Open transcript
Google just released the best model ever created. That is not hyperbole. It is not only beating every other model on the benchmarks. I've tested it thoroughly and it is able to one-shot some of the most impressive demos I've ever seen. Look at this first one. So this is a Rubik's Cube. Now it's a 3D Rubik's Cube. You can generate any size of Rubik's Cube that you want. And then let's scramble it. So it is actually persisting all the colors in the right spot. It is absolutely stunning. Now some of you may be thinking, well that's not that impressive. But I've tried this with every other model pretty much and none of them are even able to come close to getting it working. And so you can see some of the colors are missing. The rotation doesn't look right. It's not persisting the colors during rotation. But now look. Now we have a 4x4 Cube. Everything looks right. Let's solve it. And we can actually watch it solving it in real time. It is so impressive. So it's just going to take a minute. And there we go. Completely solved. And we watched it solve it. Absolutely incredible. So I wasn't sure if this was going to work. So I created a 10x10 Cube. And we can scramble it up. And we can really just see everything is persisting properly. All right. So now that it's fully scrambled, we can see every side. Beautiful. Let's click solve. And we can see it solving down in the bottom left. And this is going to take a while. So while that's going, let me show you the benchmarks and tell you a little bit about the model. Gemini 2.5 Pro Experimental. This is a thinking model. It was just released a few minutes ago. It's in Google's AI studio if you want to check it out. It is completely free and lightning fast. And I'll show you that in a moment. Now, immediately, it is number one in the LM Arena. That is essentially human voters voting on the results. Here it is. Gemini 2.5 Pro and handily beating number two, which is Grok 3 Preview. 1443 ELO score, 1404 Grok 3 Preview. Now, as I said, this is another thinking model, meaning it outputs a bunch of tokens in its thinking phase as it continues to try different things, think through different possible solutions to problems, and then finally gives you the output. These thinking models have proven to be much more capable at coding, reasoning, math, logic, stuff that basically has verifiable rewards. And let's look at some of the specific benchmarks now. Here's Gemini 2.5 Pro. Here's O3 Mini, the best of the best available today from OpenAI. Here's GPT 4.5, Cloud 3.7 Sonic, Grok 3 Beta, and DeepSeek R1. This is kind of the best of the market right now. And nearly across the board, Gemini 2.5 Pro wins. Humanity's last exam with no tools, 18.8. Second place, 14%. GPT QA Diamond, which is a science benchmark, 84% versus 79. Here's the AMI 2025 benchmark, 86.7 versus 86.5, barely eking it out. AMI 2024, five points higher than O3 Mini, although Grok 3 Beta with multiple attempts does actually win. Live CodeBench, it actually scored less. The Ader Polyglot, which is another coding benchmark, it absolutely dominated. Here's MMMU, another domination. And MRCR, which is a long context evaluation, it scores incredibly, incredibly well, up to a million tokens. And yes, this model has a million tokens. So not only is it incredibly good at coding, which I'll continue to show you with some amazing benchmarks, so stick around, but you can fit a massive amount of code into a single context window. I can't wait to plug this into my Vibe coding sessions. Here's what they say about coding performance. We've been focused on coding performance, and with Gemini 2.5, we've achieved a big leap over 2.0 with more improvements to come. 2.0 Pro excels at creating visually compelling web apps and agentic code applications, along with code transformation and editing. On Sweebench verified the industry standard for agentic code evals. Gemini 2.5 Pro scores 63.8% with a custom agent setup. All right, and there it is. It is completely solved. Very, very impressive. And by the way, I didn't even show you this, but you can actually rotate it. So we wanted to solve it ourselves. We could. All right. So here's where you can find it. Google AI Studio. Over here, you can see Gemini 2.5 Pro experimental knowledge cutoff of January, 2025. It shows the latency right there, the rate limits. Although, to be honest, I've not run into any rate limits yet, and I've used it quite a bit. Here's the token count up to a million tokens. You can set the temperature. You can give it tools like usual, structured output, code execution, function calling, and search, of course, and safety settings. Turn them all off. All right, so here's another demo that I could not get working anywhere else. Literally nowhere. Create an interactive Lego building simulation using 3.js contained in a single HTML file. The simulation should allow users to place, move, and connect Lego bricks in a 3D environment. Core features. 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. Use 3.js, which we already said. Make the simulation work in a modern web browser without additional dependencies. Contain all code styles and assets in a single HTML file. We have different brick sizes, multiple colors, accurate studs, settled beveled edges for realism, slight glossy reflection, add clicking and dragging to move bricks in 3D space, rotation controls, camera controls, highlight valid connection points, and just a whole bunch of other really cool options. Now, let me show it to you. All right, so here is the demo. So I click one. Now, the only thing I noticed is sometimes the studs on the Lego bricks appear and sometimes they don't. Right now, they're not. That's okay. So if I put the brick in the wrong place, it actually just goes transparent and doesn't let me do that. It's red. So I can't actually put it there, but I can stack them pretty easily. I can click R and rotate it. I can move the entire plane around like