room, you see his face again, and you can tell it's the same person. But if you don't want to wait hours to generate a video with the new cling 3.0 model, well, it's available inside of Leonardo. As of this week. I went ahead and generated a video using this prompt that you can see here, and this was the output. And this one did do audio and it actually generated in just like a couple minutes instead of a couple hours. I don't know what language he was speaking at the end, but at least I was able to get a video in just a few minutes. Now, before I share the next bit of news, I want to ask favor. By the time you're seeing this video, I probably just crossed about 900,000 subscribers and I'm really, really hoping to get to a million by this year. So if you like these videos and you haven't already, maybe consider giving it a subscribe. Anyway, let's move on. The company Alibaba released a new image model this week called Quinn Image 2.0. They've apparently improved the typography. It's now 2K native resolution, got much better text rendering and faster inference. We can see some examp of some of the images it generated, like this one here, this image here. I wish they gave us the prompts with it, but we've got this image here, this image here. Lots of good examples of the types of images it can generate. These are the types of images I'm the most impressed with. These like infographic style images. This comic book image here looks really good. Pretty impressive stuff. Let's go up and try it ourselves. You can try it at Chat Quinn AI. I'm going to upload an attachment. I'll upload one of my headshots here and say generate a clickbaity YouTube thumbnail using this person in the thumbnail. Give it text that says this changes everything. We've already got it set at 16 by 9 here, so it already should be the perfect aspect ratio for a thumbnail. And well, that wasn't impressive at all. It's even set on 16:9 and it didn't make a 16:9. What am I missing here? Well, if you need another image model because Nano banana or the ChatGPT image model isn't doing it for you, here's another one you can try, try. I seem to not know how to use it, I guess. So my New Year's resolution was to keep doing these AI news videos every single week this year, because I've seen your comments and I know you guys really like these videos. But I want to talk to you about what's probably a much harder and maybe more relatable New Year's resolution, which is to actually stop talking about all the cool projects I want to do and actually go and build them this year. For me personally, the hardest part is never the idea, it's not knowing where to start. That's why a platform like Hostinger has been really helpful. It makes starting on something really simple because all the different parts to a project are all in one place, and it kind of just walks you through the process of whatever you're trying to build. So, for example, you can buy a domain, build a website, set up, email and launch a business, all right, inside of Hostinger. And if you're not technical, well, that's fine. Hostinger's AI tools can literally build and optimize the site for you without any coding experience required. And then as you grow, Hostinger can grow with you. They have a ton of apps and automations you can easily add onto your existing project. You'd think that a platform like this would be pricey, and I know some other ones that can be, but Hostinger is surprisingly affordable for what you get. So this year, if you want to stop saying I'll build this one day and actually start building it, check out Hostinger at the link in the description and use my code, Matt Wolf, to save even more money. Thank you so much to Hostinger for sponsoring this portion of today's video. Google also rolled out a new model this week called Gemini 3 Deepthink. Now, this is a super advanced model and at the moment is only available to Google AI Ultra subscribers, aka the people who are paying $250 a month. But these benchmarks here, if you're interested in benchmarks, are really, really impressive. On the Arc AGI 2, which is a benchmark that tests for reasoning and knowledge, they pretty much blew everything out of the water like it's crushing their own. Gemini 3 model, Claude Opus 4.6 was pretty impressive, but I mean, this one's way more impressive. And then also GPT 5.2, it's, you know, pretty far ahead on that one as well. OpenAI's latest GPT 5.3 isn't on here yet, but so far they've only rolled out 5.3 Codex, which is mostly for writing code on humanity's last exam, notoriously very difficult academic reasoning benchmark. Once again it is the new leader, beating out GPT 5.2, Opus 4.6 and of course the previous Gemini 3 Pro. On multimodal understanding, it's beating out all those same models, albeit by less of a margin. And then on the code force here, coding for algorithms, it is beating out Gemini 