AI News: The Biggest Leap We've Seen This Year!

It's been an absolutely insane week in the world of AI. We had so much news to talk about and well, wait a second. I'm not going to waste your time. Let's get right into it. Let's start with OpenAI. We got a lot of news out of OpenAI this week, but let's start with GPT-5.5, the brand new model that we have access to inside of ChatGPT and inside of Codex. GPT-5.5 understands what you're trying to do faster and can carry more of the work itself. Basically what this means, you can actually give it less information and less details and less context about what you're looking for, and it actually does a pretty decent job at kind of understanding what you're going for and then doing that. So you get better results out of putting less information and less context into your prompts. It excels at writing and debugging code, researching online, analyzing data, creating documents and spreadsheets, operating software, and moving across tools until a task is finished. Now, it also uses significantly fewer tokens to complete the same Codex tasks, making it more efficient as well as more capable. Now, this is important because if we take a peek at the pricing, this model is going to be $5 per 1 million input tokens and $30 per 1 million output tokens. Now, if we compare that to GPT-5.4, the previous generation here, well, that's double. GPT-5.4 was $2.50 for 1 million input tokens and $15 for 1 million output tokens. So with this price doubling here, it's a Good thing that it's going to use significantly fewer tokens to complete the same tasks. Now, if you're using ChatGPT, you probably don't care about this, but if you're using the API or doing things with like coding, that's going to be important to you. Now, before we get into the benchmarks and some tests that I did, here's who gets to use it right now. GPT-5.5 is rolling out to Plus, Pro, Business, and Enterprise users in ChatGPT and Codex, and the 5.5 Pro model is rolling out to Pro, Business, and Enterprise users in ChatGPT. So if you're on the free plan, or the $8 a month Go plan, you're not getting this new model yet. It's also important to know that the API is coming soon. So while you can actually use it in Codex right now, there is no APIs as of this recording. It's not in Cursor or any of the other third-party coding apps yet. Taking a quick peek at some benchmarks here, GPT-5.5 scores an 82.7% on TerminalBench. The previous GPT-5.4 got a 75%. Claude Opus got 60%. 69.4%. This represents how good these models are at running terminal commands. And the question on everybody's mind is, well, how good is this compared to Mythos, the model that Anthropic said was way too scary to release? Well, GPT-5.5 got an 82.7, and comparing that to Mythos, Mythos got an 82%. So this new GPT-5 model is a model that performs better on TerminalBench than the model that Claude refuses to release because it's too scary. When it comes to working with operating systems, it got a 78.7%. And as you'd expect, it's really good at math and stuff. Now I'll link to the announcement post in the description as well as all the articles I share on this video. But they do show off some cool demos like this demo of the orbit around the moon that Artemis took, this cool earthquake tracker app that they developed with it, this 3D dungeon game here, and then this 3D tank game. These are all video examples that you can see on their website if you want to see what it can do. They also have an example here of it doing some financial modeling inside of an Excel spreadsheet, as well as using the browser to test an onboarding flow for like a website. Here down at the bottom, we have a few more benchmarks. We have SuiBench Pro, which 5.5 comes in at 58.6%, which is not quite as good as Claude Opus 4.7, but for whatever reason, they don't have the evaluation with their 5.5 Pro model. Now, another evaluation that I do like to look at when new models come out is the Artificial Analysis. I like this one because it's sort of a blend between a whole bunch of benchmarks, right? The Artificial Analysis Intelligence Index is a single composite score that aggregates a model's performance across 10 hard benchmarks to reflect its overall general intelligence for text-based tasks, reasoning, knowledge, science, coding, and agents. So it's sort of like one benchmark to rule them all. It takes 10 different benchmarks, combines them all together, and gives each model an overall intelligence score. And the newest model, GPT-5.5 Extra High, is now the leader on this index here. Until 5.5 came out, we actually had a three-way tie between Opus 4.7, Gemini 3.1 Pro, and Gemini 5.5. 4, but now we have a new standalone leader. So according to this analysis here, GPT-5.5 is objectively the smartest model. Now, one thing I've said over and over again in past videos is that when new large language models come out, I actually have kind of a hard time benchmarking them myself and showing you how they'll actually be better. For most everyday people, quite honestly, you're probably not gonna notice a huge difference. You might notice a slight, like, vibe change when you're chatting with ChatGPT. Like, each of these new models might get a little friendlier or a little bit more robotic, or, you know, small vibey changes. But for the most part, these models are already pretty intelligent. So when you're asking it to feed you information about something, they're already pretty good at that. When you ask it to write code for a real basic app or a basic website, they're already good at that. When you want it to compare two products for you, the previous models have been good