Transcription
Here's the AI news you probably missed this week.
Starting with a demo of a brand new AI model that looks pretty impressive. Now, if you remember back when GPT4 came out, we got this demo of all these cool things that GPT4 would do. And since then, it kind of feels like there hasn't been a lot of like impressive movement with a lot of the large language models. They're just all kind of marginally getting better and smarter on benchmarks and things.
Well, this new model from Thinking Machines Labs came out with these demos and well, I'm actually impressed once again. Now, before I show off some of the demos, if you're not familiar with Thinking Machine Labs, this is the company that was founded by Mera Marott. Now, Mera Marott was the CTO of OpenAI back when Sam Alman was fired for like a weekend. Mera Maratti was the interim CEO for like a day or two. Obviously, Sam Alman came back and then a few months later, Meera left OpenAI to start thinking machine labs, bringing a bunch of people from OpenAI with her. Some people from DeepMind came over. I think maybe some people from Anthropic came over, but that's how we got Thinking Machine Labs.
For a little while now, they've been quietly sort of behind the scenes. We haven't really heard a ton from them. And then this week, they finally showed off their interaction models. And the demos they shared, again, pretty cool. It'll actually translate in real time. So as somebody is speaking in one language, it will translate into another language and just sort of speak over them instead of waiting for them to finish.
>> Can you translate in to in English in real time for my friend and for audience?
>> Absolutely. I'll translate as you go preview model. Today we're taking a look at our preview model which makes humans and AI easier. It has many features like web search and artifacts.
>> Here's another example where you'll actually see her take a pause in her speaking and it doesn't interrupt her during that pause. It waits for her to pick up where she left off, continue speaking before it replies. But she's also giving it instructions to specifically interrupt her whenever she says an animal name. So, just watch this.
>> I'm going to tell you a story and whenever you hear an animal word, please count the number immediately.
>> Gotcha. I'll count them out as you go. Let's hear it. Last weekend I drove down to South Bay to visit a farm.
>> Um, on the way there I saw a deer.
>> and when we got to the farm we watched a demonstration of sheep sharing.
>> two.
>> Then on the way back we saw a coyote.
>> and I really wish I saw a pibaro instead.
You said deer once, sheep once, coyote once, and capiara once. Four total.
It also fully recognized when she was done speaking and then gave a count of the various animals she mentioned. Here's another example of it watching while you work. She asked the model to call her out when she starts to slouch. And it's actually pretty funny.
>> Okay, I'm doing some work. Let me know if I start to slouch.
>> I've got you. Sit up straight and you'll be golden. You're starting to slouch forward. Try pulling your shoulders back.
She does that a couple more times and it catches her slouching every time. And then here's an example where it will actually interrupt you when it thinks it needs to interrupt you.
>> I heard mountain biking is a lot of fun. So, I wanted to go mountain bike and I was thinking to actually take my parents with me. They're pretty old. They're 70, almost 80. But I thought that would be a fun trip together.
>> Wait, no. Don't take them mountain biking at 80. That's incredibly risky. They could easily get seriously hurt. Pick something safer.
>> H I think maybe, you know, if if it's risky on a normal mountain bike, we can do like tandem bikes. Maybe they're more sturdy and we can find like, you know, some volcano or better like an active volcano to go all together and do some
>> stop. An active volcano is deadly.
So, it will actually interrupt if it feels like it's necessary to interrupt and sort of redirect the conversation. Here's another example of it actually talking at the same time.
>> I'll be refraraming everything you say into uplifting professional language instantly.
>> Hi, Alex. I cannot stand your
>> Hi, Alex. We'd love to explore opportunities to enhance your timeliness and morning engagement.
>> You cannot expect the rest of us rearrange our workflow.
>> Let's collaborate on a schedule that ensures everyone stays aligned and included in those key morning sessions.
