Gemini 3.5 Flash a prova: ràpid, però no infal·lible

Relation test. Oh my. This is the best one. Oh yeah, this model is totally getting insane in the title. Gemini 3.5 flash has been released from Google and this is hypothetically a pretty big leap in performance being that it does seem to be significantly outperforming the previous Gemini 3.1 Pro model but in the flash name. So the flash model only sits beneath the Pro model. At least now it's kind of inverted but when Pro comes out, 3.5 Pro of course, next month which they did verify, it will be pretty interesting to see how it performs. So for today's video we're going to be mainly focused on testing the newly released Gemini 3.5 flash as well as the new Google anti-gravity 2.0 which I do have open right here and I just can say on first glance here this does feel far more in line with the modern agentic tools, things like Codex and basically Codex. So this is nice to see and the only initial testing I have performed so far is just saying hi to the model and it has responded naming itself anti-gravity but as we can see down here it is Gemini 3.5 flash selected on the highest thinking mode which we will be using for the duration of today's test. So before we begin, please do feel free to subscribe so I can get that 100k plaque and then I'll never ask again. Now I probably will but um at less significant interval. So let's take a look at some of the interesting things of note in this announcement post for Gemini 3.5 flash. Now really one of the more impressive things we see right here is the benchmarking capability compared to the previous state-of-the-art model of course being Gemini 3.1 Pro. It is outperforming it on I do believe every single thing we see here at a fairly impressive like gap I would say. Okay. Arc AGI 2 it is not outperforming it but seems like for a lot of the coding related things it is definitely outperforming it which I'm quite excited about because a lot of our tests are coding related. Additionally to that, it is seemingly outperforming Claude Opus 4.7 in some coding benchmarks, agentic terminal coding with the Terminus 2 harness, but not necessarily on SWE bench. So, I don't normally read too much into these benchmarks. I don't know why I'm doing that right now. I apologize, but basically the culmination of what I'm trying to say here is this does seem to be a very very performant model, especially for something that is not under like the top tier frontier naming schema, being a flash model and not a pro model. And that would just mean like this could be more thought of comparatively to something like Sonnet instead of Opus in like the Claude family of models for some reference. Now, something else that was really touched upon heavily in the announcement, at least like the Google I/O keynote for this, was its speed. They showed some demonstration, and I do believe in that demo the output speed were up to like 1,500 tokens. Right here, we can see the output speed is listed probably I'd call that like 280 or 290. So, there's also some demonstration shared here of different things that they've chosen to highlight. One of these is creating interactive animations, which I am pretty happy to see because I do have a special test design, so I'm happy to see that as a prominently displayed capability that this model possesses. Additionally to that, something they showed in the keynote was basically a bunch of sub agents. I think it was 97 sub agents they mentioned and like a billion tokens, and they used this to create an

operating system. Then they had a demo of it, which is a fantastic idea. I don't know where they would have gotten that. But also, no, I'm kidding. Cuz they made a real one, not a browser OS, but it was less visually exciting than the browser OS because a real one was just like in the terminal online. So, this looks like a pretty good front end. This looks very very Apple Logic Pro-esque, at least the previous thing did. But um really the only other thing I want to touch upon are more like specifics in terms of the actual model itself and pricing and things like that. Now, something interesting is that this 3.5 flash model is three times as expensive as the most or its predecessor flash model, which we can see right here, was always in preview. So, let's take a look at Gemini 3 flash preview. That was three times cheaper, so $3 per million output tokens and $0.50 per million input tokens for text, image, video versus $1.50 per million in and $9 per million out. Additionally, we can see the supported inputs for this model are text, image, video, audio, and PDF, which is good. It does only output text. However, there is also an Omni model that perhaps is worthy of its own video. Our input or context basically length for this model is a little over a million and the output token is 65536. So, that's just a little bit more pertinent information about like kind of what's going on underneath with this model. And with that now, we can go ahead and actually begin to start testing this in the new updated anti-gravity. So, to begin, K9 viewers may be able to guess