AI News: Claude's New Browser, Spotify Gets AI & OpenAI's New Hardware

Here's the AI news you might have missed this week. Oh, and by the way, I'm not actually at home this week. I'm actually in a hotel room here in San Francisco for Open Sauce. But I'll be teleporting back and forth a little bit throughout this video because, well, I struggle to stay in one place for too long. Anyway, let's get into it. Starting with a new update that we got inside of Claude Code on desktop. It now has an in-app browser. So just a few days after OpenAI launched their new ChatGPT a ChatGPT app that has a bunch of built-in browser features in it. Anthropic went, that's a cool idea, let's add it into our app. I mean, that's probably not what happened, but that's what it felt like. So if you're inside your Claude app, it doesn't work inside of the normal chat side. You have to be on the Claude code side. Go into one of your projects and you can hit Command+Shift+B and it will actually open a little browser on the right side. You can also hit this little browser icon in the top right and that will also open the browser here. And of course you could go wherever you want, like Future Tools. And let's say I'm editing this site right now, I can use this little annotation feature where I select elements. And let's say I want to select this button here. Well, it highlights the button over in the code on the left, and I can say, change the text to get on my list. And if I was actively working on the website right now, it would go and make that tweak for me. Another really cool thing about having the browser inside of the code window here is you can have it go and find things for you by using the browser. I did this inside of the ChatGPT app in my most recent video where I showed all the ways to use ChatGPT, and I basically had it go look stuff up on Twitter and Instagram for me. It's really expensive to use the X API and have it pull data from X for you, but if you have a tool that has browser access, it doesn't need to use the API. It can just go and scroll X for you and find stuff. So another really cool use case of having the browser over here on the right side. So Google rolled out a few new AI features this week, including the ability to connect more of your apps to search. So you can you can already connect a whole bunch of apps inside of Google Gemini. However, they're making it so you can connect apps inside of AI search mode. This AI mode that's directly in search now has the ability to connect to more apps. Now, it doesn't appear that there's like a traditional add apps section here. You just kind of give it a prompt and then if there's an app it could connect to based on what you prompted it with, it will try to connect it to that. So let's try make a grocery list for me and add it to a shopping cart for someone to purchase for me. You could see it's searching apps and then it found the Instacart app and it's asking me to connect to the Instacart app. So that's essentially how it's gonna work. You're going to ask AI mode a question and if it finds an app that will work really well to help you with that question, you can then connect to that app and sort of take the process further. So even Google Search is becoming more and more of like an AI agent for you, which is really interesting. Another new thing that Google rolled out this week is some updates with their Vids platform. Now if you're not familiar with Google Vids, It's not the same as Veo. Google Vids is like their platform that will make like slideshow presentation type videos for you. However, they just rolled in the new Gemini Omni to it. So not only will it generate what looks like a slide, you can actually use the new Omni so that it actually animates what was on the slide as well. You can also now chat to edit your videos. So you just talk through what you want changed in the video and it'll do that. But they also added this new feature into Vids where you could import yourself. You can actually drop an avatar of yourself into a video and the example they give is, you know, generate a happy birthday card and then it generates this video here. Happy 30th birthday, Sarah. Can't wait to celebrate this weekend. Now, I don't think that AI is fooling anybody into really thinking that was their real sort of video. But yeah, I like what you've been able to do in Sora and the various Meta products. You can now do inside of Google Vids where you upload an avatar and it will sort of incorporate your avatar into your video now. We got an update out of Spotify this week. According to their latest press release, you can now talk to Spotify. We're making it easier to shape what you hear, understand what's playing, and explore your listening just by asking. Premium users will see new ways to talk to Spotify across Home and the Now Playing view on mobile. By typing or speaking directly in the app, you can have a back and forth conversation to choose what's playing, learn about the music you love, revisit your listening history, and