so. I can move the camera around like so. Zoom in. Let's do a 2x4 blue plate now. There it is. And yeah, I mean, this works really well. And again, this is something where I could not get other models to do this. Let me show you DeepSeq V3 quickly. As you can see, the bricks are there, but they have these weird cylinders on top. You also can't stack the bricks. And so, yeah, it's just completely broken and there's definitely worse collision detection too. So now back to here, look how good this is. I mean, really the only thing wrong with it is the studs aren't showing. I mean, I could play around with this all day. This is what vibe coding is all about. And remember, this is all one shot. I did not give it any follow-up whatsoever. Here it is. Let's move on to the next one. All right. So next, let's try the snake game. And before you think, oh boy, he's bringing that back. This is a much more complex version. Let me read this to you. Write a Python version of the classic snake game 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. 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 boosts, reverse controls, double vision, time slowdown. Each food type includes unique animations. Evolving snake. As the snake grows, its segments become biomechanical with animated textures or shifting color patterns. Visually evolve the snake with shaders or layered sprite effects depending on the length or power-ups. Obstacles in terrain. Procedurally generate animated obstacles. Include terrain zones with varying friction or effects. Also add an AI snake opponent. And there we go. Look at that. So that's a power-up. The camera moves around. Boom. Nice effect. Time slowdown. Boom. Speed boost. And I went through the other snakes. So game over. Let's try it again. Speed boost. And there it is. Game over. So you can see this game is just incredibly, incredibly impressive. And again, this is all done with one shot. All right. And here is a simple flight sim. That's literally all I asked it to give me is a simple flight simulator. I didn't give it any more details than that. And it works decently well. I mean, it definitely works. There's not much detail to it. I like that you can see the shadow. And the shadow is kind of shimmering. Just like a shadow would at this height and with a little bit of texture on the ground. Looks very cool. Still very simple. But you can accelerate. Move in all different directions. I wonder if you can go off the map. No, you cannot. And there's a little bit of fogging, which is kind of cool. So if you wanted to create a flight sim, you can do it quite easily. Let's slow down. There we go. And can we slow to? Yeah. So we can slow to nearly a stop. So definitely some additional details needed in the prompt, but this is literally just create me a flight sim. All right. And now just for fun, I'm going to have it create me a second snake game. So I'm setting the temperature at one. So it's allowing for a lot of creativity. So the game should come out quite differently than the first one. Let's see how it does. Look how fast it is. So it's already been thinking and writing code for 160 seconds, and it's still going. And there we go. Okay. So those green pulsing food, we got one. Oh, so this is definitely hard. I got to figure out the rules. And we can see it leaves a little trail. There's my food. Here's that little blue thing. All right. So I crashed again. So yeah, very easy. One shot snake game. I can continue to add different rules and different obstacles, different power-ups. Very, very easy. All right. Next recreate the Reddit website in one file, look up Reddit to figure out what it looks like. And I clicked grounding with Google search, although I'm not actually sure if it did it or not. Here it is. So it doesn't quite look like the real Reddit website is kind of an older version, but yeah, we can see all of the posts. We have different comments. You can join different subs and yeah, pretty simple, but cool that a single sentence prompts did this. All right. Next create a single HTML file that uses 3.js to build an interactive 3D Taurus knot simulation. If you're not familiar with what a Taurus knot is, I'll show you in a minute. The simulation should include a Taurus knot geometry rendered in a WebGL scene with the following features. Use 3.js, Taurus knot geometry, default parameters, add a mesh font material with default color, ambient and directional lighting, interactivity, GUI sliders, which include radius, tube, radial, tubular, P, Q, rotation speed, X, rotation speed, Y, color wireframe, and a lot of different other really cool settings. And here it is. This is a Taurus knot. So we can move it around pretty easily. We can zoom out. We can rotate it across the plane like so with the right click. Now here are all of our settings. So here's radius. That's what this looks like. Here's tube thickness. We can make it a thick boy like that or really thin. Radial segments. So the more radial segments, kind of the smoother it looks. If we have barely any radial segments, it's more jagged. If we increase the radial segments, it smooths out. We have P windings, which changed the shape of it. We have the Q version of that. Look how cool that is. We can change the color easily. Any color we want. We can change it into a wireframe. Okay. So if we go back and we change the radial segments, you can see the wireframe is drastically different. So let's turn that back up. Let's turn the thickness down. Let's turn this and make it simpler. Look at that. Turn the wireframe back off. We have shininess. So we can see as we adjust the shininess, the light, you can kind of see it. Let me try to get it for you. There we go. We can see the shininess right there. We have opacity so we can make it transparent. We have emissive color, rotation speed. We have the ambient color. We have the ambient intensity, directional color, and directional intensity. So just overall really cool, fun to