3 Pro and Claude Opus 4.6. But they did not add the OpenAI models to this chart, so not quite sure how it compares there. Beyond mathematics and competitive coding, Gemini 3 DeepThink now also excels across broad scientific domains such as chemistry and physics. Our updated DeepThink mode demonstrates gold medal level results on the written sections of the 2025 International Physics Olympiad and Chemistry Olympiad. It also demonstrates proficiency in advanced theoretical physics, achieving a score of 50 point on CMT benchmark, which is the condensed matter theory benchmark. And we can see it's pretty much blowing away the other models. Again, it is only available on the $250 a month plan. I do have access to that plan, but I'm not smart enough to think of a real good use case. Like I don't have any good theoretical physics questions to ask it and I wouldn't know how to tell if it gave a good response or not. But we have a new insane state of the art model here. And not to be outdone, not long after Google released that model, OpenAI rolled out a new model as well. They rolled out GPT 5.3 codecs Spark. Now it's basically the GPT 5.3 Codex model that they released last week, which by the way, really impressive model. I've actually been using it for a lot of vibe coding projects. Really, really good model. What makes this one different and what makes it the Spark model is that it's using the Cerebras chips. Now if you remember like a month ago we Talked about how OpenAI and Cerebras partnered up and Cerebras essentially makes chips which make inference really, really fast. So when you type a prompt, the response you get back is like insanely quick. Now, like the Google model we just looked at, this one is only available for pro plans. So like $200 a month plans on ChatGPT, but it is actually really, really insane. So they put out this video here that compares side models. So check this out. On the left we've got the regular GPT 5.3 Codex. On the right we've got the GPT 5.3 Codex Spark. They gave it the prompt. Build a simple HTML Snake game into both of them and watch the difference in speed. I'm going to press play right now. So the one on the right, 2 seconds, 3 seconds, 4 seconds, 5 seconds, 6. And it's done. The one on the left still going. And spoiler alert, this video is 47 seconds. It takes like 45 seconds or so for the second one to go. And they're already testing the game on this right side. Mind blowing. Now this one's available right now inside of the Codex app again for the 200amonth pro users. So I'm going to go ahead and update my Codex app now down in the model selector. If I open this up, you can now see I've got GPT 5.3 Codex Spark. I'm going to give it a quick test myself. I'm going to create a new project here and I'll give it the prompt. Create a beautifully designed Vampire Survivors clone that I can play in my browser. I'm just going to hit enter and let's see how fast it generates this game. So I just hit enter and it is going. It's reading a skill not quite as fast as making a Snake game, that's for sure. It already says I finished with the playable build and it's been maybe 30 seconds. Okay, so it did the whole thing in 50 seconds. That's pretty crazy. Let's open our index HTML here and look at this. This is ready. It's called Nocturne Survivors. So let's start our run. Okay, so here's my little character that's moving around and shooting automatic and I could pick up my XP and the XP is going up like this game was made in 50 seconds. Holy crap. I mean, there's no sound effects to it and it's literally just dots. But the logic of the game is there. Let's see what happens when we level up. I want to see. Yeah. Okay, so we just leveled up. Pick one. Upgrade. Increase fire rate by 12%. Projectiles pierce. One additional. Let's do the piercing run. That seems like it would work. Okay, so we can see that the bullets are now like actually going through the enemies. This swarms are getting more and more. Well, there's more of them. Okay, we just leveled up. Okay, so gravity radius to pick up our orbs. So now it should suck in the orbs when I get closer to them. I think I made it a little too easy, but okay. Now we increased our fire rate. I'm just blown away that this was built in 50 seconds. I gave it a prompt 50 seconds later. Game ready to play. Like what? I can obviously go back and forth, give us some more prompts, but it actually did all of that thinking that it normally does. If you're using GPT 5.3 Codex, it just did it at like 20x the speed, if not more. That's wild. And I love Vampire Survivors, so I need to actually close out of that or I will keep on playing. But anyway, I don't want to go too deep on this one. There are some benchmarks here where we can see how it performed on Terminal Bench, but the main upgrade here is just the speed. It's using the Cerebras chips for the inference and that