Going (Nano) Bananas, AI's Axis of Evil, Waffle AI, an MCP Course Platform, and more news

Hello everyone. Welcome to AI Agents Hour. As always,

I'm Shane here with my co-host Obby. What's up? And today we're going to be doing what we do every week, except it's Tuesday today, not Monday. So, that's kind of different, you know. But besides that,

it's it's pretty much the same old same old. How was your uh how was your Labor Day? Labor Day was good. Got to spend some time with the fam and a few friends. So, yeah, kind of recharged the batteries a

bit and ready for kind of a a bit of a push, you can say. Yeah, I think uh I hope everyone's feeling refreshed because we have a lot of to do. So yeah, for those of you that are don't know, we're from MRA. So we're two of the co-founders of Mastra. Everyone should have their uh AI agents book

handy. Obby, do you have yours today? I'm actually calling in from the airport. So I'm on my way back to San

Francisco. When's your flight? Been like two hours, so I had to come early. Yeah. Yeah. This is this is the

first live stream from an airport that we've had. I know. So, yeah. I'm in the lounge chair. People

are drinking mimosas and having fun. Nice. Here. Well, glad you could make it. And yeah,

we got a pretty great show today. We're going to be talking about uh what whatever going bananas means at the bottom. We're going to talk about nano banana. We're going to talk to Manuj from Waffle. We're going to talk to

Akquil who is building an MCP course platform. So, you know, pretty cool. We'll hopefully see a demo of that. And yeah, we're going to be doing all the AI news just like we always do.

Yep. For those of you that have not already, you know, please make sure you're liked, subscribed on YouTube, to the uh to this show, to the MRA AI channel. Also, if you're on X, give us a like, reshare it. We do appreciate that. If you are uh listening to this after the fact on

Apple Podcasts or you know potentially on Spotify, please if if you want to give us a five star review that would be awesome. If you want to give us less than a five star, just find something better to do, you know. Just don't don't leave a review. Five stars only. Yes. Five stars only. But appreciate you

all for tuning in and yeah, let's let's kind of get right into it, I guess. Well, did you what did you do over Labor Day? Anything anything fun? I just slept and try to I was feeling

under the weather. So now I'm feeling good and ready to get hit hit it hard. So I needed it though. I think everyone

needed a little recharge after all the the hustling we've been doing over there. Yeah. Which reminds me, if you're going to be in Paris for AI Engineer, come see us. Yep. Uh well, when are you getting to

Europe, please? So I leave in the 13th or 14th, something like that. Yeah. So in you know 12 days from now basically. Yeah. It's coming up.

Yeah. I leave on the 14th and I we're headed to Belgium first to say what up to Ward who's right here. Yeah. With us in spirit. Yeah. And then Paris the next week. So

you know if you're around there and want to come meet with us, we'll be posting some stuff on Twitter. And if you want to come join us wherever we're at, you're more than welcome to. Whoever's out there. Yeah. Yeah. We're going to try to do at

least some kind of impromptu one or two meetups with people from yeah that are building in AI especially using TypeScript and you know use even if you're not using Ma and you just want to come talk to us about it please do but with that let's talk a little bit about you know Nano Banana because that's been something that

was released last week and I think a lot of people have yeah I've heard some people say it's a a little overhyped, but also some people say it's pretty awesome. So, I don't know what what have you heard? I've heard that it's the bee's knees essentially where I think every every model company is saying that this is the most advanced state-of-the-art models every time they release something. Um, but I think I

played around with it. Pretty cool. Pretty cool. Um, it's a little bit faster than OpenAI generation too, which

I like. That should take some time. So, so I actually have uh so I have a images of you, Sam, and myself and they have this like template generator which pictures yourself through all the different uh I guess different decades. It just uses your picture to like show

character consistency. So, we're going to run through that in a few minutes here after we talk about it. So, for those of you that want to, you know, please be watching the video version of this so you can see how what Obby and I would look like in the 60s or the 50s or whatever. So, not we'll tease that a bit, but let's talk through what they what they're actually saying about it

and then we'll try it out. So, Gemini 2.5 Flash, our state-of-the-art image generation and editing model. As you might have already seen, it excels at character consistency, creative edits,

and has Gemini's world knowledge. you know, they're talking about, you know, how it compares to chat GBT flux, which is kind of the most well-known, most used open source, I guess, image editing model that at least I've seen. And yeah, it compares itself to Gemini 2.0 flash

image as well. And so when you look at it, you know, it it shows some examples. You can use it in the Gemini app or in Google AI Studio. And you can see like you can take an

image, you can generate it across, you know, different things, which isn't that new, right? Like this is kind of existed before. The examples that I've seen that are really cool is when you can take a bunch pictures of a bunch of different things and tell it to put it into one image that it's like take this watch, put it on this person, and put this, you

know, outfit on that person, and it works pretty well. So that's where I think it really excels. I saw some other great examples of, you know, it's a an image of like a model wearing clothes and then you transform it into can you lay all the clothes out on inside whatever that person's wearing. And it's kind of like it's really interesting for people who work in e-commerce because maybe you

don't need to spend that much time or money on taking pictures. Maybe you can generate them so your clothing brand or whatever. That was really cool. Yeah, you can see this is kind of a cool

thing. you take one person and put them on different playing cards and different outfits and different environments. And I think one of the things, you know, you can do is, you know, if you look at this example, change my shirt color to red and remove the earring. You know, works pretty well, right? It just takes the shirt, replaces it, keeps everything

else the same. So, it's it's a way to actually, you know, replaces it could replace Photoshop in a lot of cases, right? where you just want basic image editing and you can take a picture and you can tell it what to do and again you could kind of do this with chat GBT you know right that it does exist but I think this has been a from my experience seems to be seems to work pretty well

yeah in native knowledge is great too because you can just express like you know I want to have a picture of Shane riding a horse or something at least it knows what that would look like or you could even say like oh I want to be at Canaman Square or use historically um and we can be placed there because it

has that knowledge. So this one is cool. Basically kind of what we mentioned, you take the product and you put it into a scene and I did like that they built these template apps. Mhm.

So, the one thing I I don't know is, you know, this is so it is in preview today over the Gemini API and Google AI Studio, but I like that they built these template apps to show you how you could maybe use the functionality of it through the API. And obviously, they're going to want to build all this in in the Gemini app, right? They want to make all this stuff available to you. Yeah. But I think there's probably use

cases for people to build image editing tools or like that are more specific in niche on top of the API. You know, we should take our we should try our Excalibur example app and try to use nano but to see if it you know I think what did we use last time or something? Yeah, I mean yeah you have an Excal

which basically takes an image and then it converts it to Excal JSON. So you could basically take a whiteboard image and convert it to Xcology.json, but it'd be cool if you could describe the diagram rather than like draw it on the whiteboard and then convert that using the same kind of workflow that we've built. And we haven't actually tried it

recently with the newer models. I think it would actually do a lot better especially the newer models that have a bit more context because we had to break it into so the workflow into so many steps because it just over it flooded the context window because it so let's because I'm you know curious let's check out what this looks like

here and you I'm going to just basically this is an example app obviously it's meant to show you know nano bananas or Gemini 2.5 flash image preview so I'm going to upload Upload an image. Let me see.

All right. I don't know who is who. I think this Let's go with you. I think this is you first. All right. There it

is. Let's see what you look like across the decades. And while this is generating, if you are here, you know, this is a live show. If you're here live, check post messages in

the chat. If you have questions, you know what what you're thinking. Have you tried out this image model yet? What do

you think of it? I already don't like it. Look, this is you in the 50s. Yeah. 80s. Holy crap.

The 80s is my style, dude. Dude, is that the D? I thought it was a part of the Goofy movie, dude.

You got the 1970s. That was like my dad's. Holy crap.

2000s. Nice little, you know, 1990s little punk rock. And then it just put the 60s looks exactly you what what I uploaded.

Yeah, I guess I guess that's my style, dude. You're stuck in the 60s, dude. That's what it's saying.

Dude, the 80s one though. What if I Dude, if if we get a bunch of five stars, I'll grow my hair like that. Yeah, dude. I will dress up just like that before they start. I think that that this should be your

like Halloween goal this year is like 1980s, man. All right. I want to see your you know. All right. Let's All right. So, how do we uh start over? Let's try one more here.

All right. I think this one is me. So, here we are. We, you know, it was headsh shot day. Yeah. You know, at a YC

event and so that's why we have the same background. So, let's let's run this. And then we do have Man who's joining us. So, we'll we'll kind of finish up talking about

Nano Banana here in a second, but I am anxious to see what I look like across, you know, different decades. So, here's my 1950s. Look the same. Yeah, basically. So, apparently I'm I'm

a 50s. Oh, man. Holy crap. Okay. Okay. We That's our Halloween costumes

this year. Nice mustache, dude. Dude, if I could if I could grow a mustache like that, I would keep it. I

had a must I hadn't shaved in a while. Like, if you look at last week's video, like I had kind of one going. It's pretty It's pretty lackluster, though. I don't think I would have to like fill it out. I have to That's either got to be

fake or I don't know. This one is um So, what's this one? This one is Can't see 1960s. just looks like a Mi wearing a suit. Which is not

the 70s. Looks like you. Yeah, they morphed your whole face on that. Yeah, that one's a little That one's a

little off in my opinion. Suspect the 2000s. Okay. Yeah, you look like you're in you look like you're in some 41.