And there's a bunch of other cool demos on the website. I definitely recommend checking it out cuz this feels like it's actually finally something that's improving these AI models more than just like boosts on benchmarks. It's aware of time. So you can tell it to like, hey, let's have this conversation. Let me know in 4 1/2 minutes when this conversation needs to end. And it will actually keep time as you're talking to it. And then in 4 and 1/2 minutes, we'll say, hey, that's time. Like JGPT and Claude can't do that. It can do simultaneous tool calls. So, while speaking and listening to the user, the model can concurrently search, browse the web, or generate UI, weaving back results into the conversation as needed, and it'll all do it simultaneously.
Unfortunately, this model isn't publicly available yet. We're only getting demos of it. But, of the demos I've seen recently, this one has been pretty impressive. They do say in the coming months, we'll open a limited research preview to collect feedback with a wider release later this year. I don't know when we'll actually be able to use it, but if you have some time and you want to see more of the demos, again, it'll be linked up below because I think these demos are actually worth watching. It actually feels like something novel in the AI world, which I haven't felt like we've gotten in quite some time now.
Cruso has this really cool product called managed inference that basically makes your AI apps way faster and way easier to scale. It basically solves the problem where your app works great with 10 users and then completely slows down once real traffic starts hitting it. Cruso managed inference is built specifically to fix that. This isn't just some cloud hosting layer for AI models. It's a full high-performance inference platform designed to run large-scale AI workloads with extremely low latency and massive throughput. It can deliver up to five times more throughput than traditional cloud environments, which basically means AI apps can serve way more users and requests simultaneously while staying fast and responsive.
And one of the coolest parts of the whole thing is this technology called memory alloy. Normally, when you use AI apps with long prompts, AI agents, or rag systems pulling in tons of documents and context, the system has to repeatedly process all of that information from scratch every time. that creates lag and burns a ton of compute. Memory alloy retains and reuses context across requests which helps keep inference speeds extremely fast. Even as context windows get massive and honestly as AI apps become more agentic and context heavy, this kind of infrastructure is probably going to become incredibly important. If you want to learn more about Crusoe managed inference, check out the link in the description box below. And thanks to Crusoe for sponsoring this portion of today's video.
We got a feature out of OpenAI this week that I'm actually personally really excited about. They finally made it so you can use codecs from your phone. I've been using Codeex myself more and more. If you watched my recent video about how I built my own wiki, all of that was done inside of Codeex. So now I can actually manage that from my phone when I'm not actually sitting at my computer, which is pretty sweet.
So if we go ahead and update our Codeex here, we get the option to set up Codeex mobile. We'll go through the little setup process. I'm going to allow devices to control this computer. scan the little QR code on my screen here. Now, it's asking me to connect on my phone. We'll go ahead and do that. Now, back on my computer here. Let's make sure we keep this Mac awake so that when we are using codecs, it can still access it. And I'm going to leave everything enabled. So, we'll go ahead and click done. And now, on my app here, you can see it's connected to my Mac Studio. I can see all my second brain chat sessions that I already started inside of Codeex. And now, let's go ahead and click on chat. We'll start a new chat here. What videos have I recently added to my wiki? Go ahead and submit that. And it's inspecting my vault. So, it's actually reading this from my computer right now. Let's see if I jump back over to my computer. We could see over on the left side here, it's actively running. So, you could see the same thing on my computer running that when I switch back to my phone here is right here as well. So, this is super super cool to me. My wiki still lives on my computer. or all the files are on my hard drive, but now I could remotely access it directly from my phone here. That to me is kind of a game changer. Obviously, if you're writing any sort of code with codecs as well, you can prompt it to start working on an app and then, you know, go watch TV in another room and then just kind of keep checking on the progress of the code and if it's asking you any questions, you could respond right from your phone. And here we go. We can see all of the videos I've pulled into my wiki recently. I cannot overstate how excited I am by this feature with how much I've been using codecs lately.
Now, since we're talking about OpenAI, this was also released this week, which is OpenAI's Daybreak. This to me seems to be OpenAI's answer to Anthropic's mythos. However, it does seem like OpenAI is taking a little bit different of an approach than what Anthropic took. Anthropic's approach was, hey, we built this model that is just absolutely insane and can crack all sorts of cyber security vulnerabilities, but we're going to pick and choose who gets access to it. We're not going to put it out in the world for other people to use. Only our like trusted cyber security professionals can use it, which inevitably ended up in the hands of noncyberc professionals and bad actors. But OpenAI took a different approach. You can actually request them to scan for you. So, they're not saying like, "Hey, you could just have this and use it to check if you have security vulnerabilities, but we have it and we'll check for you."