we are going to start with the browser OS V2.5. Now, because this is going to be tested throughout this agentic harness, I did check just in the regular like chat Gemini web interface and the model was not yet appearing for me, so I'm going to strictly just be testing it through anti-gravity. Okay? And we're going to see a bunch of prompts and things that are going to ask us for permission and things of the sort. Being that this is the first time I'm using this updated IDE or this model, so we'll keep in mind that some of the capabilities we're seeing here are not necessarily able to be compared to a traditional zero-shot generation where we're just using any models random web chat interface because this is functioning in an agentic harness that it has been specifically trained to pair well with. We can see here the speed is quite quite good. We have a chain of thought process that can be expanded. Also, this sidebar here will just give us a little bit of additional pertinent information and what specifically has been going on. Seems like it's still just planning out the design, so we'll see. I think it's done. Yeah, okay, it's done. It's extremely fast. All right, here is our Gemini 3.5 flash browser OS. Okay, system booted. Explore 3D games. We have these annoying notification pop-ups that appear predominantly in macOS. This is a very modern take on what an operating system should look like. With that, let's check the unspoken test. Does not have a right click. Well, that's the video. Goodbye. All right. We do have an interesting kind of dock right here. Okay, we actually do get notification feedback when we click things right here. That's for our Wi-Fi, system health. Very cool. So, when we click individual icons here, both Wi-Fi and system health, it does actually show a pop-up corresponding to what we did click. We have a clock that is the correct time in my locale. Additionally,

we do have a very like minimalist take on the icons, which I guess is acceptable. I would assume this to be some form of a start bar. Okay, Ether OS. Very interesting because I do believe the previous Grok test that was done a few days ago did also name these Ether OS. It's a common theme that I've seen among many different generations of models. With that though, we're all thinking it. Let's check out this GTA clone, which it named Ether. All right. Ether City Drift. Maybe that's not the GTA That's all right. Let's full screen this. Okay. Drive in metropolis. You know, I have some things to say. Let me turn the sound on. We'll see if it implemented any sound effects. Yeah, it did. So, here's what I'm going to say. Oh, the car model's very good. It's very neon cyberpunkish, and the actual driving in the movement is not bad. The way the camera actually like gives us the sensation of speed. The acceleration movement of the vehicle is smooth. Now, in terms of downsides, I will say there are no pedestrians in a GTA clone. One would kind of want to see pedestrians walking around. I could use some more detail on the buildings as well as streets and things like that. But, the car model is one of the better ones that I've seen. And, I'm going to actually just close this again and reopen it because I want to get a look at the specific instructions for controls. Okay, so space is to drift, which is cool. But, then we see R is to toggle the dynamic synth car radio. Oh. Yeah. No, I'm kidding. That was uh Let's do a drift with that playing. Oh, okay. Oh, okay. So, drift just gives us like a particle effect. All right. You know, it's very, very unique. So far, everything we've seen Can we minimize and reopen? Good. Has been just a very, very novel take on any of this that I've ever seen. All right. So, our next 3D game would be inevitably Cosmos Strike 3D, which is Oh, wow. That's actually kind of cool. It's got the um throbbing is not the word I was looking for. I was going to say the pulsing effect of the motor in the rear of this craft. Okay. Cool. Next up, Vibe Station. No way. Don't even. Okay, first and foremost, resize logic is good. This UI, you know, I've personally the Anthropic Opus like aesthetic with the disgusting like pale white and like orange themes, I hate it. Also like the GPT-5 aesthetic, which is like the neon blue. I don't hate it as much, but this is a novel UI aesthetic and I kind of like this. This model may be interesting for a front-end, like I mean, this in and of itself would have been an impressive result, but contained in the browser OS, we even have like a So, see the difference when we click that? It actually takes longer for it to fall back to flat, I guess, unscientifically. And then if we put that back down, so it does take effect there. That is pretty cool. We can change our sustain. Can we play with the keyboard? Okay, I don't believe we can, but that's all right. All right. Let's just go back to sign. Background rhythms. You know, it's there's no word Words are not really required to accurately reflect my enjoyment of that specific thing that was included that I've never actually seen before. This is going to get insane in all caps in the title, just FYI. Okay, matrix doesn't seem to be working, but evaluate arbitrary JavaScript unexpected identifier system. Oh, yeah, right. I don't know. I might not have

