Gemini 3.0 Is Actually Insane at UI Design
JTJosh tried coding@joshtriedcoding · 174K subscribers
- Views
- 17.1K
- 17,066
- Outlier
- ×0.4
- Median 43.3K
- Length
- 8:50
- Published
- Nov 21, 2025
- Friday
- Comments
- 55
- 541 likes
Script
How it's written
- Words
- 1,647
- Words a minute
- 186
- Reading time
- 7 min
- Segments
- 6
The hook
0:00–0:27All right, man. Gemini 3.0 is out and it's an insane UI designer. Like, it's really, really good. Do you remember that purple gradient AI slop that basically any other model generates? Well, Gemini 3 is out here crushing some of the most insane AI generated UI designs that I have ever seen. And this release was long awaited, man. So, back in January 2025, they released the Gemini 2.5 series. That was a really
Structure
All right, man. Gemini 3.0 is out and it's an insane UI designer. Like, it's really, really good. Do you remember that purple gradient AI slop that basically any other model generates? Well, Gemini 3 is out here crushing some of the most insane AI generated UI designs that I have ever seen. And this release was long awaited, man. So, back in January 2025, they released the Gemini 2.5 series. That was a really long time ago. So, let's kind of put that on the timeline here. This was Gemini 2.5, right? It has been the go-to model for basically everyone during this period because the 2.5 models were really good, especially 2.5 Pro, they were cracked, man. So, during this period up until like mid 2025, either Zonet or Gemini 2.5 Pro were the go-to models for coding. They were really, really good. In August, GP5 came out and that was probably the time up until here everyone used Gemini 2.5. And starting from August 7, myself included, people use GPT5 more and then slowly started switching away from the Gemini models. And now in November, right here, we got Gemini 3 by Google. Here we go. So people are probably going to switch over again. And the timing here from Google was kind of immaculate because 2 days ago, there was actually a new release by XAI Gro 4.1, a frontier model that sets a new standard for conversational intelligence, emotional understanding, and real world helpfulness. So, interestingly, the focus is not really on coding here. It is good at that, but
Watch from 0:00probably not the main use case for 4.1. This is not really made for VIP coding. Gemini 3 is. And one of my favorite parts about Gemini is how good it is at design, by the way. So, Gemini 3 is now the most upto-date, the best model to use for coding right after 4.1. So, really, really good timing for Gemini 3. And now, let's take a look at how good it actually is. How does Gemini 3 score? It's pretty good, but not on every benchmark. This is published by Google. So, of course, these are going to be the most favorable benchmarks here. Like, for example, humanities last exam. It's 2500 questions in a very broad range of subjects. It scores pretty high. 37.5% and then 45.8% with search and code execution. So, a whole lot better. And we're not going to go through every benchmark here, but one other one I found really interesting is right here in mathematics, right? The AIME 2025. No tools 95% with code execution 800. That's pretty similar to cloth zonet 4.5. A little bit better without tools though. Is it realistic that models are not going to have tools in the real world? No, probably not. In cursor, they do. So, it's pretty comparable, right? So, the bottom line of what Gemini 3 is. Do you remember Gemini 1 when that came out? Gemini 1 introduced two pretty pretty cool new things for the Gemini lineup. The first was the multi- modality. So it could actually understand images and understand visual context. That was one of the two big things that Gemini 1 introduced. The second one was context. So Gemini has always been really good at context, at
Watch from 1:30least in theory, right? It allows a really long context window. With Gemini 2, they introduced all the missing pieces that one didn't have. Like for example, thinking or reasoning, right? In my head, it's always kind of the same thing. But arguably more important, especially for development use cases, it's tool use, right? It actually became pretty good at using tools. Logically, the goal for Gemini 2.5 or more so 3 is to make one really good model that kind of combines both, right? It's multimodel. It has really long context and then it can also think and use tools very, very well. So that's what they're going to do in Gemini 3. At least that's the intent. That's the promise behind this model, right? With the Gemini 2 lineup, one of my biggest problems and one that many, many other people had has always been context, right? So, on paper, the context window looks great, but if you get past 100,000, 128,000 tokens or even more, it gets really, really bad at following your orders, right? And it even went so far as this right here, I have uninstalled myself. Gemini has some issues to work through. People are reporting that Gemini 2.5 keeps threatening to kill itself after being unsuccessful in debugging your code. I cannot in good conscience attempt another fix. I am uninstalling myself from this project. You should not have to deal with this level of incompetence. I am truly and deeply sorry for this entire disaster. Goodbye.
