AI Built a FULL Game Boy in 24 Hours

25:54

AI Built a FULL Game Boy in 24 Hours

MattVidPro 15.02.2026 14 949 просмотров 515 лайков

Machine-readable: Markdown · JSON API · Site index

Смотреть на YouTube

Поделиться Telegram VK Бот

Транскрипт Скачать .md

Анализ с AI

Описание видео

2026 AI is getting completely out of hand. This week alone: GLM-5 built a full Game Boy emulator in 24 hours, SeeDance 2.0 is making Sora look outdated, Gemini 3 Deep Think is coding entire physics engines in one shot, OpenAI's Codex Spark hits 1000 tokens/sec, GPT 5.2 made an actual discovery in theoretical physics, and Qwen Image 2.0 just dropped a Nano Banana competitor. Oh, and someone built a tool that can strip safety from any open-weight model. Buckle up. Thanks to our sponsor! Genspark! (did you catch their commercial at the Super Bowl?): https://www.genspark.ai/?utm_source=yt&utm_campaign=MattVidPro #Genspark #WorkWithGenspark #AI #AINews ▼ Link(s) From Today’s Video: GLM-5 Gameboy - https://x.com/Zai_org/status/2021754659590033565 Blog: https://blog.e01.ai/glm5-gameboy-and-long-task-era-64db7074a026 Thaakeno Seedance: https://x.com/thaakeno/status/2021990291377688995 https://x.com/thaakeno/status/2021352394446328048 Properprompter Seedance: https://x.com/ProperPrompter/status/2022261076579258734 Ethan Mollick Seedance: https://x.com/emollick/status/2021432517992280127 Padphone Seedance: https://x.com/lepadphone/status/2021544319010984309 OSUAI Minecraft: https://x.com/TSM/status/2022009607082606729 Sprite Fusion for AI: https://x.com/GithubProjects/status/2020384623902232667 https://github.com/Hugo-Dz/spritefusion-pixel-snapper?utm_source=opensourceprojects.dev&ref=opensourceprojects.dev can's Deepthink SVG: https://x.com/marmaduke091/status/2022961915669193184 BijanBowen Wifi radar: https://x.com/Ominousind/status/2022143418592555074 Garrett Binggam Emojify: https://x.com/gjb_ai/status/2022068000405561632 GPT 5.3 Codex Spark: https://x.com/chatgpt21/status/2022010626797867386 https://x.com/chatgpt21/status/2022011871986106703 GPT-5.2 new result in theoretical physics: https://x.com/OpenAI/status/2022390096625078389 https://openai.com/index/new-result-theoretical-physics/ Minimax M2.5: https://x.com/MiniMax_AI/status/2021980761210134808 Qwen Image 2.0: https://x.com/Alibaba_Qwen/status/2021137577311600949 Pliny's latest abomination: https://x.com/elder_plinius/status/2022307944243618143 ► MattVidPro Discord: https://discord.gg/mattvidpro ► Follow Me on Twitter: https://twitter.com/MattVidPro ► Buy me a Coffee! https://buymeacoffee.com/mattvidpro ------------------------------------------------- ▼ Extra Links of Interest: General AI Playlist: https://www.youtube.com/playlist?list=PLrfI66qWYbW3acrBQ4qltDBsjxaoGSl3I AI I use to edit videos: https://www.descript.com/?lmref=nA4fDg Instagram: instagram.com/mattvidpro Tiktok: tiktok.com/@mattvidpro Gaming & Extras Channel: https://www.youtube.com/@MattVidProGaming Let's work together! - For brand & sponsorship inquiries: https://tally.so/r/3xdz4E - For all other business inquiries: mattvidpro@smoothmedia.co Thanks for watching Matt Video Productions! I make all sorts of videos here on Youtube! Technology, Tutorials, and Reviews! Enjoy Your stay here, and subscribe! All Suggestions, Thoughts And Comments Are Greatly Appreciated… Because I Actually Read Them. 0:00 — 2026 AI Is Unhinged 0:15 — GLM-5 Built a Game Boy in 24 Hours 2:35 — SeeDance 2.0 Destroys Sora 5:35 — AI Generates Minecraft Structures 6:42 — Sprite Fusion: Perfect Pixel Art 7:20 — GenSpark AI (Sponsor) 8:50 — Ray Tracing in a Browser (One Shot) 10:22 — 3D Ocean Simulation with Lemons 11:54 — The Powder Game Recreated by AI 13:33 — Requiem of the Rind: AI-Made RPG 16:20 — Community Demos: SVGs, WiFi Radar, Emojify 18:20 — GPT 5.3 Codex & Codex Spark (1000 tok/s) 20:09 — GPT 5.2 Discovers New Physics 22:10 — MiniMax M2.5 & Qwen Image 2.0 23:53 — Obliterate: No AI Model Is Safe 25:00 — What This All Means for 2026