Can we clear? Let's see how it's UI aesthetic is. Oh, very cool. That's a rocket ship mountain. Finally, shader, which I would assume to I'm glad I didn't finish that sentence cuz you might think like only a fool would think this was going to be a paint application. Themes and accents. Cyberpunk, emerald, solarized. Solarized is kind of the cloth thing I was referring to that I'm just like, "Yeah, not a huge fan of." Live GLSL background shader editor. Oh, yeah, this model is totally getting insane in the title. This entire result could have been its own video, to be honest with you. Compile and run live background. All right, let's try something. Oh, wow. Okay, it didn't work as I'd intended. It's supposed to say subscribe to Bijan Boran on YouTube. We can see though it it did change. I'm actually going to just undo that because I don't want to cause any of like issues with the viewer cuz that was a little, you know. All this and it didn't have a right click. This was fantastic. This may be one of the best, if not the best. Uh I'm interested to know what folks think in terms of how this stacks up, assuming you've watched other videos of mine. If not, just don't comment. So, you may recall that one of the listed things that they showcased this as being good at is interactive SVGs or things of the sort. So, I have a very special test here. This is called the mid-roll ad test. So, of course, we're going to be giving this a special task where it needs to create an animated SVG explainer as well as a script that I have given this for the sponsor of today's video, oxylabs.io. Now, this is going to be creating something. We can see right here. This is a pretty intricate and heavy script where essentially it is given a folder of specific assets. So, I did just create a new project with the assets folder where there's a bunch of photos and things like that. It's going to have to go in and figure out exactly what they are. Following that, I've given it some specific things where basically the visuals need to match up to the 60 seconds or so of speech and I've given it the entire transcript of this actual mid-roll so that it can closely match the visuals to exactly what is going on right here. So, we're going to see what this does. No explanations, output it in a single SVG code block, and we'll see how well it is at actually creating those interactive animations. Interestingly, I was about to critique the model and be like, "Well, it didn't give me any follow-up or anything like that." But, I see it's actually properly adhering to the instructions, which are to only output the raw SVG code, and don't give us any explanations or anything of the sort, which is exactly what it did. So, let's take a look at this. If you're building AI agents, getting real-time data from e-commerce, real estate, or search results is a massive bottleneck. That's where the Oxylabs web scraper API comes in. It gets real-time data at scale from almost any public website using over 175 million residential proxies across 195 countries. I dropped this straight into my n8n workflow, which I did do in a previous video. There was a pretty like robust actually demonstration of this, because it seamlessly handles JS-heavy sites and scales instantly in the cloud. Oxylabs also just launched a new headless browser feature specifically for AI agents, plus a fast search API that feeds organic search results directly into our pipeline. Go to oxylabs.io/bijan for a free trial. You get up to 2,000

scrape results for free with no credit card required. Use code bijan in all caps for 20% off all Oxylabs plans. Okay. So, let's take a look at this, because not only was that a mid-roll test. Thank you to Oxylabs for sponsoring the video. However, this did a very good job of actually including some of the assets, the color schemes and everything like that. Additionally, we noticed that the animations it created in this SVG are very much in line with what the actual script was talking about, such as having proxies in countries, talking about n8n workflows and things like this. There is actually a cool video of doing a drop ship detector with this in an n8n workflow I did post a little while ago. And the colors are perfect. It does have the bar in the bottom to showcase like the timeline, and it gives us the script as well. So, this was actually pretty impressive and well done, and I like how it utilized a folder containing pre-made assets to specifically showcase like just its ability to work with existing things. So, that was pretty cool. All right, next up, we are going to do the subway test. Don't worry, but this is something I did in this morning's Quen 3.7 max video, which was perhaps one of the most enjoyable results I've ever seen. This is a front-end test, but this is a really intricate prompt which denotes a lot of specific things. The combination of what this needs Okay, so that actually started out by creating an