AI News: Every Major Announcement From This Week

Once again, there's been a ton of news in the world of AI, and I want to help you separate the signal from the noise and tell you what I think you should know, and also just some of the more interesting and cool stuff that happened. I'm not going to waste your

happened. I'm not going to waste your time. Let's jump right into it. Let's

time. Let's jump right into it. Let's

start with a company that I haven't talked about in quite a while, and that's MidJourney. A company that

that's MidJourney. A company that actually kind of made my channel in the beginning. I actually started this

beginning. I actually started this channel by doing a whole bunch of MidJourney prompting tutorials. Well,

they've just released their eighth version of MidJourney. Apparently,

MidJourney V8 is much better at following detailed directions, has a much better ability to understand your aesthetics through personalization and style references. Images are more

style references. Images are more coherent and detailed. Apparently, text

rendering works better than ever. That's

something that MidJourney has historically been like the absolute worst at. They've upgraded their web

worst at. They've upgraded their web interface, and a lot of their old parameters that you used to be able to use, you can still use, like chaos and weird and raw. They also have a new HD mode which renders images in 2K

resolution. And well, based on their

resolution. And well, based on their featured image here, it seems like they're going more for like creative images and sort of like fictional worlds and stuff as opposed to realism. Now, I

don't want to nitpick too much here, but I find it interesting that when I zoom in on this image, this lady here has only three fingers when, you know, it felt like AI models fixed fingers a

while ago. Now, we're going to do some

while ago. Now, we're going to do some of our own tests here in a second, but so far the consensus online is that this new model has not been amazing. Alex

Petrrescu here posted this comparison that he generated. If you want, you can actually pause and read the full prompt.

It's a pretty detailed prompt, but the idea here is that the man's hand is supposed to be on fire, right? Here's

what Nano Banana generated. Here's what

Cadream 5.0 Light generated. And well,

here's what Midjourney V8 created. And

it looks like not only his hand, but his entire shoulder and neck might also be on fire. To be fair to Midjourney, I

on fire. To be fair to Midjourney, I actually copied and pasted that exact same prompt in here. And these were the examples I got. Now, his fingers are really jacked up, but I mean, they're

also on fire. I didn't get the whole shoulder on fire thing. Here's another

example that I got. Here's another

example. And here's another example. So,

I feel like this person may have potentially cherrypicked like the worst version to share against these other two, but I can't be certain of that.

There are some other examples that they shared, albeit these ones they didn't share the prompt, so I couldn't actually go and do my own testing on. But here's

one of like a Ford Bronco generated in Midjourney versus one generated with Nano Banana. Here's one of someone

Nano Banana. Here's one of someone staring into the camera generated with Midjourney versus one that was generated with Higsfield Soul Cinema. This one I thought was hilarious. I think it was supposed to be like the woman sitting in

like the back of the car with the hood open, but it like put her separate from the actual car. This is the midjourney version versus what Nano Banana managed to generate. As always, I'll link this

to generate. As always, I'll link this up in the description. He did show off some other examples from his own testing, kind of showing off that MidJourney performed the worst. But when

you do generate with MidJourney, it generates four images, and we're only seeing one here. So, I mean, maybe they were cherry-picking the worst ones. Jav

Lopez here also shared some examples of testing on midjourney. Feel free to pause and see the prompts. These were

the outcomes of these like cartoony images that he generated. Pointing this

one out specifically that when you zoom in on the hand here, yeah, we thought this was supposed to be a solved problem. Here's a comparison Jav did of

problem. Here's a comparison Jav did of V7 verse, I'm guessing, the exact same prompt in V8, which yeah, that arm's definitely broken. But I'm not just

definitely broken. But I'm not just going to take these guys words for it.

I'm going to test some of my own prompts here. So, for this first test, I'm going

here. So, for this first test, I'm going to test instruction following. I'm going

to give it a very detailed prompt and see how well it follows the instructions. And I'm going to go ahead

instructions. And I'm going to go ahead and submit this. And one thing I will say that I've noticed about this model is it is extremely fast. Like way faster than version 7. I mean, it got most of the stuff right. But if we open up and

look at it, I mean, it's not horrible. I

don't feel like these keys and this mouse pad are to scale, but it's not bad. This one looks a little bit better

bad. This one looks a little bit better here. This one, it just totally botched.

here. This one, it just totally botched.

