ChatGPT Just Dropped 7 New Updates, FREE Chinese AI Models & More Use Cases!

Welcome to another week full of AI releases and there seems to be common theme here. We have a lot of Chinese releases coming at us following the Deep Seek announcement. All of them are open Source, but also ChatGPT is shipping various updates that most people didn't even talk about and even Meta AI is updating its chatbots, all in the wake of the deep sea hype. So all that and so much more in this week's episode of AI News, you can use the show that breaks down all of this week's AI releases and filters for the ones that you can actually put to work today. So let's just start by talking about ChatGPT. And there's multiple changes here, some of them regional, some of them global. So I want to start by talking about the general updates which refer to GPT4O. The first one is that they updated the Knowledge Cutoff. Now it has Data up to June 2024, meaning if you ask it about any significant events from June 2024, it will know about it. Secondly, they improved the image understanding capabilities. And this seems to also be a reaction to some of the Chinese releases that happened earlier this week. Will talk about that later in this video. There's a new QEN model that is state of the art image recognition. So ChatGPT is forced to ship some updates that they probably had in their pocket. Beyond that, the model is better at STEM subjects and I guess it uses more emojis if you prompt with emojis. So if you don't want those, don't use emojis in your prompt. And then there's two regional updates for anybody in the eu, Switzerland, Liechtenstein or Iceland. Now everybody has these updated custom instructions which make it even easier for people who are not that deep into this to customize their ChatGPT experience. The same thing goes for the video features. If you're using the desktop app and you're in the eu, Norway, Iceland, Liechtenstein or Switzerland, you can now update your desktop app. And if you update it, you should be able to use the mobile app with the video features. We covered this when it came out, now it's just available to everybody. And also there has been one more change in the interface which is the following. If you open up a brand new 4O chat, they're just rolling out this new Think button. Interesting. I was preparing the video on my laptop and here on my desktop it's not even available. So this is just rolling out, but it's essentially a light bulb but button that you can select which will activate 01 and I'm bringing this up not because there's a new button, but because the main rumor right now is that O3 mini should be shipping today, but it hasn't happened and I have to record this video at some point. So just as a reminder, all three mini will have the intelligence of O1 or Deepsea. I guess they're relatively on the same level, but it will be way faster. And then O3 full will surpass anything that we have out of the US or China. O3. There's no release date on O3 mini. They said end of January, which is now. And I suppose this is going to ship very, very soon, and that's why they're also changing up the interface. Okay, but hold up, there's even more ChatGPT news and the changes they made to the Canvas here are actually some of my favorites in recent history. Okay, so this one really went under the radar. I mean, even Operator went under the radar last week with all the deep Seek news. But I personally am a really big fan of upgrades to tools that I'm actually using and Canvas is something that I use all the time and they made it way better. So let's just have a look at this in the ChatGPT interface. And the first upgrade that we see here is that now 01 is not grayed out anymore. So the thinking model that is exceptional at larger projects that require more foresight or planning. And of course, as you might know, these thinking models also excel in coding. So now we can use this instead of GPT4.0 with canvas. So I just have to go down here, enable Canvas. And the second update here is that it can now display HTML code, meaning you can build front ends of websites with this and actually see what it is doing. This was one of my main points of criticism of the Canvas feature when it came out just quickly before I show you this, I want to point out that this was already available inside of Anthropics Cloud. Their artifact features state of the art at building front ends and showing them to you right away. It's really good. But guess what? With the use of O1 that is really good at coding and this amazing Canvas feature, that matter of fact is the best thing there is for casual users to quickly produce code is a very potent combination. So let's just give it something simple like create a dashboard to track my personal finances as a video freelancer. And when I run this, O1 will think through the different steps and alternatives here. Canvas will build and display this. So this is what we were used to it writes all the code, but when it wrote HTML code web code, it never showed us. And you needed to bring this outside of the app, which is a pain because iterating quickly is what this is good at. So let's click preview. Boom, there it is. It just works. And it built a personal finance dashboard for you in seconds. I think this is really impressive. I do want to point out that it still lacks some of the amazing features of anthropic artifacts. Like for example, this website being actually usable. None of these buttons actually do anything right here. Or the website being shareable with others. With artifacts you can build a little tool and it just works and they host it too. But for non coders, this combo of a thinking model and its ability to create websites in one shot. And if you want to just change anything, you can talk to right here, you could even use advanced voice. Or you could use one of these presets here, right? Like hey, I don't understand what's going on here. How about you just add comments and then 01 will think about adding comments to everything and just do this. And there they are. Here are the comments that explain the code to you in human language. Like this button adds a new transaction in the real app, this would open a modal or form. And to round out this segment, I want to point out that to anybody thinking, hey, I could just do this with Deepseek for free, right? Because Canvas with 01 is a paid feature inside of ChatGPT. Yeah, sure you could, but let's just try that for a second. There's still a lot of value to having a pro plan with ChatGPT which is. Wait, hold up. I was actually going to rip on Deepseek a little bit and praise ChatGPT for the amazing Canvas feature, which it is amazing. I