Welcome to the podcast.
I'm your host, Jaden Schaefer.
Today on the show, we're talking about Gemini 3 because Google has just launched it.
It has a new coding app and it has record breaking scores on a bunch of top benchmarks.
But beyond just benchmarks, it has some really cool use cases that have now been enabled.
So I'm going to dive into all of that on the podcast today.
Before we do, I wanted to say if you want to get access to all of the top AI models that I talk about on this show and you don't want to have to sign up for different subscriptions, go check out AIboxai.
That's my own startup.
You get access to the top 40 different AI models from Google, OpenAI, Anthropic.
You get 11 labs for audio and a bunch of cool image models, all for $20 a month.
There's a link in the description to AIbox.ai.
All right, let's talk about Gemini 3.
So this is just fantastic. fresh off of the press.
And it's actually coming just seven months after Gemini 25 was released which, as everyone knows, was also a really impressive model at the time that it was launched.
The new model now, Gemini 3, is, of course, the most capable LLM that Google has built.
And it is an immediate contender for some of the, I think, basically the best AI tool on the.
What I will say in most of the benchmarks that Gemini 3 does put itself up against?
To be completely fair, it's pitting itself against OpenAI 5, ChatGPT 5, because evidently they're probably running a lot of these benchmarks before 51 came out, which was just recently.
Or maybe they realized that 5.1 wasn't quite, maybe it beat Gemini 3, frankly.
So we're seeing it compared to Gemini 5.
In any case, it is a really incredible model and it's getting rolled out to billions of people around the world.
Sundar Pichai actually had a personal note in his announcement where he said that, because of the fact that they've rolled out Gemini to power the snippets at the top of Google search, that they now have over 2 billion users every month that are seen those AI overviews and are using them, and that the Gemini app itself has 650 million users per month.
We know that OpenAI's ChatGPT has about 800 million per week.
So it's gaining ground quickly and catching up with ChatGPT.
More than 70% of Google's cloud customers use ChatGPT.
Gemini and 13 million developers have built with Gemini models.
So it is getting a lot of traction.
It's not, you know, like meta AI, which I feel like is probably struggling in this regard.
It's getting a lot of traction and seeing a lot of use.
I think a more research intensive version of the model called Gemini 3 DeepThink is also going to come out to Google AI Ultra subscribers.
I am one of those and I've been able to... access and test this out.
And it's quite impressive.
I will say it's also able to generate video right inside of Gemini which, I'm going to be honest, it did have a little accident when I was testing it out.
I was like, hey, I had it like write me a bunch of implications of something that had just come out.
And I was like, make a video script for me with this.
And then, instead of making a video script, I literally generated a video of like, all of these personas standing around these charts, like talking like some marketing lingo, which was crazy for me to see, like an AI model just spit out a video for me.
That was cool.
But it also didn't do exactly what I wanted.
So that was kind of funny.
What I will say is that this is coming less than a week after opening I released GPT 5.1.
And this is only two months after opening Anthropic released Sonnet 4.5.
So we are seeing these frontier models come out at a very crazy pace.
One thing that Tulsi Doshi, who's Google's head of product for the Gemini model, said about this.
They said, with Gemini 3, we're seeing this massive jump in reasoning.
It's responding with a level of depth and nuance that we haven't seen before.
I mean, obviously they're talking their own book, but it is showing up significantly higher on a lot of the benchmarks.
So some of that reasoning power is on independent benchmarks.
They have a score of 37.4%. on humanity's last exam benchmark.
This is a famous benchmark, because people are like, when an AI model can get 100 on this it's, you know, humanity's last exam.
Obviously, they just make these like crazy names.
Humanity's last exam, for those that don't know, though, is a really impressive benchmark.
It essentially has, I think, like 60 to 100 questions that are super, super in-depth.
Now, I know people are going to like... roast me in the comments like it's not 60 questions.
It's like 72.
Okay, I don't know the exact amount of questions.
I'll be fully transparent with you on that.
It has a number of questions, which I believe is over 50.