Like, this looks horrible. This looks

like midjourney V1 or V2 level. Like,

this is bad. And then, yeah, this one probably followed the instructions the best with like the 30° angle open and everything, but hard to be impressed with what we've seen more recently with image generators. Let's test text

image generators. Let's test text generation. Feel free to pause the

generation. Feel free to pause the screen, but the text we should see is AI won't replace you, but someone using AI will. Okay, I can't help but think this

will. Okay, I can't help but think this one's actually trying to look like me, but probably not. Just coincidence. All

right, let's take a peek at these. I

mean, the text rendering is what I was kind of wanting to pay attention to here, and not great. Here's the one that kind of feels like it looks like me a little bit. I don't know why it made it

little bit. I don't know why it made it all out of focus when like I'm specifically asking for that text with my prompt. Same with this one. Like,

my prompt. Same with this one. Like,

it's clearly putting it out of focus. AI

won't replace you, but someone using AI will. That one's probably like the

will. That one's probably like the closest. But I mean, just like the

closest. But I mean, just like the details here are just mangled. Like

where is this arm going to? Why is there two arms going to the same microphone?

All right, I'm going to test something like really weird and random here. A

transparent glass elephant walking through a desert made of books instead of sand. Inside the elephant are tiny

of sand. Inside the elephant are tiny astronauts repairing glowing constellations. Sky filled with floating

constellations. Sky filled with floating jellyfish shaped like planets. Surreal

but physically believable lighting and shadows. Extreme detailed imaginative

shadows. Extreme detailed imaginative coherent composition. weird 800. I mean,

coherent composition. weird 800. I mean,

this is like weird infinite, but let's see what it does. Quite honestly, I feel like this is where Midjourney excels.

Like, you just give it the weird random craziest things you could imagine and like it kind of figures out how to do that in an interesting way, but then

when you want it to do something more specific, it's not good. They were, in my opinion, at one time the state-of-the-art image model. They were

pretty much the best you could get.

super super impressive with every image and now like I feel like they've gotten worse and maybe that's just because we've gotten desensitized to AI images and things like Nano Banana are just

like so far ahead of this now. Yeah, I

guess this is just the state of Midjourney these days. But that's not the only image generation model we got this week. We got one out of a company

this week. We got one out of a company that I was not actually expecting to see an image model from this week and that's Microsoft. This week they introduced MAI

Microsoft. This week they introduced MAI image 2. And over on the text to image

image 2. And over on the text to image arena here, it's currently ranked the third best image model behind OpenAI's image model and Nano Banana. This one

seems to have a focus on photo realism.

It's built for creatives who want images that feel like they exist in the world with natural light, accurate skin tones, environments that feel lived in. We

could see some examples here. Reliable

in image text generation. So, we'll test that. Rich detailed scene generation. I

that. Rich detailed scene generation. I

mean, it seems to have, you know, like nano banana level text generation within the image. So, let's go ahead and click

the image. So, let's go ahead and click on try MAI image 2. All right. So, I'm

going to go ahead and test a prompt down here. Feel free to pause the screen if

here. Feel free to pause the screen if you want to see the whole prompt, but I'm checking for realism, water physics, lighting interactions, and micro details. And I mean, I already think

details. And I mean, I already think it's more impressive than midjourney if I'm going to be honest. And it seems to have followed the prompt. A woman in a rainforest. Golden hour. Face partially

rainforest. Golden hour. Face partially

submerged in a shallow stream. Water

refracting sunlight into prism patterns on her skin. Tiny insects hovering above the surface. I mean, if we zoom in,

the surface. I mean, if we zoom in, they're there. Wet hair clinging to her

they're there. Wet hair clinging to her face. I mean, yeah, a little bit. And

face. I mean, yeah, a little bit. And

it's got the cinematic depth of field as well. So, I think it did a pretty decent

well. So, I think it did a pretty decent job. So, this time, let's go ahead and

job. So, this time, let's go ahead and test some text. So, a coffee shop menu board with a bunch of details. Let's see

how well it gets all these details into the image. All right. So, Orbit Cafe,

the image. All right. So, Orbit Cafe, check. Espresso, $3. Cappuccino, $450.

check. Espresso, $3. Cappuccino, $450.