use it all the time. But actually when Deepseek writes a code, you can just press run HTML and it will show you the dashboard too. So this still has way more functionalities like this or the voice mode than others. But maybe this is the reason they shipped this feature last week. And I guess now you're aware that both of these features can use A1 to write codes and then display it for you. And actually upon further review, ChatGPT cannot just do HTML but also react components, which is what a lot of these navigation nice looking websites are built in. So when I told it to rewrite the whole thing into React, ChatGPT just did that and can display it. Whereas Deepsea created multiple files with no ability to display it. So ChatGPT is still ahead on this one. Okay. And there's another update from OpenAI and this is just an announcement. I just wanted to quickly throw it in here. They're going to be releasing a ChatGPT version for governments. The main takeaway here is that this is just going to be optimized version where the government agencies can rely upon the private privacy of their AI. And as you might imagine, this is a political move in the context of Project Stargate. And as they're now collaborating with the US government, they're also going to be providing their tooling to them through ChatGPT, the government edition. All right, let's move on to the next one, which will be a rather short one. It's Perplexity integrating Deepseek into the search. They already had 01 before. It's perplexity on the Pro plan, which is also paid integrating Deepseek R1 into the search engine. They already had 01, but now all pro users can use 100 daily searches with Deep seq R1 and and if you haven't tried AI search with reasoning models, I would recommend it. So I can say something like analyze the recent DeepSeek R1 release and the world's reaction to it and it will pull together a bunch of sources and run DeepSeek R1 on top of it, think through it and present you with analysis. That includes info from all the articles that it pulled together here to get the technical breakthroughs, the impact on the AI industry and so much more. All with sources, even with the limitations like it doesn't have image or video upload, which is true and most people don't even mention that. Anyway, if you're a Perplexity Pro user, I would say using reasoning models with AI search is the best way to do it today. All right, next up we have some more new free models out of China. These are also open sourced and we've covered most Quinn models that came out previously on this channel. And every time a new model comes out we keep it pretty brief. If this is not a state of the art type of thing that I would recommend for consumers, and that's not the case with these two. So I just briefly want to update you here. One of them is called Quen 2.5. 1 million. So the big difference is it maintains the quality while expanding its context window to eight times the size of something like GPT4O. And the second one is Quen 2.5 VL, a new state of the art vision model. But I think in the vision department, quality of the model is not as important, so you're probably fine with your current assistants like ChatGPT or Claude. Nevertheless, I want to briefly talk about these because they did fully open source these and they're small models. So this comes as a 7B and a 14B, meaning you can run this on a lot of machines locally and both of them have a 1 million context window. Interesting combo. No small model that you can run locally with a big context window. Super good for a lot of business applications that will be used on top of this. But if you want a big context windows two weeks ago on this show I covered HayleyU AI's minimax chat that has a 4 million token context window and is also freely available under this link. Check out the video from two weeks ago if you want more details. So this one would be really if you want to run it locally with a big context window, which I know a lot of you actually do. So there you go. All of the benchmarks look very healthy and this is available under a fully open source license with the context length benchmarks looking excellent. This is always a thing to look at because a lot of new models have a massive context window, but it just regularly loses the context that you give it. At least some of the first large context models did that. This looks extremely good. And the other model is the Vision model. By the way, both of these are available for free now under QuenLM AI. Bit of a hard URL to type right there. I'll put it in the description as per usual, but you can give these models a try right here for free. Here's the new Vision model and there are two interesting things about this one that I do want to point out. One of them, the Vision model does allow video upload and like. And benchmark wise this thing actually outperforms everything that is out there. Look at that. Only Gemini Flash can kind of keep up on some problems, but everything else is just surpasses on benchmarks. But again, as I said, I think vision benchmarks are not as extremely important for the common user. But again, only time will show. The thing that I wanted to point it out here is that they're saying this one is really good at analyzing text, charts, icons, graphics. So not just common objects, flowers or animals. If you ask me, this is way more important, but this model will be more something that people will use to build it into other applications. And the second thing I wanted to point out, I guess there's free is the fact that it has video comprehension so you can upload videos to this and it's state of the art at video recognition. And then my final point is that they actually released a computer and mobile agents similar to OpenAI's operator. But I won't be testing that right here. We're already working on a separate video where we'll be comparing Operator to all the open source and free alternatives. Some of the other alternatives in the market, like this one, there's really no point in doing one thing with this. You want to run multiple test cases and compare it to see what works better. So keep your eyes peeled for that video in the upcoming week or two if you're interested. So yeah, overall we have another two amazing open source models coming out of China. One that's super tiny with long context and one that is state of the art ad vision and video recognition, also open source. And they're starting to release these competitor products to something like Operator. So I suppose the Chinese AI steamroll continues. But none of these chat interfaces come even close to what ChatGPT does right now. So if you want my take, I'm still using