21:08Riley Brown · Tutorial
Building a Real AI Jarvis with Cursor and GPT Realtime 2
A step-by-step build of a talking, screen-controlling desktop AI companion — no coding experience required, just long prompts into Cursor.
July 1stThe argument in one line.
OpenAI's GPT-6 Astra, especially when paired with the Codex agent harness and full computer access, closes the loop on building software end to end: it generates 3D assets, writes and tests code, controls a computer directly, and re-prompts itself until the result matches what was asked for.
Read if. Skip if.
READ IF YOU ARE…
- You already use an agent harness like Codex or Claude Code and want to know what changed with computer use and self-testing.
- You're weighing whether to learn browser control, computer-use tools, and full-access agent setups as a core skill right now.
- You're curious about AI-generated 3D assets and playable games as a new content or product format.
SKIP IF…
- You want a rigorous benchmark comparison rather than one person's excited first-week impressions.
- You have no interest in agent tooling, computer-use automation, or AI-built games.
The full version, fast.
A creator with a week of early access to OpenAI's GPT-6 Astra spent $1,500 in credits testing it inside the Codex agent harness and lays out seven reasons it feels different from every prior model. Astra turned a PDF floor plan into a finished Blender mansion and then wired that model straight into a working multiplayer shooter map, complete with a flyable helicopter, largely unsupervised. It built a niche Raycast-style desktop app in sixteen minutes, controls a computer well enough to out-draw a rival model in Microsoft Paint, and repeatedly screenshots its own work to test, critique, and re-prompt itself until the result matches the brief. The conclusion: learn Codex, and keep an agent running on dedicated hardware full-time.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 00:42
01 · Cold open
Introduces GPT-6 Astra, a week of early access, and $1,500 spent in credits testing it, then previews seven realizations to come.
00:42 – 05:38
02 · AI is better at 3D graphics than humans
A PDF floor plan is turned into a finished Blender mansion, then wired directly into a playable multiplayer map with a flyable helicopter. Also shown: a photorealistic X render, an interactive 3D engine teardown, and a part-by-part Tesla breakdown.
05:38 – 07:13
03 · Astra creates high-quality video games
Recaps that GPT-6 Astra is strong across Blender, Unreal Engine, and Unity, and predicts a coming 'vibe coding boom' where companies and individual creators build their own games and worlds.
07:13 – 12:43
04 · Vibe coding is solved
A Raycast-style clipboard and image-search desktop app gets built from a single prompt in sixteen minutes. ChatGPT's 'sites' feature bundles hosting, database, and auth so a prompt can spin up a gated lead-magnet page with its own admin dashboard.
12:43 – 15:28
05 · Computer use is insane
A head-to-head Microsoft Paint drawing test against a rival model shows GPT-6 Astra with noticeably better spatial control. OpenAI reportedly bought tens of thousands of Mac minis and Mac studios to train computer-use agents.
15:28 – 18:18
06 · Self-testing gets way better
While building the estate map, the agent screenshots its own Blender scene and in-game footage, judges the result against the brief, and re-prompts itself to fix mismatches without being asked.
18:18 – 20:18
07 · Much better with Codex
The same model performs better inside Codex than in tools like Cursor, because Codex wraps it with computer use, browser control, and full file access under the user's own logins.
20:18 – 23:59
08 · Mac Mini setups remain supreme
Closes on wanting an agent running 24/7 on dedicated hardware, new Mac Minis plus a Linux box, so Codex is always reachable, and predicts Anthropic will need a strong model ahead of its IPO.
Lines worth screenshotting.
- A single AI agent turned a PDF floor plan of a house into a finished Blender model, then wired that model into a playable multiplayer game map in about an hour, largely unsupervised.
- In a head-to-head computer-use test, GPT-6 Astra drew a noticeably more accurate likeness of a reference photo in Microsoft Paint than a rival model did.