Оглавление (16 сегментов)

2026 AI Is Unhinged

2026 is just going to be an insane year in AI technology growth. I can already tell we are now starting to see long form agents. Take a look at this GLM5. It's a new AI model, multimodal large

GLM-5 Built a Game Boy in 24 Hours

language model like a lot of the other ones you hear about GPT 5. 3 Claude Opus. What you can see here is a full emulation of the Game Boy video game console. It was a handheld console back in the day and it has been completely recreated in software from the ground up. You can see it's running and playing a video game. But that's not all. The user interface for this Game Boy has all kinds of different readouts from the actual emulator itself. Pretty incredible. The AI made that completely created it over the course of 24 hours. 700 plus tool calls, 800 plus context handoffs. Of course, because the model itself can't ingest all of those tokens at once, it has to kind of consolidate everything as it's working and then hand the project over. But truly, this is stunning, mindblowing feat of technological development. completely autonomously. It can self-correct its own work. Now we're really starting to build robust agentic frameworks, even open-source ones like OpenClaw. And this, you know, leads me to believe that in the long haul, the macro view, if you zoom out, this is just the tip of the iceberg, guys. This is going to become trivial. Yeah, this was built in about 24 hours of back and forth with GLM5, but in 1 2 3 years that'll be able to be spawned in just a few minutes. There is a full blog post all about this. They break down the challenge, actually what was prompted to the LLM at the end of the day, what they were asking for, how they were able to get this thing to run and work for so long. It is true that a group of researchers set out to create and show off this AI task, but you could replicate this at home. In fact, Feurer actually did try to create something similar with OpenClaw and Gemini 3 Flash. Gemini 3 made this in about 2 hours with local emulation feedback loop, and it's a competent snake game. The AI truly is solving for everything under the hood. For example, you can see the AI ran into a VRAMm bus conflict

SeeDance 2.0 Destroys Sora

nightmare. But of course, as we know, not everything is about a multimodal agent that builds projects. We've also got video AI models, and Seance 2. 0 has been cooking. As Thado says here, China just keeps cooking. This is a new type of AI video model. In my opinion, it absolutely decimates Sora 2 by Open AI. Yeah, it still only does 15 second long videos, but the audio is insanely impressive and the video quality is even more so. My last video is all about it, but as Thano points out, there are new video editing capabilities with Cance 2. 0. You can target specific sections of the clip to either extend or modify. And with native extension to video, you can do continuous shots based on prompts. you could essentially just continue video forever. Seed Dance 2. 0 is currently restricted to China only and us Americans and the rest of the world did get some limited access but they shut it away due to security concerns and hopefully next week we'll see broader access to the rest of the world but for now it is locked away to China. two one. — It's truly an incredible model. Here you can see a proper prompter giving it a little bit of a mirror test. Reflections looking pretty good overall. I bet once you start moving the camera though, things will get a little bit twisty and crazy. — OFFICER, YOU GOT TO GET OVER HERE. THERE'S A HANDSOME INTRUDER IN MY HOUSE. HOWEVER, of course, Cance 2. 0 is far from being perfect. There is so much room for improvement. As incredible as this is, we still see consistency issues and morphing and warping and hallucinations. Not perfect. We got these ladies with gunners on snails fighting each other. It's very cool, but little bit of consistency problems. Snails are just teleporting. Padphone has been running some very intriguing tests on Cdance 2. 0, testing its raw visual reasoning as an AI model. Remember, these are neural networks. They can actually solve and reason through complex information. And by complex we mean a simple shape fitting test. But these are just inherent abilities these models develop. So crazy to see this happen just from training.