go deeper on podcasts and audiobooks, all without leaving Spotify. This is probably a really bad idea because everybody's gonna judge me, but let's see if I can get this to work and what it tells me. All right, so I'm in my Spotify account and right in the there's like a little beta thing next to that blue circle there. If I click in there, it tells me I'm about to talk to an AI, and I can ask for some music. Now, I won't be able to play any of the music on this video because it'll get copyright claimed, but let's see if it works. What was I listening to 2 years ago? Make a playlist of that kind of music. All right, accessing listening history, filtering 2 years back, identifying dominant genres. I'm kind of scared. All right, so it's playing some lo-fi music. So here's the playlist that made, and it's all sort of like lo-fi beats. I tend to listen to a lot of lo-fi beats that don't have lyrics while I'm working, so this definitely tracks. That's what I would have been listening to 2 years ago. It's also kind of what I still listen to a lot of on Spotify. What's my most listened to song of 2025? Yep, another lo-fi song. What can you tell me about this artist and the style of music that it is? You're vibing with Toshio, an artist all about lo-fi hip-hop and cozy instrumental beats. His music is super chill, think relaxed nostalgic that riff on video game and anime themes. So pretty cool new addition. Maybe you don't want to make decisions about what to listen to. Just talk it out and let Spotify figure it out for you. All right. So this next one is something that we talked about in last week's video, but I do have an update on it. ByteDance recently released SeedDream 5.0 Pro, its new multimodal image generator, and a great place to experiment with it is over on Artlist. What makes SeedDream so cool is that it doesn't just generate images from a description. You can give it up to 5 reference images and use them to control the characters, products, colors, composition, and overall style. It also has stronger reasoning, so it can understand more complicated design instructions and create things like detailed infographics, UI mockups, advertisements, and images with actually readable text. And Artlist makes the whole process much easier. Inside their AI toolkit, you just select Seed Dream 5.0 Pro, upload your references, choose your aspect ratio and resolution, and then generate multiple options. Then you can keep refining the image with new prompts or additional references without bouncing between a bunch of different platforms. SeedDream 5.0 is available on Artlist right now, along with all the other top image and video models. So check it out at the link in the description. And thanks so much to Artlist for supporting my channel and sponsoring this section of today's video. We got a couple new large language models this week, one of them out of Thinking Machines Labs. Now, if you're not familiar with Thinking Machines, well, it was founded by Mira Murati, who used to be the CTO over at OpenAI. Back when Sam Altman got fired for like a weekend, she actually was the CEO for like 24 hours. Well, eventually she left OpenAI to start Thinking Machine Labs. They've shown off some really cool demos of some models. And finally, this is the first model they've put out there for the public, and it's actually an open weight model. So this is one that you can actually download and fine-tune and run it on the cloud. If you have a strong enough computer, run it on your computer. But when we look at the benchmarks, it's not really like a super standout model. It's a model that's really good for being an open weight model. But other than that, all of these models are better than it. GLM 5.2 is also an open weight model, but when it comes to design, it's beating out this new Inkling model by, you know, a couple models. But they do show that it could generate a pretty decent looking snake game. But we can see this is comparing it to all the other open weight models with a few of the closed weight models over here. GLM 5.2 is probably the best open weight model still, because we can see in reasoning, GLM 5.2 pretty much beats it out on every single one of these benchmarks, agentic coding, GLM 5.2 beats it out on every benchmark. Apparently it's a little bit more factual than GLM, but it's just, you know, a decent open model. Not state-of-the-art for open models, but just decent. You can play with it and test it for free over on this website called Tinker here. They have a section called Playground where you can come up here, select the model Inkling, and it's very similar to the OpenAI Playground. Let's just have it show me some beautiful code. It's decently fast, and here's the code output that it wrote, and it also explains why was beautiful and how it's elegant and not as clever, I guess. And here's the pricing if you're using it through their API, if you're interested. Now it is a 952 billion parameter model with an almost 2 terabyte file size. So I would say most