just made these things fly when you give it a prompt. We got some, you know, somewhat gimmicky updates out of Meta this week. Like this. This Animate your Facebook profile picture with Meta AI lets you turn photos and feed posts into share worthy moments. Now I'm guessing this is still rolling out. You're supposedly able to go to your home page, click on your profile picture and you'll see a button that says Animate with Meta AI. I just did it on my phone and I'm not seeing those options, not even updated. So I guess it's not fully rolled out yet. Okay, so it didn't show up at first, but then I scrolled down my feed a little bit and got it to show me this. So I guess I've got these two options to animate my profile. That's cool I guess. And another update out of Meta is you can control your threads feed with new Dear Algo feature. Now this is something that's been a trend for a while on both X and threads where people go and say Dear Algorithm only. Show me content creators in San Diego who have similar interests to me and thinking that the algorithm would actually show those to people. And I think Meta saw that people were just sort of doing that naturally as like a meme and decided to roll that out. Now you can actually tag Dear Algo. Show me more basketball and then more basketball. Let's see if this one works for me. I'll go to my Threads account here and we can see beta. Change your feed with a post. Just type Dear Algo and describe what you want to see more or less often. Dear Algo. Okay, so this seems to be working. Show me people who share cool tech and enjoy playing with AI tools. Yeah, I'll post that. And I guess that also shows up on my feed. Yep, so that shows up on my feed too. Let's see if that actually works. Maybe I'll follow up next week. We had even more new large language models come out this week, including a new open source model called GLM5 from this company, Zai. What absolutely blows my mind about models like this is how quickly open models are catching up to the state of the art models. So look, here's humanity's last exam. GLM5 with tools scored a 50.4 which beat out Opus 4.5, Gemini 3 Pro and GPT 5.2 with tools on Swebench Verify the software engineering test. It's pretty on par with the state of the art models. Now this is again Opus 4.5 and GPT 5.2. Those companies both have newer, better models now, so these numbers would be a little bit better on those new models. But. But it's really, really close. Sweep Bench Multilingual Pretty on par. Terminal Bench. Pretty on par. Browse Comp this is the ability to actually use your browser and do agentic tasks. It's beating out these other models. Now it's still a very, very large model, so you're not gonna be able to just run it on your average computer at home. But people are getting close to like the state of the art Opus model and GPT models with an open source model now. And from my understanding, if you buy two M3 Ultra Max studios with 512 gigabytes of and spend 20 grand, you yourself could have a computer or set of computers that will run this. Here's some more benchmarks if you want to pause and take a look, but as always, I'll link these up in the description below. I have a really difficult time testing and comparing different large language models between each other these days, just because the models I have access to pretty much do what I need them to do. So I've kind of run out of my own test. I just love seeing models that are getting smarter and smarter based on the benchmark scores. Also faster and faster and cheaper and cheaper to use. From what I understand, this is like 1/20 the cost of using something like Opus 4.6. This company here, E01 Research, used GLM5 to create a working Game Boy Advanced emulator. It actually worked on it for 24 hours. It says GLM5 finished the emulator with 3D graphical user interface and proper visualization. After 24 hours with few nudges and design suggestions, they gave it a system prompt and a hardware doc and then stepped back. They got it to do A meta loop where it did the work, tested, logged itself, and then advanced and continued on. So it was doing its own testing, logging the tests, and then fixing what it found during the tests. What they noticed about this model is that we're getting to a point now where you don't just give it a task, you give it a goal. The agent makes a plan, it then executes, it tests its plan, it adjusts, and then it keeps going for hours. On that loop of iteration. It's not waiting for your next message, it's seeking out paths to get to the end goal that you set for it. We're seeing that more and more with these most advanced models. They will actually build out step by step plans, follow through on their plan, run their own tests, notice when there's problems, and then fix the results. And this is an open source model now that actually does that. And then check this