Watch from 3:02And then Gemini uninstalls itself, man. And that problem is a lot better in 3.0. So I work as a software engineer. I have a full-time job. I've used Gemini 3 all day yesterday, right, to try it out and I've also searched on Reddit what other people think. Gemini 3.0 has been fixed for me at least and it's fantastic, right? So, especially the problem with context window regen. So, it gets worse the more context you have. It's gone, right? It's it's got really really good at that. I have a 500,000 token story that I had to abandon because Gemini 2.5 would process trash into chain 30 objectives together and ignore commands to stop. opened that same conversation, ask 3.0 to clean it, and so far it has not degenerated with no objective spam and a really, really good tone and style. So, the overarching consensus is people, myself included, are pretty happy with the model, right? On Reddit, talking to other developers, I feel like a lot of people like it. Let's take a look at what Gemini 3.0 looks like in numbers to back up that claim, right? in the intelligence department. This is the artificial analysis, by the way. One of the most popular LLM benchmarks because I'm I'm kind of a normie, right? There are very specific benchmarks. This is the one I usually look at because I think it's really really good at comparing the Frontier models. Gemini 3 Pro Preview is up there in Intelligence.
Watch from 4:26Higher is better and it's higher than GPT 5.1, KI K2 thinking, and some of the other best Frontier models like Rock 4. So, it's really clever. Is it the fastest? No, but nobody expected it to be. I feel like Frontier models are not really measured on speed. And still, Gemini 3 Pro is pretty far up there, even faster again than GPT 5.1. So, I think that's a really, really nice trade-off for me. This intelligence is always worth more than speed because in cursor, while the agent is doing its thing, you can just go make a coffee, man. It's really good. And then price, it's fair. It's on the more expensive side right here. Overall, the ranking is really good, man. I'm pretty happy with it. Let's take a look at the coding index really quick because it's really good at coding. And not only that, it's also really good at designing. Now, we're going to implement a landing page design together from scratch in a aesthetic like this. So, this is a project I did like half a year ago just for fun. I decided not to continue it, but I really like this aesthetic, right? The colors here, the mono font, the pixelated developer kind of aesthetic. Let's just give Gemini 3 a single prompt to get a landing page working in that same aesthetic. Here we go. Create a landing page in the aesthetic of this page. It's multimodal, so it should be able to see how this looks like. A nice green tone with the primary being gray dark monofont developer aesthetic clean and minimal.
Watch from 5:45And this is it's all right. It's all right. I think it it can do better. The biggest thing here is it does not look like AI slot man. There's no pink gradient here. It looks fine. It didn't really listen to the mono instructions that I gave it. All right, man. This is impressive. Is it a picture perfect representation of the website before? No, especially not the mono font. But this is so far removed from the AI slop you get with basically any other model. I will go as far as claiming that Gemini 3 is the best UI design model that I have ever seen. It's not a perfect website, but what it just cooked up in two shots, right? This took like two minutes to implement is really really impressive, right? We can easily follow up, tell it, hey, use the monofont more, implement the actual functionality here, which it can easily do, right? It's great at back end. This is surprising, man. I am genuinely impressed about how good this is at design. So, this website was good, but it wasn't incredible. So after this video, I did some more investigation and turns out Gemini 3 can oneshot insane websites and it turns out my prompt just really sucked. So if you spend like half a minute more than I do writing a solid prompt and attaching an image right away, being specific about that image, Gemini 3 is actually much much better than what you just saw in that demo I did in the video. But I just learned that after the video, so I thought I just tell you here, man. I really hope you enjoyed the video. That's going to be it for this one. and I'm going to see you in the next one. Until then, have a good one and bye-bye.