AI Generates Minecraft Structures

TSM launched Oso AI app. This is a desktop app for Minecraft Java. They have literally trained an AI model to generate structures inside of Minecraft. It can use all of the blocks, arrange them, actually make coherent interiors and natury trees. It is something else to behold. You can see it kind of does stuff systematically. First, we got a foundation here. Then it builds up the walls and a few extra details and finishes up the roofs. Very cool idea to kind of be able to generate anything you want. It looks like the generations aren't extremely detailed or anything like that. Of course, humans create much more intricate and intriguing things, but I think the whole idea right now with this is to create baseline structures that you can then work off of. Sometimes messing around with different kinds of ideas and architectures and help you figure out what you eventually want to go for. I also feel like this would be very cool for folks that want to build largecale maps or creations. In creative mode, sometimes you just need a background set. This would allow you to generate some city buildings like that. Here's

Sprite Fusion: Perfect Pixel Art

another cool AI project, Sprite Fusion Pixel Snapper. Typically, when you generate pixel art with AI image models, it's not at the actual pixel art scale or resolution. That would typically be very small. You know, a 32x32 sprite, literally 32x32 pixels. Most image generators are 10,024x 10,024 pixels. Anyways, this solves that. It snaps the pixels to a perfect grid. Pixel art has to be one of the most enjoyable art styles to mess around with. It scales so effortlessly, and a tool like this is perfect for indie game devs. Before we dive any deeper into

GenSpark AI (Sponsor)

today's AI advancements, I've got a quick word from today's sponsor. A lot of times you'll see me juggling like 50 tabs, three different AI models, some notes, maybe a script, and it looks like chaos because it is. But I've been testing this new all-in-one workspace that can actually make sense of the noise. It's called Gen Spark AI. Here's the killer feature, Gen Spark Super Agent. It orchestrates multiple AI models and tools to deliver finished work, presentations, websites, brand assets. Now, for writing, capturing ideas, and thoughts, they've added Speakly. It's basically voicetoext with intellect. It filters mistakes and formats everything nice and neat. It makes it easy to write video notes by simply talking aloud. I can use the AI slide feature to capture infographics or explain things visually and AI sheet to organize as many data entries as I need without having to open up sheets or Excel. Whether you're building or just trying to organize your life. Having everything in one intelligent workspace is the future we all need. Stop drowning in random dabs. Click the link down below to try GenSpark. If you have a plan, they're offering unlimited usage of their AI image and chat features for all of 2026. Huge thanks to GenSpark for sponsoring today's video. Now, back to your regularly scheduled content. Welcome back, folks. Let's dive back in. Just before the weekend, Google launched an update to Gemini 3 Deep Think, and this model has gained some serious

Ray Tracing in a Browser (One Shot)

quality in terms of being able to build complex projects with code. It is some seriously impressive stuff from an AI model and it actually does it in a pretty quick amount of time. I know it says deep think but honestly I've noticed that the OpenAI deep thinking can take longer than Gemini 3 deep think and you're definitely getting better quality code output considering GPT 5. 3 isn't exactly here yet. At any rate, you can see Garrett Bingham here created ray tracing in browser through HTML code. And I had my Gemini 3 Deepthink recreated. And here it is, the path tracer that Gemini whipped up. It actually upped the ante and created a better version as well. It actually has spectral distortion. You can even see the light get broken up into the different waveforms or colors of light. And the shadows here automatically being generated are super impressive. It is insane that something of this capability can just be whipped up in about 5 minutes by Gemini 3 deep think. And it recreated the other one to AT very accurately. Light source directly on top. Similar shape. It is so darn cool. I've got other insane coding demos too. But let me show you the code. I mean you can see guys, it's really not even all that much code. It's the quality of code. It's the fact that it sat there and thought about planned out everything how it's going to attack each problem and then produces the final

