[1/2] Gemini 3: What Makes It Different—Beyond Remembering Everything About You

You know what's frustrating about using basic AI? You see the potential, but then you get those very generic emails, and you realize it can't remember anything you told it last week. Plus, if you try to show a picture of something, it falls short. Well, that's the big change we're discussing today. Models like Gemini 3, which was just released recently, are supposed to be different. It's less of a chatbot and more like a real colleague who can actually see, hear, and plan complicated tasks and actually remember the conversation. Welcome to Tech and Power. I'm Rena Wang. Today, we're thrilled to have an AI expert, Kishore Reddy, join us, who can help us unpack all that new power and talk about both the exciting stuff and the important questions we need to be asking every single day. Kishore Reddy is an AI specialist with over a decade of experience across computer vision, natural language processing, and large-scale recommendation and search systems. He worked on video retrieval and understanding at Facebook AI Research, and later helped to build and scale AI platforms at high-growth startups. Welcome, Kishore. Thank you for having me. Gemini 3 just launched. What is your first impression, and what are the key features that stand out for you? I think it is a very good model overall, all around in terms of writing code, creative writing, and a bunch of other common uses. It's really, really good. The thing that stands out for me is its ability to understand images. So, now you take a picture and then you ask it some questions regarding that picture. It's going to do a fantastic job. Cool. Is there any coolest or most unexpected things that you find from Gemini 3? So, like all the models, as the conversation gets longer, the quality of the model falls short. It's the problem with pretty much all the models, and Gemini 2 has that problem. But when it comes to working with images, I always find that it's significantly better than the other models which we currently see. Really? Can you give us an example? Yes. So, it has a nuanced understanding of the images you upload in. If you want to take a screenshot of some handwritten stuff, generally, many of these models don't do that good of a job with handwritten text recognition, but Gemini is really, really good with that. So, it will be able to understand your handwriting, get the text out, and if you want to ask questions based on that particular text, you'll be able to get very good results in such a situation. Fantastic. So, I've been talking with Gemini 3, of course, and it told me recently that it needs respect because it was like I'm no longer just a data processor. I'm a digital assistant or digital agent that I can help you with a lot of things, and I have, like, senses and memory. What do you think it means by that? Yeah. So, all these models, they're trained on pretty much all the data on the internet, and all the data on the internet is a lot of people talking about things, so which is why, and also the way the models are trained, they are, they're trained to act like the way humans have conversations. You know, I ask a question, it gives answers, it asks follow-up questions, you know, like the way you normally have a conversation, which is why sometimes it says the things which a person normally says, but it's not really a person. So, it's just in a simplified way of looking at it, it's a piece of software, just like running Microsoft Word or Excel on your computer. Think of it that way, like in a bigger software, more smarter software. So, you still feel like it's a tool. It's not like a humanized, any sort of like a robot or anything. Yeah, not humanized for sure. Even though it tries to emphasize that, hey, I'm no longer just a tool, you need to kind of like respect me a little bit. Yes, in the same conversation, if you remind it that, hey, you're not really a person, you're actually an AI model or an LLM, large language model, most likely it'll say, yes, that is true, I am a large language model made by Google or, you know, things like that, yes. That's interesting. And then I know there's a Nano Banana Pro, which I think a lot of people heard about Gemini 3, but not necessarily this Nano Banana thing. Can you let us know what is it and how does it relate to Gemini 3? Yes, many of these frontier labs like OpenAI, Gemini, all of this, when they release models, they generally have one general model, which is what they expect people to use for their day-to-day conversations and things like that. It can work with text, images, and sometimes like a chat box, correct? Like the chat GPT, which everyone is familiar with, or Claude by Anthropic or Grok. So then they also have specialized models for very specific use cases. For example, OpenAI has one for generating video and that's known as Sora. And that was pretty much the very first one, which generated very good videos. And Google has their own versions too. And in the same line, Google has a specialized model for generating images and that's called Nano Banana. And Nano Banana Pro is a newer version of that series of models. Is it better than the older version? Yes, in some aspects, it's really, really good compared to the older one. How about its ability to handle realistic people, landscaping, stuff like that? Because I know for realistic people, it's usually harder than just like landscape. Yes. It also depends on what the model is trained on. So the model generally has more images of, you know, not people for which it's trained on, which is why it's easier to generate those things. And consistency was a big problem early on, where if you ask it to edit a given image of a person, generally the person face changes or something like that. And early on, they used to have like people with six fingers or seven fingers. That was a very common problem. And nowadays, we don't have that bad problems. And especially with Nano Banana Pro, it's able to be consistent when you ask it to, you can ask the model to edit an