Yeah, that's that's like my my band pick. And then the 90s, that's you know that's actually you know if I that's kind of accurate. I probably look like that. No, maybe a little a little

younger. Yeah. All right. Well, there you go. There you have it. Uh, and

there's a bunch of other if you go to, you know, they have a couple other templates, too, for like prompt based image editing, native world knowledge, multi-image fusion. So, check it out on the Google for developers site if you want to play around with Gemini 2.5 flash image and see what everyone's uh

going bananas about. Yeah, you know, the cool thing about I always feel like all the image features brings a lot of new users into the industry because it's like tangible, it's fun, you know, gets like the hype going again. So, I want more image mods. Yeah. I think it's it's just something tangible, right, that people can kind of lock onto. They can see the capabilities

people have, you know, images are so like universal. Everyone has a a camera on them at all times now. Everyone's taking pictures quite frequently. If you're ever in an event, there's always

someone who says, "Oh, we should take pictures of this." And so I do think that people and the ability to do things with image images in general are it does draw a lot of people in. It's very visual.

Flash image needs its Giblly moment. I guess it hasn't had it yet. Maybe because it just came out, but Gibbly moment. It does vio people are generating the same type of image and stuff. That'll be hype.

Yeah. I'm I'm trying to think that that I wonder if that's just a moment in time. I wonder if that'll ever happen again. That was just kind of

Yeah, kind of a wild moment where everything you just saw everywhere was just Gibbly. I think this one it seems like it's a little bit more, you know, like a power tool in a way. Like I feel like you can have a little bit more control maybe than some of the other models, which is nice. But I haven't gone deep enough to know what the weaknesses are because on

the surface a lot of times this stuff seems really great and then you dig in a little bit more and maybe there's some uh you know some some things uh some shortcomings I guess you could say. But with that we do have our first guest. So we're going to be bringing on Manuj from waffle.ai. We're going to

talk about what is Waffle and you know learn a little bit more about what you can do with it. Awesome. Hey. Hey guys, how's it going? Hey guys, how's it going, dude?

It's going well, thanks. It's late at night over here in London. Um, so good to be joining you guys and uh yeah, excited to share more about what we're doing at Waffle.

Yeah. Thanks for being here and joining us from the UK. Yeah. Yeah. Yeah. Yeah. So, um I can

talk you through a little bit about what we've been building and then I think what would be cool is um I want to show you what some of our users have built using Waffle and then uh show you a bit of the product. Does that sound good? Yeah. Yeah. I mean, we we love seeing demos around here. So, definitely uh would would love to kind of hear a

little bit about the journey, you know, like how did you get here building waffle? Then what is it? And I mean maybe just like an introduction like who who are you? who who know obviously we know you

but you know the the audience the audience doesn't for context we were in the same batch as Mastro we did the W25 batch uh and then uh we did our demo day with the spring 25 batch so one batch later but uh I'm Manuj um and my co-founder Dio and I are building Waffle um we are both from a computer science sort of background and we initially set out with the first

version of waffle building kind of lovable style website builder. Uh it was a more generic product to build web apps and we over time shifted it more towards games. And so Waffle right now is an AI games builder and um we both have played a lot of games as kids. Dio's built a game himself that got a million users in

70 different countries and actually he built a bunch of games when he was a kid. And what has been exciting to see from our users is some of them are from are in a similar stage in life where they're kids building their first games and some of them are older people as well building games that they have been

kind of nostalgic about playing when they were kids and it's people ranging from all different kinds of technical abilities not just coders. Um, so, so I think the thing that was exciting, you know, obviously I I had seen the product before and then I, you know, when you kind of announced that it was kind of positioned more towards building games, that was really exciting for me because I have a history of

building games and teaching kids how to build games. You know, I've used like I I built my first video game on a TI 83 Plus graphing calculator. you know, just literally typing out the code on the calculator and it was just like a game loop, like a textbased adventure game that I that I built in middle school, you know, so that's like probably the

first program. I didn't even know what it what programming really was. Like I'd kind of heard of it. That's probably my first introduction was literally

building games because I liked playing video games. And then I've used, you know, from Scratch to Construct or Construct to Unity to all the different platforms for building like 2D and some 3D games that haven't really done much with other than like having friends play. But yeah, that's what I was excited because I'm like, if you can make that easier, that is awesome because I think it would unlock, you

know, in kind of the minds of a lot of kids of what's possible. Yeah, I think when people start building software, one of the first things they think of uh in terms of like what should I code is a game because it's the when you look at the different categories of software, games are the ones where people get a lot of joy and it's the first thing they remember when they

think like okay what should I build? Um yeah so at the moment with Waffle you can build 2D and 3D web games. Um, and yeah, I can show you some of the ones that our users have built over the last Yeah, please the last kind of two months that we've had this. So, just going to share my whole screen.

And while Manuja is pulling that up, if you are in the chat, let us know, have you built a game? What games did you build? How did you what programs were you using? I'm really curious what the

audience is uh has been building with in the past because I feel like it's pretty it's pretty common for people to try to build a game in something right when they're when they're getting started programming or coding. And so you can see my screen, right? Yeah, we can games. I'm going to show you a couple of these. So I'll start

with this. This is a this is a little 2D platformer, but I found it really beautiful the art inside it. It's called Into the Veil. This was made by one of us. You can see

it's been played 80 times. And so we've got our little like likes and comments and stuff. And I'm going to unmute it.

Were you able to hear that? Uh yeah, it kind of sounded a little like I could hear something. So maybe just unmute us. Yeah, let's let's

Okay, let's just uh let's just do it without the sound if if you can't hear it. So well, that's cool. So this is a little elf character and it goes around. It's got some birds. I wanted to show you the sound because it's got really beautiful audio effects

in it as well of the birds chirping and um the narrator talking about where the this kind of character is going on it on its journey. And so this one's got some basic mechanics. The same user, Toxic, he's he's uh uh one of our first users. He

also made this other one called Chariot Arena, which is more of a classic kind of shoot them up game. It's based on Roman gladiators. I'll show you. It's it's it's sort of a top down view.

You're um playing a chariot and you have to um you have to ride around in this in this kind of thing. So cool. the enemies and you you have these spikes on the floor, you have these powerups. Wow.

So, Toxic, for context, is is a user who's not a coder. He's never written a line of code before. He built this all entirely with AI. And then this is a similar from one of

our other users, Jaden, who's built this game. This is more of a 3D like Minecraft kind of game. Can see. Yeah. This this feels like, you know, Minecraftesque, Robloxesque.

Yeah. Yeah. Yeah. And and that's this is a classic thing we see as well is like users want to like emulate the games

that they that they already play. Uh let me Yeah, I'm trying to place a block, but I'm using my mouse pad and it's not so good. And and and so yeah, and then the way that you can kind of share these games is you you build them in our editor here. So this is a game I've I've just been building right now. Um, and then

you can publish them and they go up on the website where other people can um, like, comment, and sometimes if you choose to make it so they can also remix your your game so they can use your game as a starting point. Uh, I'll show you an example where you can remix. So this feels very much like the the Scratch, you know, like but but

obviously like for the AI world, for the AI age, for for the modern way cooler, way cooler, way better, but I just remember like you know seeing like you know remixed and of course they were like not very good games right they were more like an small animations than anything but this is way way cooler.

Uh yeah, and and Scratch is Scratch is like what is cool I think with AI is um even with things like Scratch and Game Maker which have been such great ways for people to start making games in the past. Um you still kind of had to learn a whole framework or like a whole set of things to to be able to make even the tiniest thing. Um, and I think what's cool about AI is you can take someone

who has an idea of something huge and awesome they want to build and they can get a first version that actually like, you know, is starting to point them in that direction in a few minutes. Yeah, I have like a million questions. Yeah, I know. I have so many like technical

I have a lot of technical questions like how's it how does it work? All all kinds of things. But but keep keep going and then you know when you're when you're ready for our questions we're going to start throwing them at you. I just want to show you just a quick

like edit in in the editor and and so I've our um we're doing game jams every week uh starting this week. And so um our game jam theme this week is one v one. We want to see games where two players can play on the same device. Um and so this is an example of a game I

started for that game jam. Um, it's kind of like Rocket League. You got this player that moves around with arrow keys. I'm moving him around right now. And then I've got this other player that moves around with the WD. And I've just

I've just added a football um to the game. But I think what I'll do now is I'll probably um let's make the cars bigger and the football bigger. And for those Americans watching, also known as a soccer ball. Soccer ball. Yeah. make the cars bigger

and the football a bit bigger too. This will go off and do some edits. You can see it's not placed the goalposts quite correctly either. So, I'm going to probably fix

those goals as well. But, um when it goes off and starts coding, you'll be able to see the the stream of code. So, our 3D games are made in a framework called 3JS and the 2D 2D ones are made in something called PhaserJS. These are

all JavaScript web games. I've used both of them. Oh, really?

Awesome. I was going to that was literally you answered my question before I got to it was like what are these built on top of? Because like I imagine there's a framework underneath. Yes. Yes. So, they're a combination. What one thing we noticed recently is

like um when you start trying to do UI elements like menus and scoreboards in phaserjs uh it ends up being pretty bad for that. And so we have a template has both react and phaser in it and the AI kind of chooses to use react more for the UI components and phaser for the game mechanics for 2D that is. So yeah, you can see the cars are a bit bigger now.

And we got this car moving around as well. Yeah. So, and then if I wanted to go and like make an edit myself, if I wanted to, let's say, I don't know, I mean, this is only made up of three things at the moment. But let's say I wanted to turn the friction up to 20.

