Inside Google’s Gemini 3 Flash – One-Shot Apps, Hand Tracking, and Crazy Speed

Well, Google has a new model and it's really good at building apps quickly and cheaply. So, this is an app that someone's built where he's using the camera to track his hand in in 3D space as well. I see those details of like the pixels that light up and oh man, that that's really cool. Pinched strength. So, it's actually registering his hand gestures as well. So, really really cool. Uh another one or another set of cool examples is Google tried where this is again using multimodal and this is the thing about the flash models. It's really good at quick multimodal things that this is assessing video at the same time but I think what's happening here is actually giving you strategic advice based on the current where you are. So, it's really good at like oneshotting things and especially compared to how cheap it is, right? Yeah, exactly. And and we had a brief chat about um all the all these different models and what they mean and what they do and how you can use them and you know, it's hard to really summarize it, but we we'll try our freaking best. The naming confuses me. So, can you tell me like the names? The latest version of Google models is three, right? So it's Gem their their model is called Gemini and it's called and the latest version is level three. So we've just seen Gemini version 3 Pro which is like the hardcore researchers model. They also have a everyday use model called flash. So it's Gemini 3 flash. I see which is probably more comparable to GPT 5.2 you know everyday fast models whatever. They also have light. So they have three. They've got that three-stage thing. Light, flash, and pro. I think they have deep think as well, which is their mega mega mega one, but we won't get into that. Um, and they haven't released version three of Flash Light yet. It's still I think that's still on version 2.5. So this is basically their everyday model. Now, when it comes to practical use cases, as you say, it's really good at oneshotting something. They probably gave these they gave these apps maybe a couple of follow-ups but realistically they don't do well certainly in coding they don't do well with a lot of follow-ups like you know change this do that and then do that. That's you know Claude is just without a doubt the best coding model that I've come across. Um I actually don't think Google have released a coding model specifically for coding. They they have put flash in Google AI studio. They've put flash in the Gemini CLI. It is available in them, but the real world applications of this stuff speaks for itself. And it's not they're not their models aren't really all that good for develop like codebased coding. Now, I already feel like this is already getting quite confusing even just talking about it. The point is is that um this is your everyday model. Gemini is really good as your everyday. The flash anyway is good. good with for your everyday stuff. So, I hope that clears things up, but it all comes down to your independent usage of the model. If you want to code with it, you go ahead and code with it. If you want to do some research with it, go ahead and do some research with it, you know, and and see how you feel about it. But don't don't expect these models all to be great at at everything. And we'll get into benchmarks maybe next week, but there's a lot to be said about the actual benchmarks because actually Flash is benching really, really well. In code, it's benching really, really well, which is fine for the oneshot, but we don't oneshot things. We we develop it takes time, you know. Um, so anyway, just a bit of some of the boring stuff. You've seen what it can do. Actually, let me get some more examples on the uh on the old screen. I mean, it's just showing the speed at which is it. I think this is uh Gemini 3 Pro and then this is or vice versa um coding something. So the speed and the token usage a lot more efficient there which is really good. Um this is like I guess like another oneshot app that it's kind of built. It's you know you can click on all these sections of an image blah blah blah. That's cool. Anti-gravity as well. They're obviously pushing it in anti-gravity. Vertex, Genti, Gemini Enterprise, uh, Gemini CLI, Android Studio, Gemini app, and AI mode and search. So again, this will probably be the the mode the model they use when you Google search things. Thing to be said about it though, 90% hallucination rate. It hallucinates like a So how how how reliable you know again you always need to do your 91% hallucination like 90% of the things he says includes hallucination yeah some level of hallucination reliable you know just a small correction here it's if it doesn't know the answer to uh your query then rather than say it doesn't know 91% of the time it will make up a response to that query. the uh price is is against is 5050 50 cents in three $3 out. It's definitely usable, but at the end of the day, you are the human at the end of this chain and uh it will be up to you to make sure it hasn't hallucinated on you. It's interesting that, you know, Gemini 3 Pro kind of shook open API, Open AI's boots, right? And it's like crap. They're trying to they it's rumored 5.3 garlic they're trying to release and then like Google then casually release another model that's actually again benching really well and has some real world practical use cases that are relevant to its you know what's intended application. Um man I've written a post I think I think open AAI are done. I reckon they're falling off. I think they're falling off. It's really difficult to be in their position. I and I watched a video on how they are becoming Facebook or they tried to become Google but they are becoming Facebook. They are owning like crazy amount of data and your connections and like very sensitive data about you the way Facebook and Instagram and WhatsApp do. Um but they tried to to own so much of like the whole vertical but they they haven't been able to really do much even though they have Atlas they have they they're trying so many things and we'll talk about uh more stuff in uh in the other episodes and uh parts of the show but it is difficult to be in their uh in their place especially with the amount of money that they are Promising. Yeah. Yeah, for sure. This is part of a larger conversation on my show, Command AI, which we stream live every single week. We discuss the news and all things related to AI in the world of design and web. Catch us next week and join in the banter.