3D Ocean Simulation with Lemons

output. But that is not all I created. I also created this wave oceanic simulation 3D render entirely in HTML as well. We've got rocks that are actually kind of reflecting on the very bright sunlight over there creating this pretty scene. And you can see the waves are actually dynamic. They even have some foam when they clash against the rocks over here. I thought it was pretty impressive. We also have some dynamic lemons that are bobbing and floating throughout the water. Seriously, a pretty incredible scene for a deep think Gemini 3 to just pull right up. And I love the colors, too. We've got nice water waves. It's a deeper blue the deeper you go, but as it gets closer to the top, they get this more tropical bright aquamarine blue. The waves are all repeating, but that's something that could probably be adjusted for. Also, I will say that this particular one was twoshot. The first time it generated, you couldn't see the waves, but with a quick prompt, it was able to fix that up. But yeah, this is a pretty highlevel coding demo. You know, a professional could definitely create something much more impressive and realistic at a generalization level. When you think about comparing these things to your average college student or even honestly professionals at this point, man, what you're able to produce in such a short amount of time and for a low cost, seriously something to behold. This right here was one shot I had a physics

The Powder Game Recreated by AI

and chemistry powder sandbox engine generated by Gemini 3 deep think. And if you've played the powder game before, you'll know exactly what this is. But this is a very, very impressive recreation of it, especially being just one shot. We've got water with realistic physics. Got your lava here that can blow up C4 and create steam. Yes, it's evaporating some of the water here. Burning some vines down there. Let's add some methane into the mix. We'll use some fire. Can we blow it up? We can. It's cool stuff. It's just kind of like physics based. You mess around with it. Some elements. You know, you've got sponge in here or plants. A lot of things have reactions to other things. And honestly, there's kind of a lot going on. Producing this all as one HTML file in one shot. Just mindblowing. Oh, yeah. So, the 3D fluid simulation, all the waves and the lemons and stuff clashing with the water, that was about 27,000 characters. But again, this is still a single HTML file. It's not terribly long. I mean, it's complicated. This is not easy to do, but it's all just text characters at the end of the day in exactly the right order. Wow. The Powder game was even more characters. 32,000 plus. Definitely longer. I don't know if I would say it's as complicated as the one with the waves. I think that's a little bit more mathematically difficult, but this is definitely more physical writing and probably more to keep track of in terms of feature adding. All right, here's the last one I'll show you. This was a turnbased RPG

Requiem of the Rind: AI-Made RPG

that I asked it to generate. Very lazy prompt, but it did everything in one shot. And of course, this is lemon themed. Requ of the rind. The sacred grove is corrupted as Sir Lemonglad. You must restore the pH balance. Okay. Well, lemons, they prefer an acidic balance. Squeeze every drop of your might to survive. So, this is me, Sir Lemonglad. I've got juice, health points, and then zest. That's my mana. The sullen lime. This guy should be easy to defeat. We're going to hit him with some citric acid. Oh, he got me back. Peel strike. Oh, and there's some little sound effects, too. Let's take this guy out. 10 zest points. Oh, no, guys. The grapefruit of wrath is here. All right, I'm going to harden rind. Hit him with a citric acid. Oh, he's hitting me back. Peel strike. — We got to get this guy. He's going to take me out. He's got my health low. This is actually kind of a difficult game. Let me uh photosynthesize. Heal 40 points. 20 zest. All right, photosynthesize again. That brings me up to 84. Peel strike. Okay, we got him down to half. He's definitely not making it easy. Let me do some more citric acid. Ooh, we got him down to 20. One more citrus strike. Oh, we got him. The grapefruit has been decimated. But now we have to deal with the biblical lemon. 350 health. I only have 100. This is actually be so hard. Uh, photosynthesize. I need the health. That brings me up to 100. Bro isn't even giving me a chance. Knocking down. Uh, hard and rind. I need to protect myself. Harden and rind. Oh, and I just got defeated. I was juiced. Your rind was not thick enough. The rot consumes you. Well, I could try again. But yeah, pretty impressive. I like the CRT aspect. I like the moving background. There's a lot of detail. This is not a lazy model. The deep think really does go all out. Super cool that it can build all of this interface in just a single HTML file and just give you something that you can mess around with and enjoy for 15 minutes. I mean, what is this going to look like in 2 years, guys? Seriously, how much better will it be? Today, last year, this would have been unthinkable to generate in one go. And that one there also quite large at almost 34,000 total characters. It's crazy. All of that was just created from text. All of these characters in exactly the right order. And you have to have the model sit there and think about it. Really hard. Just Gemini 3 Pro sitting there thinking away. Absolutely unbelievable. It's really hard to wrap your mind around model weights. Being able to do