at that as well. So it's really, really difficult to tell that sort of everyday ChatGPT or Claude user, here's a big noticeable difference that you'll get out of these models. But the one thing I keep coming back to is how much better they're getting at doing more with less. So for example, you can give it a really, really basic prompt now, and it's much better at helping get you an answer to that really basic prompt. But that also means that the more information you give and the more context you give into your prompt, the better it is at that stuff 2. So it's gotten better and better at doing more with less. But if you do give it more, it's also way, way, way better at that stuff. And we're seeing that happen across every new model that comes out, is it's getting better at understanding what you're looking for even when you give it less and less information. What you're seeing on the screen here is actually still ChatGPT 5.4. This isn't the new model. I said, help me build a plan to be healthier. Super vague. Not a lot for it to work off of there. Here's what it gave me. I'll give you a practical starter plan for the next 30 days. The win is sleep a little bit better, eat a little bit cleaner, more consistently, and make it automatic. Here's your 30-day baseline. Sleep: protect energy first. Food: simplify, don't optimize. Movement: daily movement, not daily punishment. A simple strength plan. And it built me like a little workout plan here. Cardio: use it for energy recovery and stress. Your health day checklist. What to avoid. If you read this, it feels fairly generic, like this could be a health plan that it would just give to anybody. Now I switched the model to ChatGPT 5.5. Now this just says ChatGPT up here, but if I click on configure, you can see that this is using the 5.5 model. I have this set to latest and it defaulted to 5.5. When I asked this one, help me with a plan to be healthier, it seemed to have dug into the context of past chats, understood more about me, and gave an answer that was much more fine-tuned to me. More energy, less fat, stronger arms and shoulders, better focus, and a system that works even when YouTube news events get chaotic. Already much more personalized just in its sort of preamble. Your biggest nutrition issue historically is not that you eat too much junk, it's that you often skip meals, under-eat protein during the day, then eat most of your calories at dinner. That's 100% me. I have had conversations with ChatGPT about that in the past, and it pulled that out when I asked this prompt. When I scroll down here, your workout plan, your ideal workout days are Tuesday, Wednesday, and Friday. Mondays and Thursdays busier, so those become movement days. It knows that I record these news videos on Thursdays and publish them on Fridays. So it's like, hey, take it easy on Thursday. That's the day you spend all your day making this video. It even went on to give me a travel plan because it knows that I travel a lot. When traveling, don't try to follow the full home routine. And it gives me a routine to follow when traveling. So again, same exact super basic prompt. Help me with a plan to be healthier. Same prompt here. As I gave with the 5.4 model here, one was ultra tailored to me. The other one with the previous model was fairly generic. Now you can just imagine that if I gave the new model a prompt that had tons of detail and tons of context, how much more detailed and absolutely relevant to me this response would be because it already made a pretty good response without all the detail. So next up, I wanted to quickly test like a really small coding example, but using the same premise. I'm going to give it very little context and just see what it gives me. And this is the prompt I gave it: Create a beautifully designed interactive website that describes and visualizes what your model, GPT-5.4, is best at. I want to see what you do well in your work. This is the one I gave to GPT-5.4. You can see down here in the prompt box. Sometimes the models don't actually know what model they are. So I did in parentheses tell it that you are GPT-5.4. But other than that, I— used this identical prompt with GPT-5.4 and GPT-5.5. I just changed the model that was in the parentheses. And here's the website GPT-5.4 generated. We can see there's a little bit of like an animation going on in the background. When I move my mouse, there's a little bit of motion. Strongest when the work is real. I tend to do my best work in the messy middle. Repo-aware coding, long context synthesis, tone-sensitive writing, tool-assisted research, and the deliberate follow-through that turns a vague ask into a finished result. Reads before changing, acts with structure, verifies against reality, adapts keeps tone without flattening it. We have a little bit of like margin issues here. This doesn't look amazing, but it's not horrible either. I actually don't mind this design at all. So Capability Atlas— this design's a little weird how it overlaps some of the text here, and we can see it sort of scored itself ambiguous asks to action. That's what we were just talking about with 5.5. 