I'm Tired of AI transcript

It's very funny how the internet works. We're now in the AGI era. >> Wow, look at these characters. >> I created Call of Duty. >> This was the lava lamp that I created. >> People do amazing things. >> You take whatever is the latest model. You make an FPS demo. You create New York City in 3D. And people go, "This is AGI." And this process spawned AI cringe lords. You probably keep hearing about AI second brains. >> This is my second brain.

This is everything that I know. Normally this training would cost a pretty penny and I have charged for it before but >> I need you to go to my free school community. The link for that is down in the description. Their second brains and metric tons of games slot. >> People are creating Fortnite like games. >> This is really impressive >> and I've never felt [music] so confident in AI. I created Call of Duty and I'm sure there is someone at OpenAI or Anthropic looking at this right now and thinking [ __ ] it.

Let's add more Bloom next time. For the longest time, I had this feeling that the models are not really improving. So, I wanted to do a reasonable comparison between the models for you guys that doesn't really involve any 3D slop. And recently, I came across a nice little problem in my codebase that might just be a decent example to try. I have this project I've been working on for a long time. And in that time, I've made some decisions that I knew were questionable in terms of performance, but at a time, I decided to delegate this work to my future self.

And so, welcome to the future. In this video, for your entertainment, I'm going to present the Frontier models with the same exact performance problem, and we will try to see if they will earn the price of tokens they charge or if they are all just full of [ __ ] I wanted to split the test in two parts. First, I presented the models with the problem and I asked them to come up with a conceptual solution. But then, I also let them write some code.

And I wouldn't be the professional content producer that I am if at this point I didn't say the results might surprise you. So the models I tested are Fable 5.1 and Astra, but also Codex 5.6 Soul and Terra, Opus 5, Sonnet 5, Grock 4.6, and even Gemini 3.7 Flash. For each model, I used a separate git work [music] tree to isolate them from one another to make sure that they can't read each other's output. The prompt that I sent was deliberately vague to avoid giving the models any particular direction.

I was trying to get them to fix a performance problem. So I literally asked them, can you improve performance in this part of the code? >> Oh, I can deploy 10,000 agents. >> Think about the fact that AI can now create highquality 3D multiplayer video games. >> And yet they all gave me kind of the same solution. They used different words and some of them wrote a long answer, some of them wrote a short one, but it was all kind of the same idea, which was very strange.

And what's even worse is that this idea wasn't even that great. We will look at the code as well. But right off the bat, my overall impression was that the cheaper models produced better code overall. And Fable 5.1, for example, was straight up bizarre. But before I let them write code, I put all the proposals from all the models in one folder. and I asked Gemini 3.7 Flash being the cheapest model of the bunch to make this single page document that I can read all of them together.

Now, I don't remember the exact prompt that I used, but I must have said something like, I want to review them all in one place. And Gemini understood that it should also give them a score. And it ranked its own output the highest because it was the longest. And Fable 5 scored the lowest because it was the shortest. But let's ignore that for a moment. The problem that I'm trying to illustrate is that regardless of what these models cost, they gave me the same idea, probably because it was in the training set or whatever.

And moreover, as I said, this idea wasn't even that good. So, in the rest of the video, I will describe the problem, show you what they proposed, and then give an example of a much better solution that's not really that hard or even complicated. And then finally, I'll try to give my two cents as to what is happening with these models overall. So, what was this performance problem? And as a quick intro, my app is a utility app for the Mac that has a bunch of different features and you can customize a fair bit in the settings.

The settings are then stored on iCloud, so all your devices are in sync. And the performance problem has to do with this syncing process. At the moment, my app stores in about nine different JSON documents. And I know that sounds oddly specific, but they just really correspond to different individual features of my app. And I split them into separate documents to avoid sync conflicts while giving each document a specific purpose like window manager settings, appearance settings and you know you get the gist.

So we have these nine documents and every now and then for each one I query iCloud to fetch the latest version. And while this is beautifully simple, the main drawback is that these downloads happen even if the documents never change. And I want to point out this is the actual performance problem. Most people don't change settings often. It's more of a set it and forget it type of thing. Once the keyboard shortcuts become part of my muscle memory, for example, I'm not going to change them.

