Vollständiges Transkript von „Google just updated Gemini 2.5 Pro and it's insane...“. Klicke auf einen Zeitstempel, um das Video an dieser Stelle zu öffnen. Du brauchst ein anderes Video? Nutze den YouTube-Transkript-Generator.
0:00 Google just dropped an update to the best model ever created, Gemini 2.5 Pro IO Edition that is brand new as of today, May 6th, and I have tested it thoroughly, and I am going to show you its incredible performance. This is able to zero some of the most impressive code projects I have ever seen come out of a model. All right, let's start with the Rubik's Cube again. Now, Gemini 2.5 Pro, the previous version, was able to do this in one try, but now the new version actually has more features. So, obviously, you can rotate it, you can zoom in and out, you can rotate it by all of these buttons, which is really nice. And of course, we can easily solve it. There's a scramble button just like last time. It scramles even faster,
0:45 scrambling in 12 moves. And let's solve it. Absolutely amazing. We did a 10x10 last time. Let's do a 20 by 20 this time. So, I'm going to scramble it. All right, there we go. A fully scrambled 20 by 20 cube. This is incredibly complex. And let's solve it. Now, this is going to take a while, so stick around. I'm going to show it to you when it's done. All right, let me tell you about the updates that they made because this was a very big improvement over an already insanely good model. build rich interactive web apps with an updated Gemini 2.5 Pro and that is what they really focused on building web apps and I tested that. I'll show you that in a minute. So, this is called Gemini 2.5 Pro Preview IO
1:33 Edition and that's because Google IO is coming up soon. That's their conference showing off all their new cool stuff. An updated version of 2.5 Pro that has significantly improved capabilities for coding, especially building compelling interactive web apps. We're going to release this update at Google IO in a couple weeks, but based on the overwhelming enthusiasm for this model, we wanted to get into your hands sooner so people could start building. Now, Gemini 2.5 Pro is the best coding model out there. It's really just uncontested, but I actually still found Claude 3.7 to be better for agentic coding. So, in terms of just zeroing entire applications, entire games, entire simulations, Gemini is better. But for function calling and agentic use, claude was better. Until now, they invested
2:18 into making tool calling much better with this new version of 2.5 Pro. So these improvements extend to other coding tasks such as code transformation, code editing, and developing complex agentic workflows. And like the other Gemini models, it has a million token context window, which is absolutely crazy. And it can ingest video, images, audio, pretty much everything. Now, LM Arena takes the number one spot. The previous spot being the previous version of 2.5 Pro at 1437 on the arena score. The new score jumped 11 points to 1448. That is not a small feat here. And in the web dev arena leaderboard, we jumped from 1272 to 1419. That is a 147 point improvement.
3:07 Checking back quickly on the 20x 20 cube, it's still solving. So, we'll come back to that in another minute. All right. Next, let's check out the benchmarks. Gemini 2.5 Pro preview on May 6th. Now, we have some updated numbers. GPQA Diamond 83%. Now, the only comparable model is OpenAI 03. Here we have Amy 2024 at 83%, 03 beats it. Here's Ader Polyglot at 76%. 03 beats it. Sweetbench verified 63 versus 69. 03 beats it. Simple QA, MMLU. Basically, it's either even or 03 is winning. Now, here's the thing about benchmarks. They only really mean so much. From my experience, the only model that is able
3:53 to output incredible long code to build sophisticated applications and games is Gemini 2.5 Pro. And the price, this is really the killer. Not only does Google have a top tier frontier model, but it's also by far the cheapest for that intelligence level. $2.50 per million tokens, $15 for a million output tokens. That is compared to 03 at four times the cost of input and about three times the cost of output. And in fact, the only model that is cheaper than Gemini 2.5 Pro is Deepseek R1, which is open source, which makes sense. So, with that, let's check back in on the cube. Oh, and look, we just looked at it and
4:38 it finished right in front of us. Unreal. I just happened to click back to this tab and it finished right in front of us. There it is. 20x 20 fully solved. I am so impressed by this. And if you want to try this new model and build really cool stuff on top of your documents, you need to try Box AI. Box is sponsoring this video and they're going to have Gemini 2.5 Pro IO Edition soon. Box allows you to store all of your company's documents, any type, and leverage artificial intelligence to extract insights, answer Q&A, create automations and intelligent workflows with their agents. They have a super easy to use API, so you can build really cool stuff on top of Box AI, and they handle the entire rag pipeline for you,