5.4 was also good at that in comparison to the model that came before it. But again, as new models come out, that is one of the things that they are getting better at. Tool use follow-through, long context synthesis, verification under Change and Tone and Collaboration. And if I hover over these, this side sometimes changes. Maybe I have to hover over a specific spot, I don't know. The code is a little wonky on this one. This was just a one shot and we could see it's sort of rotating through this but not consistently. But moving along, Interactive Workbench, we can see it gave us some charts. Reasoning Depth, Verification Speed Under Uncertainty, Change Restraint. I don't know what this is down here, it's just like colored lines with no actual like context to what they're for. GPT-5.4 is usually strongest when ambiguity is real, stakes are visible, and there is some path to verification. Drag the sliders to change the working conditions. So if we crank up ambiguity here, we could see it sort of adjust these scores. It's better at reasoning. And as we adjust these sliders, we can see that this is actually like an interactive fit score that it gives us. And then down here, the habits behind the output: read first, pick a path, verify claim, keep the human in the loop. So the site design, actually not too horrible. There is definitely some weirdness with the design, but a couple more prompts and you'd something that looks decent. Cool. Now let's take a look at what GPT-5.5 generated with the exact same prompt here. Completely different design style. It says, messy intent into working shape. GPT-5.5 is strongest when the task needs judgment and execution together. Reading context, making a plan, building the thing, checking it, and explaining the trade-offs without flattening the nuance. Again, messy intent. If you don't quite understand what you're telling it to do, it will kind of go and figure it out for itself. It built this little visual here. Read, build, test, design, explain with some lines connecting them. Context first, clarity, craft, verification. This isn't interactive. This appears to just be like a single image here that it generated. We scroll down here and a lot of this interaction stuff works a lot better. Look, if I move over this, you can actually see that this interactive part works pretty well. This is actually hiding something underneath it. So the design's a little wonky, but it gave us a little sidebar that we can click. So if you can't actually click on the node behind it, the sidebar lets you click on it. But as I click through these, you can see our capability profile over here changes, our graph in the center here changes, the little like nodes here change, and we get a different description as we click around. So the strengths here are code-based navigation, implementation craft, debugging under uncertainty, interface design, research synthesis, and collaboration memory. Tune a task and see where I lean in. The sweet spot is not only answer generation, it is calibrated action, deciding what to expect, when to implement, when to verify, and when to ask for sharper constraint. And we can see here we've got some charts here with sliders. So prototype, if I crank up the ambiguity, once again we have a fit score. Both models decided to generate a fit score independently, which is interesting. We up our context size, up our quality bar, and now we've got a 99 fit score for prototyping. Let's go to research. We could drag this around and see how it adjusts our score as well. So very, very interactive. And then down at the bottom, very similar. What we do has a rhythm. Read before moving. Build the usable thing. Verify with evidence. Keep keep human in the loop. Once again, comparing it to this: read first, pick a path, verify the claim, keep the human in the loop. Two separate models, very similar outcomes, two completely different designs. So again, to sum it all up, this model is better at coding tasks and terminal tasks and math and science and all the things the benchmark says it's better at. But where I feel like most people will notice the difference is that doing more with less information— you can give it a vague prompt, you don't need to prompt engineer anything, and it will kind of infer what you're looking to do and help you get to that result, even when you kind of don't know what that end result should be. And then when you do actually give it more details, well then it's even better. Warp also just came out with a bunch of new features, and these are actually really interesting if you're using coding agents. One of the biggest updates is that Warp now has universal agent support, which basically means you can run things like Claude Code, Codex, OpenCode, all inside the same environment without changing your workflow. What Warp does is it turns your terminal into this agentic development environment where you can run multiple agents side by side, see what they're doing, and manage everything in one place. So you can have Claude Code writing a feature, Codex debugging something else, and you could literally monitor all of it with these vertical tabs that show you things like status and directories and branches in real time. They also added this code review loop feature that's built directly into the terminal. So now you can actually review the agent's code, leave inline comments, and have the agent update the code instantly, all inside the same terminal. And then there's this updated notification system. Usually when you use these agents, they run in the background, but you kind of need to babysit them to know when they're done or if they need input. But Warp now has this unified notification system across all your agent sessions, so it can ping you when something actually needs your attention. Instead of you constantly having to check. These are some great new updates coming out of Warp that seem to be really focused on making life easier for developers. So if you want to check them out, head to warp.dev through the link in the description box. And thank you so much to Warp for continuing to support my channel and sponsoring this section of today's video. OpenAI also