Vanilla latte, five bucks. Cold brew,

$4. Avocado toast, $7. Open daily, 7 a.m. to 4 p.m. Wi-Fi password, stay

a.m. to 4 p.m. Wi-Fi password, stay curious. So, yeah. I mean, it it nailed

curious. So, yeah. I mean, it it nailed it. It did the exact thing it was

it. It did the exact thing it was supposed to. Let's go ahead and test a

supposed to. Let's go ahead and test a weird one. Here's the prompt for that. A

weird one. Here's the prompt for that. A

highly detailed product shot of a transparent glass sneaker. Inside the

shoe is a miniature ocean ecosystem with waves, coral reefs, and tiny fish swimming. Condensation forming on the

swimming. Condensation forming on the inside surface. Soft studio lighting.

inside surface. Soft studio lighting.

Hyper realism materials. Apple product

photography style. Pure white

background. Ultra sharp. I mean, yeah, that did a good job. We've got our miniature ocean inside of our transparent sneaker. It's even got a

transparent sneaker. It's even got a little bit of waves in there like I asked. Tiny fish swimming coral reef.

asked. Tiny fish swimming coral reef.

And it's an Apple product photography with a plain white background. I mean,

we got two image models this week. And

in my opinion, one is more impressive than the other. If you've been following my AI news videos recently, you know that AI agents are a huge priority for all the big tech companies right now.

Some people have been scrambling to buy Mac minis to run their agents on, so it's on a separate computer from their personal files. But if you want an even

personal files. But if you want an even easier and cheaper option, you could run your agents on a VPS. So, this platform called Hostinger has a whole pre-built system that spins up a VPS specifically

to host OpenClaw agents on. You just go to hostinger.com/mattopenclaw.

to hostinger.com/mattopenclaw.

Choose a plan. Use my code Mattwolf to save some money and then click deploy.

Openclaw is pre-installed through their Docker template, so it autodeploys in minutes. After checkout, you set up your

minutes. After checkout, you set up your environment variables. You confirm, and

environment variables. You confirm, and then it launches. Connect WhatsApp with a QR code and your agent is live running 24/7 in the cloud and not on your

personal laptop with access to all of your sensitive documents. So instead of hearing about everyone's super cool OpenClaw projects, you could build your own without spending a ton of money or

setting yourself up for security nightmare. And the VPS is great because

nightmare. And the VPS is great because it's always on so you can have your agents running in the background all the time or like while you sleep. If you

want to create your own openclaw agent without the cost or risks, check out the link in the description below. And thank

you so much to Hostinger for supporting my channel and sponsoring this portion of today's video. We also got a new vibe design feature that came out from Google this week and it's for their product

Stitch. There's a new AI native design

Stitch. There's a new AI native design canvas which to me looks very very for lack of a better descriptor Figma. It's

an AI native infinite canvas that gives your ideas room to grow from early ideiations to working prototypes. You

can also create what's called a design MD. MD is a markdown file. It says you

MD. MD is a markdown file. It says you can easily extract a design system from any URL or use the new design MD, an agent-friendly markdown file to export

or import your design rules to or from other design and coding tools. So, you

create like a markdown file which is kind of like a skills file and something like cloud code where you explain the design you're looking for and any new design you make follows the same

aesthetics based on this design markdown file. They also added voice capabilities

file. They also added voice capabilities where you can speak directly to the canvas so you can make real-time updates like give me three different menu options or show me this screen in different color palettes. It almost

sounds to me like they're designing this to let your agent work inside of it. It

uses MCPs. It uses skills. It uses

markdown files. You know what else does all that? Things like OpenClaw, things

all that? Things like OpenClaw, things like Clawude Code, the various agentic tools that are out there that just go and do things on your behalf. So to me, it seems like they're designing this so you can go to something like OpenClaw

and say, "Hey, design me a new website and give me three variations so I can pick between them and your agent will go and do the design work for you and then load it into like a Figma like design

platform." So I'm over at

platform." So I'm over at stitch.withgoogle.com.

stitch.withgoogle.com.

We can choose different models for the design. We can build mobile apps or web

design. We can build mobile apps or web apps. Let's have it design a web app

apps. Let's have it design a web app here. There's design systems. So,

here. There's design systems. So, there's already existing I guess this is like color schemes in here. Let's try

design a website that helps curate and filter all the latest AI tools the same way futuretools.io does. Hopefully, I

don't regret this. All right, so we can see it's designing over here. And while

we're waiting, now's a good time to let you know that I just redesigned futuretools.io.

futuretools.io.

Probably not as good as what this is about to do, but I think it's more useful and easier to navigate. Okay. So,

it looks like it gave me like a color scheme and it's asking me to approve it and I think it's basing it on the existing color scheme of future tools.