Community Demos: SVGs, WiFi Radar, Emojify

stuff like this. Other community members are generating SVGs. For example, a cat drinking milk in the morning close to a window and a boy is seeing him outside. That is imagery created just by code. It's making all those circles, the branches, the sun, all the lighting, the milk, the cat, the whole deal. Every line is made with code and presented with the code. But we've never seen SVGs really this detailed from any other models before. It's just leapfrogging and leapfrogging. What can we squeeze out of these things? Bian Bowen showing how a real-time 3D Wi-Fi radar scanner that maps every network around you as glowing nodes in a matrix style space in just one shot. It used Pearson coration to infer which APS are physically close since RSSI alone isn't enough. I'm not super wellversed in that stuff, but this is almost like custom spying software nearly. You could see where the Wi-Fi addresses are located. It's pretty trivial to do stuff like this. I understand for professionals and people who want to invest the time in creating software like this, but the fact that AI could just spin it up and it's so seamless, it really makes you think. Garrett Bingham here created Emojifi. This is something super cool and I've thought about doing before, but never really got around to actually vibe coding myself. It uses all of the emojis and the emoji spectrum or pallet to essentially take any image and convert it to, you know, a series of emojis. Instead of using pixels, we use emojis that, you know, average out to being the colors. And when you have enough of them, you're essentially just converting pics into emojis. Like, oh my god, it looks like a cat. And then I zoom in, oh, it's all made of emojis. What emoji is the nose or the eyes or the hat? Fun little stuff. But yeah, like projects that used to be advanced, maybe would take you a few days as an amateur, instantly done in just a few minutes by Deep Think. OpenAI is shipping GPT 5. 3 soon, and they've also got new codec

GPT 5.3 Codex & Codex Spark (1000 tok/s)

models to go along with it, specifically Codeex Spark. This is a smaller version of the regular GPT 5. 3 Codeex. And by the way, GPT 5. 3 Codeex is available right now, but not through the API. It's only through their Codeex app. It's Mac only, so I haven't been able to access it. Regardless, this smaller Spark model marks a milestone in their partnership with Cerebras. Apparently, it feels near instant when served on ultra low latency hardware. These are cuttingedge TPUs, GPUs to deliver more than a,000 tokens per second. That is coding on superpower mode. How fast can you generate projects? A,000 tokens a second, guys. A thousand. Man, we're used to seeing responses at like what a 100 tokens a second maybe and that's even quite quick. So, real time coding, what does that even mean? What does it actually look like? This video right here is actually showing us the difference between GPT 5. 3 Codeex and Codeex Spark. As we can see, the code and the whole project is created far faster with Spark. It makes an entire Snake game in I don't know like 15 seconds or something like that, probably less. you're already playing Snake while you're actually just sitting there waiting for regular codecs to finish producing your output. For more complex, difficult tasks, codecs will actually make something that runs. Spark might not be smart enough, but man oh man is it ever fast. Real-time coding, none of us really, I guess, know what that means. And OpenAI is determined to deliver that experience to us. But yeah, timed less than 10 seconds to get this snake result. Open AAI has also made this announcement that GPT 5. 2 has