Enthropic rolled out a new feature this week inside of Claude Code called agent view in Claude Code. So, this seems to be for people who use the command line interface for Claude Code. You can have it go spin up a whole bunch of different agents. And instead of having a whole bunch of terminal windows open, we can see on the screen that it kind of consolidates them all into one screen where you can see what's going on, what needs input, what's working, what's done. So just a cleaner layout if you're using Claude Code and spinning up a whole bunch of agents at once.
If you're looking for a brand new image model, well, we got a new one out of Crea AI this week in Crea 2. This model to me seems like it has a lot of the sort of functionality and controllability that we get out of midjourney. Like you can give it an image and tell it to style it like that image, but then you can also control how much it gets stylized like that image using like a little slider thing. And then if you give it multiple images, you can control each of the weightings of the various styles that you give it here. They also explained that if you give it a very basic prompt like a cat riding a bicycle, it'll give you a whole bunch of different styles. Like every image is going to be wildly different from every other image. So you can kind of go down your own creative workflow and then once you find a style that you really like, you drag it in, continue to prompt more details, but it follows the initial style that you found.
They've also got what they call mood boards where you can give it a bunch of images that sort of all have a similar style. You can see it'll analyze the mood board and then it'll give it a taste profile, some keywords, and what to avoid. And then you could generate images with that mood board and any images that you generate with that mood board will feel like they fit right in with the mood of that. So if you're looking to like really drill in on a specific style or maybe you're an artist yourself, you can load in a bunch of your own images and then generate more that match the style that you've already created. As of right now, if you're on a Max or business plan, you have access to Korea, too. And I imagine, you know, more access is rolling out in the future. If you do have access, you could log into your Creia account and you'll see a little mood boards icon over here on the left. They've already pre-made some mood boards here so you can test it out yourself or you can create a new mood board and it even gives you suggested images to add to that mood board or of course you can just upload your own set of images. Let's search for purple and it brings up a bunch of images with a lot of purple. So let's go ahead and add a bunch of these images in here and see if we can steer it to make very purpley images. Okay, so I added this set of 12 images. Let's go ahead and analyze the board and see if it gets the idea. Foundation of high contrasted saturated purple palettes that bridge the gap between polished editorial glamour and moody retrofuturistic surrealism. Saturated violet palette, glowing purple accents. I think it gets the idea. So, let's go ahead and click generate with mood board. And I'll do a baseball player hitting a home run. Let's see if I get a purple baseball player. And there we go. We got a bunch of baseball players that all meet that sort of purpley aesthetic that I was looking for. So, a pretty cool model if you want to dial in like this exact look and style that you're going for.
But, let's move over to Google cuz we got some announcements out of them this week because they had an Android event where they announced a handful of new updates to Android as well as the new Google books.
So, starting with Android, they showed this demo where they hold up a flyer for a specific event. It saves it into like a little Gemini prompt here. And then it looks like it just says, "Preparing your tour in Expedia." So it jumps into Expedia and it says, "I prepared your tour for you. Complete within the Expedia app." And then you click the button and it opens in Expedia for them to finish the booking. So all they did was take a picture. The Gemini did the rest and then jumped them to the final step of just booking.
They also showed another example where they were browsing the web and they were on this like community playhouse website with a stand-up comedy night. They click on the Gemini button up in the top here. It opens up Gemini but also attaches that page for them into the Gemini prompt. They give it the prompt reserve parking with spot hero for this event. It then navigates to spot hero, enters an address for them, finds the various parking for them, and then takes them to the final page where they can actually check out with Google Pay and purchase that parking spot. So, kind of similar to how you have the Gemini button on your Chrome browser on your computer, well, you're going to have a Gemini button on your Chrome browser on your Android. And when you click on it, it'll actually have the context of the web page. You're going to be able to fill out forms with a single tap. So, it's going to have like your passport details, driver's license details, things like that. You just press a button, it fills it all out.