Watch from 7:10Full transcript
- 0:00
All right, man. Gemini 3.0 is out and it's an insane UI designer. Like, it's really, really good. Do you remember that purple gradient AI slop that basically any other model generates? Well, Gemini 3 is out here crushing some of the most insane AI generated UI designs that I have ever seen. And this release was long awaited, man. So, back
- 0:21
in January 2025, they released the Gemini 2.5 series. That was a really long time ago. So, let's kind of put that on the timeline here. This was Gemini 2.5, right? It has been the go-to model for basically everyone during this period because the 2.5 models were really good, especially 2.5 Pro, they
- 0:42
were cracked, man. So, during this period up until like mid 2025, either Zonet or Gemini 2.5 Pro were the go-to models for coding. They were really, really good. In August, GP5 came out and that was probably the time up until here everyone used Gemini 2.5. And starting from August 7, myself included, people use GPT5 more and then slowly started
- 1:02
switching away from the Gemini models. And now in November, right here, we got Gemini 3 by Google. Here we go. So people are probably going to switch over again. And the timing here from Google was kind of immaculate because 2 days ago, there was actually a new release by XAI Gro 4.1, a frontier model that sets a new standard for conversational
- 1:23
intelligence, emotional understanding, and real world helpfulness. So, interestingly, the focus is not really on coding here. It is good at that, but probably not the main use case for 4.1. This is not really made for VIP coding. Gemini 3 is. And one of my favorite parts about Gemini is how good it is at design, by the way. So, Gemini 3 is now the most upto-date, the best model to
- 1:44
use for coding right after 4.1. So, really, really good timing for Gemini 3. And now, let's take a look at how good it actually is. How does Gemini 3 score? It's pretty good, but not on every benchmark. This is published by Google. So, of course, these are going to be the most favorable benchmarks here. Like, for example, humanities last exam. It's 2500 questions in a very broad range of
- 2:04
subjects. It scores pretty high. 37.5% and then 45.8% with search and code execution. So, a whole lot better. And we're not going to go through every benchmark here, but one other one I found really interesting is right here in mathematics, right? The AIME 2025. No tools 95% with code execution 800. That's pretty similar to cloth zonet
- 2:27
4.5. A little bit better without tools though. Is it realistic that models are not going to have tools in the real world? No, probably not. In cursor, they do. So, it's pretty comparable, right? So, the bottom line of what Gemini 3 is. Do you remember Gemini 1 when that came out? Gemini 1 introduced two pretty pretty cool new things for the Gemini
- 2:48
lineup. The first was the multi- modality. So it could actually understand images and understand visual context. That was one of the two big things that Gemini 1 introduced. The second one was context. So Gemini has always been really good at context, at least in theory, right? It allows a really long context window. With Gemini 2, they introduced all the missing pieces that one didn't have. Like for
- 3:10
example, thinking or reasoning, right? In my head, it's always kind of the same thing. But arguably more important, especially for development use cases, it's tool use, right? It actually became pretty good at using tools. Logically, the goal for Gemini 2.5 or more so 3 is to make one really good model that kind of combines both, right? It's
- 3:31
multimodel. It has really long context and then it can also think and use tools very, very well. So that's what they're going to do in Gemini 3. At least that's the intent. That's the promise behind this model, right? With the Gemini 2 lineup, one of my biggest problems and one that many, many other people had has always been context, right? So, on paper, the context window looks great,
- 3:53
but if you get past 100,000, 128,000 tokens or even more, it gets really, really bad at following your orders, right? And it even went so far as this right here, I have uninstalled myself. Gemini has some issues to work through. People are reporting that Gemini 2.5 keeps threatening to kill itself after being unsuccessful in debugging your
- 4:14
code. I cannot in good conscience attempt another fix. I am uninstalling myself from this project. You should not have to deal with this level of incompetence. I am truly and deeply sorry for this entire disaster. Goodbye. And then Gemini uninstalls itself, man. And that problem is a lot better in 3.0. So I work as a software engineer. I have
- 4:34
a full-time job. I've used Gemini 3 all day yesterday, right, to try it out and I've also searched on Reddit what other people think. Gemini 3.0 has been fixed for me at least and it's fantastic, right? So, especially the problem with context window regen. So, it gets worse the more context you have. It's gone, right? It's it's got really really good at that. I have a 500,000 token story that I had to abandon because Gemini 2.5