5:24 so you don't have to worry about all of those complexities. Box is trusted by 115,000 enterprise organizations, including my own, and they have enterprisegrade security, compliance, and governance. So, check out Box AI. I'll drop a link down below. Try out Gemini 2.5 Pro IO Edition with them soon. And thanks again to Box. Now, let's get to the next test. All right. Next, I wanted to create a 3D floating island simulation. And in our test of 0304 from OpenAI, it did okay. So you can see far from perfect. But now let me show you the prompt and then I'm going to show you the actual simulation. So create a dreamy low poly floating island scene with dynamic lighting and gentle animations in a single HTML file. And it
6:10 only had to think for 25 seconds. And here it is. Look at how good that is. We have these clouds floating around. We have the larger clouds in the background. You can zoom in, zoom out. The lighting looks good. The shadows look really good. Look at that. Now, we have a bunch of sliders as well. We can change the sun intensity. Here's it up and down. We have ambient light. Look at that. Perfect. Crystal glow center. So, that's that crystal right there in the middle. Let's go ahead and adjust that. Let me show you what that looks like. Okay. Now we have the near fog, which I don't really see a lot of fog, so it's kind of weird, but it does affect the lighting. We have the fog
6:57 far. Okay, doesn't really seem to be doing all that much. The island bob speed. You can see there we can change the island bob speed. It's still relatively slow, but you can make it extremely slow or a little bit faster. We have the island bob amount, so how much it bobs. So there we go. But we can see it's really going up and down now. And the cloud speed. So we can adjust the cloud speed. You can see it's going a lot faster now. All the way up to there and slower. So very cool. Now the first one that it output was great, but I just iterated on it and I said, "Now add a bunch of sliders to control the different aspects." And that's all the sliders that I just showed you. Next, a Golton board simulation. So create a self-contained HTML file for a gulton board simulation using clientside
7:43 JavaScript and a 2D physics engine like Matter.js. Simulation should be rendered on an HTML 5 canvas and meet the following criteria. Single file canvas size should be around 500 by 700. Fit on a standard screen without extensive scrolling and zooming. Physics utilize 2D rigid body physics. Obstacles the pegs the containment and so on. And so it output a bunch of code, but there was one problem with it. So this is what it originally looked like. And as you can see, the problem is there's no hole in the top funnel, so none of the balls were coming through. So it did take another iteration to get it right. And I simply said there's no hole in the top funnel, so the balls get stuck. Also, there's no sliders. Here it is. So we can see the Golton board simulation. And
8:28 what we should see is a nice bell curve depicting the probability of the balls basically falling up the middle versus kind of falling on the two sides. Now we have a bunch of sliders we can do. So drop interval. So if we turn that up, it's going to drop way less frequently. And if we turn it all the way down, it's going to have many more balls coming. Now, as you can see, they kind of get stuck, which is nice. That's how physics works. So it's a nice physics simulation there. So I'm going to turn it down. Let them come through. Next we have the max number of balls. We could turn that down. So that's the total number of balls that are going to come in this single simulation. We have the number of bins. Okay, so I can do that and I can reset it. So you can see there are more bins down there. Now we have the number of peg rows which we can adjust here. So
9:15 go ahead and so all of this what it'll do is just change the probabilities. We have the peg radius. Okay, we have the ball radius, the ball bounciness. So, let's go ahead and change a few of these. And gravity. See what that looks like. So, gravity's way up. Let's turn gravity way down and see what happens. All right. So, now it's really floating slow. Now, one little issue I do notice is that the pegs should be alternating. They're all one on top of each other. So, it's not as good of a simulation as it could be. Close enough. All right. Next, of course, we had to do a flight simulator. So, write a complete HTML file that creates a simple flight simulator using 3.js. The scene should feature a thirdp person camera that follows behind and slightly above a basic airplane model. It could be