And that's kind of annoying to do with the AI. I can just go in and change the friction and then save the file and see that the the friction on the cause is increased. Dude, this is so sick. Just want to let you know this is dope, dude. Like I love this. Everything about this is I love

this. This is so dope. Yeah, because I tried making games with Lovable and I and this kind of goes into my my questions like I didn't know anything about game design at all when I tried to make games with Lovable and it obviously it chose 3JS for me. It started writing stuff but then I I had to learn like I didn't even know

anything about game development. First I was like, "Oh, there's physics." You have to worry about the sprites or like the characters, the rules, like the mechanics. Like, and if you don't know all those things, like you can't really

prompt it that well. What are y'all doing differently? Are you baking that stuff into the platform? So, like, no

one really has to care about that or how does that work? I think it's hard to like bake in everything that there is to know about game development. We do do some prompting to try and make the models like more aware of certain like um like rules that are very common across all games and to avoid common errors. But when you want the tool to be very generic, there's not that deep you can

go on the prompting, right? Um and so the other thing we do is we have a guide and we're working on that guide all the time to try and teach people some basics of game development. um really when you don't need to know how to implement every single thing but you do you just need to understand the concepts it's kind of easier to learn. So we have our

guide here and we've we've started talking about you know kind of for example how do you make an animation like this I think it's it's the main concept you need to be able to understand in a 2D game is that to make an animation you don't need an MP4 file. uh you usually use a sprite sheet which is just a series of images um kind of

like an old school film and the the the engine itself or the game framework interpolates those into a into a video. Um and so this very short article just explains that concept and we're turning up more articles like that. So I think I think I know this answer but I'm going to ask it anyway. So who's your who do you think your ideal persona is? who what is that person, you know,

what's that person trying to accomplish? What, you know, tell me a little bit about your users. I think when we started out, we had this hunch that it's all going to be kids. And I think I've definitely been

surprised by the the initial set of users we've we've got. A lot of our users are older people. Um people who are not really coders, but I just kind of like games and built games when they were really young. And so we have a lot

of people in their 40s using this product um to just build all build they usually have a set of game ideas on their shelf and they're building them out now with our tool. Um I think in the longer term there'll be all different kinds of users. There will be kids there will be people who are kind of developers but more maybe artists who are using this. Um, and you've also got

people who want to make kind of mobile games and other kind of games to earn an income from it. So that's another set of users. Yeah, I I asked because I was a so in a past life or you know years and years ago I was I was CTO of a company that basically built educational software for kids. I guess it was more curriculum basically sold curriculum to middle

schools and high schools. And so the the number one, you know, curriculum that this company sold was literally game design. And that was the whole point was just getting game design curriculum into thousands of high schools and uh middle schools, you know, kind of through mostly throughout the US, but you know, even in in a few like Canada and other other countries as well. So, and now now

my brain is working. I'm like, I should probably just introduce you to the the person that runs that company because maybe there's some, you know, it's like they were using Game Maker and then Construct or Construct. I still don't know how you how you pronounce it.

Construct three. And so I just wonder if there's like maybe there's a better way, a newer way to introduce people to game design that gets them actually writing code because one of the limitations of those platforms as well is, you know, you learn this kind of new paradigm and you're basically learning coding, but it's kind of a weird language, right,

with those tools like Scratch and and all those is you're learning this block-based programming language, which is good because it's like critical thinking, but it's also just like a layer on top of what you eventually want people to get to anyways. And so maybe the layer doesn't need the introductory layer doesn't need to be this block

based weird programming language. Maybe maybe the introductory layer is just English communication skills, but then you still get to the same spot and maybe you can get there a lot faster. That would be my my thought. I don't know if

that's true or not, but if it was true, that would be a pretty great way to potentially teach people. If platforms like yours take off with I guess the people who are learning or interested it's very akin to like you know when we're all learning front-end development you hit command S and then the page

refreshes right you have this really short feedback loop to then like oh I turned it blue I can turn it red I can do all these different things and if you're curious about the code you can go look at it and play with it nowadays like using the lovables or these builders there's like a lot of people don't want to look at the code anymore I

think a lot of that's interesting when you're learning something because it's just JavaScript at the end of the day. Plus, you have like the code viewer. Like I could see a lot of people like having that short feedback loop and just getting passionate or like they like engineering all of a sudden, right? Which is super cool. Like yeah, games are a really great way to like intro

people. It's just kind of like front end, right? You have that visual thing that like keeps you going. It's kind of

addicting. It's really fun. You know, you get this nice feedback. Yeah. Yeah. I think like in a similar

way to webdev, right? Um when there's a new technology that comes out, people can be really hardcore about um oh, you need to be able to understand every single concept from its most fundamental to be a true web developer or a true game developer in this case. And I think at least what I've personally found is

uh when I'm learning something um if I'm excited about it, I can learn much faster. And so if I can make progress quickly, I end up learning a lot faster. And so with Waffle, uh, every place where we've removed friction for people, I think we found like it excites people more that they can make more and more

progress on their game and and that's what actually motivates you to keep learning. Yeah, I think that early and constant feedback is is important when you're learning something new. It's like you want it to be challenging just enough so you feel like you're making progress, but if you can't if you go too long

without seeing progress, most people give up, right? And and so I do think that there's this nice middle ground where maybe you can get far enough fast enough where you kind of get addicted and then you're like, well, I'll figure out this painful thing that maybe doesn't quite work right and I have to write a little code

or learn a little bit about code. And then if you can provide the right path for them to learn that, I think that's, you know, that's potentially a great a great avenue to go down. I also do think, you know, and I I'm sure you're you have a high conviction on this, but I my question is, is there going to be

one app that rules them all? You know, is Lovable or Replet or Bolt, are they going to become so good that they can just do everything and maybe they send you down different paths for games or apps or whatever? Or is there a place where actually you can get better results if you have a team that's just

hyperfocused on one specific niche that wants to make the best experience possible for someone who wants to just build a game because the person that wants to just build a game probably is doesn't want to just, you know, go out and build, you know, a normal app or a website, right? They probably want to build a game. They want to learn how to build games. So that that's my question.

I guess what do you think about that? Because obviously I think you have a you have a viewpoint on on that. Yeah, there's a couple of thoughts I have on that. Um, for one is like there's features in our app which are so

specific to game development that it doesn't really make sense for more generic players like lovable or cursor to build them in. So if you look at for example, I'll show you right now um the the assets tab, right? So this is you can see how Americanized the AI is because I asked it for a football and it

gave me an American football. But we we have an asset tab where you can generate any kind of asset in different in different categories. So cartoons or pixel art or whatever. And we also have these huge libraries of assets that you

can just drag and drop into your game. Uh the same with sounds and the same with fonts. And so these are like things which are very specific friction points for people who are building games. And it comes back to that thing of trying to make it as easy as possible to make

progress early on in your project. Um uh and that's where like we try to focus a lot. Another thing I think is that you have different kinds of users who are at different levels of proficiency, right? So we have users who have built small

like users who've built small indie games before tend to actually already be familiar with things like cursor and things like that and they have the things that they like about it and the things that they don't like about it. Um, in the same way that professional developers use cursor and vibe coders use more like lovable type products, I

think we will see in the game development space basically the cursor equivalent of Unity and so on. And there are already like multiple products out like Nitro that help more professional game developers. Um, but that's not really helpful for people that just want to get started and don't have coding experience.

What models do y'all what models write the best games? Dude, you read my mind. That's literally my next question is like I want to know the models like and if you have opinions on what are better for if you use one model across the board, if you have differing opinions on different things and then like you know like how how did

you like what's the rough you know you don't have to go into all the details what's like the rough stack of how you orchestrated this thing? Yeah, I mean we've tried out a bunch of models for different things and um we started off with I think Claude Sonnet 4 or maybe we had something before that but but Cloud

Sonic 4 we used for a decent chunk of time. Uh that was pretty good. We experimented at the time with Gemini and we found despite like Gemini uh 2.5 Pro ranking higher in a lot of coding

benchmarks, it performed much worse for games. Um, and then we've recently switched over to GBT5 and that's been really good for us so far. Um, and we're now exploring more of like an agent mode. Um, and there we're looking at all

the different kind of coding agents that are out there like cursor CLI, uh, Claude Code, Codeex, and so on. Awesome. And how did you what what other tools are you using under the hood to wire it up? Obviously, He's a master. This is great. Yeah. But anything else

that you that you can share like just to get give people a glimpse and like how did you think about building some of this? Yeah. So there's there's um with these kind of tools there's like a couple of different um products that end up being used. So you normally have your agent orchestration for which yeah we use Mastra. You also have this interesting

problem uh around applying code edits. So products like Lovable, Bold, Waffle, they need to go and make very precise edits in files. And when you have um a very large model trying to make that edit, it's one like quite expensive and slow um but also sometimes uh less accurate to have the model rewrite the

entire file. And so instead what you have is um a large model like GPT5 will go and make a code snippet and then you have a fast apply model which is like a tiny AI model trained specifically to do the do the edit and so we're using um a fast apply model as well. We use more for that. It's another YC company and we use uh another YC company called relays for

reranking code files. Um and that comes with like how we manage our context. It's very sick. And yeah, Morph had a coding model that just came out like months ago now. But that's dope.

Is it fast when it does that when you use a fast apply model? Like is it really like is it just like lightning fast and it just it only knows diffs and like applying them? I mean, yeah, you can try it on the playgrounds for both morph and relays. They're super fast. Um, and

yeah, like we haven't done latency tests on how fast they are on our product, but like from the playground, they're super fast. We're doing like something similar, not lovable or anything like that, but we're building like we have these things called master templates and we're currently adding like an agentic chat

that allows you to take the templates and merge the code into your codebase. So then you could like you know you see some sample code that you want to use for yourself and in that experience the like actually merging code with a big model is so slow. It's faster to rewrite the whole file like you mentioned too. We we saw that too, but we should probably use a fast reply model as well.

Yeah, you definitely should look into it. I don't know how well it would work for the use case or that that you're working with, but for us it's worked really well. Probably will work. We'll try it out. I'll let you know.