So, I wanted to see if the AGI can figure that out. And I would expect AGI to be able to identify that we're downloading data over and over, even if it doesn't change often. And it's not really difficult to deduce that it doesn't change often because it's keyboard shortcuts and things like that. But no, literally all the models were like, "Bro, I got you. You're downloading these documents individually by primary key.

What you can do is write a query that returns all of them at the same time. Even Astro gave me the same exact idea. >> And I used this ChachiBT work session that went for 34 hours. They completely failed to see what the actual problem was, the water usage of ChachiBT. >> But I didn't give up. I let them write some code as well. And this is where the differences between the models actually started to show. It seems like in this era of AI hype, we've forgotten what good, valuable, and secure software looks like.

Today's sponsor, Interext, is just that. Internext provides secure cloud storage for all your files and data. And what sets it apart is that it's an open-source, independently audited, and end-to-end encrypted solution. If you've ever worked with cloud file storage, you know that it's very far from simple. costs easily pile up and with traditional cloud providers, the provider may hold the encryption keys, meaning your files can potentially be accessed or disclosed under certain circumstances.

Interext implements postquantum zero knowledge encryption that gives you absolute control over who can see and access your files and data while at the same time it has all the conveniences of a traditional cloud. secure collaboration and sharing, advanced access controls, automated backups, scheduling, and the ability to mount your cloud drive as it were a local disc. It's also made in Europe, and it stores data across European data centers.

It's got a CLI, so you can manage and access your drive from the comfort of a bash script. Interext also includes a range of privacy tools and services, including photos, mail, anti virus, cleaner, and a VPN browser extension, giving you more ways to protect your data and privacy without relying on the usual big tech ecosystem. One thing that caught my eye is that they offer lifetime deals. Imagine a fully redundant, backed up, and managed cloud drive for a one-time payment.

That's pretty unusual in the cloud storage space. So, if you're thinking about where to host your files, check out Interext using the link in the description. And that link actually gives you an even better discount than what's on the website. But all right, so for the coding part, I wanted to give the models a fair chance. So, I wrote a prompt that leads them in the right direction conceptually, and all the models have to do is code.

So, at this point, it's a good time to show you one idea that I had to make this a lot more efficient. And honestly, I was hoping that the models would arrive at something similar. So we have these JSON documents and we have iCloud. So imagine that we add one more document that acts as a dirty flag essentially. So when you modify one of these, you also flip this flag to indicate that the document changed. So when you know which document changed, you can effectively stop listening for changes on all of the individual documents and only listen for updates on the flag itself.

And it's just one solution. I'm sure you can think of other solutions as well. But the problem is the models can't and they are supposed to be AGI. >> We're now in the AGI era. Astra has really hit something that I'm like, okay, I think this is pretty reasonable to call it AGI. Right off the bat, my favorite implementation was actually GPT 5.6 terra. And here is why. Without looking at the pull request line by line, it is still pretty easy to tell that the changes are focused on areas that matter.

And on the other hand, Fable 5.1, for example, rewrote a lot of the existing code without any reason. And some of the logic and variable names it added are super convoluted like boolean called remote changed behind our back. That's the variable name. It's so tiring to watch these hype videos about how good the models are and then open a pretty simple part of my codebase and Fable starts writing variables like remote changed behind our back. >> It's it's pretty nice actually. >> I'm I'm so tired man.

But you must be thinking, well, [music] if Terra was good, then Soul and Astra would be amazing. Both Soul and [music] Astra produced significantly more slop, very similar to Fable. Now, Astra wrote tests, and you might be thinking, see, that's why you're paying for Astra. Gemini 3.7 Flash also wrote test. Does that mean Gemini is AGI now? Felt like about the timeline to AGI. if you really squinted at it and you're willing to scale up and build massive supercomputers, spend the hundreds of billions of dollars, that kind of thing, that maybe it' be 10. >> But that is going to do it for this one.

If you guys now want to see since you have an AIOS set up, how you can control everything with your [music]