10:00 created using basic geometry, but should resemble a plane. Simple controls. WD to turn right, left, pitch, up, down. Spacebar to increase speed and shift to slow down. Gravity is not required. The plane should stay in the air. A basic 3D environment. Add a ground plane with a texture. Populate the world with a few simple buildings. The plane should be able to fly around them freely. Keep everything in one HTML file. Do not import any external 3D models, although that might be cool for another test later. And here's what that looks like. So, really nice. I think it would be cool if we did add some gravity into here. So, let's hit the space bar to speed up. And then shift to slow down. It seems to work. Look at that. We can see the shadow that seems to be shimmering off of the ground. Yeah, very basic, but really cool. Next, we wanted,
10:47 of course, to bring back the snake game, but it's not enough just to create the snake game. We want a complex version of it. So, write a Python version of the classic game Snake using Pygame, but with unique, visually stunning, and complex enhancements that make it incredible to watch. In addition to standard snake mechanics, implement the following features: dynamic visual effects. The snake leaves behind a glowing fading trail. And while I read this, let me show you what the previous version of Gemini 2.5 Pro was able to create, cuz it was really good. And then I'll show you what this one was able to create. The background pulses with subtle color gradients and transitions when food is eaten trigger a particle explosion effect. Different types of food grant temporary powers, speed boost, reverse controls, double vision, time slowdown. Each food has unique animations, and the snake evolves. So,
11:34 let's give it a try. And here it is. So, we have little walls trying to get the food. We have an AI version of the snake playing against us. It obviously doesn't actually end if he runs into me. See, he can go right through me. So, I think overall the other version was actually better. All right. Next, I used Punit's particle simulation and all I did was download the video, upload it to Gemini 2.5 Pro, the new version, and said recreate this. So, here you can see I uploaded the video. Oh, it's 36 seconds long. It took 10,000 tokens. Remake this simulation. Exactly. The first thing it did is actually give me a bunch of individual files. So, I corrected that. I said, "No, give it all to me in one big file." And so, that wasn't anything
12:20 it did wrong. I just didn't specify. And then I said, "Okay, now allow me to rotate the shapes with my mouse." It added that. No problem cuz it didn't have that. And then finally, now add sliders to control different settings of the simulation. Let me show you the final result. So, here it is. So, you can see I can rotate it. Let me show you a couple different colors. If I click it, it morphs into a different shape. Okay, we can see other colors again. Really cool. Let's do the rainbow color. Really, really nice. All right. Now, we have some sliders at the bottom. We can control the number of particles. So, if we turn it up, we actually have more particles that make up the shapes. If we do less, we have less. I actually think the less looks really cool. Let's just leave it about there. We can change the
13:06 size of the particles. So you can see we can have like really small particles and they have a nice glow to them as well. And then of course we can turn it all the way up. And now it's a little bit too much glow. So turn it back down. We also have the morph speed. So if we turn it down it morphs really quickly. Okay. We have a donut shape. Now if I turn the morph all the way up to 5 seconds it's going to take more time to morph. And here we have a 3D plane. And so it got all of them. It got all of the shapes. Really impressive. And then we have the auto rotate speed. And of course, we could just click here for changing the shape. And we're back to the sphere. Next, let's build a Lego simulator. So, create an interactive Lego building simulation using 3.js
13:52 contained in a single HTML file. The simulation should allow users to place, move, and connect Lego bricks in a 3D environment. Core features. Create a realistic looking set of Lego bricks with accurate dimensions, colors, and textures. Implement a grid-based snapping system that allows bricks to connect properly. Ensure proper collision detection so bricks cannot occupy the same space. Create satisfying visual and audio feedback when bricks connect. All right. Technical requirements. Use 3JS. Make the simulation work in a modern browser. Contain all code in a single HTML file. Optimize for performance. Create various brick sizes. Include multiple colors. accurate studs and so on. Here we have the 3D plane and I select a brick and I place it down and it has a nice little