- OpenAI reportedly bought tens of thousands of Mac minis and Mac studios to train computer-use agents, while Anthropic rents its Mac minis through AWS.
- The same model that built a Blender scene opened it, took a screenshot, judged whether the result matched the brief, and re-prompted itself until it did, without being asked.
- A creator built a working desktop clipboard-and-image-search app, modeled after Raycast, from a single natural-language prompt in sixteen minutes.
- ChatGPT's 'sites' feature bundles hosting, a database, storage, and authentication, so one prompt can produce a hosted lead-magnet page with its own gated admin dashboard.
- A full playable multiplayer map, including a flyable armed helicopter, was generated from a single PDF and added to an existing game in about an hour.
- The presenter argues vibe coding is 'solved' in the sense that the knowledge required to build working personal software has dropped to zero, not that the resulting software is polished.
- The same underlying model performs noticeably better inside the Codex harness than inside other tools, because Codex wraps it with computer use, browser control, and file access under the user's own logins.
- The presenter's takeaway from watching AI generate photorealistic 3D assets on command is to get into 3D printing, since turning a digital asset into something physically useful is now the remaining bottleneck.
Computer-use agents can now build, test, and fix their own software end to end.
WHAT TO LEARN
GPT-6 Astra paired with the Codex harness turns 3D generation, game building, and self-testing into one continuous loop, which makes an agent harness with full computer access the highest-leverage tool to learn right now.
02AI is better at 3D graphics than humans
- Turning a floor plan into a usable 3D building no longer requires 3D modeling skill: describe the source material and the shape you want, and let the agent iterate in Blender.
- Photorealistic renders and interactive 3D teardown explainers are now achievable from a single prompt, which shifts the real bottleneck to turning that digital asset into something useful outside the screen.
03Astra creates high-quality video games
- Full playable multiplayer games can now be assembled from natural-language prompts across Blender, Unreal Engine, and Unity in a single afternoon.
- Expect creators to start shipping small playable worlds the same way they currently ship PDFs, courses, and communities: a simple game or interactive world is worth testing as a new lead-magnet format.
04Vibe coding is solved
- A very specific personal tool can take one prompt and under twenty minutes to build, so the bar for justifying a custom app for yourself has dropped to near zero.
- Hosted pages with a built-in database, storage, and login mean a single prompt can produce a working lead-magnet page with a gated admin view, cutting out separate hosting and backend setup.
05Computer use is insane
- Computer-use accuracy, an agent seeing a screen and operating it like a person would, is improving fast enough to be worth re-testing every few weeks if you dismissed it as unreliable before.
- The scale of infrastructure being built for computer-use training signals it's a strategic priority for the major labs, not a side feature, so expect faster gains here than elsewhere.
06Self-testing gets way better
- The most reliable way to get a long, multi-step build right is to have the agent screenshot and evaluate its own output against the original ask, then let it re-prompt itself instead of manually checking every step.
- This self-testing loop is what turns one ambitious prompt into a finished, working result over an hour instead of a rough draft you still have to fix by hand.
07Much better with Codex
- The same underlying model performs differently depending on the harness around it, so evaluate a new model inside the tool built for it before judging it in a generic chat interface.
- An agent with full computer access, your logins, your API keys, your files, moves faster because it skips the copy-paste and re-authentication steps a sandboxed agent can't avoid.
08Mac Mini setups remain supreme
- Running an agent harness 24/7 on dedicated hardware you can reach remotely removes the 'I can't use this while traveling' limitation of a laptop-only setup.
- If a workflow depends on a coding agent, treat always-on remote access to that agent as infrastructure worth setting up once, not a nice-to-have.
Terms worth knowing.
- Codex
- OpenAI's agentic coding tool, which gives a model computer use, browser control, and full file-system access so it can build and test software with less human oversight.
- Computer use
- An AI capability that lets a model see a screen and directly control the mouse and keyboard to operate software the way a person would.