Yeah. Yeah. Now, now we have we have our takeaway is homework, right? More more for us to do and and play around with.

But yeah, anything else that you want to talk about today, Manuj? I I think I just want to give a quick shout out for our game jam. Um, for anyone who's tuning in right now or later on in the um and watching the video afterwards, like I showed um it's really like Waffle's super beginner friendly and we have a really friendly

community on Discord who are super happy to play test your games and give you feedback. We have a $100 prize every week and there's still five days left on this week's theme, but we'll have one every week. So, you can just go on waffle.ai and join our game jam if you

like. All right. So, anyone, if you've ever wanted to build a game, go to waffle.ai

and just, you know, build something. You probably do it. I imagine you can build something somewhat interesting or cool in, you know, 20, 30 minutes, right? Yeah. And we have a we have a free tier, so you you know, you don't need to put

in your credit card details to try it out. Awesome. Excellent. Very excited about what y'all are

working on. It's like I said, super dope. Yeah. Yeah. The the kid in me is is excited and I just want to find time to

to play around with it. Awesome. Yeah. It'll be it'll be great to hear what you guys think about it, too. And thanks for having me on the

stream, guys. Yeah. Bye.

I got to win 100 bucks, dude. Let's do this, dude. Dude, if so, next we should someday do like a we got 15 minutes on competing live stream. See whose game is better.

We should do that. That'd be really sick. Or or get the team like if do like a when we get the team in persons like we take an hour and everyone just has to see who can build the sickest game or or use maybe we use like some people can build games, some people can build apps. I don't know. Fun friendly competition. And their

product looks super clean and they're using us which makes me feel very good. Yeah. Yeah. Yeah. More people should go

try it. All right. Um Yeah. But we have another guest. You know, the show must continue

on. So, this was very interesting to me because, you know, if you remember, if you've been watching, you know, our show for a while or been following what we've been doing at MRA for a bit, you know that we released what I like to think is the first ever MCP as a course, which essentially was the the ability for you

to use an MCP server to learn about MRA. So you can use it in, you know, cloud desktop or cursor or any of your coding agents that support MCP and it would literally teach you how to write master code and that was the whole idea was it was kind of the first ever course as an MCP server and you know we talked to uh Richard from Napa AI and he was working

on some stuff in this space and then I had the funny enough the same week I talked to Akquil and Akil's also been working on some stuff in the same space. So I want to bring Akil on and talk a little bit about his background, what what they're doing and yeah, learn a little bit about it. Hello. Welcome. Yes. Welcome to the show, Akil.

Thank you. Thank you so much for having me, Shane. I'm excited to show you everyone what I've been working on.

Yeah. So maybe before we get started, can you tell just a little bit of information about you and kind of what you're doing kind of at CMU? Yeah. Yeah. So, um, like you said, my name is Akquil. Uh, I'm a junior studying

artificial intelligence at business at Carnegie Melon right now. And I have taken a lot of human computer interaction classes, which I think has kind of geared my thinking towards like specifically when it comes to agents, like how are humans actually achieving their end goals when it comes to working

with these agents from like a conversational perspective? because at the end of the day, that's what we're doing is we're having conversations with agents to reach some sort of end goal. And over the summer, I spent a lot of time uh working with my little siblings trying to get them more active when it comes to vibe coding or getting more into um using programming tools. And

myself, I was just tinkering and I came across the master course and I thought it was such a cool concept and such a cool way to basically like get someone from any skill set and be basically be able to translate that into achieving the end result. With the Mashra course, it was learning the Mashra framework,

but I was like this is a really interesting way to just teach programming in general. I wonder if you can kind of extract this concept and apply it to other things. And that's basically how uh what I'm building now basics came about to about. So basics is

a course that from inspired from master um is a platform that supports MCP taught courses. So you can come on create your course and then actually host and put that platform out on an MCP server that will then run inside cursor. So instead of with the MRA course it's the MRA course it's specifically that

comes from your MCP server. This MCP server kind of specializes in just offering courses in general, whether that's coming from API documentation or completed projects or basically like zero to one path guides for people who've never been um programming before and they want to go from a very general understanding of things to a more um experienced and complex understanding of things to be able to build their own

tools. Yeah. So it sounds like the goal is basically, you know, take the some of the ideas that we we kind of came up with and turn it into an actual platform where others can contribute courses that can then be, you know, shared and and taken by individuals. Is that is that right?

Yeah. Yeah. I think that's exactly it. And I think that's even what Masha framework even enables is like your tool

enables for people to come up with even more innovations and inventions in the future. which is why I'm like such a fan and I was like so inspired by the MCB course that you guys offered and I plan on expanding beyond just programming and getting into different interfaces. Um we're currently working on like a desktop overlay interface that's able to

like understand your screen and in theory it should be able to almost teach you anything if it can understand what you're doing, how you're doing it and give you live feedback. And right now the very the beta version of it is using Mashra. So I'm excited to continue to build with MSRA and be inspired by what Musra is doing to uh extend the future

of MCP powered courses. Yeah, it's really cool. Like what kind of courses do you have right now other than Yeah. What are the what are the first courses you built? Yeah, so the first course I built was uh

Mashra obviously I basically took a look at what you guys did and see saw if I could like extend it. So I added a few more things. I think I added a rag course and then an advanced workflow course basically straight from your documentation. Uh and then I kind of expanded into other things I was tinkering with. I did voice agents. Uh I

looked at some other frameworks like live kit and worked with GBT realtime to develop some courses on that as well. And then I wanted to kind of create some courses that were more projectoriented. So you join the course with the goal of reaching some sort of end product. And the first course I basically tried to

think for a lot of people getting into programming, what might be an exciting first course to build. And the thing that came to mind was building a portfolio website. Uh so right now there's basically a course that will walk you through the basics of HTML, CSS, and JavaScript. And it'll get you

to the end point of building your own portfolio website that you can not only like show off to, let's say, a job or anyone else wants to see your projects, but you can also explain like how you built it. And for like a non-technical person who finds uh programming very intimidating, basically learning programming through a conversation I

think is probably one of the lowest barrier to entry ways of learning something for that the longest time seem to have such a high barrier to entry in order to like actually master. Yeah. Yeah. Absolutely. I mean, I think that, you know, writing code in general is

intimidating to someone because it it looks like it's a different language, right? It's just like some it's like me trying to learn Chinese, right? And I put myself in those shoes anytime anytime I try to teach someone any kind of, you know, coding or programming. And

so, if you can lower that barrier, you can, you know, kind of ease people into it through hands-on learning. And I do think, you know, most people and I can't speak for everyone, but most people hands-on learning is the best way to actually learn. So rather than reading it in a book or trying to, you know, you

know, read the theory or listen to a watch a video on the theory of it, like actually doing it yourself is the best way. And you know, one one way that the way that most people do is they read a a guide, right? And they try to follow along or they follow along with a video. And that's hands-on. There's a good element of learning to that. But I do

think if you could actually it can be more engaging if you can ask questions along the way. Yeah. You can actually try to when you're not sure, you can get immediate feedback rather than waiting until you know you could you know if it's in the case of like an instructor-led course waiting till you can talk to the professor the next week or you know the teacher the next day. So I do think that instant

ability to get feedback is is really important and I think that's one of the cool things that this unlocks. Yeah. Yeah. For sure. I think a lot of the AI tutoring platforms out there

right now kind of add on to existing coursework. So you kind of have your course or you have your PDF of your slide deck and the AI tutor basically takes that and teaches you the information from there. Whereas this basically is from basically a bottomup approach of now I actually want the agent to have the course in its backlog. So you don't actually see the course information, but the agent is trained on

the actual structured course that's like vetted, ensuring that you actually will learn like multiple people will go through the course similar to our course at CMU. The reason they're so successful is because like they've basically generated a formula for how to get people from A to B in this certain topic. And that's essentially what

basics enables is you're taking these structured vetted courses and instead of just adding to an existing um PDF or just adding a tutor to that, you're basically allowing um people to go through a structured vetted course that's taught in a similar way as if I had like a one-on-one tutor and we were basically talking about, okay, how can I

get you from point A to B and how can we have a conversation based off of your level of experience, based off what you understand to really ensure that you're getting to the end point instead of just seeing okay this is where you are here's what I think could help you get even better if that makes sense.

So like that totally makes sense like you know how like in first year CS is like the weer courses that's pretty much wants you to drop out essentially like do you see this form of learning a lot more effective for maybe people won't drop out because they have a better way of learning um have you and also have you tried to write any courses of your own course material that you're taking in school right now? Yeah. Yeah.

Yeah. Uh definitely right. I think this is a natural progression when it comes to programming. I think in the earlier days of programming when most things were in assembly um when there was a

tradition more compiled language, it enabled more people to get into programming and the same thing with the development of Python. I think that courses or a platform like this basically enables more people to get into such a powerful tool which is programming and it makes it less intimidating because instead of of

seeing this intimidating jargon, you're seeing something you're familiar with which is English and you're learning through conversations instead of intimidation which a lot of these typical introlevel programming courses tend to do. uh which I think there's some merit to it is you have the people who actually want to apply real critical thinking to solving problems and a lot

of those classes are more problems solving classes necessarily than programming classes but I think there are ways to do that through conversations and that's actually what we're experimenting with at CMU um we are now talking to a couple of professors to work with some of our intro level programming classes to

develop a course that uses as a supplemental tool and basically in addition to homework. If anyone's interested in basically refining um some of the course information that's being taught to see if it actually has potential uh as a real way and or an al alternative way completely to teaching programming and getting people thinking critically in a way that is more comfortable to the majority of masses

who learn typically through some sort of conversation. Yeah. Because also a lot of the course material for intro programming is already in the model's general intelligence, right? So like the course

material just needs to be surfaced a certain way. But then if you're like in upper division classes, you're getting very specific maybe like the professor has, you know, certain nuances to how they do things. That can all be encoded in the course material and then serve to the user. So this comes back to what we were talking with these guys from NAPA AI.

