← Back to video archive

Airdroplet AI summary

Gemini 2.5 Pro is a coding GENIUS

April 2, 2025Matthew BermanAI score 95142,943 views

Watch original on YouTube ↗

AI-generated summary

Gemini 2.5 Pro: The Free AI Coding Genius?

Google's latest AI model, Gemini 2.5 Pro, is making waves, and the big news is that it's now free for everyone to use through Google AI Studio and the Gemini app. This is a pretty big deal because it's being touted as potentially the best model currently available, especially when it comes to coding and creative tasks. The presenter seems genuinely impressed, calling it the best model they've ever tried and highlighting its high simulated IQ score.

The summary below breaks down what makes Gemini 2.5 Pro stand out, based on numerous examples shared in the video.

Key Capabilities and User Creations:

Here’s a rundown of what Gemini 2.5 Pro can do, illustrated by real user projects:

  • Availability & Access:
    • It's completely free via Google AI Studio and the Gemini app.
    • The rollout happened fast because Google wanted to get this powerful tool into users' hands quickly.
    • Mobile access is also becoming available.
    • Like many popular AI services, expect rate limits due to high demand.
  • Technical Specs:
    • Features a massive 1 million token context window. This means it can process and remember a huge amount of information at once – like an entire video transcript or large codebases, way more than most other models.
    • Scored a simulated 130 IQ on a Mensa Norway test benchmark, placing it above competitors like GPT-4.5 (around 100) and Grok (much lower). The presenter feels this score reflects its high performance.
  • Coding & Development Prowess:
    • Office Simulation: A Google engineer created a complex office simulation, complete with little Sims-like employees, meeting areas, and even a ping pong table, just by prompting the AI and iterating.
    • Drawing to App/Code: Users are literally drawing interfaces or objects on paper, feeding the image to Gemini 2.5 Pro, and getting functional code or 3D models back.
      • One user recreated the Google AI Studio interface from a hand drawing.
      • A Google researcher turned a drawing of a cake toy into a 3D model using OpenSCAD (a type of code for 3D modeling) and then actually 3D printed it! This workflow from simple drawing to physical object is mind-blowing.
    • Game Creation: Even people with zero game development experience are building games:
      • A 3D Flappy Bird clone was created iteratively over an hour using "vibe coding" (more on this below).
      • A fully playable Galaga-style arcade shooter was generated instantly using Pygame (a Python library for games).
    • Web Development: Created a stunning liquid metal shader effect (think shiny, flowing metal) for an Apple logo based on a research paper, seemingly in one go. You can even try the demo yourself.
    • App Development: Generated a complete, functional iOS drawing app using Swift UI (Apple's framework for building interfaces) from a single prompt.
    • Blender Integration: The team behind the Klein VS Code extension successfully generated their 3D logo directly within Blender using Gemini 2.5 Pro, after another leading model (Claude 3.7 Sonnet) failed.
    • Physics Simulations: Capable of generating complex, interactive 3D physics simulations using three.js (a popular web graphics library) from text prompts. Examples included demonstrating electromagnetism (solenoid) and Einstein's General Relativity.
  • Content & Analysis:
    • YouTube Timestamp Generation: The presenter successfully used it to generate near-flawless timestamps/chapters for a YouTube video by feeding it the URL. It processed almost 300,000 tokens (well within its 1M limit) and watched the video frame-by-frame for accuracy, not just relying on the transcript.
    • Personal Intelligence Agency: One user built an open-source system that scrapes news sites 24/7. It cleverly uses the cheaper/faster Gemini Flash model for initial summarizing and then employs Gemini 2.5 Pro for the final, high-quality daily briefing. This shows a smart way to use different AI tiers based on task complexity and cost (when using APIs).
  • Vibe Coding: This term pops up frequently. It refers to the iterative process of building software by having a conversation with the AI. You give it a prompt, see the result, give it feedback or ask for changes, and repeat until you get what you want, often without writing much code manually.
  • Comparisons:
    • When creating a spinning hexagon animation with bouncing balls, both Gemini 2.5 Pro and GPT-01 Pro produced good results, but Gemini's was considered slightly more realistic, while GPT's had more 'bounce' effects.

Overall Impression:

Gemini 2.5 Pro appears to be a remarkably capable and versatile AI, particularly strong in coding and understanding visual input (like drawings or video frames). Its free availability lowers the barrier for experimentation. The examples show it can handle complex tasks, generate creative outputs, and significantly speed up development workflows, even for non-experts, through methods like "vibe coding." The presenter is clearly enthusiastic and uses it daily.