Sure, let's go with it. Okay, not bad.

So, yeah, very, very Figma-like. We can

see what the design looks like here. And

it looks decent. I mean, it's pretty close to what I've got on my site. Let's

test this live mode here that's in preview.

>> Hey, what are we designing? Can you take what is currently selected and give me three different color scheme designs?

>> Got it. On the home/discovery screen, let's explore some vibrant palettes to make it pop. Working on it? I'll

generate three exciting variations for you to check out.

>> So, I can see down here, it created a new color scheme. And if I click right here, it actually zooms to one of the new designs that it just made. It looks

like it just finished another design.

So, if I click around, this is a second color variation. Not only is it a

color variation. Not only is it a different color variation, but it's like a completely different design as well, which is kind of cool. This is the original design that looks very close to what Future Tools looks like. And then I

scroll over here, and these are like two different color schemes, but also like two pretty different designs. And

apparently, it just did another one.

Let's see. Where did it put it? Oh, and

then here's the other one. I don't quite understand how it's deciding how to lay stuff out. Like, why is these way over

stuff out. Like, why is these way over here? these close together, but let's

here? these close together, but let's move them all close together and we could compare different potential designs side by side. And it just designed all of these websites for me.

Four potential design ideas for future tools here. This is a really cool

tools here. This is a really cool project. If you haven't already, check

project. If you haven't already, check out stitch.withgoogle.com.

out stitch.withgoogle.com.

I'm really quite impressed because now what we can do is we could literally take this design and throw it into codeex or claude code or cursor or whatever your favorite vibe coding app

is and literally give it the screenshot and say code this up for me and it will do the rest. Honestly, those tools where they probably struggle the most is in

design and this tool here seems to be pretty decent at it. So, you design with a tool like this, get your color schemes dialed in, and then either pass it off to a developer or pass it off to a vibe

coding tool to develop it out. I

approve. Google actually calls this vibe designing cuz it is very similar to vibe coding, just uh trying out new designs.

But since we're talking about vibe designing and vibe coding and well, Google, let's talk about this one next.

Google introduced a new full stack vibe coding experience inside of Google AI Studio. I can't help but think that it

Studio. I can't help but think that it was pretty intentional that the same week they released their Vibe designing tool, they also released their Vibe coding tool. Cuz it seems like you'll be

coding tool. Cuz it seems like you'll be able to just pull what you designed in Stitch over into the new Google AI Studio coding experience and say, "Code this up for me." From prototypes to production, build multiplayer

experiences, add databases and authentication, use modern web tools, connects real world services, easily pick up where you left off. So, pretty

cool. Let's go over to Google AI Studio real quick. So, describe an app and let

real quick. So, describe an app and let Gemini do the rest. Well, why don't I just grab one of these designs here that I just made. Let's take this bright one here. Okay, let's download I downloaded

here. Okay, let's download I downloaded that screen and I didn't even realize it, but it actually already coded that up. I just downloaded it and it gave me

up. I just downloaded it and it gave me a code.html, a design.md, and a screen.png file. So, it actually coded

screen.png file. So, it actually coded this up already. I mean, it gave me like a sample. None of the buttons work or

a sample. None of the buttons work or anything. So, I'm going to jump into our

anything. So, I'm going to jump into our Google AI Studio here and toss these three files in our code.html, design.md,

and screen.png. Although, it does say design.md is unsupported. Let's just see

design.md is unsupported. Let's just see what happens. Let's build a functional

what happens. Let's build a functional version of this site. All right. So, I

tried to copy and paste all three files over the code. HTML, the screen.png, png and the design, but it kept giving me an error and saying it couldn't read the design file, which I find interesting

that these markdown files are kind of what all these agentic systems run on, but Gemini can't read them. That seems

like a little bit of an oversight. I

have a feeling that will be temporary.

But let's just go ahead and paste in code.html screen.png and say, let's build a functional version of this site, which of course is the site I just designed over in Stitch. And would you

look at that? It built a version of this site. I need to reload this so you can

site. I need to reload this so you can actually see what it looked like when it first loaded in. So check this out. I'll

click launch. And all of this like animates in when it starts. Like that's

pretty cool. We've got hover effects when I click over the various tools.