- Vibe coding
- Building working software purely by describing what you want in natural language, without writing or reading the underlying code yourself.
- Blender
- A free, open-source 3D modeling and animation program used here to turn a floor-plan PDF into a finished 3D building.
- ChatGPT sites
- A feature that lets a single prompt generate a hosted web page with a built-in database, storage, and login, without separate hosting setup.
- Self-testing loop
- An AI agent's habit of reviewing or screenshotting its own output, judging it against the goal, and re-prompting itself to fix gaps before calling the task done.
Things they pointed at.
Lines you could clip.
00:10
“I am genuinely shook by how good this model is.”
07:25
“Vibe coding is just solved... you can just spawn software on command.”
22:45
“Why wouldn't I want a computer that runs Codex 24-7?”
23:28
“I think we've crossed the era into kind of the AGI era.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
You probably have already heard by now, but OpenAI released Astra or GPT -6. And I'm going to be honest with you. I am genuinely shook by how good this model is.
So I had access to this model for about a week before it came out. And since the release, I've spent around $1 ,500 worth of credits using it. And I've been using it for everything.
I built a Call of Duty game in just four prompts. And then on the fifth prompt, I made it multiplayer and immediately started using it on stream with 20 different people. This model, GPT -6 Astra, simply feels different.
And in this video, I want to talk about the seven reasons why I feel this way. So let's hop over to the board. And so the first realization that I had while using this new model.
is that AI is better at 3D graphics than humans now. Straight up. And many of you have probably already seen the demos regarding Blender.
But let me show you what I was working on. So yesterday, I found this PDF online of this estate. And this shows this massive house on this estate.
And then it shows the entire estate. It even has a floor plan for this estate. So I thought...
Why don't I just pass this whole PDF to Codex and why don't we just use the brand new GPT -6 Astra inside Codex, get it to Control Blender and create this house right here or this estate. This is what it created. After about 20 minutes, it went off and created this inside Blender.
It created this house and like look at this shot right here. Like look at this shot versus this shot. Maybe it got the color slightly wrong, but look at this house right here.
But I didn't just have it create the Blender file. I said, once you're done creating the Blender file and you've checked it and it looks good, create a new map in our Call of Duty game. This new map should be called Estate.
I want you to make this a new map that you can play in the same format as the other maps that we've created in the video game that we created yesterday, which is Call of Singularity. Then after an hour, it sent the blender scene, but it also sent this. This is our Modern Warfare game that we created called Modern Singularity.
There was a map on here called Estate, and I could play it. Here is the estate that we've created. Oh, come on.
There it is, but I want to show you something That is the house that it created on blender and it immediately added it to my Call of Duty game and it even added this helicopter I can fly in this helicopter And I can shoot people in it and they can shoot me out of it, too It's pretty overpowered, not gonna lie. Oh, they shot me out of it.
Just to fully illustrate my point, I do want to show you a few more examples. This is a render someone posted on X created by Astra. It's not just that kind of video game look, which is what I was going for.
It can create photorealistic 3D renders. Another trend that I've been seeing on Twitter is people are using Codex with Astra to... create these educational 3D websites.
So this is like an engine and someone wanted to learn how to use the engine and it created this realistic 3D representation. I'm not sure how accurate this is because I'm not an expert in engines, but I found this to be extremely interesting. This person actually took apart an entire Tesla part by part so you can see how they all fit together.
I mean, look at how cool that is. And it can show you every single part. And you can actually zoom in on each part.
It tells you what each part is and what it does. And you can go back and forth. between the car being put together and all of the materials being spread apart.
And so that is my main realization. And it is the most viral realization that we've had collectively about this model, which is GPT -6 is a baller at using Blender, Unreal Engine, Unity. It is great at creating digital 3D assets.
And one of my takeaways that I thought during this whole thing is I want to get into 3D printing. I think that is a massive opportunity right now and something I'm going to dabble in on the side for some of my side projects. I'm fascinated by this because.