rolled out an image model called ChatGPT Images 2.0. This is the next evolution of the image model that we've had access to already. You know, the one that everybody used for the gibbly-ification of images and the one that got quite a bit better at adding text in images. Now it got a lot better and by most people's accounts, even better than Nano Banana. When it comes to image models, I really like to go to LM Arena to see how people rank them because this is pretty much based on taste. LM Arena, you give it a prompt, it gives you two images without telling you which model gave you each image. You tell it which image you like liked best, and then that sort of blind taste test is what informs these scores right here. And for the longest time, Nano Banana dominated, which is also known as Gemini 3.1 Flash Image. Well, here we are now, and GPT Image 2 is now the top dog with quite a big jump too, right? The score for Nano Banana is 1271, and if you look at all the models below it, they're all kind of clustered in the 1100s to 1200s. This one jumps all the way to 1500. Here's an even better look without my dead blocking it. Interestingly, their website announced it fully in images. The whole thing is written up on top of images here. So as you scroll through their website, you can read about this model in the images that were actually created with this model here. They also have a classic mode. So if you just wanna read a normal blog post, you can click over and do that. Some of the standout features for this new model are the ability to render out dead text. The images feel less AI-generated. It's accurate across languages, and it uses world knowledge to fill gaps for you. So this is something we got in Nano Banana where you can say something like, generate an infographic with the human body and point to where every bone is, and it will use its knowledge and even search the web to find more knowledge to then inform what goes into that infographic. Well, now you can do that with the ChatGPT image model as well. Images 2.0 is our first image model with thinking capabilities. It can search the web for real-time information, create multiple distinct images from one prompt, and double-check its own outputs. But this is an image model, so there's no real big point in reading about what it does. We're much better off looking at what they've done with it and what other people have done with it. So for example, a screenshot of ChatGPT in a browser in Mac OS X. The user types draw me a dog. ChatGPT draws an ASCII dog. The front window is ChatGPT, but the desktop is quite messy with lots of windows open, example a terminal. They're all in the background. And this is the image that it generated. We got ChatGPT in the foreground, we got a terminal open in the background, a whole bunch of other apps, and it looks like a Mac OS X home screen. If you look close enough, you will see a little bit of weirdness, like our two emails and the Slack logo looking kind of like the wrong color. But for the most part, it looks pretty good unless you're looking close. It looks like what it's supposed to look like. I'm creating a magazine page with the theme of visual polyglot, and then there's just a lot of details. I'm I'm not going to read this whole prompt. Feel free to pause it or go read the actual article. But here's the actual image that it generated. Create everything at once. And then it just has a ton of stuff in collage form. A lot of them are things that were prompted in this initial prompt here. Anatomical drawings, old maps, climate charts, engineering schematics, transit signage. So we know this model can take in a lot of information and get all of those things into the image that it generates with very impressive results, quite honestly. Again, there's some other examples here. You've got your of rice, Science magazine page here, a handwritten essay, but I will link to it so you can take closer looks at all of these if you want. We've got a comic book page, an Indian bookstore, another Chinese comic, a Korean advertisement, typography poster, some realistic images that, I mean, actually look pretty dang realistic. There's not a lot of giveaways here that these aren't real. So definitely check out the actual blog post if you wanna see a bunch of examples. But instead of doing a bunch of testing, I went and found found some really cool examples that other people from around the internet have generated, and I'm going to show off some of those first. This one was really, really impressive to me. Leon Lin here showed that he created a 360 equirectangular image, so it almost looks like an image that was shot with like an Insta360 where you have a full 360 image that you can keep circling. And it obviously has Sam Altman and Jensen Huang and Tim Cook and Elon Musk all within this 360 radius, something that would probably never actually happen in real life these days. My buddy Riley Brown here posted this video on X that I thought was pretty cool. I'm not going to show the whole video, I'll link to it below if you want to see it, but what he did was he asked it to generate an image of the book Good to Great. The barcode on the book should scan to the actual book. So of course we can see here that it generated the book with a barcode on it. Riley then used a barcode scanner on his phone here to scan the barcode, and look at that, it actually took him to the book Good to Great. So the barcode that the image model generated was the actual barcode for the book. He then did the same thing with the book The Intelligent Investor, scanned the barcode, and it went right to the book's page here. And then to prove that it