And they're going to be covering some things that are very highly technical in very specific niches.
So it's like if you're an expert on how to you know, like wood filler using a certain type of tree when you're trying to build a table right, like just really nuanced bits of information, or you're Perhaps if you're a physicist and there's a specific field that you study quite in depth, you might know the answer to you, know you have a PhD, you might know the answer to one specific question, but you're not going to know.
But most people are never going to hear about that or even know what you're talking about.
Or let's say you're a historian and you've studied, like a specific, you know area in Greece from a certain time period and a certain artist and a certain style of art that most people and most AI models do not know.
The answer to
These are the kind of questions that are on humanity's last exam.
Very niche, very specific questions that if you're an expert on something, you might be able to get one of the questions on this exam right.
But you'll never be able to get 100%.
No human basically can.
And so the fact that they were able to get 37% on that is impressive.
And I will also say that I believe they got that without using any tools.
So it's just the raw model answering the questions.
If they add tools like calculators and different, like you know, physics kind of model tools, you could do a lot better.
But Without any tools, this is impressive.
The previous high score was held by ChatGPT 5 Pro, and it was 31.64.
So to go from 31 to 37, that's a big jump.
You know, 0.6 almost, 0.5, 0.8 is a good jump.
They also hit the top of the leaderboards for LLM Arena.
LM Arena is awesome.
It's basically a human benchmark.
It's going to measure user satisfaction.
So what that means is it's going to throw.
It'll give you like a question and then it will have the response generated by Gemini 3 and by GPT-5 or another model and it will compare them side by side.
And it's a blind test.
So you just look at the responses and you pick which one you like the best and then it will say okay, you know, This is the chosen model, and it will compare you to all the other models against a whole bunch of people.
And so getting a good score on that basically means in a blind test, people prefer your model.
So they were able to score a high score there, which was really impressive.
And according to Google, the Gemini app is... picking up a ton of steam.
Like I mentioned earlier, they're getting a ton of new users.
I mean, over 650 million monthly active users is a lot.
Beside the base model, Google also said that a Gemini powered coding interface called Google anti-gravity, which is going to allow you to have multi-pane agentic coding.
So this is going to be similar to a Gentic ID is like warp or cursor 2.0.
Specifically anti gravity is going to combine this sort of like.
Basically it's gonna have like a chat GPT style interface or I guess Google or Google Gemini style prompt interface.
And then it will have a command line interface and a browser window that can show what you're actually generating.
So you're chatting with it.
And it's, it's, you know, writing code, but then also creating what you have built.
I've seen some really cool demos where people are creating 3d models and images.
Google itself had an interesting demo, just to give you an idea, where they wanted it to create like a nuclear fusion simulator for the specific type of thing.
And it was not just making like this cool shape as the nuclear fusion, you know reactor, was like turning up its power, but it was also like making the pitch, the frequency that the nuclear or the yeah, the nuclear reactor was like creating, and the pitch was going up as the dial turned up.
So anyways, I bring this up to say like it was creating these 3D objects and it was also creating sound with them.
And it was, it was just, you know, making these really complex things you don't see other places.
The CTO of DeepMind, which is Corey said quote the agent can work with your editor across your terminal, across your browser, to make sure it helps you build that application in the best way possible.
Overall, I think this is, honestly a very impressive tool, I highly recommend you try it out.
And you got to get the Gemini Ultra Pro tier in order to do that.
But it is a fantastic model.
And, like I mentioned, it can do a lot more than just responding with text.
It's creating 3d models, it's creating videos.
Of course, there's amazing image models, Gemini is coming out swinging.
So I'm really excited to follow some of the advancements and what they're able to roll out.
A lot of exciting stuff coming down the pipe.
Hey, thank you so much for tuning into the podcast today.
If you enjoyed this episode, make sure to leave us a rating and review wherever you get your podcasts.
And also make sure to go check out AI box.ai.
If you want to get all of the top AI models in one place, you don't have to pay subscriptions for everything.
You can just go test out the latest from everyone over there.
Thanks so much.
And I'll catch you in the next episode.