Creating digital assets is no longer like hard or a scarce skill. You can just do it. Now, how do you actually turn that into something that looks cool in real life or is useful in real life?
That is what I would focus a lot of my attention on. If you're kind of looking at this and you're like, okay, what do I do with this information? 3D printing.
Because it's so good at these things, I would definitely get into 3D printing. And then I would also just focus on using this tech for learning. And then I want you to really think about this deeply, right?
Think about the fact that AI can now create high -quality 3D multiplayer video games in a matter of 30 minutes to an hour, right? People are creating full Rocket League clones and customizing it to make the game more fun, the exact feature that they wanted. They're just creating it.
And people are creating Fortnite -like games. I created Call of Duty, right? I could have put a lot more time into it and I might still do that.
But you could just create these games. And so I think that this is going to lead to the next vibe coding boom. I think there will be a platform out there that focuses only on games.
And maybe it's just a hosting platform, right? On ChatGPT, you could just at mention hostmygame .com and it just hosts the game there. And it's free to host your game there.
And you can maybe create some in -game. currency where there's transactions. And I think that a company to figure out how to allow people to host their games and then sell things within that game, they will become billionaires for sure.
It's a little bit like Roblox in a way. But I do think there should be no limitations on what you could create. And then also the next thing is like, I just think creators are going to start creating games for their audience the same way that they create PDFs, courses, communities.
All of these things are very like low fidelity. Like imagine if you could just create a video game or a little world people can go and interact with. It's kind of like the metaverse, but I think if it's just a link people can go to to play a game to enter the creator's world, I definitely think this could be a huge trend in the future.
So the third realization that I had while using GPT -Astra that made me feel the AGI was that vibe coding is just solved. And I'm kind of, I'm serious about this. Right.
You can just spawn software on command. The level of knowledge that you need to just create your own personal software is zero. Let me show you an app that I created this morning when I had like this minor urge to create an app.
So the problem that I had while I was at coffee this morning is I was making this presentation and I got sick of going to Google images and then taking on images and then looking up AI video game Astra. as a visual, and then I would like copy it. And then I would, I would come to back to the tab that I need to be on.
And then I would paste it in my Excalibur diagram or whatever presentation that I'm working on. And so I have this idea for an app that would look a lot like Raycast. And so Raycast is my clipboard manager and it's my app finder.
It's a very cool app. I love how it just kind of pops up when you need it. And so I just went to Codex, right?
I just said, I want to create a desktop app that I can use via shortcut. And it's a Raycast style app. That lets me have Google images and remove the background and all of them that I copy or save get saved to the saves folder.
And then I just like had a bunch of ideas for the app. And then literally after 16 minutes, it created this app right here. Command shift space.
I can open this up and I can type Codex logo and notice here, if I press enter, it searches for my. the Codex logo and I could copy this right here and we could paste this in whatever it is that I'm working on. Here's the Codex logo.
But once you put it on Excalibur, you can't remove the background. So I wanted the app to be able to remove the background just like this. I also wanted to be able to search multiple things so I could go open AI logo and I could do a semicolon Microsoft logo, semicolon Bing logo, semicolon Google logo, enter.
And this would automatically create four Google image searches. So I can search a bunch of things at once. And so while one is loading, I can move to the other one.
And then I also added this AI feature, which does that through AI. Microsoft logo, Santia Nadella, Sam Altman, Dario Amodai, Steve Ballmer, and Gavin Jensen Huang. All separate searches.
So I can just paste this in and I can press command enter. And this will use AI to automatically create the right searches. So it'll search Satya.
So if I need a ton of images for whatever it is that I'm working on, right, I can remove the background of Satya. There we go. Copy it and paste it on the board here.
And because I saved or I copied it, it automatically gets saved to my clipboard. or it automatically gets saved to this file here. So ever in the future, as soon as I start typing it, it'll show up here.
