Deepseek Will Read 10X MORE From Images!

a picture finally says a million words or at least that's what Deepseek thinks in their new DeepSeek OCR model. So this is really interesting. It is just a paper. It's not they haven't um it's there's some work that's gone into it. So you do you know what OCR is? Object content recognition or something like that. It's it's a way of abstracting text from images. Trying to build a model because this is actually really hard. like you can kind of do it inside. I've tried it in Gemini. I've tried it in chatbt. It's kind of okay at doing it, but it's not great. In this app that I've built, I have actually resorted to using Python libraries that do it, and they're heavy. They're heavy libraries. These are like 4 gig, you know, just to just to do a bit of object recognition or whatever, right? Really cool. If they can do it well, amazing. Love it. But this what this paper is talking about specifically is they're exploring the idea of actually embedding tokens in images that then can be extracted from the image. And not only that, but you they they're getting 10x ratio on its token usage. So with a an image that's say a,000 tokens, they're getting 10,000 return on that tok like they can extract an equivalent of 10,000 tokens from that image. [snorts] So it's a really interesting way that you could put you know hidden messages or or whatever it is. you just store text in an image and it's a way of providing and if we talk about skills which we were just doing you could just provide an image inside of that thing or whether it's um an actual traditional sub agent that or MCP that takes up a lot of context you provide the instructions in that image and the model just knows how to abstract tokens from that image and again it's just a much more efficient way of doing things. So it's really cool. It works by like breaking up the the images and then obviously just having tokens representative of each section of the image. Um, and you can download it now from hugging face which you can actually use. You can't use the the the context optimal compression. So that that abstract that aspect is still being researched but you can download the the actual Deep Sea Go off hugging face right now and use it. So I might actually use it see how good it is. Um, it is saying it's got like a de decoding precision of 97% which is Oh, there we go. There's the 10x ratio. So, yeah. Um, interesting study. I'm feeling like Deepseek are really they're not really driven by finance. They're they're kind of really thinking out outside the box when it comes to like technologies and and what AI can do. Feel like all of the other companies are chasing profit. They're chasing money. Whereas Deepseek, they were the first ones to come out with a training. I don't know how much their training cost, but it was like it was a mere million or something like that. Whereas training of other library uh other models, they they're in the they're in the billions or millions, I don't know, a lot. They found out a way to train a model in using using a fraction of the costs. And then here again, they're actually putting forward a paper that's just rethinking how we use AI, how we use models, and what those models can do um in a really interesting way. So, yeah, I know they're Chinese, but that's I'm not saying that as a racist term. I'm saying it's like it's something to think about when it comes to the government and all that kind of lots. Yeah, there are things that we don't know. For example, we don't know if it's actually costing them as much as they are saying. they are not as transparent and we we don't know if it's true but presumably it's true. It's very um impressive. I'm I'm led to believe that they are because here's the thing and I don't know the ins and outs of the Chinese government but it basically the Chinese government has a hold of any company privately trading company or whatever it's called in China. So they have to obli if if China wanted them to put in something into their model or something in their data they would have to do it or if they the Chinese government wanted access to the data they would have to give it to them. Whether they're doing that, I don't know. But, um, so far, I mean, they seem like, you know, it seems like it's a fairly I mean, this is this is just a great uh example of how what they're doing to actually contribute. Like, I don't think they're collecting data by putting together a research paper. But anyway, that's by the buy. 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.