implementation plan for itself. I didn't tell it to do anything. I just pasted the entire prompt in. So, it seems that it made the determination on its own that this should start out with an implementation plan, or I have it in a setting that I neglected to realize. But, this is to create a beautiful front-end for an immersive landing page called Ravioli Rosso, a fictional premium pasta sauce company. The background should have a bunch of like particles, ravioli shapes, and things like this. This was a very, very enjoyable result to take a look at, and will give us some insight into the front-end capabilities of this model. It's just insane how fast this model is. That was 2,640 lines of code, and it's genuinely going to take me longer to find the actual directory to place this in than it than it did to generate the script. All right, so I just made it move that into the home directory because it was This is very interesting because this is such a similar result to the 3.7 max I got, which shows that a much stricter prompt like this definitely produces more consistent results across many different types of models, which is kind of like why I like doing more open-ended. The Okay, secret sauce. This is what I wanted to see. Okay, it's I mean, it's funny, but it has like a weird card effect on it. We do have the background which has the floating ravioli and things like this. Your pasta has been waiting. All rights reserved. Slow cooked with passion. Okay. Again, this is a very weird test. Sauce viscosity, watery fast flow, thick and heavy. Now, this could have effect on the specific like UI background effects, and it does. So, if we put it on boil, things move fast. Slow cook, things move slow. Active ingredients. So, the background is actually adjustable based off of what the user selects on this card. Again, taste the sauce. Okay, good. That just brings us down there. Watch it simmer. Okay, cool. So, this was Again, I think I was more enthused with this in this morning's video when I tested it, but I wanted to try like a

random front-end prompt that we've not really done ever. So, this is definitely not in the training data of any model. Next up, it's time for the combination beautiful static subway scene {slash} FPS game, as well. This I am going to be very, very excited to see how it does. WASD is to move, space to jump, mouse to look around, left click to shoot, R to reload. Okay, something I'm mad about is I can't actually get to the pane on the right that would allow this to be less dark, which is kind of frustrating. I also What was it to shoot? Left click. So, sadly we're encountering some issues where the shooting isn't working, and the lighting is just way too dark. I'm going to say though, there's like a low hum of electricity that it implemented here. That's actually good background sound. I don't always get that. I don't notice any enemies though. So, I started off giving it some decent feedback, and then I realized that I had left caps lock on, so I just stuck in that theme. We'll see what it does. Okay, it's already brighter. Okay. Escape is to free the cursor to access the slider settings. Unfortunately, pressing escape just brings us back to this. You know what? That's all right. We're going to iteratively Actually, no, we're not going to iteratively improve this. I'm angry. There's no visible weapon model. I can't shoot. The actual like cracks in these bricks I don't think I've ever actually seen that before. So, I'm seeing things here that are promising, including the rust on the pillars. This is a very eerie. It has the flickering light. Oh, there was sound effects. I apologize. This tab is muted, but we're going to need some additional fixes. So, Oh, mode selection ex plore versus FPS combat. That's on me, but it's also on the model because I can still can't get to that menu. Something I do want to make mention of just based off of one, the animation test we did, and two, what I see right here. In the animation test, we told it just give us the SVG without any additional explanation or anything. It did that. In this one, the way it went about making this is very interesting because the initial prompt here was basically two prompts combined. So, before we would do the beautiful static subway scene, and then we would also, following that, turn it into an FPS game. Because of that, it actually gave us two different selectable modes here. So, the explore one is fitting with the original prompt, and then the FPS combat selection actually turns it into the game. So, this seems to follow instructions very literally, which I do like to see. I should unmute this tab. Let's see what's up. Oh, it's a shotgun. Can it Oh, okay. I I hear the enemies, but I don't see them. Oh, this is creepy. The creep factor of this with the sounds and things are are good. Look Get away from me. Okay. Yeah, well, that will happen when the weapon does nothing. Look at the reload effect. It actually has the gun tilt like it would in real life if you were reloading one of these styles of of weapon. Okay. There's good and there's bad. All right. Yeah, great. So, last chance here. It's still getting insane in the title just because of that browser or rust result, but we'll uh we'll see what's up here. No, it's still that's so frustrating. All right, there were pieces here and for a flash model I would say that still shows a large level of promise. I honestly think non specifically denoting that it should be contained within a