This is actually filtering what was there already. And we have a functional

there already. And we have a functional website from just one prompt. This is so good. Let's see. Does the dark mode

good. Let's see. Does the dark mode work? No. Okay. So, it didn't build out

work? No. Okay. So, it didn't build out every single feature. Like we don't have dark mode. Submit a tool doesn't do

dark mode. Submit a tool doesn't do anything yet. But it actually coded up

anything yet. But it actually coded up the design. It animated a lot of the

the design. It animated a lot of the design. It made these filters

design. It made these filters functional. Let's see. Sort by newest

functional. Let's see. Sort by newest first. The sort dropped on doesn't work.

first. The sort dropped on doesn't work.

Some of the stuff needs to be fleshed out. Like it needs to know what happens

out. Like it needs to know what happens when you click into these things. But of

course, you jump over to Stitch, you design what these internal pages look like, and then you bring those designs back into AI Studio. And like you've got the full loop here. You've got a really

good platform for actually designing websites or web apps or games or whatever. And then you've got like a

whatever. And then you've got like a vibe coding tool where you can just drag and drop the designs it made in and work with it until it's functional. Like

that's pretty cool. And if you do want to get into the code and mess with it yourself, you've got the option here to jump into the code as well. Honestly, I

think Google's kind of killing it with these two new updates. But since we're talking about Google, one last update I want to share from them real quick, and that's that they're rolling out the personal intelligence to more people. So

the personal intelligence allows you to connect your Gemini app where you actually chat with it to more of the Google ecosystem, things like Gmail, Google Photos, your calendar, things like that. And when you chat with

like that. And when you chat with Gemini, it actually can reference those other tools and help provide personalized response based on all of the stuff that it knows about you. And

before this feature was only available if you were on one of the paid plans like a pro or an ultra plan. But now

personal intelligence is available in the US for AI mode in search and it's starting to roll out in the Gemini app and Gemini and Chrome for free tier users. So you're going to be able to

users. So you're going to be able to have access to all of your personal information inside of your AI chats even if you're on a free plan now. All right,

moving on. Let's talk about where pretty much all of the attention was this week and that was Nvidia GTC. That's their

annual conference where they show off all of the stuff they've been working on. There's a big expo hall. They do

on. There's a big expo hall. They do

like interviews with leaders in the AI space. I got to see some insane panels

space. I got to see some insane panels while I was out there and it felt like pretty much everybody from the AI space was in San Jose at this event. Now, most

of the announcements they made were a little more focused on data center buildouts and enterprise and stuff that I don't really feel like most people are going to pay that close attention to.

But there was a handful of things that I found interesting that I thought I'd share with you because you might find them interesting as well. Starting with

Nemo Claw. If you're not familiar, this year has been all about OpenClaw, the open- source project that basically made agents that can do almost anything available to pretty much everybody. And

a huge portion of Jensen's keynote and a ton of the talks at this event were all about OpenClaw. Well, Nvidia released

about OpenClaw. Well, Nvidia released something called Nemo Claw, which is a sort of like bundling around OpenClaw.

It adds extra security layers. It lets

users install Neotron models from Nvidia and basically provides a bunch of extra Nvidia focused optimizations and capabilities around your OpenClaw. It

also makes it really really easy to install OpenClaw. there's like one

install OpenClaw. there's like one command that you run in your terminal and it installs OpenClaw plus all of this like Nvidia stuff that provides that additional security because so far OpenClaw the biggest issues people had

around it was security and privacy. You

can see here Nemoclaw installs in a single command adding security and privacy to run secure always on AI assistance from the cloud and on premises to Nvidia RTX PCs DGX stations

and DGX Spark. I actually had them install all of this stuff for me on my DGX Spark while I was out at GTC. So, if

you've been hearing a lot about OpenClaw, but you've been too scared to set it up, well, look into Nemo Claw because Nemoclaw a makes it easy to set up and b adds those security and privacy layers that most people have been

worried about. They also announced

worried about. They also announced DLSS5. And what this essentially is is

DLSS5. And what this essentially is is basically a new feature for like game developers that they can bake into their games to have it essentially like upscale the quality of their games. Now,

they announced this and they got like a ton of backlash from the gamer world.