wasn't reading the ISBN or any of the numbers, that it was actually reading the barcode itself, he took the image, blacked out all the numbers, and then scanned the barcode again, only for it to go right back to The Intelligent Investor. That was super impressive. You should really watch Riley's video. It'll be linked below. Justine Moore here showed this really impressive looking comic book image. I really like the aesthetic. Like the colors look really, really good, but appears to do really well with these comic book pages. Gabriel here on X, I don't know what his prompt was here, but if we take a closer look at this image, it's like everything, I don't know, '90s maybe, with just a giant collage of like everything from the '90s or early 2000s here. Again, I don't know if he asked it for all of these things like Blink-182 posters and Monster Energy drinks and Game Boy Pockets and PS2s and all of that stuff, or just like give me a collage of everything from this era. But either way, he got an image that has a lot going on in it. Bo Wang here showed an image with a complete cell image by ChatGPT. And I'm not a scientist, so I actually couldn't tell you for sure if what it's pointing at is actually accurate, but it's a pretty good looking infographic if it is. I mean, I'm pretty sure that's the nucleus. That looks right to me. Plasma membrane, like this looks like a membrane to me. Again, I'm too dumb to know if this is actually what it's supposed to be, but it looks really impressive. Riley Goodside here posted this maze that they had it draw. School worksheet with maze, 32 wide by 48 tall, rows and columns numbered, no unreachable rooms, solution in blue pen, name field reads ChatGPT Images 2.0 in cursive at top. So we can see it's got the ChatGPT Images 2.0, it drew the whole maze and solved it. Chetoslua shared this 10x10 grid image that they generated, and it looks like they might have the prompt in the alt tag here, so feel free to pause if you want to read their entire prompt. It's right there on the screen. But here is the image that it generated. A 10x10 grid. AI models and agents. We can see a bunch of models. Robotics. A bunch of robots. Semiconductors. Networking and smart devices. Biotech. I mean, it's all pretty good looking. And yeah, impressive 10x10 grid. Mark here generated this image of a newspaper with a lot of dense text. I'm not but the news might be accurate. I mean, I obviously know the unveiling ChatGPT Image 2 is accurate here. Not sure about the rest. Maybe it is. It all feels like it could be, I guess. But again, the amount of dense text that made it onto the image and it's all coherent and readable. That's what impresses me the most about this. Underwood here posted this movie poster for Echoes, which looks pretty good. Viral Ops here generated 2 images, one in GPT Image 2 and one in Nanobanana Pro. Apparently this is the Nanobanana Pro version. And this is the GPT Image 2 version. If I'm being totally honest, I actually think the Nano Banana one looks slightly more realistic. Like here, her face looks a little too shiny and fake. I don't know, what do you think? I actually think Nano Banana Pro did it a little better. Yuchen Jin generated this image of, uh, Sam Altman and Elon Musk offering you a cigarette. So overall, fairly impressive. It does feel like a pretty decent leap over what we were getting out of Nano Banana. And you can use it directly inside of ChatGPT. You can either tell it to generate an image, or you can specifically click the plus and click create image. If you create an image, it actually gives you some ideas that you can potentially play with, like a blueprint poster or animal infographic, anime icon designs, newspapers, tarot cards. Create an infographic that best shows and describes who I, Matt Wolfe, am. All right, let's see what it created here. Who is Matt Wolfe? AI YouTuber, tech explorer, creator, educator, full-time YouTuber and tech creator based in San Diego. Covers AI tech tools, no-code, and futurism. Known for making fast-moving AI news and Tutorials, easy to understand, breaks down the latest AI tool. I mean, yeah, this all looks good. These numbers aren't super up to date. I've actually crossed 900,000 now, thank you very much. Newsletter is a little bit higher than that, not too far off, but yeah, I mean, they're pretty close. So all in all, a really, really fun new model. It's kind of getting to a point where I feel like image generation is fairly solved, like there's not much we can't do with it. I'm sure we're gonna get some like new viral trends out of this that I haven't predicted yet, kind of like how we got the Studio Ghibli style out of the model that came out, but this is probably one that I'll do a much deeper breakdown on. I love to get new tools like this and find like 50 ways to use it and then really deep dive in it. So that might be a video I make in the future. All right, now let's move on to Anthropic because they had some really cool announcements within the last week as well. This first announcement actually came out last Friday, but because I record these on Thursdays and publish them on Friday, it didn't make last Friday's video, but it's actually one of my favorite things Anthropic has released in a little while now. And that's Claude Design. Claude Design lets you collaborate with Claude to create polished visual work like designs, prototypes, slides, one-pagers, and more. It uses their vision model, Opus 4.7, and is available to Claude Pro, Max, Team, and Enterprise. Some of the examples they mention in their blog post are things like realistic prototypes, product wireframes and mockups, design