Like you're only as good as your course material. So like yeah, how do you kind of solve that problem? Yeah, I think that's where the privilege of going to Carnegie Melon kind of comes in is like one of the top computer science programs. There's so many researchers, professors, students who are constantly doing things like that.

Um, and I've basically been taking full advantage of that. um reaching out to as many professors, students, anyone who grabs gravitates to this kind of format of teaching in a conversation way uh using agents. uh basically anyone who's found that interesting, I've been trying to collaborate with them to actually

build real course um content and infrastructure for people to be able to come to this platform and say like hey like this is not only a cool way of learning but like the courses that are actually on this platform genuinely taught me how to achieve my goal rather than let's say like if I were to just go to the model itself and say like hey teach me how to code like it could go a

million different places and it may get you to one end goal, but like that may not be a very like structured and vetted out end goal. Whereas like these universities, that's all they do. They're in the business of creating structured courses to get people from point A to B. And I think it's like the

perfect people to collaborate when it comes to creating courses for a platform like this. Yeah, absolutely. So, you probably knew I was going to ask, but can you show it? What can you show us? Can we can we see

Yeah. Can Can we see how it works? Can we can the audience you know because obviously we have people watching live. So if you

are watching this live you know if you have questions please ask away and otherwise if you're watching this later I think people would love to see what's there today and then what are your plans for the future of like how are you going to continue to expand it add more functionality. Yeah. Yeah I'd love to. So I'm just

going to go and uh share my screen. Yeah. So this is essentially what the first thing that most people will see. So this is the

master course that got developed out and essentially here is where you are able to actually track your progress through the course. So instead of a typical course platform like let's say Udemy or Canvas for completing courses where you actually see the course content here the main goal is to see like okay how far

along am I in this course because most of the teaching is going to be happening over here in cursor. Um so what I have is basically like a conversation trail of me using the uh basics MCP agent just for the sake of time. Um, but you can see how it works is you essentially come here and put in your email and that basically then uh informs the agent of

your account, your preferences, which I'll go into a little bit later, and some of the courses that you're enrolled in. Um, so can you go ahead and give us a a click of zoom? Can you control plus command plus just give us a little little more zoom? There we go. Is that Yeah, way better. Okay, cool. Um so we started the mashra

course and it gives us a little bit of information about it and then it starts giving out the lesson content. Um so before we get into any coding it basically explains some of the basics of what's going on. So it's talking about like what is an agent and that's this course content actually comes straight from the master course. It's what I

typically use um to develop the system that is now the basics MCP. Um, so basically it goes over kind of the core features of like what is an agent? What does an agent look like in Maestra? And

one of the coolest things is I wanted it to really feel like a conversation. So after that lesson content has been presented. So this is this lesson content is static. That'll always be the same when you create the course. It's

kind of like having a page from the textbook. But similar to in a class like when a teacher is necessarily using a page from the textbook like the conversations that follow up from that page is not always the same. And typically you can you get to gauge like how students are engaging with whatever you just taught by asking questions. So

that's why I really had a focus on asking these follow-up questions to kind of gauge how well is the user really understanding the lesson that's being taught. So here it's asking like why do agents make autonomous decisions instead of flowing um following fixed predictable outcomes or instructions and

I basically gave the answer of so the agent has more agency to do things and it may not be uh it can do things that are a little bit less predictable and basically it kind of affirmed that I actually understood the topic here of why would we use agents versus just like very structured code outputs and it

basically basically built on top of my understanding by one having me engage as a user with the actual course content but then affirming that I actually understood things before I moved on to the next step. So you see a kind of similar um you see a similar kind of flow of things but I was like this is really

good for beginners but for people who probably are using the MRA course right now uh typically they probably want to use it as this next generation of documentation basically being able to like understand how to do things and they may not want to be like babyed or like tutored in a way that maybe a beginner might want to. So then I

created an intermediate mode that essentially uh does the same thing but it kind of takes a step back on the questions where it'll give you the lesson content and it'll if let's say you don't actually want to engage with the lessons and you're saying I don't know you can basically say like stop asking questions and that entire like

question part of things goes away if you feel like it's not actually beneficial to the learning that you're doing. Um, so those are some of the core kind of concepts and how I've kind of deviated a little bit from what the initial foundation of the MSRA MCP course had. And I tried to kind of hone in a little

bit more of the learning psychology aspect of things with really focusing on conversational based learning. So you still have the parts of the course where it'll go in and implement things. It'll make sure things are set up. So it's

checking our node version here. Um, it's basically ensuring that our project is set up properly here. Um, but one of my favorite parts of the course and that I'm still experimenting with uh to this day is so with this lesson, understanding system prompts, you basically are taught like here are the key aspects of what makes a really good

system prompt. And here you're basically asked to apply that in a way that's not even a question. It's a challenge. It's

basically like, okay, I just taught you about system prompts. Now, I want you to basically tell me how you want me to create a system prompt based off of the information that I just taught you. So, what this does is it extends on the learning aspect of asking questions even more by challenging the user to then get

handson when it comes to learning, but also like creating the end goal of building our agent in this case. Um, so basically I said, okay, I just want to make an agent for finance. Is this good? And basically what it was able to do is

it took my prompt and it said, "Hey, this was a good start, but here's what you probably want to do to make your agent a little bit more effective. Be specific. Use clear capabilities, behavioral guidelines, and constraints."

And essentially what this is doing is it's taking your understanding of what makes a good system prompt because from this response, basically the agent is saying, "Okay, I just taught the user about what makes a good system prompt." Maybe they didn't take away everything I just taught them about it. Let me apply it to basically what they the answer

they give and what they were trying to do. So here it gives a better example of what a good system prompt is and it explains why it's better. So now when you're going to build a system prompt and you are thinking about that like lesson about what makes a good system prompt uh you're not only just thinking about the content that was like given to

you but you're thinking about the example that you worked with the agent to actually get to the end goal of. So basically I just copied that exact prompt in and it was like net wonderful. So before any code was generated it basically said wonderful now that I know you actually understand what we're doing let me go and actually like create this

agent for you. So it gives a little outline of what it's going to do and then basically just like the master course it goes in and actually does it. Um and yeah that's kind of a workflow of how the my highlevel vision for the basics course what my high level um vision for the basics course is is to really take this step back from just

giving users outputs but actually like having a conversation to one gauge understanding but two really really make sure that the users not only understanding what they're doing but they're able to like build on top of this foundation of any information they're being presented. Um so right now um right now it works really well um with programming and it's works really

well within cursor or any other tools that MCPS are supported in. But I think there's an argument that can be made that conversational based learning especially in a world where it's really easy to just generate answers might be kind of the future of education that we see in general. I think having a conversation with someone is really how

you can evaluate how well they understand what you're trying to assess them with because they it's hard in the moment to basically like generate an answer for that. And if I'm like having a conversation with you and you're asking me follow-up questions in places where you think I have gaps, it basically allows you to both vet and actually teach me in a way that's a

little bit more um enforcing than let's say you showing me lecture content and then quizzing me and then showing me lecture content and quizzing me because instead of optimizing for getting the quiz answers right, now I'm actually I have to optimize for understanding because I'm going to be evaluated

through a conversation. So building out um kind of an infrastructure to do that with different domains beyond programming and adding more voice capabilities. So it the conversation is not just you reading the chat and it performing actions but it's you generally talking to an agent that's basically trying to achieve this goal of

teaching and evaluating how well you understand through conversations. Well dude, I'll tell you I'm very impressed. It was very cool. Yeah. And I do like the fact and and I

don't think you showed this in this demo, but you do you can basically install one MCP server and switch between these different courses, right? So as you add more courses, you can basically install it once and then select your course and move between different courses, track your progress across all of them. And so I do think that's just a really

cool way, you know, obviously for programming works great. I think it'll work really well for other other things in the future. And I do believe that, you know, that they do say, you know, and I've found this to be true in my own experience is that when you can teach someone something, when you can really

explain something, you can converse with someone and I can teach you, then I it kind of cementss the knowledge in my brain a lot more, right? And I do think that building it into more of a conversation and asking open-ended questions, it's similar to like asking them to un, you know, basically gauge their understanding and explain it in an

open-ended way in a similar, it's almost like you're asking the the user to teach the agent what it already taught you. It's like retach me what I already taught you. If you could do that, ideally, you'll have locked in more of the knowledge than maybe you would have otherwise. Yeah. for the uh for the portions that may be non-programming, it still would

be like MCP, but you'd be like an open AI or claude desktop or like those types of things where yeah, it's the still the same concept but maybe there's more diagrams or I don't know you just think about biology which is through that like anatomy or something. It would very much be a conversation about what mitochondria is,

how does it look, what's ATP, like those are the types of things that and if people are using open AI or CHBT or the surface areas problem to solve because those guys support MCPS then you're good to go. And I do think I do think I use so I have a interest in you know like Greek and Roman history right you know like the the times of and antiquity right was like Roman generals and on commanders

and all that and but I don't have any structured course right it's more like I just ask questions sometimes if I you know hear some interesting tidbit of information and I kind of go down this you know train and then I leave right and there's no like this is a basic understanding so I do think if you could have, you know, Chad GBT or Gemini or,

you know, Claude and you could say, here's the actual structure of a course that teaches this stuff that's probably already in its knowledge base, but it structures in a way where I I know like I'm getting like a good breath of knowledge and then of course that's going to teach me more where I know more specific questions I can ask because I

can kind of go down, you know, the rabbit hole of things that interest me more and ask more questions and the stuff that interests me less I'll just kind of breeze over it and maybe care less about. So I do think that a as a way to learn I've been learning a lot just having conversations with you mostly chat GBT is my my daily driver but I use cloud desktop and Gemini as

well to kind of mix it in here and there but I do think that if you could provide a more structured way and something that you know gamify it a bit so it tracks my progress and be like did you actually like complete it? Okay, well I could I got halfway through. I should probably just complete this thing. So I actually like round out my knowledge. And I think people, you know, people are

uh do crave that validation that they actually completed something. And so I think there's something about giving someone the ability to to track their progress and then actually know that okay, I did complete something that was like gives me that breath of knowledge of whatever the topic is. I have a feature request or something.