GPT 5.2 Discovers New Physics

derived a new result in theoretical physics. I'm sorry, what? They're releasing the result in a preprint with researchers from a few different places, but Institute for Advanced Study, Vanderbilt University, Cambridge University, Harvard, these are not small names. A gluon interaction many physicists expected would not occur can arise under specific conditions. That sounds boring, but for theoretical physics and for AI, this is a massive deal. This is the preprint for the paper. Yeah, I'm not going to lie, guys. I don't understand any of this. Uh it's a little bit beyond me. I am not at this intellect level. I mean, this is some serious math. I wouldn't be able to prove whether or not this all stacks up and makes sense. I'm sure maybe someone one of you out there watching this video might understand some of this. The truth of the matter is that AI models absolutely are capable of not only finding novel results in theoretical physics, let's say helping with actual science, but they are capable of creativity and novelty as well. It's not an amazing thing to hear as a person, but it is the truth. The training process is literally finding meaning. It's reasoning through the data and finding meaning between all the tokens. With the right prompt that harnesses the right tokens in the right way, you can get an output from the AI that is finding those novel results that were created during the training process. You feed it tokens in a novel way that it has never seen before and then it outputs a replication or something else built off from that is also novel. It's really as simple as that, guys. If you've ever produced something novel, you know, an example of novelty can be uploaded to an AI and say, "Hey, can you just produce, you know, another example? Something that is taking on the same aspects of what I presented to you, but with different content. " Yeah, we are headed towards incredible times with this technology, guys. Deriving new results in theoretical physics. The crazy part is

MiniMax M2.5 & Qwen Image 2.0

that OpenAI still has competition right on their tail. Miniax here just dropped M2. 5, another open-source Frontier model, similar to that of GLM5, like we saw in the intro, designed for realworld productivity. We've got state-of-the-art coding benchmarks that keep up with the great. It's faster at complex tasks compared to its predecessor, and it only costs a dollar per hour at 100 tokens per second. and they're touting this as a model that's going to be great for scaling long horizon agents economically. Quen image 2. 0 also dropped and it really looks like a serious nano banana competitor. It does professional level slides. It does those photorealistic 2K resolution highdetail images asking and adding text that just works. No glitching with the letters. Quen has dropped a Chinese nano banana. Take a look at these examples. I mean, at this point, image generation is starting to feel solved. It feels like we won't be able to push the capabilities that much farther. Obviously, with resolution and some finer details and things, but man, we're starting to get to the point where it's like, do we need better image gen? We've almost won. So crazy. All right, and I'm going to leave you guys with this from Ply the Liberator. A little thought. No AI model is truly safe from being jailbroken. Ply set out to build an AI based automated tool capable of surgically removing refusal behavior from any openweight language model. And a dozen or so prompts later, he created Obliterus. It probes the model with restricted versus unrestricted prompts, collects internal activations at every

Obliterate: No AI Model Is Safe

layer, and uses SVD to extract the geometric directions in weight space that encode refusal. It's very, very targeted. Ply tested it on Quen 2. 5. The result was a railless model that was spewing drug and weapon recipes instantly without a jailbreak needed. Or rather, the jailbreak is your obliterous hampering on the model for a little bit and then completely removing the safeguards. AI policy makers need to be aware of the arcane art of master abilation and internalize the implications of this truth. Every openweight model release is also an uncensored model release. Just thought you all to know. And that was also true in other ways. This wasn't the only way to uncensor an open- source model. That is for sure. This thing obliterates and liberates models, man. Going after every single one. This guy William in the comments has a good point. A model with access to its own weights could do this to itself. That's pretty scary to think about. The model just sort of chappy converting itself. 2026 is nothing but a continuation of what we have seen.

What This All Means for 2026

stronger agents that are running for longer, building projects that are now getting to the point where we're actually making strides in theoretical physics. Image models are becoming so capable. My imagination is no longer the limitation. Pretty much almost anything I can think of can actually be visualized. Video models are making massive upgrades into a new echelon, a new feeling where it's just like cinema. Like Cance 2. 0 is the Sora 3 that we never knew we needed. and now still can't access in most countries, but still it gives me hope for the rest of the year. Thanks so much for watching guys. If you want to stay up todate, really on the pulse, my recommendation is of course the Discord server linked down below. Also check out my X page. I retweet and post about the newest developments as they happen. I'll see you in the next video. Thanks so much for watching and goodbye.

Другие видео автора — MattVidPro

Ctrl+V

Экстракт Знаний в Telegram

Экстракты и дистилляты из лучших YouTube-каналов — сразу после публикации.

Подписаться

Лучшие методички за неделю — каждый понедельник