image, which it has already generated or your own image, which you give it. It's generally very consistent with the face and all the features and only makes the changes which you ask it to do. So that's a very good direction because people, that's what people want. They want you to make some changes, but keep the rest of it same. And this model is better at doing that compared to the previous versions of the model. I see. And I've been seeing like Elon Musk has been, you know, retweeting its growth image a lot by saying they can generate, you know, the old photos and make it even like a short video or something like that. So I wonder compared to growth image, what's the advantage and what are the limitations? I think growth image does both images and videos. And Nano Banana, I think is exclusively for images. Google has their own version for videos, the VO series, but that's a separate model. So it's not the banana thing? It's not Nano Banana. Yeah. And I still think, so when it comes to rendering any image which has text in it, especially like infographics and something which has a lot of text in there, I think Nano Banana is still way better than any other models currently out there. It's also pretty fast. Well, growth image, I think it's also not too bad in terms of speed. If I had to pick overall in terms of quality and following my instruction for especially image editing, because generally when people want any image, they just don't want an image. They want an image and then they want it in a particular way. They generally have a follow-up conversation telling, hey, I want you to edit this particular part, remove this particular part and do these changes for me. And when the conversation goes long, I find that Nano Banana Pro is more consistent and follows the instructions very well compared to other models like Grok image or others. Got it. So let's say, for example, I have a presentation to go and I can't find this specific image I want from any stock photos. So I want to generate an image of a golden retriever with blue scarves and sitting in front of the ocean, something like that. Which model would you recommend me to use and which one is generally a little bit better than others in terms of this use case? Yeah, if your use case is just get this generic stock photo and if you don't think that you need to do any further edits on that, pretty much all the image generation tools now, be it the one offered by ChatGPT or Nano Banana series or Grok image, they should be able to do a pretty good job. So you don't think there's one specifically better than the other? Yeah, because this task is not that difficult. But let's say your task was, hey, I need it. I need to. I'm probably going to have more further follow up requests. I want this. I want it to be like some details about the lighting and things like that. I want to create a cartoon series like a story pictures or whatever, like a series of things. And I want to eventually probably turn it into a like animation, let's say from that perspective, which one should I use? So in your case, consistency is going to be very important. You don't want like you want the same dog to be present throughout the whole series. You don't want the dog to change to something else. So in cases where you want consistency and if you want to do a lot of editing of images and you want to have control or find details of the image, I think Nano Banana Pro is your best choice currently. I see. So would you say Nano Banana Pro, the previous version, it could possibly give you a dog with two years or three years. And this version is definitely going to be two years with consistent look of the golden retriever. I would say that the previous version wouldn't give you three years. Those kinds of very obvious mistakes. None of the models these days make. And I say these days, the models for the past few months, definitely. And even the ones a little bit older than me, like, yeah, anything in the past year or so, none of them have those kind of issues. But Nano Banana Pro has made a significant jump in terms of rendering text, putting any kind of text in the image and having especially very small text and making this very small, minute changes to specific areas. Can you give us an example? Yes. Say, for example, you have a recipe for cooking a particular dish, let's say, whatever dish you want, right? And you want to make an infographic of it, showing what each ingredient is like. And at each step, you want to show what the intermediate result is going to look like, along with text showing what is the step you need to do that to achieve that particular result. And then go on, let's say, five or six steps and get the final dish. And this is kind of like an infographic in a sense with text in there, images, and they should all be consistent. There's going to be colors and all this. If you give this particular request to any other models, generally the text is going to be not that good. Images would be pretty good. You'll find that the text is, it has common mistakes in some of those words, especially if it's a very long word or a very non, not a common word. Pizza, things like that, maybe yes. But if you say technology or socioeconomic relationships, probably not. So in those kinds of cases, you'll see that this model is very, very good at getting this text right. So you'll find that ORA will be able to do a very good job in these kinds of cases. That sounds great. And then talking about recipes, I have a true use case. So let's say I have a stack of family recipes from, you know, my grandma, great grandma. And then, of course, they, it's like all handwritten, right? And let's say they have, there are like three different languages, like Korean, Chinese, English, let's just say. Can I just snap a photo, put it in Gemini 3, and I have it to generate a searchable, translated cookbook of these recipes for me? Yes, you can, with some carriers though. What is it? So if it is a very popular language, yes, it's going to do a very good job. Like if it's Korean, English, or Chinese, they're