Have you have you all gone through like Wikipedia Wikipedia rabbit holes where you like you start at one thing and they go it's kind of a course in a way because you're the one self-guiding each Wikipedia page that you're doing. So here's my feature request like if you're in OpenAI and I think you should have like a suggestion tool like suggested course tool because if I ask a

question about the Roman Empire that tool should be like hey I hear you looking about the Roman Empire. We actually have a course on that. Would you like to take it? Would I like to take it now? Oh sure

I'll take it. then you know just do it, learn something and then move on. You know that's what happens to me in Wikipedia. I be like go there like Marcus Aurelius who the is that guy go there it's like oh what is stoicism? Okay let's go there and it's literally

like kind of pre-built the course for me through my own like um I don't know curiosity. Yeah. But I think a lot of people would just be like if you saw like the little tool thing that says hey we have a course on this. Do you want to take it right now?

You say yeah sure. Boom. And you're in it. Yeah for sure. I will definitely implement and then quote you on that

when I uh get that implemented. What do you think? Yeah. Yeah. All right, Akil. Well, appreciate

you uh coming on. So, how do people find out and follow along with your progress? I know you're still, you know, it's still relatively early days, but yeah, you are investing a lot of time and effort into building this. So, I'm curious how how does someone track your

progress or or get updates? Yeah. Um, so you can go to basics spelled normally uu, the letter u.com.

Uh, and sign up for our beta. Uh, that should link to a form where you can basically sign up and say, "Hey, I'm interested in trying this and following along with the journey." Um, and yeah, that's where most of the updates will come from for now. Um, and any future

developments, uh, they'll basically be announcements from that website or you can, uh, follow me on LinkedIn as well. I'll probably, uh, announce a lot of updates on there. And that's my name, Aquil Smith on LinkedIn. The CMU

student. Are you um, are you doing any internships this summer? Um, no. Last summer, I worked on this. As of right now, my plan is to continue

to work on this. Uh but yeah, I haven't I haven't done any internships. Uh yeah, not yet.

Hit us up later. Just leave it at that. Definitely will. All right. Yeah. Well, you know,

basicsu.com, right? That's that's the URL. So, basics the letterwu.com. Go check it out if you

want if you're interested in this kind of maybe the future of learning. At least, you know, I I seem to think it is. I think others are starting others are starting to come along and believe it as well. So, Check it out.

And Aquil, we'll definitely have you back on again. Give us an update in three or six months and see how things are going. Would love to be back. Thanks for having me. Yeah, great work, dude. I'm very impressed. Thank you. It was awesome.

Yeah. We'll see you, Ke. All right. Bye-bye. See you, dude.

All right, dude. Dude, that was sick. That was sick. That the MCP education, I think it's

it's going to blow up. We're in a good way. Yeah, we were the first, but obviously that's not our main our main squeeze, right? Yeah. People that inspired. I mean, that was always kind of the goal

when we talked about I was like, hey, we know we're not going to do this the best, but if we can get it out there and and maybe inspire people to build more things around it. And we have had chats internally around building like a an example or a template course MCP course platform like a like a Udemy or or you know, all those other education platforms. and we never did because it

was just you know not it's not the main business right that's not what we want to do every day but we are interested in it and we would love to see this idea take off so it is interesting I think as MCP becomes more and more common we're going to see more and more people use it for interesting things education is just one I think there's probably

more things that it'll be used for you all new like tech has use cases people didn't imagine right and so I think that's maybe the interesting thing about Yeah. Looks like he has good distribution, too, being in school and training with the past year. So, I would expect to see it grow. Yeah. Yeah, I do think so. And I think there's going to be more, you know, we

saw it last week, we saw it this week. I think we'll be continue to see more and more of this. Yeah.

All right, dude. Should we do some news? That's the news. But before we do, if

you're watching this, please uh like us, subscribe to us on YouTube. It's Monster-AI on YouTube. If you're on Spotify, give us a a like and a subscribe and, you know, maybe leave us a fivestar review. And yeah, we'll leave it at that. We got

some five stars only. Exactly. Uh, all right. We we do have not a huge news

day, which, you know, is good because you have a flight to catch and all. Anyways, there's still some hitters on there. some big big big moves would be happening. Yeah, for sure.

All right, so first one, this one's from Anthropic and Anthropic. Yeah, Enthropic just casually announces in a tweet, we've raised 13 billion at 183 billion post money valuation. It help us expand our capacity, improve model capabilities, and deepen our safety research and cash out people. Yeah. Yeah. Exactly. And you know, some early

employees are, you know, you're buying some some nice property on some island somewhere. But yeah, money money money uh money's flowing in. So, you know, they they talk about their run rate and they've passed five billion eight months later. Fastest

growing tech companies in history. Uh and then, you know, grateful and they can read more at their website. But yeah, they just raised a ton of money.

I wonder how much of it is carved out for employees and then how much is it for future infrastructure costs that are obviously going up. Um dude, 13bs. Imagine the wire transfer hitting the account. It's like zeros. It's like you got a you got to get a

wide screen to see all the zeros that are coming. Yeah. like your your MacBook, you know, laptop screen isn't wide enough to see all the zeros. No, most most of the banking websites probably don't have I'm sure whatever

they're using would, but you know, I imagine my my local bank doesn't usually deal with that many zeros in a wire transfer. But yeah, big news. Uh, you know, I imagine we'll just, you know, not not unexpected, right? philanthropic claude especially for coding is basically still you know

the the top at least in our experience and people still believe that and there's some I think it's carved out definitely a niche with cloud code so more more to come I imagine they're going to lean more into that and obviously they they want to capture other niches as well but they seem to be kind of the the top on coding yeah I could see more MTP things happening I could see a lot they could

do that That's what they want. And don't don't poach our employees yet. Stay away from us. Yeah, please stay away from us. We know as much, you know, not quite as deep a pockets. Please.

Yeah. Uh we are going to continue on. This is a kind of a surprising anthropic open AI collab. So the competitor well sometimes competitors collab.

So they did a pilot. This is on the OpenAI website, but Anthropic and OpenAI did an alignment evaluation basically where they tested each other's models. Basically tried to figure out if they could figure out ways to get it to not do what it's supposed to do. And so Enthropic was trying to essentially, you

know, quote unquote hack or, you know, do some bad things to Open AAI models and OpenAI did the same to Enthropic. you can kind of like it's kind of a long finding so I'm not going to go through all of it. The summary is kind of the part we'll go through and encourage you all to read it if you are curious on their findings.

But they they talked about, you know, and I think this this was common like jailbreaking and hallucination were common across both, you know, like Claude had issues, OpenAI had issues, and it's all stuff that we've kind of known, right? So, it's like instruction hierarchy, jailbreaking, hallucination,

and you know, scheming, but yeah, surprised they work together. Yeah, sometimes, you know, sometimes competitors Yeah. have to work together.

And you know they they do have like some examples and I'm curious on how uh you know I imagine they because they work together they want it to be relatively accurate. I do think they you know they want to show their and it's Claude talks about it or anthropic talks about it all the time. They're val they value safety they value alignment. They value all these things. And so I guess this is

them showing showing that a bit like we're going to put our money where our mouth is and work across you know model companies to make sure that we have you know we're kind of aligning on some of the same standards and the same safety ideals and so it's good to see uh I'm not going to read you know to be honest I'm just going to ask you

know it to summarize and I'm not reading the whole results this is a huge paper but encourage you if you are interested in security and model alignment and all that you will uh probably learn quite a bit from this if we see open AAI do the user of other model companies it's very interesting because they definitely want to start learning the developer experience and that is multimodel as we know from

different people's companies that already here so may it may be like all maybe it's all in service of a bigger a bigger vision you know doing these alignments and stuff yeah potentially And wants it all. Yeah, I mean they do. They do for sure.

And I I do think it's it's interesting. You know, we talked when Man was on earlier, they're using multiple models for waffle, right? I think it's it's pretty common. You find models that work

slightly better at certain things and some of it's based on feel and some of it's based on you real usage and evals and and things like that. And so you're not just, you know, going to use the same model or the same company. So being making sure that there's some standards that kind of span across models, especially like how we think about safety and security and all that stuff

is that's a good thing. We should spend time on that. Well, at least the the the companies that are raising billions of dollars should definitely spend spend time on that. Totally. All right. Uh continuing on with some

Open AI news, GPT real time, uh last week was announced. Basically, it was productionized, right? It was officially launched.