Video transcript

Open transcript
This office simulation was built by a principal engineer at Google using Gemini 2.5 Pro. Check this out. So we have all these desks, we have little meeting areas. All of this is simulated with these little Sims-like guys. We have a ping pong table right there and we have kind of the lunch coffee area over there. Jana Dogen goes on to say, the office is quite cramped, but it started more dystopian. And that's what we're seeing here. So her original prompt, this is what she got out of it. And then she asked it to be more like Google. So more colors, modern look, playful elements, table tennis, et cetera. We have a number of active projects, the productivity rate, the number of employees. You can start meetings, you can add projects. So just a worker simulation. Here she added a library and a couple of couches with an open air meeting area. And she's calling this a board game table. And then she also said that the micro kitchen is hell. And it looks like people are going to be waiting a very long time for their meals, but very cool simulation. All right. And check this out. This is a hand drawing of Google AI studio. And all you do take a picture, upload it to Gemini 2.5 pro. And there you go. This person has successfully recreated AI studio from just a hand drawing and a single prompt. And speaking of Google AI studio, Gemini 2.5 pro, the best model on the planet is completely free to everybody. And quickly, let me just show you that announcement. Gemini 2.5 pro is taking off. The team is sprinting. TPUs are running hot and we want to get our most intelligent model into more people's hands ASAP. We decided to roll out Gemini 2.5 pro experimental to all Gemini users. So not only is it available in AI studio, but now it's available in the Gemini app itself, which I downloaded for the first time in the last few days specifically to use this model. And by the way, Google Gemini with canvas is available to everybody for free. And this was two days ago. We're working hard to bring Gemini 2.5 pro to everyone on mobile too. And it's already there. They do have rate limits. It seems like all of these model providers are just not able to keep up with the demand. And not only is it the highest quality model, not only is it the best coding model I've ever used, but just like it says right here in this recreation, it does have a million token context window, which is more than any other model that I've seen. And you might be asking yourself, okay, why would I ever need a million tokens? Well, let me show you one of the use cases that I've used it for. So right here, I literally just copy pasted my YouTube video in. And actually what's nice is there was an update recently where you can paste in a YouTube URL and it'll automatically upload it. My video took nearly 300,000 tokens. So only less than a third of the total context window available to me. And I said, give me YouTube timestamps chapters for this video. Now the first one was terrible, but I realized I had the wrong model selected. I had Gemini 2.0 flash. Then I used Gemini 2.5 pro experimental. And after lots of thinking, here are the timestamps and it was nearly flawless. And remember, it's not just looking at the transcript of the video. It is actually watching the video frame by frame. And that's how it was able to be so accurate with these chapter markings. And that's not all. Apparently Gemini 2.5 pro has a very high IQ. Check this out. So this is by Mark Kretschmann. Gemini 2.5 pro scores 130 IQ on Mensa Norway. And we can actually see all of the models against the Mensa test. Now, surprisingly, Grok is double digit, which I thought it was going to be higher than that. GPT 4.5, just barely eking out a hundred. And then we start getting the thinking models up here. Here's a one, oh three mini high, oh three mini, oh one pro. And at the very top at 130 IQ, we have Gemini 2.5 pro experimental. And that would make sense to me. It is definitely the best model I've ever tried. Let me show you some other examples of what people have been building with Gemini 2.5 pro. So this is actually by Klein, which is a VS code extension that allows you to do a Gentic AI assisted coding, vibe coding, really cool product. Definitely check it out. And we tried creating the Klein logo with Blender MCP using Claude 3.7 Sonnet when it launched. Wasn't worth sharing. Gemini 2.5 pro delivered on it. First try. Impressive. So let's take a look. Remember, this is an MCP. So a tool that plugs directly into Blender. You literally just type out in plain English what you're looking for. And there it is. It is flawless. This is exactly what the Klein logo looks like and easily integrated directly into Blender. Ilian built something really cool with Gemini 2.5 pro. So for a dollar a day, a personal intelligence agency that delivers daily briefings, it scrapes hundreds of news sources 24 seven uses Gemini 2.0 flash for all the heavy lifting reading, summarizing analysis, but more importantly, Gemini 2.5 pro to write the final briefing. So this is actually a really good point. You don't need the best frontier model for everything. Now, if it's free and it's fast, both of which Gemini 2.5 pro are go ahead and use it. But for the API, it's not, and you're going to be paying. And so you can use Gemini 2.0 flash for a lot of the stuff that any model can do, and then put all the heavy lifting off to Gemini 2.5 pro or the other frontier models. Now, let me show you what this actually is. So it scrapes hundreds of news sites, just like he said, writes up briefs about all the major topics and it is open source. So he provided all of the code. If you want to get this set up yourself. All right, check this out. No way Google deep minds Gemini 2.5 pro just one shot at paper is liquid metal shader. And there it is very beautiful liquid metal Apple logo. And you can change all the different settings. You have the background color, refraction amount, edge, pattern blur. We have the liquid amount, the speed, the pattern scale. And yeah, it is absolutely gorgeous. And if you want to try it yourself, liquid-metal.versell.app. Easy, easy, easy. Okay. We can play with the edge. There's the