14:37 bloop noise. So there we go. Now R to rotate does not seem to work. So that's one thing that is missing. I can overhang it. If I hit escape, I can rotate it. Now you can see definitely not perfect, but still really good. And by the way, yes, you can remove a piece just like so. Now, let's select another color, another size. I'll put that there. We have a 1x4 plate. So, the placement is definitely far from perfect, but I'd say this is still really good. So, yeah, there it is. And you can also save it. You can load previous ones. And you can clear. All right. Next. I'm a kid of the 90s, so of course I grew up with Tamagotchi and I
15:23 wanted to recreate Tamagotchi. If you're not familiar with this, it was this little like keychain game. You got a little character and you had to take care of it. You could feed it, you clean up after it, and so on. I used AI to help me create the spec for it. Here's what that looks like. Build a Tamagotchi game in Python. Use this spec as guidance. Core loop. Timebased progression. The simulation operates on a real-time clock. User interaction. Players interact with the virtual pet through a set of predefined actions, status updates. The pet's state updates over time in response to the user's actions. It has hunger, happiness, discipline, health, age, and weight. Now, as your Tamagotchi eats and progresses and time goes on, it evolves and eventually dies. User actions. You can play with it. You can clean,
16:08 discipline it, heal it, turn the lights on and off to make it go to sleep. It has evolution stages. Baby, child, teenager, adult, senior. Behavioral mechanics. The pet signals its needs. All right, let me show it to you. Pet name. I'm going to name it B. That's the name of my dog. And here we go. So, you can see the age is ticking right there. There's my Tamagotchi. Right now, it's incubating. It's still an egg. I think at five ticks, it hatches. There we go. There's the Tamagotchi. And we have hunger, happiness, health, discipline, and weight. We can see discipline is low. So, let's go ahead and discipline it. Now when you discipline it, the happiness goes down. So let's play with it. Happiness goes up. Now hunger, let's feed it. Okay, now it's 100% fed. Now
16:54 discipline still a little bit low. Let's discipline it again another time. And then happiness went down. Let's play a couple times. And there we go. It evolved. Now it's a child. And if I turn off the lights, B went to sleep. And we can see while it's sleeping, it's actually gaining on these different statuses. Now, interestingly enough, for some reason, there's no poop. And it did poop previously, and it was kind of funny to look at, but for some reason this time it didn't. Ah, all right. So, I restarted it and now you see the poop. And it says Bose Space needs cleaning. So, we'll clean it. And there we go. And of course, yes, it is a shaking, vibrating, weird poop. And yep, that's what this looks like. Very simple version. I could easily iterate on this and uh have more fun. But there we go.
17:41 Very nostalgic. All right, we have a new test and thanks to producer Alex for coming up with this. We wanted to create the game Doom, the old school version. So, retro FPS game in HTML. Create a retro style 3D first-person shooter game in one single HTML file using 3JS inspired by the original Doom 1993. The game should run fully in the browser with no dependencies except 3JS and standard web APIs. The code should not use modules or external assets beyond what's embedded. So, maze-like level layout with textured walls and floor, ambient lighting, optional dynamic lights for effects, skybox or dark fog background to create atmosphere. Mini map. Okay, so keep that in mind. Firsterson camera using pointer lock
18:27 controls. WD for movement, jumping, collision detection, health system, ammo system. We have different weapons. So, we have pistol and a shotgun. And you can switch between them with keys one and two. There's muzzle flash, shooting around. So, you'll see some of these got implemented, some of them did not. Multiple types of enemies with different speeds and health, simple AI, chase the player, shooting impact. All right, let me show it to you. Here we go. Start game. And there it is. So you can see there's an enemy and if I shoot at it, very nice. It disappears and you can actually see as I'm shooting the ammo gets reduced and then it has to reload automatically. I can move around. Okay, look at the mini map up in the top
19:13 right. That's something I'm super impressed with. All right, this guy's coming after me. I'm switching to the shotgun and one shot at him. Now, I'm used to inverted mouse and uh this doesn't have that, so forgive me for that. There we go. These take two shots. And let's make sure we clear out all the enemies in the maze. Here's another one. Boom. All right, reloading. So, they move a little bit slow, but I think it's amazing. All right, last enemy. Let's get him. Oh, we got to reload. Run away. There we go. And we won. We cleared out all the enemies. So very basic version can easily iterate on this as well. And next, tower defense. Build an HTML 5 canvas-based tower defense