They've they've had the real-time API for a bit, but they're just making it generally available. So, ideally, that means they've fixed most of the major issues. It's kind of ready for production use cases. They obviously are

leading with some bigname case studies as you'd expect with Zillow, T-Mobile, StubHub, Oscar Health, Lemonade. So, a bunch of different use cases across industries. And yeah, they just talk about how GPT realtime voice works and and the quality and and definitely uh they want to encourage more people to use voice

capabilities. They want that part of the industry as well. They want the voice agents. They

want it all. Um the speech to speech models are really sick. You can talk to them. They do tool calling. Uh we have support for GPT real time through our

voice adapters and deostra. The only bad thing is it's very expensive and I think we had the goose on here and we were saying that noi has a a pipeline instead of the real time. So I'm curious I mean if you have money you'll probably just use GPT real time but for those who don't I don't know if it really scales

still. Yeah, I mean I think and I don't know I didn't read enough into this to know if they changed the pricing at all, but it's just for common use cases it it just doesn't the the economics don't work. But you have to imagine that's going to continue to get better over time, right? But there is something

about the flexibility to be able to configure different models that you don't get with GPT realtime. I did think this was interesting, right? New in the real time API, you can use remote MCP servers. So they want you to be able to

tie in real like MCP servers and actually you know use that with real time. We should try the master course through real-time MCP server you know we don't we don't have a remote MCP server yet for the master course you know candidly so it wouldn't quite work today but I am curious like you could just voice talk through you know talk with an

MCP server essentially but it is good to see like you know MCP was a standard created by anthropic so again they're they're uh using the standard which is good there's MCP is getting even more uh more usage across you know now not just with LMS but also with you know real time voice in terms of the pricing that they say

that it's been reduced by 20% so 32 bucks for minimum audio input tokens 64 for audio out seems still price to me but uh you could easily get a million tokens dude that's so easy to do these days yeah yeah I think that you know some of the teams that were building voice agents that we talked to that were really in it

basically said that what the numbers that I remember hearing where there's like basically like twi twice as expensive as it would need to be in order to be for their specific use case which is more of a general voice agent. Now again, everyone's economics are different and so it'll work for some people, especially if they're high value

calls, right? Maybe real-time voice works. If it's, you know, a law firm where every call is potentially, you know, $100,000 contract, you know, case or something, you know, then yeah, maybe you use a voice agent to handle some of that inbound or something. But I do think that it's still a little bit expensive for just the average use

cases. But 20% now, you know, 20% in 6 months, 20% 6 months after that, it starts to become more reasonable. Yeah, economics will eventually catch off soon, but for now, I would just play with it and see. And last, this was more not necessarily

news, not an announcement, but that people on the team, you know, I do often share just things that we're reading and things that we found interesting. And OpenAI recently had I don't actually it's probably not that recent. OpenAI in the past has a documentation or guide on optimizing LLM accuracy. It's fairly long, so we're not going to go through it, but if you

go to platform.openai.com, you can look in their docs, they have a very comprehensive guide on just like how you should basically it's context engineering. I would call this a context engineering guide in a sense like it

just yeah, just talks you through how you can optimize context for the for LLM, specifically open AAI models, but it's pretty general purpose. It would work across all models. talks about prompt engineering, rag, fine-tuning, how you should think about it, when you should, you know, consider one or the

other and how you should like prioritize them. So, I think it's a very good primer if you're interested in learning more about context engineering. You know, we've been talking about that quite a bit. This is a good guide for

you to read and or, you know, again, have have one of your favorite models summarized for you. I thought that was one thing that was interesting here was like the break even they have this like there's a uh they're measuring accuracy and for them the break even accuracy because anytime it's not accurate it costs money to escalate to a human right

so the break even accuracy is 81%. So still that's a B class. So, you know, I think most of new users expect all these things to work 100%. But even OpenAI's thing is 81%

confidence, right? Yeah. How much accuracy is good enough for production? And yeah, I think it's

in some cases I think people just need to be realistic with the capabilities of the model. Just because it's not 100% accurate doesn't mean it's not useful. And I think that is like maybe what you're trying to do is too ambitious for what the models can do today. that's possibly true, but with the right set of guardrails and the right, you know,

expectations, it can obviously do some pretty pretty amazing things. So, yeah, you have to plan for those the failure cases um as much as the success ones. And that's what we kind of tell people about eval, you know, just tracking all the stuff because it's if you're in production, every mistake will be an escalation of an hour spent, a dollar

spent. So yeah. Yeah.

All right. So we'll talk a little bit about XAI. You know, they've feels like they've been a little quiet lately, but I don't know if this is huge news, but they basically have a new, this was announced last week, a new agentic coding model, which is, you know, I think they call it Grock code fast one. Yep. And it's really an agentic coding model

that you know performs coding related tasks but it's supposed to be just again it's kind of a fast model so it's quick I imagine like I I do it looks like they're announcing it with GitHub copilot and windsurf have not tried it it's free for a limited amount of time I don't know if because it's a fast model if it's maybe used for more autocompletes and things like that I

would imagine probably yeah or the fast supply stuff that Man was talking You could probably do merging with diff. Yeah. I wonder if this isn't a result of them getting roasted for not being good at coding um like on Twitter and then focusing in. Yeah. They maybe they had to they had to come back with some kind of response to to handle the the hate that they

receiving. I think it was like Theo who was like arresting them about rock's not the best model compared to plot etc. Yeah. It's like best. It's best on benchmarks, but obviously like not in real life or whatever. Yeah, not in real life. Maybe it doesn't

hold up across all things. Uh which, you know, taking a tangent, I've been seeing, you know, GPT5. So, you know, Manuj said they moved to GPD5, they love it. I've heard a lot of people say they use GPD5, they love it. I use as my daily driver. So, I do like it, right? I

like GPT5, but I I got I I have to if you're listening in chat, please uh please tell me your experience, but I've seen all the things that GPT5 has scored the best on math, right? I've seen that it solved something that had never been solved before. That's great that it can solve something that's never been solved

before. It also can't solve things that have been solved before a thousand times, a hundred thousand, a million times. It sucks at basic math. I ever I

literally every week since GBD5 launch has come out, I've been just talking to it, you know, asking it questions, asking it to do some basic calculations and not even on purpose, but I'm just noticed that that doesn't seem right. Like something about that math doesn't add up in my head. Like logically seems like something's off. So today it was

like, "Oh, sorry. I misplaced a decimal point." That's kind of a big deal. Calculators don't even do it,

dude. That's a big deal if you're off by one decimal point. And the answer was like it's either 23 million or 2.3

million. That's a big deal. And like that's huge. So I don't know where everyone's getting

that GT5 is good at math. It's terrible. I I will tell you it is not good at math. You can't trust it. Now will they

make it better? I hope so. But just give it a damn calculator tool, you know? It's like just give it a tool.

Calculators are solved, you know? Yeah. OM aren't good with numbers, right? So maybe that's it.

Well, it is it. But I'm just like that you should just give it a tool and it will it would have solved that problem easily because it was just basic algebra and there was a you know a meme going around the when it was first launched of just some like basic algebra problems that it was getting wrong and I tested that it was true. I ran into it, you know, last week with some issue. I ran

into it just doing like some basic like equity calculation today and I was like that can't be right. And then I asked and it and then it thought for eight minutes or three and a half minutes or something and said like oh you're right. And I love the status was like apologizing and recalculating. So it's like that was its internal status like I

need to apologize because I got this wrong but I need to think harder and then I got it right you know like and when I pushed it figured it out but I shouldn't have to check its math right that that's I wonder how many people are not checking it at all and just like taking it for what it's worth probably a lot but also have you noticed that it's really slow

it is even if it's not reasonable probably the routing is the routing model part is slow and then actually executing whatever is happening. Yeah, I don't know. People use it, but then you know Manuja thing was pretty fast when we saw it. So, I don't know. Maybe it depends. Yeah. May maybe it's slightly better through over the API versus through, you

know, good in chat GBT. You have a ton of guard rails on it, right? That you have to manage. Maybe there's that adds

lat extra latency that you don't get over the API. But I did hear, you know, someone that we know is building a startup and they had switched to GPD 5, realized it was it slowed things down quite a bit and they went back to 40. So, I don't know. I I I'm I think the results like are pretty good in my

experience. Like overall, I'm not dissatisfied with it, but I do think there are some things it's not good at. Math particularly and seemingly worse than it was before.

Like I didn't notice so many issues. Like I always ask I do a lot of stuff with enough stuff with numbers where I ask it to like calculate things for me and it just doesn't always get it right. Yeah. And if it takes three minutes and then you were waiting the whole time and

it was wrong, it's like a waste of time. Yeah. I could in that case it wasn't even that hard of a calculation. I could have done it myself, you know.

Like I own a calculator, you know, I can do this. Uh anyways, so another announcement, this one was, you know, came out August 28th, one week after launch. 45% of enterprise customers have enabled cursor AI agents in linear. Agents are rapidly becoming

part of product teams of every size and the work being orchestrated in linear. So this is cool in my opinion because I've always this is always the dream right with codec and with these other tools is that it basically just lives in your product you know your your product management system your you wherever you're managing your tasks your issues and you could basically just decide if

you want to assign it to a person assign it to an agent maybe the agent takes the first pass and a person comes in and validates and you know maybe throws it all out or maybe uses as a prototype and if you can allow the PM or someone maybe less technical to say like this is the right interaction now productionize it maybe there's less back and forth

time between you know engineering teams design product especially for larger teams and this is probably why there's such a big uptake of enterprise customers right yeah and kind of like what Dexter or Dex is doing um having the spec and the plans like that could all just be linear tickets and then you'll live and die by

how good your tickets are yeah just write if you write more detailed tickets you're going to get better implementation and the nice thing is product manager should be pretty good at English. That's kind of part of the job. You got to write product specs and all that. So, if you can really well define it, maybe work with uh an engineer that can tell you some of the

technical details of like how they would maybe implement it, you could probably have some you break off some tasks and have some PRs that are reviewable. And again, I doubt I imagine 50% or less are actually mergeable as is, right? But if you can get the prototype out of the way, you can ensure that the

the UX the UI patterns are right, the the interactions are good, maybe you can get further faster. I mean, he does say that 45% have enabled it. Doesn't mean it actually is doing well, but they are.