- 4:56
would process trash into chain 30 objectives together and ignore commands to stop. opened that same conversation, ask 3.0 to clean it, and so far it has not degenerated with no objective spam and a really, really good tone and style. So, the overarching consensus is people, myself included, are pretty happy with the model, right? On Reddit,
- 5:17
talking to other developers, I feel like a lot of people like it. Let's take a look at what Gemini 3.0 looks like in numbers to back up that claim, right? in the intelligence department. This is the artificial analysis, by the way. One of the most popular LLM benchmarks because I'm I'm kind of a normie, right? There are very specific benchmarks. This is the one I usually look at because I
- 5:38
think it's really really good at comparing the Frontier models. Gemini 3 Pro Preview is up there in Intelligence. Higher is better and it's higher than GPT 5.1, KI K2 thinking, and some of the other best Frontier models like Rock 4. So, it's really clever. Is it the fastest? No, but nobody expected it to
- 6:00
be. I feel like Frontier models are not really measured on speed. And still, Gemini 3 Pro is pretty far up there, even faster again than GPT 5.1. So, I think that's a really, really nice trade-off for me. This intelligence is always worth more than speed because in cursor, while the agent is doing its thing, you can just go make a coffee, man. It's really good. And then price,
- 6:21
it's fair. It's on the more expensive side right here. Overall, the ranking is really good, man. I'm pretty happy with it. Let's take a look at the coding index really quick because it's really good at coding. And not only that, it's also really good at designing. Now, we're going to implement a landing page design together from scratch in a aesthetic like this. So, this is a project I did like half a year
- 6:43
ago just for fun. I decided not to continue it, but I really like this aesthetic, right? The colors here, the mono font, the pixelated developer kind of aesthetic. Let's just give Gemini 3 a single prompt to get a landing page working in that same aesthetic. Here we go. Create a landing page in the aesthetic of this page. It's multimodal, so it should be able to see how this
- 7:03
looks like. A nice green tone with the primary being gray dark monofont developer aesthetic clean and minimal. And this is it's all right. It's all right. I think it it can do better. The biggest thing here is it does not look like AI slot man. There's no pink gradient here. It looks fine. It didn't really listen to
- 7:24
the mono instructions that I gave it. All right, man. This is impressive. Is it a picture perfect representation of the website before? No, especially not the mono font. But this is so far removed from the AI slop you get with basically any other model. I will go as far as claiming that Gemini 3 is the best UI design model that I have ever
- 7:47
seen. It's not a perfect website, but what it just cooked up in two shots, right? This took like two minutes to implement is really really impressive, right? We can easily follow up, tell it, hey, use the monofont more, implement the actual functionality here, which it can easily do, right? It's great at back end. This is surprising, man. I am
- 8:08
genuinely impressed about how good this is at design. So, this website was good, but it wasn't incredible. So after this video, I did some more investigation and turns out Gemini 3 can oneshot insane websites and it turns out my prompt just really sucked. So if you spend like half a minute more than I do writing a solid
- 8:28
prompt and attaching an image right away, being specific about that image, Gemini 3 is actually much much better than what you just saw in that demo I did in the video. But I just learned that after the video, so I thought I just tell you here, man. I really hope you enjoyed the video. That's going to be it for this one. and I'm going to see you in the next one. Until then, have a good one and bye-bye.
Description
DAMN BRO THEY COOKED FR FR
--- crazy gemini 3.0 ui designs i found on twitter
https://x.com/eter_inquirer/status/19...
https://x.com/ryanvogel/status/199082...
-- my links
twitter: https://x.com/joshtriedcoding
second channel (in depth videos): @Joshtriedupstash
github: https://github.com/joschan21
thanksfor watching, appreciate ya 🐐
More from Josh tried coding
×1312:00:00Build a Complete Digital Marketplace with Next.js 14, React, tRPC, Tailwind | Full Course 2023
Josh tried coding571K views
×1212:00:00Build a Complete E-Commerce Shop with Next.js 14, Tailwind, React | Full Course 2024
Josh tried coding509K views
×4.44:41My Favorite UI Library Keeps Getting Better
Josh tried coding191K views
×3.76:58Fetching Data Doesn't Get Better Than This
Josh tried coding160K views