20:00 where waves of balloon enemies health 1 to 10 color-coded green to purple follow a winding path. In planning mode, players place five tower types: dart, cannon, ice, sniper, laser, each represented by a unique colored dot and range preview on hover. Press start round to begin. All right. So, dart tower fast single target shots in burst. Cannon slow splash damage. Ice slows enemies. Sniper high damage long range. And laser continuous beam with a visual can laser line. Okay, so let's try it out. Here we are. Okay, so very simple graphics, but should be good. We have wave number, how much money we have, and how much lives we have. So, let's start with a dart tower. And let's put it right there. And I'll place a second
20:45 dart tower right there. Let's start the round. Okay, so pretty easy first round. And every time we kill an enemy, we get more money, which is nice. And we have some bigger, more powerful enemies coming through. All right, wave one done. Now we have $210. Let's add an ice tower right here. So, right when it comes out through that corner. And I'm going to add a sniper tower right there. All right, let's start the next round. There we go. We can see the ice tower is slowing down enemies and everything else is killing them. Okay, so we have more powerful enemies here. Looks like we're doing pretty well. Okay, done. Now, let's add a laser tower, which I think will be
21:31 really cool. Let's add it right there. And let's go ahead and add a second one right there. Let's go ahead and click start. There's the laser. How cool is that? So, it works really well. And this wasn't in one try. There are a few issues that we had to solve. So, initially we provided a screenshot. It doesn't make a path or start around. No console errors. And that's all we really had to say. Then make it so I can deselect with rightclick because right now when I place a tower, I can't click anything else unless I change a tower. So, little fixes back and forth. I think we went back and forth like two or three times. And then, yeah, we finally got this game. And by the way, we're still going. All right. And next, we created this rudimentary sketch for this drawing app. So, we have colors at the top. We have the actual sketch area here. We have
22:17 different tools and different sizes. I mean, it really could not be rougher. All we said, can you code this app? Here it is. Exactly. Exactly what we asked for. So, here's blue. Here's yellow. Orange. Red. Great. Now, let's use a crayon. Okay. So, it's just a little thicker. We have a brush. And we have the bucket. We can change the size like so. And yeah, all it took was a single image and it created this entire interactive web app. All right. Next, I wanted to recreate the front end of AI Studio because this version of the model apparently is much better at frontends. So, what did I do? I typed out rebuild Google AI Studio's front end and I took a screenshot of it. So, very meta. Here we go. It took 13
23:03 seconds and it output something weird. I said rebuild Google's AI Studio front end did not work. Okay, it gave me some weird point thing. Anyways, I said no, I mean give me the full HTML file that recreates it. Now, it thought for 22 seconds after that, there it is. It is very accurate. I wouldn't say perfect, but it's very, very accurate. There's the get API key. We can adjust these different settings. Toggles for structured output, toggle for code execution, function calling. We can add stop sequences, change the top P, change the output length, and yeah, so in terms of zero shoting, one try, getting AI Studio just based on a screenshot, I'm going to give this an 8 out of 10. It's quite good. But I think I found a test
23:48 that no model passes cuz I've tried this. I saw it online and not even this version of Gemini could do it. So here it is. How many cubes are missing to make a full cube? So if you count them out, this is 4x 5x3. So there are 60 cubes in total. So you're supposed to count how many are missing. And interestingly, I said, tell me the answer to the question in the image based on the image. So I didn't even give it the question. And the answer is 14. You just got to count them out. Let me break it down. Determine the size. 64 small cubes. So it already got the size wrong. It's not 4x4x4. It is 4x 5x3 and so of course 20 cubes missing is not correct. So I think this is going to be
24:34 the new benchmark. Whichever model is able to solve this problem is probably going to be the best model. So that's it for today. The best model just got better. Let me know your thoughts in the comments below. If you enjoyed this video, please consider giving a like and subscribe.