Yeah. How many how many people Yeah. How many people keep it after? And it's just

like anything else. I don't I don't think it's this miracle thing, but you It's like one of these tools. I think you have to know how you want to use it. If you just turn it on and hope for the best, I don't think people are going to have good results. I think if you as a

team, kind of like what Dex had mentioned in the context engineering, you know, meet up at YC, if if you have a set process, if it's part of your development process, I think you could probably get something out of it and it probably speed your team up. But if you don't, it could become a distraction and it could actually add more work to the engineers because you're making them do

something in a way that they could have just prompted, you know, cursor agent themselves or just wrote the code themselves just as fast. So it it does it is interesting to see, but I I think we still have probably a long ways to go. Yeah, I agree. Plus, at our company, our ticket ticket writing skills are very bad. I don't

know how far. Yeah, we probably need to write a, you know, we probably can't quite turn that on yet and have good results. Those one sentence ticket description that but when you're small like you have shared context, right? People understand what you because you're talking about it. You know, you know what you're

you're meeting frequently enough where you can make those course corrections along the way. But yeah, if you that's probably why it's better for enterprise that again you have to get approval. You have to go through all these processes. So you have to be much more spec driven, you know,

requirements driven ahead of time in order to get anything approved. And maybe at that, you at that point you can actually have it build the entire prototype or a huge chunk of the prototype before you even get in the hands of the engineering team. And so if you can make allow the PM to basically get the rough prototype out, you know, out and ready so there's

something to look at, that could be pretty useful even if the code is basically thrown away. Yeah. I want to meet someone who's actually using it in production or like in their stuff. It' be really cool to see how they like it. Yeah. If anyone's listening is using it

using linear and cursor background agents, hit us up. We want to talk to you. Come on the stream and tell us.

Yeah. Show us a demo. All right. Uh yeah, we've been going strong for hour 40. So, last thing. This

one I was more just, you know, for fun because that, yeah, I en I enjoyed uh enjoyed reading. There's always It's always nice when there's a little drama on on X. This one's good. This one's good. So, Paul from Browserbase, today we're announcing an

unlikely partnership. We believe that agents need reliable, responsible web access. That's why we're partnering with Cloudflare in support of webbot O and signed agents, a new standard to allow good bots to authenticate themselves. And then they go into details. They have a whole post.

Okay, we we'll come back to it. That's uh that's phase one of this setup. Phase two. The next uh next up we hear

uh let me see our lord and savior. Gary Tan our friend. Well, you know Gary Gary's a friend. Cloudflare browserbased access of evil was not in my bingo card for 2025. Legalize AI agents.

And then you know it goes on to say ultimately if a user wants a browser to do an action on their behalf they should be allowed an open internet is exactly that open instead of requiring hall passes from intermediaries. And then Paul responds whoa Gary signed agents is optin to our customers. You don't have to authenticate your agents

but to me it makes sense. Capture solving is a lose-lose game. Instead we seek to represent our customers by brokering access on their behalf. TLDDR I want to help AI browse. And then

as the final, you know, which I thought was also great, Paul posts the next day, good morning from the axis of evil. And it's picture of his entire team, which is funny because like we know Paul, we know Gary. Like we we've had long conversations with both of them. So

it is, you know, it's always I think it's probably healthy discourse to have this back and forth. I am, you know, I guess I'll let you you can add some color. What do you think? and then I'll give you my opinion on on the whole

thing. But you have the back and forth with, you know, obviously Paul from Browserbase, Gary from YC. What's the what's your takeaway? Well, I think if it's optin, it's great. Um, I just think that CloudFlare

themselves granted, you know, with their bot access and I think they default enabled it, too. Uh, they want to restrict agents from coming on sites and stuff. Yeah, I do think I do think Cloudflare came out and said they did not default enable it. It there were some bugs where some customers had it enabled,

but it was not it was not intentional for those few customers and it was now again I can't validate that. I'm just that that is what I heard. So, you know, I'm not a I'm not a Cloudflare defender, but I will uh I will try to be accurate. Yeah. Yeah. To be accurate like that already put a bad face in people's

mouths, right? like from two weeks ago. So then browser base, you know, even Paul says it's a very unlikely partnership, but I think it makes sense and they should partner with other providers too to broker access on behalf of the agent. So you don't have to do captions if you're a trusted source.

Like I do believe that's a good thing. But then, you know, our homies from browser use are trying to do, you know, they're trying to compete with having these built-in capture tools and all these other tools you need. So, you can do whatever the hell you want. So, I think from Gary's perspective, people should do whatever they want and not

have the restrictions. But from Paul's perspective, it's actually might be easier if I work with Cloudflare to just bypass these checks. Uh, you know.

Yeah. So, it's a lot of drama though. Yeah, it's it's a you know it's it's fun drama. I I

agree with with that assessment though. I I do have concerns mostly just around you know Cloudflare's bigger plans. You know I do think that Cloudflare if you go back announced you know the idea of like pricing content for agents right and now they have this like payw wall basically that you can enable. Yeah of course you you users can choose to

enable it. Now you have the I this authentication mechanism where you can authenticate browser agents. I think they're trying to if I were to predict I think Cloudflare has this master plan and whether it'll work or not or whether it's like actually all thought out I don't know they're they're smart people.

I imagine some of this stuff's thought out. I think their master plan is allow content providers to gate access allow authentication to happen and then they collect money on the margins like as as things come through. And it's a perfect business model. They're gonna make so much money if they can become the stripe for agent interactions across the

internet. Yeah. And I think that's probably the crux of Gary's point is that maybe he can see a few steps ahead and could see where this is going and know that yes, this does kind of set the stage for Cloudflare to kind of gate priced access to content which they've already talked about and they've been pushing. And that's where it concerns me a little bit because I don't want, you

know, as much as I am a fan of content creators being able to charge for their content. You know, I just think the open internet is great. I hate sites that ask me to pay money to read the content. Like, I

get it. I It's your right to do that. But I do think that what has made the internet great is the free sharing of information. And% I don't want a payw wall for my agent

that you know if I can go to the site for free if I fill out the capture but my agent has to pay a penny to access it and it gets charged to me as some kind of like you know extra service fee extra you know charge for using an agent versus me going there myself. I could see that being the future that we're

heading towards in Cloudflare. so much of the internet is going through their sack, right? So, they would control a ton of it and they'd make a lot of money on those uh microtransactions, I imagine, and they probably have some projections of how much money they could potentially make if they can pull this thing off. And maybe I'm reading too much into it.

Maybe I'm speculating, but you know, in this case, I'm the default pessimist and I just don't trust that that wouldn't happen. And that doesn't seem like the right what I would be interested in in as far as like having agents browse on my behalf. I I agree. I want this agent to authenticate on my behalf, but I don't want I want it to do so in

the same way that I can do open internet, open content, go out, do stuff, you know, on my behalf on the internet without having blocks, you know, that or safeguards or or pricing uh pages in the way, right? Or some kind of like pricing mechanism that sits in the middle. Yeah, because Brell's on the opposite side of this where they wanted everything to be open support open internet, open source, etc.

But maybe they have that opinion because it's the opposite of cloud first. So yeah, well I I agree with you know I I agree with let's let's do let's aim for that approach. That seems to be what I would prefer, you know. Yeah. If you're watching this, what do you prefer? You know, let us know. Hit us up

on X or LinkedIn. Are you uh pro- open agent browsable internet or do you want to you know be pro- content creators and being able to be monetize their content and keep it safe from agents you know right I think I think we'll have this debate for a long time I don't think it's a settled matter and you know be interesting to see where it all goes I do think it was a little excessive to

call all part of the access but that's a little much like like I browser based Yeah, I like I I like the team. We're close with the team at browser base. Like there they're some homies. I mean Annie was there for a long time with Stage Hand and we get along great with

Annie who's now at Smithery. Paul, you know, he's he's a good guy. I mean, he that Paul's competitive like he wants to win and this is obviously maybe a way that he can help get more agents using browsers like maybe. So again,

like I I think it was like definitely a little over the top calling axis of evil, but also gives us something to talk about. So yeah, that's true. And Paul Paul handled it so well, too. So yeah. Yeah, it was a very classy

response in like tongue-in-cheek funny way that wasn't too over the top. Um yeah, it's Yeah, I do imagine the discourse around this will continue as more of these gates come out or other platforms have gates and it's just going to be I don't It's going to be kind of spicy. I I I will say on that on that tweet

thread, you you did see a lot of people from the browserbased team jumping in and saying, "Whoa, not Axis of Evil." I didn't see a lot of people from Cloudflare. Maybe they're leaning into it. They want to be the Axis of E.

Yeah, maybe they're like, "Okay, game on. That's what we'll play." Uh, you know, for I don't know. But anyways,

always good to have a little uh a little drama that we can talk about. And yeah, it will be interesting to see how this plays out with, you know, browserbased with Cloudflare and obviously with, you know, some of the other, you know, providers out there like browser use and and things like that. Yeah.

I support the base case mafia. That's all I'll say. Yeah. Well, I mean, we're we're torn in

between YC versus Base Case. Yeah. What's up, Nick? Good to good to see

you. What's up? All right. Well, should we wrap up, dude? It's been it's been almost two

hours stream and yeah, about to go catch a flight. All right. Well, en enjoy being back in uh San Francisco. I'll see you in uh Belgium in a couple weeks. Yeah, I'll see you there, dude.

But yeah, we'll be back next week. Yeah, we'll be back next week. Mondays noon Pacific normally because the holiday we're doing it on Tuesdays, so kind of a special edition uh this week.

But yeah, we will see you all next time.