people aren't going to be running it on their own, you know, consumer-grade GPUs. This is one you're going to be using in the cloud most likely. So if you're going to have to use it in the cloud anyway, you're probably going to pick one of the stronger models right now. I don't know. But if you're curious about how it performed on Bench here. Well, yeah, that's what it thinks Gary Busey looks like. I don't think that model's hitting the top of the leaderboard anytime real soon. Then on the opposite end of the spectrum, we got this model called Bonsai 27B from Prism, from Prism ML. And this one's a 1-bit model, meaning that it's been like super, super compressed down so that it could actually run on a phone. The model itself is about 4 gigabytes, but it's got 27 billion parameters, which is pretty much the largest model you're going to be running. On a phone. This isn't gonna be like the most intelligent model you've ever used, but it'll probably give you like decent quick answers on things. I tried to get it to write code for me and it just couldn't do it. I tried to run it through Bucybench and it flat out worked for 20 minutes and never actually ended up generating an SVG. And I ran it 3 times and got the same result. So it's not gonna be like a coding agent for you. You can however demo it over on Hugging Face. You click the demo in the article and then jump to Hugging Face and then you can load model here, and we can see it says the entire 27 billion parameter model runs entirely on your device. No data ever leaves your machine. Let's get it to generate a logic puzzle here. And it's actually really, really fast. 94 tokens per second, which is funny because it basically asked itself its own logic problem and then solved the logic problem that it asked itself. You have 3 boxes labeled apple, oranges, and mixed. Every label is wrong. You may take 1 fruit from 1 box without looking inside. How can you correctly relabel all 3 boxes? Now I wanna see if this is actually running. So if I click generate, Generate code, Google Chrome Helper. It's not showing any memory, so I don't know if this is actually running locally or not. I can't tell. I don't think this is actually running locally. When asking it to write code, this is what it gave me. I mean, it doesn't look amazing. Again, it's a small, very, very heavily compressed quantitized model, but it's designed to be able to run on your phone. So this is probably not going to be your go-to if you are running on a desktop. And we got another AI model this week out of Moonshot AI called KIMI-K3. And this is a model that, well, they haven't opened it yet, but they did say they were going to open the weights this month to this model. And when they do, it looks like it will be the most capable open weight model there is. So if we look at some of the benchmarks here on the coding benchmark DeepSWE, which is actually a really solid software engineering benchmark, we can see that it scores just behind 5.6 SOL and Fable in that benchmark, but it outperforms GPT-5.5, Opus 4.8, and GLM 5.2. And GLM 5.2 is definitely the leader right now in open models. On TerminalBench, it actually outperforms Fable-5. And this seems to be a theme with a lot of the benchmarks for this model is that, well, it's right up there with Fable-5 and 5.6, and it beats out 4.8 from Opus and GLM-5.2 and GPT-5.5 and pretty much all of the best models outside of the 5.6 and Fable model. It's kicking their butt. And this is going to be a model that's open weight. This is the first open model to reach 2.8 trillion parameters. So not one you're probably going to be running on your local computer, anytime soon. But if you wanted to run it on like a cloud GPU and fine-tune it once they open the weights, you'll be able to do that with this model. Now, this is a model that came out since I've been here in San Francisco for Open Sauce, so I haven't actually had the opportunity to play with it yet. I haven't even benchmarked it on BoostyBench yet, but definitely stay tuned for that. Maybe next week's news video will take an even deeper dive into KineK3. All right, I have a handful more things I want to share with you, but nothing I really need to go too deep on. So let's jump into a rapid fire. So this week, Groq Build was open sourced. This is SpaceX AI's coding agent and terminal user interface. They're open sourcing it so that people can build off of it, but also explore what it's capable of. And sure enough, here it is on GitHub where you can sort sort of fork it or, you know, do whatever you want with it. Groq rolled out a new feature this week. They added automations into Groq. Now, this is something that both the Claude app and the OpenAI ChatGPT Codex, whatever you want to call it now, all had access to, where you can basically tell it on a certain schedule, run this task. You know, a lot of people are calling these loops. You can now do more of those loop-type things inside of Groq with these automations. So you can have triggers like daily triggers or even