So I can use an existing one that I've already had. Anyway, super niche app, but I made it for myself. And now it's a keyboard shortcut that just lives on my computer, right?
This is a desktop app. This would have taken me a long time. And I created it in a single prompt using Codex.
And this trend is just going to continue. The trend of being able to just spawn whatever software you need for your business or for your personal life is only going to exponentially get better. OpenAI and Anthropic and all these companies see the value in creating it.
And that's why they're adding. This sites feature, they've already added the sites feature, which adds hosting database, storage baked into it, authentication. All of these things are baked into sites.
I can just create a sites and it will host it on chat GPT on like chat, gbt .site. And then you can share it with your team and they could, and you could require that they sign in with their team account and you can just spawn apps really easily. One thing that I made clear on Twitter yesterday, I was talking about sites publicly is I was saying that I think one of the best use cases right now are creating these lead magnets.
So like I wanted people to be able to sign up for a call of singularity and very easily I can just require like I can put up this site. I said, hey, create in the style of my modern singularity game. Please create a site where people can upload their information so I can message them.
And then it created this. It created this really quickly, hosted it on ChatGPT sites. I also specified that I wanted an admin dashboard for only my email.
And then it created this. And now on the website, I can just click on this right here. And if I sign in with my account, I can see the admin dashboard.
Only I can see the people who've entered the information. So you can create these secure sites and you can give yourself admin access. Like whatever your GPT account is, just make yourself the admin.
You're the one who can see behind the scenes. I think this trend is only going to continue as like you can create any app you want, just host it on chat GPT sites and share it with other people, right? That's how you should get feedback.
And so this is just a really easy way to vibe code and get it in other people's hands. And pretty soon people are going to be vibe coding full games directly on chat GPT sites, which I think is going to be pretty cool. Okay, my fourth realization.
Which left me feeling the AGI when using GPT -6 Astra was that computer use is simply just insane. It is accelerating so quickly at how good these models, especially when you use it inside codecs, are getting at controlling a computer. And I thought this test was actually really telling.
So this guy pinned Fable against GPT -6 Astra, put a photo of himself, and then opened up Microsoft Paint. And had both of them draw him. And I think this is sped up like 10x or something.
But the point is, look at the difference between GPT -6 Astra and Fable 5 .1 at computer use specifically.
And so right now, like, it is using computer use. And if you go to like... If you go to ChatGPT, if you were to use this at computer, please control my computer.
This is a skill or a plugin that's built into Codex. It can just fully control your computer, which is actually really useful when you're using it on the remote. So if you're controlling Codex from your phone.
which you can, which you can basically control the desktop app from your phone, while you're away, you can get it to fully control your computer. And this is getting better exponentially. And we're starting to see OpenAI, at least it seems like they're pulling away from Anthropic a little bit on the computer use side.
And like, look at this. So it used Microsoft Paint for a period of time. And, you know, this is about average.
You know, this is actually above average from two weeks ago if you got AI to try and draw you from computer use. But, like, look at how much better it is now with GPT -Astra. It is insane that it can just control a computer.
It can have that level of spatial awareness and create – it can, like, recreate an image. That's pretty insane. And when I say that this computer use trend will only continue to get better, right?
These models will only continue to get exponentially better. I mean, look at this. OpenAI reportedly bought tens of thousands of Mac minis and Mac studios for reinforcement learning and training computer use agents, while Anthropic rents Mac minis through AWS.
Both Anthropic and OpenAI have... Tens of thousands of Mac minis and Mac studios running all the time training AI, evaluating how good AI is at controlling a computer. And this kind of leads me to my next realization that I've had while using GPT -6 Astra.
And it's that it is getting significantly better at self -testing. Meaning when you ask it to do something, no matter how long it is, it'll do the thing. It'll evaluate how good it is at doing the thing.
It will then decide what changes it needs to make. And then it will make changes. And then it will re -evaluate the thing over and over and over again.