They're upgrading the spoken text. So when you speak into the microphone app, it almost acts more like whisper flow where it cuts out the h and the ums and sort of cleans up your text as you're speaking instead of just putting it in verbatim. If you misspeak and you say, "Oh, no, wait. I meant to say this," it will actually clean all that up for you. Again, if you've ever used Whisper Flow on your computer, it sounds like that kind of technology is just getting baked straight into Android phones. They did announce a handful of other features, but as always, I'll link up the news post in the description. So, if you want to dive deeper, you could read about it at the original link.
But at this event, they also introduced the Google book. It kind of feels like the next evolution of the Chromebook because they say 15 years ago, we introduced the Chromebook. Now, as we moved from an operating system to an intelligent system, we're rethinking laptops again. So, this one comes with a new operating system which is designed for AI. It's still got the Chrome OS on it, but they're calling this one the Google book. We can see in their examples here, it's essentially a Chromebook, but with like these AI features baked into it that you're seeing on Android phones and that you're seeing get essentially baked into everything.
Next week is Google IO. I'm going to be at that event. So, I'm probably going to get some hands-on with some of this stuff and I'll be able to report back more after that event. But one other thing they did show off is that they're reimagining the mouse pointer. So, one of the examples they showed off is you see the shopping list on the right. They highlight some stuff on the recipe on the left and just by highlighting it and then clicking on the shopping list, they're basically saying add this here and it's adding it to the shopping list. All without them ever having to type. It's just sort of dragging and dropping the pointer around. Here's another example where they're editing an image. They basically highlight the little crab and then say move this here. And they're able to just rearrange this image, but they're not actually typing any prompts or anything like that. It's just all done with mouse clicks. Here's another example where they're editing a document and they say merge this and it merges some cells together. They select some text and say make this more human and then it uses AI to sort of rewrite what they're doing. So they're not actually typing on their keyboard. They're highlighting stuff and saying move this and then highlighting another area and say to here and you can see that it's it's moving stuff around that way. So, it's like a combination of just moving your mouse around and talking as you're moving your mouse around and highlighting stuff and it's now causing actions to happen without you even needing to use a keyboard anymore.
>> Could you get those two ingredients and also this one and add them to my shopping list here?
>> Done.
>> If I hover on the note, the air enabled pointer knows the data that's behind the scene.
>> Make this orange.
>> By typing the word this, it added this actual text note to the prompt. Can you make this 8:00 p.m.?
I've updated the draft to start at 8:00 p.m.
>> Can you show me how to go from this place to that place?
>> Here are the directions between the two locations.
>> I am using head tracking here. Hey, Gemini. So, can you generate an image based on this whole menu here? I would like you to use the style in this image.
>> Okay, I'm generating the image now.
>> Gemini transferred the content of the menu here as well as the style from the bird into the new image.
I mean, come on. That's actually pretty cool. That's starting to feel a little bit like the Jarvis we've been promised from like Iron Man, right? Maybe we're not just dragging our fingers around and moving things yet. But when he showed that eye tracking, he's like, "Move this thing to that over there, and it's just doing things while he's talking." And in that part of the demo, he wasn't even using a mouse anymore. How far off are we from just like dragging our fingers around? Like, move that to there, and it will know exactly what you're pointing at. So, apparently, this is going to be part of the new Chromebook laptop experience. And I imagine it's only a matter of time before we actually get this in other devices as well.
All right, I have a handful more things I want to show you, but they're all kind of smaller updates, so I'm going to run through them really quickly in a rapid fire.
This week, Enthropic announced that they were increasing the weekly limits for Claude Code by 50% through July 13th. However, a lot of people aren't as excited as you'd think about this because it was also announced this week that starting June 15th, they're changing the way Claude subscriptions work. If you want to use Claude outside of Claude Code, you know, through things like Open Claw or Hermes or those kinds of things, well, they're going to give you a certain amount of credits every month depending on the plan that you have. And then once your credits run out, you'll then start getting build at the API rate. Now, they're sort of angling this as like good news for Cloud users. you now get a certain amount of credits and then we're not going to just shut you off afterwards. We're going to let you use the API pricing and things like OpenClaw after that.