single file kind of did a disservice here, but on the flip side of that we got to see this function a bit more agentically where it had to do multiple iterations over an initial result with multiple different scripts associated to it as well. So, next up we're going to be doing some C++ in the form of the self-contained C++ skate game with the California boardwalk aesthetic and the first thing we see right here is the model is just checking the environment to see what specific things are available in terms of actually being able to build, compile, and run this game, which is good and the fact is from when I pressed send with that prompt to when it actually went ahead and tried to initiate checking the system was extremely quick, which is pretty cool. All right, it seems to have assessed at least what it needed to on the host system and it's now going to create its implementation plan and then whatever it did I'm just going to click proceed. I will take a quick peek at it right here. And again, all of this is happening so quickly. It even like what the heck? The entire application will live in this main CPP file. Like Okay, I didn't expect a flowchart for the actual C++ orchestration here, but I'm impressed with that. I very much like that and it just happened so quickly. All right, so I had it move the end generated result just to a simpler location. So hypothetically, we should be able to get our result here. Okay. This Oh, it's I don't even care. Yeah, on first glance, it's not working great here, but this is a very very very promising result. The way the Look at the way the Oh, it's so synthwave. It knows me. Is there personalization on with this? I don't think there is. Look at the pedestrians. Yeah, they don't have arms, but they don't need to cuz this is synthwave and it's like a game. Um they should have arms, but I mean, the buildings, surf shop, um burger shop. It's so like synthwave. Say like What's up? With this. And I'm going to do it sarcastically cuz I personally enjoy it more. And that's kind of partially You may be like, "Dude, you're a tool." Like when you tell the model like feedback like this, but partially it is also genuinely to see if it can infer and understand like, "Oh, okay, there's an issue wrapped in this like kind of you know, whatever you want to call that. Sarcasm. Very good. Excellent. Okay. I could I could have used some movement on the skateboarder though. Oh, the arms are moving. Look at the way they're swinging. All right. Wow. All right. Well, uh-oh. Okay, good. All right, let's just try a kickflip. Oh, that's weird. Okay, so it's Oh, that was cool. All right. The sound is is, you know, it's loud. That guy just got flexed on. Got showed up. All right, let's see if we can grind. Jump onto rails to grind automatically. Yeah, I'm trying. It's like semi-snapping somewhere, but it's not like fully um you know, but just like this view right here, this is very promising, especially for a non-frontier model. And what I mean by that is this is not the top dog in terms of the model. That will be 3.5 Pro, which now I'm very, very excited to see. And this really seems to like this like cyberpunk aesthetic or neon, being that the browser OS was like this as well. So, very interesting. All right, this is one that's going to be difficult. It is the Jerry's apartment from Seinfeld test, where it needs to create an accurate 3

JS model of the apartment from Seinfeld. This is going to involve a few different things, such as like understanding spatial layout from photos. It is inevitably going to be searching for reference photos online, which it is right here. And this is something that I've never properly gotten a result here that is actually true to the layout of the apartment. Though so far 5.5 in Codex, GPT 5.5 did the best by a mile in this test that I've done. So, we'll see how this does just in terms of layout and detail and what it chooses to include. Okay, that logo is is proper. Do I have to like hold down on it? Okay, so it started role-playing. Listen, you imbecile. Good. Okay, so that was a pending message. Okay, and it just went back into role-playing when I apologized to it. I'm just going to leave that one as it is. Good. It did. It made a bass scent in the style of like Seinfeld. Interesting. This is closer to the actual layout. Now, I could definitely use a bit more light. But, it has potential. Now, it Oh, wow. It's interactive. Okay. Hold on a second. This is the only result I've seen that actually Let's go into explore mode. Very good. Okay, yes. This is the only one I've seen that actually put a screen on the vintage Mac. To that point, it actually did it in like what the system 6 system 7 OS on this system would actually look