out. This is the final result. They're actually able to load a ROM on here and play Game Boy Advanced games on this. And it's got this whole graphical user interface and it's even movable in like 3D space. Look at this. They can even move the controller around. This model essentially built that autonomously over 24 hours off of being given a goal in some documents, and it just went and built it. But we also got another model this week in Minimax M2.5. We can see on the Swebench Verified and Swe Bench Pro, it pretty much falls in line with the other state of the art models. On the multilingual version, it's actually beating the other models. Models, it comes in second behind Anthropic on that browse comp, the ability to go and use a browser on your behalf. So another really, really impressive model out of Minimax. And their goal with this was to make it a really inexpensive model to use. Our goal in designing the M2 series of foundation models is to power complex agents without having to worry about cost. We believe that M2.5 is close to realizing this goal. We're releasing two versions of the model M2.5 and M2.5 Lightning that are identical in capability but differ in speed. 1 runs at 100 tokens per second and cost $0.30 per million input tokens, $2.40 per million output tokens. So if you were to run this model continuously for an hour, it would cost you a buck. If you run it slower, the price drops to 30 cents. To put that in perspective, you can have four M2.5 instances running continuously for an entire year for 10 grand. We can see the pink here on this chart is Minimax. These are the other Frontier labs here. And we could just see quickly it caught up to now it's pretty much on par with the anthropic opus 4.6 model. This is based on the Sweebench Verified. So needless to say, we got some insane new models this week. Between the new Gemini Deepthink model, which is like doing theoretical physics now. GPT 5.3 Codex Spark. Those guys really are good at naming, aren't they? We've got one of the fastest models we've ever seen. With Zai's GLM 4.5 we've got a really, really sol, state of the art model that's actually open source. And with Minimax M2.5, we've got an insanely smart coding model that's really, really cheap. Yeah, it's been a good week for LLMs. Again, don't really know how to test them or compare them because all of the models that I've already had access to kind of do what I need them to. We're just getting faster, better, smarter models that are at like another level where I don't even need how smart they are anymore because they're already doing so much cool stuff. Stuff. But some of these models are going to be really cool to test in like agentic workflows when you're using your OpenClaw bot or things like that. All right, so those are all the main stories I wanted to share with you this week, but I do have a few more that I want to run through quickly. So let's jump into a rapid fire, Starting with some news out of OpenAI, they officially started testing ads in ChatGPT. Now I did a whole breakdown video about this earlier this week, so if you want to hear my full thoughts on ads, I do break it down, but essentially the TLDR is. I think there's pros and cons, right? The pro is having ads makes it more accessible to more people because you can offer free plans. But the downside is there's like this slippery slope where it's very likely that over time chat GPT is going to optimize for people engaging in their platform and clicking on the ads more often in which sort of incentivizes them to make the ads less and less transparently outside of the responses. That scares me. But again, I did a whole breakdown about it. I don't think it's a horrible thing, but I also don't think it's an amazing thing. Yeah, Just go check out that video if you want to know more about my thoughts on ChatGPT's ads. At the end of last week, there was a rumor going around that OpenAI was going to show off their new wearable product during a Super bowl ad. And this ad even popped up here that shows Alexander Skarsgrd here looking at some sort of shiny device with like these metal earbud looking things in his ears and he's kind of like poking at it. You've probably seen this ad by now. If you haven't seen the whole thing, it'll be linked up in the description as always. Well, OpenAI's Greg Brockman said that this is fake news, and company officials told Business Insider that the ad wasn't real. It's not OpenAI and it's not connected to us at all. What we do know is that Skarsgrd himself confirms that the viral ad he is featured in is authentic. So this isn't some sort of like AI generation of him doing it. It's for an actual product. But according to OpenAI, it's not one of theirs. In fact, we did get some details this week about their device and, well, apparently it's delayed now until 2027. OpenAI's original stated goal was to ship