edge amount, the pattern blur. Okay. So let's make it not blurry. That kind of looks weird. We'll add a little bit of blur. We could play with the liquid amount and you can upload any image as well. All right. And I know we've all seen flight simulators thanks to levels kind of making it popular, but here's a post by Cedric. I built Vibejet using Vibe coding and over 15 rapid incremental iterations in one Gemini advanced chat session. Gemini 2.5 pro is wild. Let's take a look. So fighter jet physics looks good. The clouds look pretty good. We have a giant robot right there. We have the speed up in the top left, the number of players, one altitude, a nice little blimp. Yeah. So it looks pretty good. 15 iterations to get this is really nice. I actually like how the jet gets smaller when you're kind of boosting and going faster. That's a nice little touch. You see little houses on the ground, some trees, some buildings, although not a lot of detail in the buildings. And yeah, definitely the mountains don't look fantastic, but decent overall. I'd say the model for the actual plane looks really good though. The code is completely open source. So you can check it out. This is called Vibejet. And not only did he provide the code, he actually walked you through exactly how he iterated which Gemini 2.5 and put it all together. And so if you want to learn to Vibecode, not only watch some of my videos about it, but check this out as well. I'll drop the link in the description below. Now here's a post by Flavio Adamo. And this is a comparison of that very popular kind of spinning hexagon and with floating balls on the inside. On the left, we see Gemini 2.5 pro on the right, we see GPT-01 pro. I actually think they both look pretty good. The GPT-01 pro definitely has more of a bounciness to each of the balls and has nice little effects when they do hit something. But in terms of realism, I would actually say Gemini 2.5 pro is the winner, but both are actually good, just different. Here is a one shot physics sim by Render Fiction. The prompt, demonstrate electricity and magnetism using a simple example, like a solenoid. Create an animated 3.js scene to depict a charged copper coil and the associated magnetic field. Use three TSL, sprites, bloom, et cetera, dot, dot, dot, HTML file. And here that is. So you could see Gemini 2.5 pro right here, all the code. And here is the actual output. So the coil, we see the magnetism. I think that's what that is running through it. And yep. So a current simulated by moving textures flow through the copper coil, generating a magnetic field. Blue particles. Yep. Use mouse touch to orbit, zoom, and pan. Very nice. Here's another one by the same person. Demonstrate Einstein's theory of general relativity using 3.js. Create a wireframe mesh with a high density grid. Place one large emissive yellow sphere in the center of the mesh. Animate one blue sphere in orbit. Deform the mesh via displacement according to sphere POS. And there that is. We have plane size, plane segments, orbit radius, orbit speed, gravity strength central, gravity strength orbiting, gravity fall off. And I'd say this looks pretty darn good. Very nice. And how cool is this? This person went from a drawing on a piece of paper to an image to a 3d model and then actually 3d printed it. Look at this. So here we go. We have the image, a birthday cake character with legs type thing. Then Gemini made an image from it. So very simple image. Then it made a 3d mapping of it, which then was used to do the 3d printing of it. So you can literally now just draw anything you want and have a character created that is a physical embodiment of that rough drawing in the real world. Really cool. So here's the original drawing. Create a 3d rendering of this three-tiered cake toy. Make it cute. Put on white background. That was the first output. Then create OpenSCAD code for this three-tiered cake toy. Make it cute. Keep it simple. The model needs to be 3d printed. Make sure things are connected, et cetera, et cetera. And here's the thinking process. And here it is. I mean, I'm, I'm just blown away by this. It makes me want to get a 3d printer and actually just draw something and then go print it. And there it is. There is the little 3d cake toy thing. I don't know what to call it, but yeah, it looks really cool. And by the way, this person, Fei Xia, hopefully I'm pronouncing your name correctly, is a staff research scientist at Google LeapMind. Next, we have Bi Yang. This is the progression of a 3d Flappy Bird game. I made using Gemini 2.5 pro and Sonnet 3.7. No reason you can't use both. Wrote zero lines of code, zero experience with games and graphics, vibe coded over one hour with a tool that integrates with vanilla VS code. Here's kind of the first attempt. Here's the second attempt. Remember this is 3d Flappy Bird. Here's the third attempt. So now we get some speed. We can actually have some sounds going and then the final version right there. So pretty darn good. And you can even create iOS apps. So check this out. Create a fully functional Swift UI drawing app using the pencil kit framework. All right. All of that code was output. That's just one shot or zero shot. And here's the code. And here's the app. It actually looks almost identical to the art app that I think everybody has on their iPad. Incredibly, incredibly good. This is just a beautiful app created in one shot with Gemini 2.5 pro. Next, we have Galaga. So we built an AI agent and asset to create Galaga and in seconds and generated a fully playable arcade shooter. How it works. Pi game to create the game, generates saves and runs the game instantly powered by Composio's file and shell tool built with Gemini 2.5 pro. And it's open source, by the way, if you wanted to check that out. And here's the game. Okay. So simple game, but still very fun. And it was created in one go. So we are in the era of vibe coding. So that's it. Gemini 2.5 pro continues to blow my mind. I am building things with it every single day. Drop a comment below on what you're building with it. Let me know. And if you enjoyed this video, please consider giving a like and subscribe and I'll see you in the next one.