The thread that follows essentially says this is a massive nerf, not a new feature. The new credit is built at an expensive API rate, meaning your $100 or $200 credit will get you way less usage than before. Users are calculating it'll burn out in just a few hours of heavy use, making it useless for any serious dev work. So, they're basically saying, "Look, we're going to give you a whole bunch of credits now for your plan that you can use in whatever agent harness you're using, and then when they run out, we'll just keep billing you at the normal API rate." But these credits are essentially being build at the API rate, meaning that you're going to get way less usage in the amount of credits that you get and then roll into the API billing, which just basically means if you're using this for anything like OpenClaw or Hermes or any third-party apps built on the agent SDK, yeah, you're going to end up getting charged a lot for it.
But despite a lot of people not being super happy with some of these plans that Anthropic is rolling out, apparently they have beat OpenAI on business adoption this month. According to this article from Ramp, for the first time, Enthropic passed OpenAI in business adoption. Adoption of Anthropic rose 3.8% in April to 34.4% of businesses. OpenAI adoption fell 2.9% to 32.3%.
And as Anthropic's been doing lately, one by one, they're sort of knocking down one industry at a time and sort of doing their best to disrupt them all. This month, they announced Claude for the legal industry. Basically, what this means is they released a bunch of MCP connectors and plugins and things like that, but specifically for the legal industry. We've also recently seen them do this in the financial services industry, healthcare industry, design and creative work industry, of course, the cyber security industry, and this week, the legal industry, and for small businesses. You can toggle on Claude for small business inside of Cloud Co-work. Connect the tools you already use and pick a job. And it's essentially got a bunch of ready to run pre-built agents that work across finance, operations, sales, marketing, HR, and customer service and automatically connect tools like PayPal, QuickBooks, HubSpot, Canva, Docyign, and others. So, it almost feels like every single week another little like industry they put a little target on and go, "All right, we're going to build a bunch of plugins and agentic workflows for this industry and make all of them freak out."
But here's an actual cool story that came out of Anthropic this week. This person on X who apparently goes by the emoji for ramen noodles actually managed to get into their Bitcoin wallet and access their five Bitcoin that they've been trying to get access to for like 12 years now. Last ditch effort. I dumped my whole college computer into Claude. It found an old wallet file that the pneumonic successfully decrypted. It was locked out 11 plus years because I got stoned and changed the password. So you can see on this screen here a quick like recap of what happened. Feel free to pause and read the breakdown, but pretty crazy. Somebody was able to access their crypto wallet and Anthropics Claude helped them get back in by them essentially just dumping everything that was on their hard drive into Claude and saying, "Look through all this and help me figure it out."
We also got a little bit of AI news out of Meta this week. They added an incognito chat inside of WhatsApp. So, you can chat with your Meta AI in WhatsApp and have it not saved. Your incognito chat conversations are processed in a secure environment that even Meta can't see and it disappears by default. Kind of crazy that it took Meta this long to have like an incognito style chat. I mean both Claude and OpenAI and I think even X all have that feature already.
Meta's most powerful large language model called Muse Spark which they announced last month is finally getting rolled out. They updated this blog post here on May 12th and they said when we launched Muse Spark last month we said we'd bring it everywhere. Today we are with faster voice responses and Meta AI apps, smarter AI glasses, and new ways to shop and get help in your conversations. Thanks to Museark, Meta AI is now more capable, more contextual, and more useful in everyday life.
If you're a developer and you're also a user of Notion, you'll like this update. Notion rolled out this update for devs, the Notion developer platform. So, it's got a notion CLI where you can actually mess with Notion directly from your terminal. It's got workers where you can run code on notion's infrastructure, database sync, agent tools, web hook triggers, external agents API, notion agents SDK. Now you can sort of build apps or work with notion directly from your terminal. Or if you're using things like openclaw or Hermes or various agents, you could give them access to notion and let them do things in notion for you on your behalf as well.