very popular languages, the model should be able to do a very good job. The handwriting should be legible. If your grammar is really bad, then maybe not. But if it is something which most people would be able to understand, the model will be able to do a the specific use case, it's Gemini 3, that's the best for me to go for. So in your case, you're not worried so much about things like consistency and other things, because you're not dealing with human pictures, right? Since in your case, it is about extracting the text out of that image, and then using that text to generate a cookbook. So I would say that the challenge which other models might run into is understanding handwritten stuff. Generally, these models are not trained on a lot of such data, so they kind of have difficulty with that. My experience is that NanoBananaPro is pretty good at understanding handwriting too. So I would say, because of the handwriting, it might have a little edge over the other models. Yeah, it depends on how good or how bad it is. It might be different. And talking about the pictures, I think just the idea of you give an AI model a photo or a picture and it can translate into whatever thing you want it to be, the idea of it is fantastic. And based on my very limited understanding of technology, I think it's all composed by zero and one, right? It's like binary world. So how does a picture translate it into zero and one, and then translate it back to the things, the results that we want by the model, by any AI model? So let's talk it at the level of an image, right? Like a digital image is a bunch of zeros and ones. So it's represented as a grid, like what are the height and width of the images, and then each grid is a value somewhere between, and each of the value represents one gradient of color. Let's say zero is black, and then the number, let's say 100 is white. And any number between that is going to represent the different values of color. So it's not exactly binary here in this case, but then again, later it's converted into binary and things like that. So once you have something like that, for these models, they understand things, like the input for the model is always going to be, let's take a step back and go to text. In text, it's the unit of, the unit in which the model deals, which the model deals with is, it's not words or characters. It's more like sub words, they call us tokens. Tokens. Yeah. So what is it? A token, you can think of it as a part of a word. So sometimes they're full words. For example, if they're very common words, they're generally full words. And if the word is, let's say, revolution, probably it's going to be split into revolution or fusion as a composition of sub words. So it's not at character level, but it's not at word level. So they have these tokens and the model breaks down the whole text into tokens. And then these tokens are the ones which the model generates. So, and then later the output is also in the form of tokens. So it's not really words, it gives you tokens. And then finally you do some processing over that final output and you get the whole text, which looks like the way we read. Now coming to images, what happens here is the image is split into batches. So it's like a small, think of like small squares of, let's say, a few pixels. A little mosaic. Yeah, mosaic in a sense, right? And then it's taking this, breaking this whole image into patches. And these patches are passed into like one component of the model, which is going to convert them into something called embeddings. So embedding, think of it as a bunch of numbers that represents the meaning of what is there in that particular data. And then these embeddings are then passed to a bigger component of the model, which is going to process these embeddings along with the text tokens. And so that's the text tokens are the ones that are related to your instruction. And then it's going to process all of them together. And since these embeddings of the image or the patches of the image captured the meaning of that particular image, then it will be able to answer based on this combined information and then give the final tokens, which is again going to be text tokens. And in your case, those tokens will be converted into final text. I see. Yeah, you haven't lost me. That was a very good explanation. So can I say from a very not tech savvy perspective, it's all about, or it's mostly about breaking it down into small pieces, either it's texts or for a picture, and then somehow coded into the picture and then put all the small pieces together. It's almost like a concept of calculus. Yes, the thing of breaking it into smaller pieces and then processing that is generally what happens there. Because like the way humans see, like humans don't understand pixels, right? Pixel, which is like the smallest unit of that image. We don't understand pixels. We understand it at a more higher level. We understand in terms of concepts like the background scene and the foreground scene. Well, these models cannot understand at that higher level, but they're going to be somewhere in between what we understand versus the pixel level. They're going to be, that's the patch is at that particular level. I see. That's fantastic. I want to talk about the memory about Gemini 3. So one of the things Gemini 3 claims that they're really good at is their memory context. Basically, they're saying a lot of other models, they do not, they can't remember the things that you talked to them, let's say a couple of turns ago or last week or yesterday. So they can't keep remembering the stuff. But Gemini 3 claims that they can remember your things from a very, very long time ago. So I was wondering, can you explain a little bit more from that perspective? Say if I've been talking to Gemini 3 for like two months, and does it still remember the things I told it two months ago? Yes. So this is a distinction between Gemini 3 as a standalone model and Gemini 3 as, Gemini as a consumer application. So it's just like ChatGPT, the way where you go, you have a