People basically saying like, why are you reskinning existing games instead of just leaving them the way the game developers and designers made them? But

Jensen reiterated multiple times that this isn't just like an upscale layer on top of existing games. The game

designers actually have control of like what level they want to allow this tool to be used. So game developers who want their graphics to look a little bit better than what they're capable of

designing can set this up and improve the output of their graphics if they choose. And the people who are using

choose. And the people who are using Nvidia GPUs, if they don't want this DLSS5 turned on, they can literally flip a switch and turn it off and just not

use it if it bothers them. Nvidia also

announced their plan for space computing with Nvidia Space 1 Vera Rubin modules.

Now, these are literally what they sound like. They're data centers and GPUs

like. They're data centers and GPUs designed to go into space. Now, no

timeline was announced for this because, well, there's still quite a bit of issues with sending data centers to space, but the main one being is that GPUs generate a lot of heat. The sun

generates a lot of heat, and they haven't figured out a way in the vacuum of space to dissipate the heat off of these. You actually have to like

these. You actually have to like engineer ways to sort of push the heat away from these things. And that's

something engineers haven't quite figured out how to do yet. But, they are starting to build the infrastructure and the GPUs that can go to space. they just

don't necessarily know how to keep them cool yet. And probably the biggest

cool yet. And probably the biggest talking point from Nvidia GTC was Jensen's announcement that through 2027

they expect to have $1 trillion in GPU sales.

>> I see through 2027 at least $1 trillion.

I am certain computing demand will be much higher than that. Now, I went to a private press Q&A session where people were able to ask questions to Jensen.

And the question that kept coming up was like, where did that trillion dollar number come from? And Jensen confirmed that these are basically purchase orders. Like, companies have already

orders. Like, companies have already said they will buy these GPUs if they're available for us to buy. So, he claims that that trillion dollar number is based on existing purchase orders.

Meaning that Nvidia still seems like it has a lot of growth left. For context,

they've did like 500 billion in the previous year. So, like they're talking

previous year. So, like they're talking about doubling that number. Again, those

were the sort of big takeaways from GTC.

There was a whole bunch of other announcements, but they were more focused on like very sort of like niche areas. They showed off some stuff for

areas. They showed off some stuff for like the automotive world and the biotech world and and things like that.

But the things that I think most people are talking about are the Nemo Claw, the DLSS5, the Nvidia and space stuff, and the fact that they're projecting a trillion dollars in chip sales through

2027. Okay, I've covered a lot of stuff

2027. Okay, I've covered a lot of stuff already and there's still quite a bit more I want to get through. So, instead

of breaking it all down like really slowly, let's jump into a rapid fire and I will rattle off everything else that I think is interesting from this week.

Let's start with the new large language model news cuz we did get quite a bit of it. Although for most people it's

it. Although for most people it's probably going to feel fairly marginal.

So this week, OpenAI released a new smaller version of their GPT 5.4. They

released a mini version and a nano version. These models are designed to be

version. These models are designed to be slightly faster, slightly cheaper, but also not quite as intelligent. We can

see when it comes to SWEBench for coding, they slightly underperform their state-of-the-art model here. With

Terminal Bench, again, they underperform their state-of-the-art model. And well,

I mean, pretty much all of the benchmarks, they're going to underperform slightly from the bigger version, but they're fairly close and they do it faster and cheaper. It also

feels like with every new announcement out of a large language model, they're talking about agent use cases. GPT 5.4

Mini is a strong fit for systems that combine models of different sizes. So,

they're basically saying that these models are good for use with your agents. And you want these smaller,

agents. And you want these smaller, cheaper models for your agents because well, your agents are using a lot more tokens. If you're creating something

tokens. If you're creating something like an open claw, you're not only using tokens when you're sitting in front of a chat, it's actually operating behind the scenes and doing stuff all the time for you, which uses tokens, which you

probably want to be slightly cheaper than using the frontier state-of-the-art models. The mini version is also fairly

models. The mini version is also fairly on par with 5.4 full model at computer use. And once again, this feels like a

use. And once again, this feels like a play to get you using these models with your agents instead of the bigger models because these are going to be, you know, good enough for your agents. Claude this

week made a million token context window available in their Opus 4.6 and Sonnet models. So, if you're using Claude just

models. So, if you're using Claude just over at claude.ai, you can paste in massive amounts of text now and it's going to understand all of that and be able to sort of chat about it. And

you're also going to have much longer conversations because you're going to run out of context a lot slower than you used to. The company Mistraw released a

used to. The company Mistraw released a new model called Mistraw Small 4. This

is actually an openweight model, so you can actually fine-tune it if you want and run it on your own servers. And we

can see here on these various benchmarks that when given reasoning, that's what this little sort of hashed version of the chart is, it performs close to on par with models like Claude Haiku, Quinn

380B, and Quinn 3.5122B.