ChatGPT and I'm not even close to switching over to anything like this. All right, next up, I want to briefly talk about Med AI because they shipped a very interesting update here. It's not available in the U and I'm not even going to bother testing this right now because in our preparation, Daniel on our team actually uses Med AI regularly and he tested this for us and basically what they did is implement a memories feature that you do not get to opt out of. Meaning as you can see from this marketing image here, it picks up different memories that you can then delete, but the collection of these memories and is automatically enabled and they don't even give you a choice. So when Daniel went ahead and he asked Med AI to tell him everything that Med AI knows about him, but after he talked to it for a while and gave it some info, and then he turned around to ask what Med AI knows about him, it told him that it doesn't have any information. So the features seemed to bug out on him on its first use. But when he asked where should he be eating today? It actually had access to location and gave him custom recommendations. I don't know, this is just a very brief test, but that doesn't seem to be a smooth integration at all. And additionally they don't exactly tell us what data they're using here, except of the fact that it will also consider your Instagram viewing history, your Facebook profile, Your Facebook locations and other profile data. This could potentially be a really wide net and you don't even get control over what you want to give it. I mean, I'm not sure I want my personal assistant to be customized on my Instagram reels viewing history. I think this is the right direction for an assistant, but having some level of control feels right, especially if you're shipping this to billions of users. So I don't know. We'll keep our eye on this one. So. So my guess would be that over time this is going to be better, but right now it's a bit hard to recommend. Okay, next up, we have another open source model from Deepseek AI, the makers of R1 that everybody went crazy about over the past week. And this is actually an imaging model. So what R1 could do is reason over ideas and produce really high quality outputs. What it could not do is intake or output images. And Deepseek actually just released the open source image generation of their own Janus Pro 7B. I don't want to beat around the bush for too long here. I think the best thing we can do is just open this up and run our typical test prompts through it that if you're a follower of the show, we like to throw out all new models. All right, so on this hugging face interface, we can just give this a shot. We scroll down to the text image generation, put in one of the prompts that we like to test with every model. Wait, hold up. What? Good lord. What are these images? What are those? I wanted to go old school for my first day. Oh my God. This is really bad. Ha. This looks like something out of 2020. Okay, wait, wait. Surely this must be wrong. Let's throw easy one at it. Aerial photography. Okay. These models are really good at that. Because in this day and age, every second tech enthusiast seems to have a drone. I actually have two myself. FPV one and a normal one. So they have a lot of images to train on. So let's see how it does. What the heck are these? Okay. No, this is just terrible. That's the right thing, right? I just got to double check. I'm on the right link here. Janus Pro 7B. The very first space that uses the model. Yeah. So okay, they did an imaging model, but this makes the old stable diffusion look good. Okay. Seems like Deepseek nailed it with the reasoning model imaging, not so much. On to the next one. And this is not really an application, but I like to filter all the AI news that comes out for the things that you can benefit from today as a consumer. That's the whole premise of the show. And I think this story meets a criteria. Basically, the United States Copyright Office that in many ways will influence the laws for the rest of the world finally published something on artificial intelligence and how that relates to copyright. I'll link the entire official document below if you want to have a read for it. But the TLDR goes something like this. Anything that is purely AI generated is not copyrightable. Your unique prompting style does not make it a unique piece of art. Now, if you're starting editing on top of it or use the AI generated assets as a part of some other piece of work, that's a different story than it's covered by traditional copyright. But if you generate something, it's not copyrightable whatsoever. Finally, some clarity from a government office on this. I mean, this is the common sense conclusion that most people came to. That's official though. Link to the document below. Let's see what's next. And there have been two more models on the open source front. I just quickly want to shout them out. I don't think anybody will be using them on a daily basis. But I wanted to mention Europe's Mistral Small 24B. And I don't know, I guess I'm doing this gesture because I'm a big fan of European companies. Jesus. This continent is just so beautiful. But in terms of technology, not exactly ahead of the game. And I wish that changed. I just don't know what to do about it. You might have seen one of these European innovation memes with the bottle caps as what Europe is doing. It is really funny because there's just a lot of truth to it too. Hey, but we have Mistral doing good stuff over here. There's also two Loo 3405B. I don't know why, but I find that name kind of funny. Two Loo. Two lo. And there's two benchmarks where this actually tops the charts. So it's a fine tune of llama 3.1405 B. There's some good benchmarking results, but I don't think for consumers there's anything specific you would want to go to this for. Just wanted to quickly bring it up because you could use this. And that's everything we got this week. I personally have been playing a ton with Operator and I just wanted to shout out the video that we made on all the use cases that we have found so far. I thought that was a really good one. So it's linked the screen right now, and I'm going to be spending my time moving forward playing with those kinds of tools. I really think that with Operator we cross some sort of threshold now where it's worthwhile to invest time into these tools. And now they're finally worth your attention as they bring some roi and they will only get better. So with that being said, I hope you found something that was interesting to you here. And I will see you next Friday in the next episode of AI News that you can use. Have a wonderful day.