like. We have a mini map of the floor plan. The heater is there. The floor is there. We do have the doorway. And it just plays sounds whenever we The iconic gray-green couch. Audio controls. It This thing really likes putting in like synthesizers. I don't know what to say. Paradox Visualizer. Oh, okay. Paradox Visualizer. People say the hallway can't actually exist in this model, so it actually wrapped like the wrongness of this. So, it's showing us the hallway as it's drawn would actually intersect with the kitchen. And it called that a paradox and visualized it. This model's intelligent. That's what I'm going to say here. We have cereal boxes in the fridge with a That is so That's just odd. All right. Well, this was just the physically impossible 3D replica. This was very, very unique. So, next up, I'm going to go to ultimateplay.com, which is my own website where you can play, browse, and upload 3D generated single file HTML games for models. Now, the reason I am here is because there is a specific prompt that I'm searching for. So, I was looking for this. It's called low poly GP. It was probably one of the best game results that I've ever seen. It was made with GPT 5.5 Pro, and it took about three tries or four tries to get it to where it ended up, and it was really just a stellar result. But, I'm going to have Gemini now take a stab at doing this, just for reference, I suppose. I will showcase quickly the actual like result that we received here. I mean, so, it was one of the finest results I've seen, though it did take a few tries. So, it is actually fair to compare this to whatever 3.5 Flash will spit out agentically, but keep in mind that was GPT 5.5 Pro, which is 20 times as expensive, I believe, for the output tokens than 3.5 Flash. So, keep that in mind. I don't know, it just wrote, "This approach is proving to be a highly effective way of managing racing physics." Exclamation point. So, I do like the enthusiasm it has for for its own work. I share that, as well. That's

frustrating. It has that same blocker issue that we noticed with Seinfeld's apartment. So, I'll have it fixed that in a non-sarcastic tone. I'll just accurately describe the issue. Um it likes its own work, which I guess is good. Oh, wow. It loves putting synth music over stuff, too. This model loves music. I do see a floating track, so all right. Okay, and they're just like >> All right. All right. It has a mini map. The opponents look cool. If the ground look better and the lighting look better, it would be nice. The actual like showing of the RPMs and the speed here are cool. Obviously, one of the main issues here would be that this camera is just totally totally wonky. And unfortunately, this happened with the GP 2 Pro model test as well. The ground having actual elevation is messed up because we can't actually see the track, which is letting this down. The bike model is very low poly, but we do see the wheels are actually spinning. So, I'm not going to have it going like totally fix this because I've seen enough, but it it has the changing camera view. It does have the wheelie like logic or something of the sort. Even though everything's kind of inverted here, it's interesting just I don't quite know what to say about this. It definitely shows promise. Some of the individual elements here could produce a very very nice result. I'm now going to do the flight combat simulator test because so far we've had some interesting results when it comes from like web graphics and 3D things where some of the assets are really interesting. Some of them just don't quite go together right. So, I'll be very interested to see how it does with this, which is a more traditional test that's been around for a while. I want to mainly see how well it does on the plane models and things. All right, here's our flight simulator game. I will turn the sound on. There's probably going to be some sort of synth music if we can go off. Yep. Oh, okay. I still hear Moto GP Arcade. Okay, so there wasn't. Now I look like a tool. All right, let's just take a look. >> It's not funny, but like Okay. Mouse to pitch and bank. All right. I mean, I'm trying like Oh, it's inverted. Well, it's actually not it's I was inverted. Let's just leave it at that. Okay. So, the plane models actually seem nice. The thing is it's very hard to control. This has such a GPT-5X looking UI to it. All right, let's let's just try this again now that I know how to fly it. This this model has a fantastic grasp of of sound design, I would say, comparatively to all the others I've tested so far, specifically with making games and things like that. Okay, uh maybe the jet shouldn't have had like the propel I'm done with this. All right, so we're doing the historic 3D printer simulation test because this is always fun to take a look at, especially for new generations of model. So, here's our 3D printer result. Okay. So far this looks pretty nice. It is a core XY. The grid on the base plane looks good. A little dark, but again, kind of like what we saw with the flight simulator UI, there's just a ton of additional UI options here. Okay, it's going to print a calibration cube. Chamber heating. That's not a cube. Is it? Yeah, it is. It did a skirt around it or All right, this is this is good. It's a little hard to see some of the plastic. Right as I said that. So, as I was saying, it's extremely