Also, when it comes to coding, we can see it outperforms GPT OSS, Claude, Haiku, Quinn 3, and is about on par with

Quinn 3 Next 80B and Quinn 3.5122. And

then in math we can see it's pretty much about as good as those same models. If

you use cursor for coding you got a new model this week called composer 2. This

is the model designed by the cursor guys and it's very very optimized specifically for coding. We can see it off over here where it isn't quite as

good as GPT 5.4 high or medium but it's pretty close but for a lot cheaper. So

you can see this axis here is how inexpensive it is to run this model inside of cursor. So you're getting better than Opus 4.6 level of coding.

Slightly not as good as GPT 5.4 level, but at a much lower cost. So pretty

impressive cost to sort of coding ability ratio there. Here it is compared to the other models on Terminal Bench.

And you know, if you're like me and you use cursor to code a lot of stuff, this might be the model you want to set as your new default model where you only switch it over to 5.4 when maybe

Composer 2 gets hung up because you're going to save a lot in token costs.

Miniax released a new model this week called M2.7, which claims to be self evolving. Once again, it's designed to

evolving. Once again, it's designed to perform well powering AI agents. And

they even specifically shout out Claude Code, Kilo Code, and Open Claw. So,

another one designed for agents, it seems like. Now, a lot of the models

seems like. Now, a lot of the models that have come out of Miniax, I believe, have been openweight models. This model

is actually a proprietary model from them. So, that's an interesting shift

them. So, that's an interesting shift from one of these Chinese labs that tends to do a lot of openweight models, but here's where this one stands out. By

autonomously triggering log reading, debugging, and metric analysis, M2.7 handled between 30% and 50% of its own development workflow. model optimize its

development workflow. model optimize its own programming performance by analyzing failure trajectories and planning code modifications over iterative loops of a 100 rounds or more. So once again, we're

seeing this loop where the models are being used to train the next models to get even smarter and then that model is even smarter and training the next model and we're seeing this loop where these

models are getting smarter and smarter and smarter because they're using the smarter models to help create the next smarter model. We also got a open-

smarter model. We also got a open- source model called Mamba 3. This one

isn't apparently using the transformer architecture, which all the other models we typically talk about are. This one is a state space model, and according to this VentureBeat article, it's effectively a high-speed summary machine

for AI. While many popular models like

for AI. While many popular models like the ones behind chat pt have to re-examine every single word they've already seen to understand what comes next which gets slower and more expensive the longer the conversation

lasts. The state space models maintains

lasts. The state space models maintains a compact everchanging internal state.

It's essentially a digital mental snapshot of the entire history of the data. So this will be something

data. So this will be something interesting to follow along to to see if more models start to adopt architectures like this. Claude Co-work got a new

like this. Claude Co-work got a new feature this week called dispatch.

Again, I'm sounding like a broken record here, but it sounds like they're trying to get closer to that like open claw experience. One persistent conversation

experience. One persistent conversation with Claude that runs on your computer that you can message from your phone and you can come back to find finished work.

My favorite reply is from Immod here.

You should call it Claudebot. Manis,

which is Meta's recent answer to Claudebot. They recently acquired Manis

Claudebot. They recently acquired Manis after Cloudbot's sort of explosion.

Introduced My Computer, which is Manis, but on your desktop. Sounds awfully a lot like Claudebot. Manis can execute command line instructions in your computer's terminal. This allows it to

computer's terminal. This allows it to read, analyze, and edit local files as well as launch and control your local applications. So, Claude Co-work is

applications. So, Claude Co-work is starting to model Claudebot, a company that they probably could have acquired if they didn't like try to cease the system from using the name. Perplexity

built Perplexity computer and now Manis has Manis on your desktop. Andre

Carpathy, who used to be at Tesla and then was at OpenAI and then left to do his own thing, recently released this US job market visualizer to show which jobs

are most at risk of AI sort of taking your job. We can see that the green is

your job. We can see that the green is growing jobs with a positive outlook and the red is declining jobs. If you want to click around in here, of course, I'll link this up, but we could see things like cashiers are going away, general

office clerks, bookkeeping, customer service representatives, child care workers. Interestingly, I wouldn't have