triggers that do things like watch your inbox, and when you receive an incoming mail, it will fire a trigger based on seeing that incoming mail. So if the Groq harness is your preferred go-to harness, well, you now have this option as well. So ChatGPT improved their search functionality inside of the app on web, iOS, and Android. So basically, if you've had any conversations with ChatGPT in the past and you want to be able to pull up that conversation again they've really, really improved the ability to search. So from the sidebar, you can now search chats, projects, images, and documents in one place, and you can use filters to further narrow your results. It seemingly is going to get a lot easier now to find exactly what you're looking for among everything you've done within ChatGPT. Claude got the ability to access your 1Password credentials now, which I think is pretty cool. Also a little bit scary because like if you have bank account details and stuff like that in there, now Claude has access to it. So like probably probably not an issue right now, but like, you don't want it to go purchase stuff on your behalf because it has your bank details and credit card details. I don't think it'll do that yet. What makes it really cool is if you ask it to go do something for you that requires it to log into a website, it doesn't need to, you know, manually ask you to go and log in for it. It can just use your 1Password and then log in for you, which I think will be pretty handy given they go over the top with security and make sure it doesn't log into stuff that it shouldn't be logging into. Now, the Claude agent can't actually see the credentials. So it'll use 1Password and log in, but none of those passwords are actually going to get sent back to Claude or be trained on or anything like that. And if you've ever wanted your AI agent like OpenClaw or Hermes or Codex or any of them really to order your DoorDash for you, well, now DoorDash has a CLI, a command line interface that you can give access to any of your preferred sort of agent or AI harness. You can now order DoorDash directly from your agent, search stores, find the best deal, checkout, and more. Apparently it's on waitlist for US and Canada, so not totally rolled out yet, but pretty soon you'll be able to go to your Open Claw and say, "Hey, order me a burrito." This is an update to a story from last week where I showed you that Meta was allowing you to tag anybody on Instagram and it will generate images of them. Well, there was a ton of backlash from creators and from actors and just from tons of different people. And well, this feature didn't even last a week and Meta went and just rolled it back. So now you can't do that anymore. The newest version of Siri has started to roll out now, and so far the feedback on it has been pretty good. One of the coolest features is that you're actually gonna get AI on your Apple Watch if you wear an Apple Watch now too. Now Siri has been on your Apple Watch since the beginning, but you know, who uses Siri on their Apple Watch? But now that it's got the AI features built into it, well, better AI features, supposedly it's really useful now. Not something I've had the opportunity to test yet, but again, most of what I've heard about the new Siri. People have really been liking it. Anthropic introduced Claude for teachers. So if you're a verified K-12 educator in the US, you get free access to premium Claude capabilities, a library of teaching skills, and a direct connection to evidence-based curricula. So hey, if you're a teacher, go use that. Claude's expensive. Google's rolling out some new features because they're celebrating 25 years of visual search and innovation. They have this cool blog post here where they sort of show a timeline of Google Images and how it's evolved, but they're rolling out some new features soon where Google Images Images is going to look more like Pinterest. It's gonna sort of learn what you save and what sort of images you're looking for and suggest images directly to you. They're also rolling out the ability to create new images right in search. So you just go to google.com and prompt an image to generate, and you'll be able to do it straight from the regular website now. Well, not now, but coming soon. NotebookLM is getting a rebrand. It's now gonna be called Gemini Notebook, which, you know, everybody kind of knows the name NotebookLM at this point. And likes it as a tool. So changing the name might confuse some people for a while, but still the same product that everybody loves. It's just now called Gemini Notebook. So yeah, there's that. OpenAI finally launched their hardware. The new hardware is a keypad for Codex. It's a collaboration with the company Work Louder. So basically I think Work Louder made it and OpenAI threw their branding on it and sort of dialed it in for Codex. If you head over to openai.com/supply, this is like the OpenAI store where you can buy stuff directly from OpenAI. Click on shop and you can see the Codex