explorations, pitch decks and presentations, marketing collateral, and frontier design. But this one you can use directly inside of the Claude website. So if I go to Claude.ai, over on the left, I've got a new tab that says Design, and we've got our new design page here where we can get going with this. And there's already some pre-planned things up here like prototyping, building slide decks, or from template, one of the templates being animation. And that's something they didn't even seem to mention in their blog post. But probably my favorite use case of this is actually the animations. So when I first got access to this, the first thing I tested was having it redesign the Future Tools website. I gave it the prompt, I want a fresh design for my website at futuretools.io. That's it. That was the whole prompt. It did ask me some follow-up questions like scope, fidelity, direction, vibe, all this sort of stuff. And you can already see the design here. If I click on present mode, I can go full screen and you can see the design that it gave me. In my opinion, it's pretty nice looking, but also a little bit busy. I actually like this idea of showing how the site's changed, almost like a stock chart. So you can see how much the tools have like grown or shrunk in votes. I don't quite know how to implement that yet, but I actually think it's kind of a cool idea that I hadn't thought of. I also like this ticker here that might show like the newly added tools or the tools that are getting the most love on the app recently. But it made this cool animated interactive website that could be a complete like redesign, rebrand, rethinking of what Future Tools actually is. And I literally just gave it the one prompt, reimagine Future Tools. Saying that, I've seen some other websites that people use Claude Design design for, and they all have this exact same feel. Like, I saw this post from Matthew Berman here, and he had Claude Design create a slide presentation for him. And I mean, look at this design here. This is the slide presentation it designed for him. This is my Future Tools website. Very similar aesthetic. I mean, this is a dark mode, but if I switch it to light mode here, you can see it's like a pretty dang close look. He also shared this redesigned website that it created for his Journey platform, and again, it's it's even got like this little bar right here across it, similar to this ticker bar that it created for mine. So it does seem to have a very specific aesthetic that it really leans into. But again, my favorite use case is to have it actually create animations, because it can create like very basic level, almost After Effects style animations. So I was speaking on a panel at NAB earlier this week, and I was going to show some animations that I created during that panel. Unfortunately, I didn't get the videos to them in time and wasn't able to show the animations, but And this was the animation I generated to show what Claude Design is capable of when you tell it to go make me an animation. I told it to highlight Las Vegas on a map and then zoom in on Las Vegas with an animation and put the words Las Vegas. And it did that. I had it generate like a scene of the halls at the convention center and put NAB 2026 and it animated that onto the screen. I then told it to make me some bar graphs that show who shows up to Vegas every year. And it made this. Now, interestingly, the animation animation just showed these bar graphs pop up instantly, but if I scroll through the timeline, you can actually see they were animated. So I don't know why they just instantly popped up this time, but if I scroll here, these do actually animate onto the screen. So if you're looking for that like After Effects-y look with your animation, you can do that with Claude Design. And then I had it generate a bar graph for me about how much AI is mentioned at NAB each year, and we can see how many booths were related to AI year over year. And it animated this little line graph for me, and then it puts some text on the screen. AI, the defining story of NAB 2026. Again, very, very basic animations, but this is something that I gave it just a couple prompts. It spent a couple minutes generating this, and this is something that used to take, you know, maybe an hour or two in After Effects to get this same style. I really, really love that we can do that in Claude Design. Now, if I needed something a little more advanced with a little more detailed graphics, I would still hire somebody to do that. But if I just want to put like a quick bar chart or a quick line chart on my screen. This does a great job for that. There's another thing Anthropic released this week that I thought was pretty cool as well. To be honest, I haven't put it totally to the test yet, 'cause again, I was at the event most of this week, but they released this new live artifacts feature in Cowork. It creates dashboards and trackers connected to your apps and files. You open a dashboard anytime and it refreshes with the current data. So it sounds like you can connect this dashboard to Excel sheets or CSVs or things like that. And if the CSV updates and you open this dashboard, well, the dashboard then updates with the new information. So if I open up Claude here and I make sure I have Cowork selected, you can see there's a new option for live artifacts here. I'll click what needs my attention and see what it actually builds for me. We could see it actually created a prompt, create a live status page of what needs my attention today across my connected tools and keep it updated. It's telling me only Figma is connected That's thin. Let's just do build with Figma