visible in terms of the plastic. We have camera view nozzle. Okay. I like it. Uh I don't necessarily know what to make of that. We have a bed cam. Equally is perhaps, you know, slicer path view can be toggled on and off. The movement of the nozzle here and things like that. This is probably one of the most realistic nozzles moving across a gantry for a core XY that I've seen. This is exceptionally well done. We can change the material to ABS, PETG, PLA. That's very, very good. Let's try it with higher infill density. Okay, we can only go up to 50, but we're just going to want to see if it is visibly more infill. It's definitely more of a crosshatch pattern. Let's do the nozzle cam here, which we may regret. This is showing some definite like potency to it. All right, let's check the other. Industrial double gear circle. I like how it named everything like very like CAD like, like technical. Oh, no, it's actually making a gear. All right. Well, it didn't follow instructions there, so fail. No, cuz we just say make a circle and things like that. All right, that's pretty cool. What was the final shape that it made? Okay, a triangle. Good. And it's getting narrower as each layer goes up, I do believe. Yep. Yeah, and it's doing such a nice job of that as well. Well done. Very, very well done. This was Oh, it even has a spool in the Wow, this is quite a detailed model. Let's just try one more thing. I just want to see if the spool spins, cuz sometimes they do. It is. Look at that. It is. It's hard to see, but the spool is actually spinning. Very well done. All right, I've been dying to try this since we first saw the musical edition to the browser OS. This is the drum kit simulation test, and it must also include four tracks that it will auto play in different styles. All right, so here's our drum kit simulation test. Oh my This is the best one that I've seen, period. Even if it doesn't work, this model is so like it loves music stuff. This is I mean My biggest complaint with this is they're always laid out just horrifically. This is basically perfect for all intents and purposes. Unreal. Even like some zebra wood shells. Look at the the studio mixer snare drum. We can actually tune the specific drum sounds. This is just like This is Nothing has compared to this. Let's make sure it works, but Yeah, it works. Does the kick pedal work? Okay, it's kind of inverted, but I'm going to give it a pass on that. Also, the UI elements kind of wrapped down a little, but I mean Okay, I'm going to turn this all the way up. Phil Collins test. Or it should be Well done. And look at how the synth tweaker changes depending on the specific one we've just had. Yeah, it does. It's hard to hear, but I think I Yeah, listen. So, that's changing the actual like pitch. Wow. All right, let's try auto play. >> Can I play with it? Yep. And look at that like All right, funk break beat. Oh, yeah. Yep. Yes. Jazz swing beat. Well, we should start with that symbol. Okay. Yes, but I'm going to stop that to preserve my ear-ies, my hearing. Yeah, that is all right. Best one by like not even a comparison, truthfully. So, I'm doing something that I have not done in a little while. This is a multimodal coding test. So, we're giving it this hand-drawn portfolio wireframe. In addition to that, we've given it the prompt to turn this into a

portfolio website that would get someone hired at a high-end AI company or something. It should also implement real machine learning concepts. This is very interesting. So, it seems like it's actually Oh, look what it's doing. It's prompting Nano Banana to actually create some of the assets for this specific site. So, it actually just went and did that for the specific photo of Stevie Slappis as it is listed in this portfolio. Wait a second. Did it actually Oh, no. Okay, it did Stevie Lappis. All right. Yeah, none of them is have ever really gotten it fully right, but that was cool that it was generating assets with Nano Banana where it would need them. All right, let's take a look at our wireframe to portfolio website. This is I need a second. So, first and foremost, I think the interesting thing to notice is that it actually generated the asset where it needed. Oh, look at that. It's actually running like a an image scan over the profile picture. Okay, it did properly animate the squiggle that was there for Stevie Slapass. So, the argument could be made it did kind of get that. We also have some like circuit-looking things or perhaps a pattern or pathfinding. Okay, we have an attention heat map right here. Look at that. Oh, what? Okay, so it hyperlinked these specific things to actually change the attention heat map. This is it tokenized like the Okay, I wouldn't call it full like tokenized cuz it's just word by word, but Okay, this is like a nice attention map bio. Neural Sandbox