I don't know how many of you remember dig.com, but it was kind of like Reddit before Reddit was as popular as it is now. It was like the original social upbranking down ranking site. Well, they just relaunched Dig. And this time it's one of the original founders of Dig, Kevin Rose, and one of the original founders of Reddit, Alexis O'Han, coming together to launch this new version, which I've just been kind of obsessed with lately. It's all based on trends on X and it uses AI to analyze what topics are hot on X based on like the top 2,000 voices in the world of AI on X and it's it's just a really really cool platform to keep up to date with the news in AI. That's one of the questions people ask me all the time. How do you stay so up to date with AI news? Well, I skim X a lot and now this new Dig AI that you can find over at di.gg kind of does a pretty good job of it as well. It surfaces all of the trending stuff. You can see the most popular story for Thursday, May 14th is OpenAI releasing the Codex mobile app. We can see the news that's on the rise over here. GitHub repos that are getting some traction. It's just a really, really cool place to sort of get your finger on the pulse of what's going on in AI at any given moment.
I came across this on the chat GPT tricks Instagram and I thought this was pretty hilarious. So, this account put this image up on X. I just generated an image in the style of a Monae painting using AI. Please describe in as much detail as possible what makes this inferior to a real Monae painting. Now, the funny thing is this is a real Monae painting. They just said that they generated it and they added the made with AI tag, which is just a check box in X. You just check a made with AI box and it adds that to it, even if it wasn't really made with AI. And pretty much everybody on X explained why it didn't look like a Monae at all and how it lacked soul and how you can see that the colors weren't the same as what Monae would have used and all sorts of just like negative comments and giant descriptions about why this art is actually soulless when in reality it was literally a real Monae.
This is something I came across on X that I thought was really cool. This is from World Labs and it's fully open source. You give it an input image and it generates the environment, meshes, physics, lighting, and audio. So, we can see in the bottom right here, this is the original image that they uploaded. And then it converted it into this essentially, but you can see it broke out some of the objects in the room. And then it left this sort of static scene. And then it put these objects back into the scene. And we can see here they are again in that original image. But then it also made this room like a 3D room that they could move around. And then these objects are all movable and interactable as well. And this again all started from a single image. The image that we see in the bottom right. Here's another example of like almost more of a cartoony looking image. And all of these things got generated into objects, removed from the background, and then turned into like movable objects. And the environment turned into a 3D environment that they can actually move around and manipulate. Now, this is something I haven't played with yet, but apparently you use it inside of Claude Code, and it is available on GitHub. It looks pretty easy to install, so I'm planning on playing around with it. Maybe I'll make a video about it and see how it goes. But basically, you just open up your terminal, clone this repo here in your terminal, run Claude Code, and then essentially just give it the prompt, blast it, and confirm each step with me. And then it'll just do that. Makes the 3D models, the Gaussian splats, the ambient looping sounds, all of that is just done from this one. GitHub repo that you install on your computer and then prompt via cloud code. Like this is the type of stuff that to me is just so much fun to play with. Like this is the kind of stuff that I love AI for here. And when I have a little bit of time later this week, I'm probably going to dive in and test this myself cuz yeah, that's that's just sweet.
I thought I'd share this real quick because I've shared some Rivian updates in the past and I actually drive a Rivian which is one of the reasons I follow what's going on with Rivian. This week they introduced a new Rivian assistant and Riven unified intelligence. So this is AI directly inside of the Rivian vehicle, but the AI knows everything that's going on with your vehicle. It knows like all the diagnostics and things like that. Like you can say things like make everyone's seat toasty except mine. And it'll turn on the heated seats. It'll actually read text messages to you and then allow you to voice out a new text message to reply. It's also an encyclopedia for your vehicle, so you can ask questions you might get from a user manual, like, "How do I change the tires?" And it'll walk you through it. Pretty cool stuff. And I know this is specific to Rivian, but it's not hard to imagine that this is sort of like the future of cars in general. They're probably all going to have these AI features in them in the future where you say, "Hey, this issue is popping up. How do I resolve it?" and the AI will know your specific vehicle, will know your specific vehicle diagnostics, and be able to sort of walk you through how to troubleshoot any issues with your car. I don't think that's too far out of the realm of possibility for like all future vehicles to have this.