play with, experiment. You can add as many different settings as you want. You just continue to prompt Gemini 2.5 Pro. All right. Next, I wanted it to build an ant farm simulation. So create a beautiful, interactive ant farm simulation using 3.js in a single HTML file. The simulation should mimic the classic toy ant farms from childhood. Transparent side view, flat 2D plane, but rendered in 3D for depth and beauty. It should look like a clear plastic container filled with sander gel. By the way, I did not write all of these prompts in such detail. What I actually did was take kind of a very basic one-line prompt with my idea. And I asked an LLM, specifically Chad GPT, to essentially expand and add more detail to it. So that's a really good way to take your prompts and make them much more precise and prescriptive. All right. So ants digging tunnels, moving food, and interacting. Side view perspective, like a slice of earth inside a toy farm. Render transparent plastic walls. Animate ants digging, carrying food, interacting in real time. Add particles or subtle effects to enhance realism. Use smooth animations. And I actually ended up at the very end saying, convert it all to Python, actually. And it did. So here's the original ant farm game that it gave me. And I can't really do anything. You can kind of see the green food right there. I can't control the settings, but it is a nice simulation. But I wanted to be able to control things. So I simply asked it to do that. And here's the second version it came up with. Now we have the ant simulation. I can actually change the number of ants in the simulation. I can change the dig speed. We can change the food rate. So we can add a lot more food. And you can see they're kind of heading towards the food. And we can also change the time of day, which is kind of interesting. We can adjust the substrate between sand and gel. And yeah, and we can just watch the ants go. And just like always, we could add breeding and we can add death and we can add other things. So just a really cool, really easy one shot way to do this. All right. Next, create an interactive simulation of a virus attacking cells in a bloodstream. The environment should visually represent flowing blood with red blood cells, white blood cells and viruses, including the following features. Simulation elements, red blood cells, passive cells that the virus targets and destroys, white blood cells, defensive units that detect and attack viruses, various types of viruses, aggressive, stealthy, fast replicating. And we have some sliders, virus settings, number of viruses, virus replication rate, virus type selector, and so on. And I'll show you that right now. So this is the original simulation that came out. We have the red blood cells, we have the white blood cells and the little purple stuff right there. This is the virus. Now down below, we have all of our settings and we can see how it'll change it. So if we turn up the number of viruses and we reset, we could see it's a lot more. Let's turn it up even more. And there it is. We can even increase the replication rate. So now the viruses are going to replicate at a much higher rate. We can also have standard, aggressive, stealthy, and fast replicating. So let's go ahead and reset one more time. And actually let's do aggressive. Oh yeah, look at them go. Okay. But let's give our white blood cells a chance. Let's increase the white blood cells, increase their movement speed, and increase the detection radius. Now we can see the white blood cells are doing very well attacking all the viruses right there. We also can increase the blood flow speed. We can increase the simulation speed, increase the red blood cell count, which requires a reset. So there it is. And we can adjust the camera settings. Now this is all really cool, but I wanted it in 3D. So I literally just said, give it to me in 3D. And here that is. So there are the red blood cells. The purple are the viruses. We can zoom in. We can kind of watch the white blood cells go after the purple viruses. Zoom out. We can move it across the plane like normal. Let's increase the number of white blood cells. Boom. There they are. We also have all of the same settings as before. Let's do aggressive viruses. So we could see the viruses are now aggressive going after the red blood cells, killing them when they reach them. And the white blood cells are trying their hardest. We can add turbulence like so. Let's increase the blood flow speed. Red blood cell count. We can increase attack cooldown. Okay. Let's say it's a low. Boom. Look at that. Okay. So let's add more viruses and see what happens. Yeah. Just absolutely destroying the red blood cells. We can also increase the average lifespan of a virus. So they just last longer. They continue killing more red blood cells. So yeah, another impressive demo. One-shotted. All right. And the last one, a fun one. Alex, the producer created this one. Here we go. This is a surgery simulator, kind of a toy one. So we have a scalpel. Let's go ahead and make a cut. It opens up. We can make a second cut right there. And you can see we have the precision score, the stability score, which is actually decreasing. Let's make another cut. And what if we make a cut like this? Yeah. So our patient is not doing well. Let's suture them up. Okay. There we go. And when all of this is done, the stability will increase and stabilize. So kind of a fun little game, but again, just one-shotted. All of these are more or less one-shot. And when I say more or less, sometimes I just add some features to them. I didn't actually ask it to fix any code. So yeah, this is by far the most impressive coding model I have ever seen. I don't even test it against all of my previous rubric tests anymore because those are too easy. Let me know what you think of the new benchmarks. If you enjoyed this video, please consider giving a like and subscribe, and I'll see you in the next one.