the ChatGPT powered device before the end of 2026, but in the latest filing the company said its first hardware device won't ship to customers before the end of Febr. So yeah, likely that ad wasn't for an OpenAI product, because why would they make an ad for a product that they now don't expect to ship until 2027? So it's likely for something else. It's just nobody has said what it's actually for yet or what that little silver product was. And in the final bit of OpenAI news, they wanted to remind you that several models are being deprecated and are going away from ChatGPT. So GPT5, GPT4 oh, which is definitely a fan favorite of a lot of people. GPT 4.1, 4.1 mini and 04 mini are all being deprecated as of February 13th, the day this video is going live. And there is a lot of people sort of freaking out about it because a lot of people still really love 4.0 and well, I'm not going to make fun of them in this video, but if you want to go to this post on Twitter, I will link it up and you could go see the people that are not happy about that. We learned this Week that VEO is coming to Google Ads. They already rolled out Nano Banana Pro and now they're going to let people make video ads with AI. So I guess you upload a picture of your product and it will make a VO video of, you know, people actually using that product for your ads. The company Korea AI introduced a new workflow called Prompt to Workflow. It's like this node based builder that looks similar to like a comfy UI workflow. To me this kind of workflow just kind of confuses its things. But for a lot of people, they love this sort of node based workflow where it looks like you're, you know, building out an N8N style thing but for your videos. Not really a fan personally of that kind of thing, but I do understand a lot of people like that kind of visualization for this type of video. Runway is also rolling out a new sort of workflow feature called Story Panels. The workflow allows you to build a catalog of shots with consistent characters, locations and styles all in a few simple steps. Again, another sort of node based visual workflow builder where you give it a bunch of different images and then it connects those images and turns it into a video, but helps you keep consistent characters and scenes and backgrounds. A lot of these companies have made like the very simple version of an AI video generator where you just give it a prompt to get what you want. And now they're starting to build out more complex workflows so like power users could get in and really dial in the video that they're looking for using images in this like node based system kind of thing. Jumping to Anthropic news A couple weeks ago they announced Claude Cowork, which is essentially Claude code, but it runs on your desktop. It's like a simplified version. It could clean up file folders and do pretty much anything you can do with like terminal commands. Well, it was only on Mac when that first came out and this week they made it available for Windows. So if you're a Windows user and you've been wanting to use Claude Cowork, it is now available to you. It's available on all Claude paid plans. So if you're a free user you don't get access. But it's if you pay any amount to Anthropic to use Claude, well you should get access to it. Now we got some news out of Waymo this week as well. They're rolling out their sixth generation Waymo driver. This sixth generation Waymo driver delivers significantly higher resolution, so I guess it should you know, see things a lot better and make decisions better. It's also got a different look. It looks more like a van. I'm trying to see if I can find a picture of it here. So yeah, this is what the Waymo 6th gen looks like. Quite a bit different than the old like Jaguar based models that they were using before, but they have a lot more technology built into them. It's been a busy week. I think I covered it all. I do record these on Thursdays and put them out on Friday. So if any news came out on Friday that I might have missed, I'll make sure it makes it in next week's video. Hopefully you feel more tapped in, more looped in with what's going on in the AI world. Things are not slowing down right now. I'm just now hitting 900,000 subscribers on this YouTube channel which is absolutely amazing. I never ever thought I would grow a YouTube channel this big. I never set out to grow a YouTube channel this big, but now that it's growing at this pace, I really, really want to hit that million milestone. So if you haven't already, I would really, really appreciate a subscribe and a like on this video. It really helps my channel. It also makes sure that more videos like this show up in your feed and you stay super, super tapped in. Thank you once again for tuning in and nerding out with me today. This was a fun one and yeah, I don't ever know what to say at the end. So bye.