workers. Interestingly, I wouldn't have guessed that one. Some of the jobs that still seem to be pretty safe, cooks, home health and personal care aids, software developers, quality assurance analysts and testers, electricians,

plumbers painters carpenters construction labor. Still lots and lots

construction labor. Still lots and lots of jobs that are on the rise, but there are definitely jobs that are taking a hit. And according to this, there's a

hit. And according to this, there's a lot more jobs becoming available than jobs going away, which is an interesting finding. Another job that could be going

finding. Another job that could be going away in the near future, Uber driver, Uber is going to invest 1.25 billion into Riven as part of a new robo taxi deal. So, it looks like Uber is

deal. So, it looks like Uber is deploying 10,000 autonomous Riven R2 vehicles to compete with like Whimo and like the Tesla Cyber Cab and things like that. But if your job is one that's

that. But if your job is one that's going away, don't worry. You can always do door dash tasks. Door dashers have new ways to earn on their own terms by completing short activities like taking photos of a dish or recording everyday

tasks. So like you can go to restaurants

tasks. So like you can go to restaurants and take pictures of the dishes and things like that for the restaurant and earn some money. But one thing that's interesting is if I scroll down here, there's one little paragraph that I find

interesting because it's related to AI.

We're also piloting a new standalone app where dashers can complete activities like filming everyday tasks or recording themselves speaking in another language.

This data helps AI and robotic systems understand the physical world. So,

they're also going to pay dashers to basically help them get AI training data in the real world. So, again, some jobs like entry-level type work might be

going away, but new entry-level type work is starting to pop up. whether

you'll still get that same level of pay at the entry level to be determined. But

this is something I've been saying in past videos kind of over and over again that this is where I see things going.

AI companies need data to train on and if jobs are going away, well, maybe these AI companies will actually start to pay people to help them provide data

for them. I've sort of argued on

for them. I've sort of argued on podcasts and things like that that this could be what happens as opposed to like a UBI type thing where the government just gives handout checks is that maybe

some of the big tech companies will come in and pay people to help them provide data. They need the data. More and more

data. They need the data. More and more people are getting concerned with just giving up their data for free. So, they

will likely resort in the future to paying to actually collect that data, providing new income streams that didn't exist before AI. Anyway, just a random rant. Maybe I'll rant deeper about that

rant. Maybe I'll rant deeper about that in a future video. I've definitely

talked about it a lot in past videos and on podcasts and things, but that's one of the like sort of core beliefs I have around AI is that a lot of these companies are going to pay for data as all of the money sort of rises to the

top and all of these big corporations sort of get more and more profitable as a result of AI. Maybe some of the ways that money gets redistributed is through paying people to help them collect more

data. I don't know. Just random

data. I don't know. Just random

speculation, you know, save this for later. Let's relook at this in a couple

later. Let's relook at this in a couple years and see how that played out. But

that's sort of my prediction of where this will go, which is good and bad and has its pros and cons, and I'm not going to go too deep into that cuz this video is already way too long. But that's what I got for you today. If you like staying

tapped in on all the latest AI news, but you don't want to have to try to keep up with it every single day with the fire hose and flood of new AI stuff, maybe like this video and subscribe to this channel because it'll make sure videos

like this show up in your feed. And my

goal with videos like this is to keep you looped in once a week with everything you need to know in just one video breakdown. All of the news that's

video breakdown. All of the news that's important to the most amount of people, plus some of the stuff I find kind of interesting. That's my goal. Separate

interesting. That's my goal. Separate

the signal from the noise and share it with you so you don't have to be totally overwhelmed on a daily basis. Again, I

record these on Thursdays, publish them on Fridays. So, if there's any news that

on Fridays. So, if there's any news that came out late Thursday or Friday that I missed, it'll likely be in next week's video. I'm doing my best to keep my

video. I'm doing my best to keep my finger on the pulse and then turn around and help you keep your finger on the pulse. So again, if you like that kind

pulse. So again, if you like that kind of stuff, I would really appreciate a subscribe. I'm getting really close to a

subscribe. I'm getting really close to a million. I'm hoping to hit it this year

million. I'm hoping to hit it this year sometime. And uh just clicking that one

sometime. And uh just clicking that one little button would really help my channel a lot. So again, thanks again for tuning in. Really, really appreciate you hanging out and nerding out with me today. And uh yeah, that's what I got.

today. And uh yeah, that's what I got.

Hopefully I'll see you in the next one.

Bye-bye.

All right, that is a wrap for the intro.