Creator Micro is here and available for sale for $230. I do have one on order, so I'll let you know what I think about it once it arrives. But this is what it looks like, a little keyboard specifically dialed in for using Codex. I ordered the clicky version and as soon as it arrives, I'll give you my thoughts on it. But apparently OpenAI is maybe working on a new device that's like a screenless smart speaker, kind of like an Amazon Alexa sort of thing. This is more like in the rumor realm than it is actually confirmed, but apparently OpenAI is developing a mobile screen-free smart speaker to serve as a human-like AI companion that can control smart home appliances and tap into the capabilities of OpenAI's ChatGPT. This also comes as OpenAI just got caught up in another lawsuit against Apple, where Apple is suing them, basically saying that people that used to work at Apple moved to OpenAI and brought like trade secrets and hardware software examples with them. OpenAI claims that that's all unfounded and not true, but you know, another legal battle with the AI companies that we'll watch and see how it plays out. But some of the rumors are that maybe this new device has some ideas that might have come from Apple originally, but who knows, we'll see how it plays out. I'm not gonna speculate. This is all still just rumors about what's coming out anyway, nothing actually confirmed by OpenAI. New York New York became the first state to enact a data center moratorium. So there's going to be no new hyperscale data centers set up in New York for up to a year. The order blocks new environmental permits for data centers over 50 megawatts, which the governor's office says will give the state time to come up with the regulations needed to protect residents from rising energy prices and environmental impact. We'll see if other states follow suit with this. Speaking of data centers, I saw this article on The Verge called "Would You Host Part of an AI Data Center in Your Home?" Now, this is actually a concept I talked about in previous videos, but Sunrun is actually implementing it. They're sort of like mini GPUs, mini compute systems that they're distributing to people's houses, and then AI companies can tap into the compute that's running in people's houses, and the people that have these little mini data centers will get paid for it. So Sunrun is offering to pay its customers to put its compute units in their homes. They're launching a pilot program for a new distributed AI compute program that will place numerous compute nodes equipped with Sunrun solar and battery storage systems. Customers will be compensated for participating in the pilot program. So it almost sounds like a crypto mining kind of thing, right? It's sort of a different alternative to building big data centers is everybody has little mini data centers in their house. I wonder if you could run your own local AI models on it. That'd be kind of cool if like, all right, I'm distributing it, but also I want to run a model myself and send it to my own little mini data center. I wonder if like you'll be able to do that. I don't know. I'll be following along and seeing how that plays out. And finally, I wanted to share this. I am actually doing free workshop hosted by Teachable on July 22nd. So in just a few days at 1:00 PM Eastern time, I put together a whole presentation called Deciding What to Automate with AI in Your Personal and Work Life. I'm basically gonna break down what I think is worth using AI to automate, what's not worth using AI to automate, and then giving you tons of examples of things that are automatable and things that are not automatable. I think it's gonna be a really fun session. You'll get a lot of practical workflows out of it. It's totally free. You can just sign up at the link. There'll be a link somewhere down in the description to go sign up if you want to check it out. Again, July 22nd, 1:00 PM Eastern time. We'll go for about an hour. My presentation's really only like 35, 40 minutes, but then we'll just hang out and do Q&A for like 20 minutes. Hopefully I'll see you there. All right, so after last week where we got all sorts of new apps from OpenAI, a new voice mode, new models, Fable extending how long people can use it, just all sorts of news and rollouts. This week feels like a fairly quiet week in comparison, but still, there's never a dull week in AI, and this roller coaster just keeps on moving forward. Maybe like this video and consider subscribing to this channel. I'm getting really close to a million, hoping to pass it within the next couple of months here. I would really, really, really appreciate it, and, uh, it'll make sure more stuff like this shows up in your YouTube feed. But once again, thank you so much for hanging out. With me. Thanks so much to Artlist for sponsoring this one. It's always a blast to nerd out with you on all the cool AI stuff that's coming out. And, uh, again, really, really appreciate you, and hopefully I'll see you in the next one. Bye-bye.