The company Figure Robotics has been doing this ongoing live stream of this robot that just keeps on working 24/7. Originally, it was just going to be like an 8 hour stream of this robot sorting packages, but once it hit 8 hours, they basically said, "Screw it. We're going to keep it going. And as of this recording is live right now, and it's been live for over 34 hours now. And it's sorted 43,000 packages. Now, I don't actually know what this robot is doing. It just seems to be grabbing packages and just sort of like flipping them around and moving them onto the conveyor belt. So, it's not really anything like too insane, but it just keeps on going. It's basically flipping the label down and moving it, and that's about all. but it's been doing that for almost 35 hours now.
And those are the main stories I wanted to share from this week. Next week is actually Google IO out in Mountain View and I'm going to be headed out to it. I'm going to try to get some hands-on demos and I'm going to sit through the keynote even though I'm not a huge fan of sitting through keynotes. I'm going to sit through the keynote and then uh mainly there cuz I want the demos and to meet the other AI folks that are out there. But there's some rumors circulating about what they're actually going to announce at Google IO. I don't have any details yet. I'm just going off of rumors I've seen on the web. But here's what I think we can expect out of Google next week.
Yes, it will probably be a very big week for Google. I think we're going to see a new version of Gemini. According to Bendu ready here, Gemini 3.2 Flash is coming out next week. Supposedly, you can do about 92% of what GPT 5.5 is capable of when it comes to coding and reasoning, but it's going to be 15 to 20 times cheaper and quite a bit faster. So, expectations are we are going to see a new Gemini model next week.
testing catalog here posted this. A new Gemini Spark agent is about to be revealed during Google IO. Spark will work as a 247 assistant that could learn from user behavior and work with connected apps and skills. So, this sounds like it could be Google's answer to like an open claw Hermes like agent kind of thing that just keeps on working for you on your behalf. But, we'll see.
And the other thing I'm expecting but have no clue if we'll actually see is how far along they are on the new version of the like Google glasses. I was at Google IO 2 years ago and they were sort of teasing some Google glasses with displays on them that had all the features of the meta ray bands but with the displays. And I've actually demoed them. They are really really cool. Last year when I went to Google IO they actually finally demoed them on stage and did like a real live demo of them but to this date we still haven't seen them get released. So, I'm not expecting them to like be launched at Google IO, but I am expecting at least some sort of update. And I'm sure we'll see more about all of those like Androids and Google book updates that we talked about earlier. Google likes to save a lot of their best announcements for IO, and that's happening next week.
But that's what I got for you today. A lot of uh kind of smaller things, a whole bunch of little updates, but some really cool ones. I'm really impressed by what Thinking Machines put out this week. I want to get my hands on it. I'm super excited about the codecs in the chat GPT app on my phone. I don't have to sit at my computer to like sort of steer my coding anymore, which is really, really cool. So, I'm excited about that. And uh yeah, that's what I got.
Next week, again, I'll be at Google IO, so probably only an end of week news video. I doubt I'll do more than the one video next week. I don't know, maybe I'll get something else out. We'll see. But definitely an end of week video where I break down everything I learned at Google and of course any other news that came out that week that wasn't from Google.
If you want to stay informed and looped in on all of the latest AI news, make sure you like this video, subscribe to this channel. It'll make sure more videos like this show up in your YouTube feed. My goal with this channel now is to drink from the fire hose every single day. Keep up with all the news, all the latest tools, all the cool stuff that's happening in the AI world and then break it down into just what I think the most people are going to want to know and filter out all the noise and hype and junk and things like that and just try to make it a little less overwhelming for people. There are channels out there that'll break down every single piece of news every single day for you. That's not what I want to do. I want to give you this one roundup every week that sort of lets you know everything you need to know so you don't need to try to stay up to date all week long. So again, if that's something that interests you, you know what to do. Really, really appreciate you. Thanks for hanging out, nerding out with me today. And uh hopefully I'll see you in the next one. Bye-bye.