only for now and see what it does. I don't actually use Figma that often. And it built like a real simple dashboard, but there's not really much in it cuz again, I don't really use Figma. But if you had it connected to your Gmail or your Google Calendar, your Google Drive or something like that, you can have it actually pull in information and details you might need, I guess. Again, something I haven't spent a lot of time playing with, but I thought was pretty cool. We did get some additional large language models this week. I'm not gonna go too in depth into any of them, but I do wanna shout out that these new models are here to use, starting with Google DeepMind's latest model, Deep Research Max. This is an autonomous research agent. As the name implies, these models are really good at doing deep research. We can see from these various benchmarks here that this new model is pretty much the state of the art at all research-type tasks. Alibaba released two new models this week, Qin 3.6 Max Preview, with improved agentic coding capability over Qin 3.6 Plus, stronger world knowledge and instruction following, and improved real-world agent and knowledge reliability performance. Here's the benchmarks on that one if you want to pause and take a quick look. Unlike many of the other QIN models that have come out, this one is not open weight. They do say in their blog post that this is their next proprietary model. However, that wasn't the only model they released because they also put out QIN 3.627B, which is their latest open-source model. This one they claim is outstanding at agentic coding. It surpasses QIN 3.5397B A17B strong reasoning across text and multimodal tasks, supports thinking and non-thinking modes, and it's fully open. Here's the benchmarks on that one if you want to pause and take a peek. Keep in mind that it is comparing it to pretty much other open weight models and Claude 4.5 Opus, which is, you know, two generations behind on Claude. And we got another model out of KIMI this week as well, KIMI K2.6. Once again, an open source coding model. It's good at long horizon coding, motion-rich front ends like videos in the section, WebGL shaders, Three.js, things like that. Agent swarms, it can run 300 parallel sub-agents, proactive agents, claw groups. And here's what the benchmarks on this one looks like. And in some of these benchmarks, like Deep Search and Humanity's Last Exam, it's actually beating out Opus 4.6 and GPT-5.4 Extra High. So this model, actually being an open weight model, beats out some of the state-of-the-art models that were state-of-the-art as of last week in a lot of important benchmarks. All right, there's just way too much to talk about this week. I have just a few more things I want to talk about, but I'm going to blow through them really quickly in a rapid-fire section. Starting with even more news out of OpenAI, they introduced another model this week called OpenAI Privacy Filter. It's a state-of-the-art model for masking personally identifiable information, and they're actually making it an open-weight model, so you will able to fine-tune it for your needs if you need to. This is a small model with Frontier personal data detection capability. It's designed for high-throughput privacy workflows and is able to perform context-aware detection of PII, personally identifiable information, in unstructured text. It can run locally, which means that the PII can be masked or redacted without actually leaving your machine. As far as availability, well, it's open, so they put it on Hugging Face and GitHub, and anybody can go use it. They also released ChatGPT for clinicians this week, and it's actually free to use for verified clinicians in the US. This is a version of ChatGPT designed to support clinical tasks like documentation and medical research so clinicians can focus on delivering high-quality patient care. It appears to be pretty fine-tuned for the needs of clinicians, and I think it's pretty cool that OpenAI is making this widely available and freely available to the people that will actually need it. We also got more news out of Anthropic this week, like new connectors in Claude for everyday life. It's basically just more things that you can connect directly into Claude, things like AllTrails, Instacart, Audible, TripAdvisor, Intuit TurboTax, and more. If you go to Claude AI slash directory, you can see a list of all of their connectors, and as you'd expect, it's getting bigger and bigger. These are all different apps and tools that you can connect directly with Claude and interact with them directly through the Claude interface. If you're on a Claude Pro or Max plan, Claude is now available inside of Microsoft Word directly. You can learn more about that by going to cloud.com/cloudforword. Microsoft's Copilot rolled out more agentic capabilities in Word, Excel, and PowerPoint. Copilot can take multi-step app-native actions directly in your documents, worksheets, and presentations. And their blog post here, which again linked up in the description, has examples like that for Excel and PowerPoint if this is something that interests you and you want to dive deeper on it. I know a lot of people who watch my channel try to avoid X. I skim it and try to filter through the noise so you don't have to, but X did just roll out a new feature called Custom Timelines. This feature allows you to pin a specific topic to your home tab with support of over 75 topics. It's powered by Grok's understanding of every post with the algorithm's personalization, meaning every timeline is made just for you. So essentially it's using Grok to create a timeline for specifically the