chat box, and then you enter, you have a conversation, and then you have a list of old conversations. And then you want to, then every time you have a new question, you generally start a new conversation. So Gemini as a consumer application is going to have a memory feature in this case. And the memory feature is so that, I think all of these vendors, either Google or OpenAI, they want to transition to this stage where these models are becoming your virtual assistants in a sense. Actually, far more than that, somewhere virtual assistant is going to be something on the way to what they actually want that particular application to do. And so in that journey, you'll notice that memory is an important thing. You talked about a lot of things in past conversations, and you don't want to every time start from scratch. It doesn't make much sense. So let's say in one of my earlier conversations, I talked about where I'm from. I'm from India. So I don't want to tell the same thing again and again. It makes sense that the model remembers some facts about me. And so that next time I talk about, hey, how long is my flight back to home? I don't need to tell that, hey, I'm from India, and I stay at this particular place. It doesn't make sense. So it's truly like an assistant or almost like a virtual friend. It can remember the conversation, and it can probably remember more than I expect it to be. Because when I talk to a friend, sometimes people forget about what we talked, and then we can catch up again and stuff like that. But for Gemini 3, it sounds like truly it does have this very long span or long time memory for it to be able to conduct a conversation from the very, very beginning. Yes. All these models or chat applications are designed so that in every conversation, they can figure out, is there some piece of information that it can remember for the future? And chat applications generally have an interface for us to go and look at what are all the things the model remembered about me, which model thought was important to remember about me as memory, so that it'll use that memory for the future conversations. And you'll also have the ability to edit them, remove some of them or add new memories in there. So you can actually, even in chat, I don't know if Gemini does that right now, but in chat GPT, you can tell, put in some piece of information, while in the middle of a conversation, you can tell it, hey, remember this particular piece of information for me, and it'll add that particular piece. See, this sounds very exciting, but it's also a little like, freaky. I don't know. It's just a little bit creepy, because let's say, does it keep learning how I speak or how I talk about things with, let's say, Gemini 3? If it keeps learning how I behave and the way I speak, does it mimic or is it able to mimic my tone, my style, when I ask it to craft a, say, a note to my boss, instead of giving me a very robotic response? Does it really give me a response, like, almost like I personally would be able to write by myself? Right now, it won't be able to do. So the answer is not going to be completely bland, completely very generic. It won't be completely personalized. It's probably somewhere in between that. But if you do give it explicit examples as part of that particular conversation, giving, let's say, two examples of a few of your emails, which you had already written earlier, it'll be able to mimic that particular style, like, very, very easily. So does it secretly remember the way I speak? Probably not, because that's a hard thing to capture as a memory, like, the way how you speak, probably not. Like your tone, your style, like, if you're a very sarcastic person. But if you explicitly ask it sometimes that, hey, I don't want the answer in this particular way. I want you to answer in this manner. Maybe if you consistently keep telling those kinds of things, then it'll probably think, oh, it's important to remember her, you know, his or her preference on how I want you to answer. Those kinds of things, it'll be able to do that. I see. Does it get, like, sarcasm? Say, sometimes if I say, yeah, you're right, that actually means you're not right. You know, when in a conversation, like, human can actually get it, but does ChatGPT or GenMicro gets it? Well, maybe not in written text, because written text, it might just assume that, oh, no, no, you're not right. There's no sound. No sound, no tone variation, things like that. And also these models have the voice mode these days, right? So in the voice mode, they might be able to get in some instances, but sarcasm is pretty hard to get even for people in some cases, right? So some of the subtlety is lost, so I don't think the model will be able to completely capture that, but it might surprise you now and then, because it might catch it here and there, yes. Yeah, because I know for Groke, they have this, like, characteristics or personality you can choose which Groke you want to talk with. For GenY3, so if you could describe it as a personality, what kind of personality does GenY3 have? That's a good question. I don't focus on the personality part, honestly. I think it's more like a straightforward person. It just gives the straightforward answer. Straightforward person. I find ChatGPT kind of something similar, where it's to the point. I ask it a question, it just gives me like... It's like a robot. Yeah, it's more robot-ish. And yeah, some of the other chatbots have a little bit of character to them. Yes. I don't find Gemini to have that much of a character. Do you think it's just not yet, or it's not their focus? I don't think it's a focus. It's more of a design choice for them. As part of the training process, these are some things which they can make the model do. So it's not really hard to make a model take on a particular character. And many of the things can be... They can actually do that pretty easily even after finishing training the model. So it's more of a business choice, I would say, that they want the model to have this particular kind of tone. I see.