classification engine. Okay, we have our live loss value being updated. So, this is emulating an actual training process right here. It's possible it's training it. We have the network architecture. We have the data set pattern. We have the activation. Let's just This is cool. If we click these, it actually affects this. This is a very, very fitting result. Sender identity. So, our actual contact card. Latent space explorer. That is cool. That's crazy. So, if you actually search things that show up, it actually gives you like additional feedback here. Run inference pipeline. This was cool. And it was interesting because it did a good job of it, but it kept it pretty realistic to the source image that we gave it, which if we scroll back up all the way, I mean, it genuinely kept this almost a mirror image of what we got here. The only difference was the contact card was on the right side of the page instead of the left, but this was a very, very interesting transposition of that wireframe into something with actual like ML concepts integrated into it. Very, very neat. And the fact that it made the asset that it needed, which is going to be pretty cool. So, I think overall that is going to conclude our first look and test of Gemini 3.5 Flash with anti-gravity 2.0, which is how we ran this thing entirely. We did have it on high thinking settings for the entire duration of the test. And now let's just take a peek at least at some of the results that were really quite like, "Whoa." Definitely one is going to be the browser OS. I would perhaps go out on a limb here and say, "Oh yeah, now that I go back at this and remember some of the things it had." This was definitely, I believe, the most impressive result that I've seen. The audio synthesizer just had me completely sold. This as its own result would have been awesome, but implementing it just randomly into the browser OS was really, really, really quite cool. And then the modifiable backgrounds and things like this. Very, very cool. This Ravioli site

was I don't know what to make of this cuz it's just such a weird prompt, so we're going to skip that. The Seinfeld apartment thing was interesting because it was weird in some ways. Like, I didn't like it in some ways in terms of like it was kind of dark and the detail wasn't like super there, but then it actually went ahead and called out like some lore, like the paradox of that physically being impossible to have this laid out with the hallway there to the point where it drew a wireframe around the paradox. The Mac actually had a little screen. I forgot about the sounds. This model loves music. The Mac had a screen on it which actually showed like an old System 6 OS, which was fascinating. The 3D printer sim was very good. The plane game and the motorcycle game were not as impressive as some of the other results we received. Namely, this Logic Pro X drum kit. This is the best result I've seen to such a high degree that I'm actually surprised by it. This is in a league of its own comparatively to every single other model I've ever tested this with, which is basically all of the big ones. So, this is very, very interesting. The portfolio website was obviously very good. The SVG test where we had to create the mid-roll ad for the XVG was really well done. It did properly utilize the assets I had given it, which was nice to see. Additionally to that, although it was not like fully perfect, our skate game was pretty darn cool with the funny sounds. And it just had a very cool neon like synthwave aesthetic to it, which this model seems to very much be a fan of because I didn't specifically denote any of those like styles, but it did it for both the browser OS and for this. The water's cool, the sun, like this was neat. So, in terms of like feedback, I guess I would say I noticed it is very, very fast and it really does seem to like music. I also noticed a few times that while watching its chain of thought as it was going through things, it seemed very, very pleased with itself and I don't mean that as a critique. It's just funny where it was like, "I'm going to update the walkthrough MD to mention this elegant fix." It's like, "Okay." And then it had like an exclamation point. So, it was interesting that. I really think it was cool that it created the assets it needed at least in this one specific example using Nano Banana. That's going to be an interesting and very useful feature depending on like site design. I can't wait for all Shopify stores to basically just be doing stuff like this if they're not already. So, make of that what you will. It's a cool model and being that this is Gemini 3.5 Flash, 3.5 Pro is definitely going to, assuming the improvement follows to the Pro model. I mean, this is not the top-tier model. This is the Flash model, which is akin to like the Claude Sonnet kind of. So, when Pro comes out, it's going to be very, very exciting. With that, that's going to wrap up our first look and test of Google Gemini 3.5 Flash. If you have any questions, please feel free to leave them in the comments or in the Discord, and thanks for watching.