niche topic you're most interested in. Earlier on, I talked about how Claude Design could do animations. Well, it looks like HeyGen just rolled out a feature called Hyperframe. Hyperframes that uses Claude Code to write animations. So you can prompt inside of Claude Code using the Hyperframes skill and it will actually create MP4s of some really cool animations that you can use in your videos. So again, if you need really simple After Effects-y kind of animations for your videos, you have another tool at your disposal using this new Hyperframes feature in Claude Code from HeyGen. We also had another image update this week, this one from Ideogram. This one allows you to use videogram with custom models. So you can train your own model on your images and every new generation will follow your art direction. So we can see in their demo here, they upload a very specific style of image here. It trains on that style. And when you give it a prompt, it then creates an image that follows that same style. And we can see a whole bunch of images in that same Peter Rabbit style that they gave it a handful of images on. Here's another one that looks like, you know, Sharpie art and another one that kind of looks like the Red Bull commercials. So if I head on over to ideogram.ai, click on models and the plus button, you can upload between 15 and 100 images and it will train a new image model exactly on that style. Now I haven't had time to play with this one yet, but it is something I'd like to deep dive a little bit further on in a future video. The biggest news a couple weeks ago was Anthropic's Mythos, the That was so big and so scary that Anthropic said we can't release it to the world. Well, it got accessed by unauthorized users and people that weren't supposed to get their hands on it apparently got their hands on it. My opinion on this one, well, if you're gonna go around and market your model as this is the greatest model the world's ever seen, but it's too powerful and too scary for anybody out there to use it, that just makes people wanna get their hands on it even more, especially the bad actors. And well, bad actors are going to them to go and figure out how to get their hands on it, which is what they did. Apparently Anthropic is saying that there's no evidence that the access is impacting any of its systems, so I don't know if it's been used for anything too nefarious yet, but apparently people that weren't supposed to access it did get access to it. Also this week on the Core Memory podcast, Sam Altman shared exactly how he feels about what's going on with Anthropic.

It is clearly incredible marketing to say we have built a bomb, we're about to drop it on your head, we sell you a bomb shelter for $100 million. You need it to like run across all your stuff, but only if we like pick you as a customer.

So yeah, I don't think there was any debate there on what company he was referring to. And finally, there was a half marathon in China this week, and 4 different robots finished in less than an hour. Here's a quick clip of one of those robots approaching the finish line. I know this video kind of looks like it's being sped up, but it's not. This is the real speed of this robot. That I believe runs the half marathon faster than any human is able to run the half marathon. Oh, and by the way, there's a whole bunch of awesome videos from this event that show different robots from the marathon. Some of them are pretty amazing to watch, like this one, which looks like a drunk person. And oh, there they went. Here's another one with like a head with legs at the marathon. Another one with like a little reindeer dude and a little pink dude in the background. This robot can't seem to figure out how to get over some tape on the ground and then basically explodes. You got this little guy here who's going in the wrong direction. I mean, some of these robots are really, really awesome to watch. I wish we'd have a robot marathon here in the US. I would go watch it in person if we did it. It would just be really cool. But that's what I got for you. I know this video's long. I know this was a really big week. There was a lot that happened. I wanted to talk about all of it and try it all. Some weeks are just gonna be like this where there's way more to talk about than other weeks. I really, really appreciate you hanging out and nerding out with me and taking the time to learn AI. My goal here is to sift through and keep on top of all of the AI news throughout the entire week and just drink from the fire hose, read it all, learn about it all, talk to people, try the tools, understand what's going on, and then condense it into one video for you at the end of the week so you can get totally looped in. And I will filter all the signal out from all the noise and the hype and just share what I think you'll probably want to know. And that's my goal with this channel. I also learned that it's in bad taste to ask you to like and subscribe. I am getting close to a million subscribers, but since I'm not supposed to ask for a like and a subscribe, I won't. But, uh, I really appreciate you tuning in and it would be great if, uh, you wanted to see more videos. There are things you can do to make that happen. Again, I really, really appreciate you. This is so much fun every week to make these videos and stay on top of it. It is a full-time job to keep up with, but hopefully I'm adding some value to you and helping you stay more tuned in and looped in with less overwhelm. Again, that's my goal. Thanks again for hanging out with me. I hope to see you in the next one. Really appreciate you. And yeah, that's all I got. Bye-bye.