And we've been talking about this for a long time, right? Even back to like the early Cloud Code days, it would actually be able to do this, especially when vibe coding. But it's just getting a lot better at it now.
I realized this when I was using it to create something in Blender. Remember earlier when I had that Blender game get created? I was creating that Blender game and I wanted to put it in my call of like that entire map into my Call of Duty game.
And the AI went off forever and it would create the map. And then it created this image, right? It had Blender open on the computer and it took a screenshot.
It looked at it. It evaluated it. And it says the state now builds successfully as the third map on the game.
I'm extending the vehicle system to helicopter boarding. And then it started playing the game. You can see here it started a match.
It was moving around the map as one of the characters. So it started playing a game, got in the helicopter, started flying the helicopter to test it for the game. And it was just playing these games.
and making sure that the map looked good. Here you can see it's checking in on the house. And so we asked it to take that PDF, turn it into a 3D model, and then turn it into a video game.
And at every stage of that process, it was checking its work by actually playing the game and then viewing the images and evaluating it. Is this supposed to be... Or is it supposed to look like this?
And the house was actually like really white at the beginning and it didn't look like the right texture. And so it basically saw that and then it re -prompted the AI. It's basically re -prompting itself in this loop.
And I've never felt so confident in AI than I do with this model. I know that it's just going – if I specify the end result that I want and it is realistic in the sense that it can be done on my computer, the AI will just work until it gets there. Unless it hits like a hard paywall, maybe you haven't paid for a software that it needs, in which case it will tell you to pay for it.
End result, the agent just iterates, iterates, iterates, tests, tests, tests until it reaches that end state. And it's truly – It is insane.
It is playing a video game that it created. Anyway, I just think we're at an insane point in the world of AI. And then my final two realizations that I have about this model kind of go hand in hand.
The first realization is that this model is much better with codecs. I don't know if it's much better, right? You can use it in other tools, like you can use GPT -6 inside Cursor for now, right?
Something about using it in the Codex app makes it significantly better. This model is made for the Codex harness, and it's made for the tools that are available inside Codex. You know, Codex has a ton of useful tools, like computer use, like...
browser use, right? This browser, it's able to spin up like whatever it is and run it locally in the computer. And then it can fill out forms really quickly.
It can test websites really quickly. It's made for this application. And so all I will say for this kind of realization here, I do not think there is a more important tool.
Like this is the number one thing that I would learn right now. I would learn how to use Codex. Because the UI is so nice.
It's like ChatGPT, except... It can fully control your entire computer. It can do anything you want as long as you have full access on.
And that's what makes this so powerful is it has full access over your computer. It can just use Blender when it needs to use Blender. It can view Blender if it needs to view Blender.
It can take screenshots, send it back to Codex. It can create a game. It can build a website.
It can edit files. It can delete files, remove files, do anything on your computer. It has full access, and that's what makes it so powerful.
And the fact that you're signed in to all of your applications on Codex, like especially GitHub and tools like Vercel, it has all of my API keys on the system. So it's able to just do things much quicker because it's running on my computer. It has full access over my computer.
I'm signed in on all of my browser tabs because it's my computer. And that brings me to the final one and the biggest. realization that I had while using GPT -6 Astra with Codex is that the Mac Mini setups remain supreme.
I just bought a Mac Mini. In fact, I just bought another Mac Mini right here. I'm going to be using this.
as well as this Linux machine, I'm going to get Codex running on these computers because I want to have access to Codex wherever I go. Because right now I use Codex on my laptop and whenever I'm traveling or moving around, which is quite frequently, I can't use Codex on my computer. And so I've been talking to some of my really good friends, my really smart friends, and we are going to talk about and make a video this week on the absolute best setup to run Codex remotely and the other agents too.
Like you can run Anthropic, you can run Grok, you can run all of the different super apps, I like to call them. on your Mac mini and you can control them from your phone. And if you have an iPad or your MacBook pro, you can actually remotely control your Mac mini.
If you leave it running at your house so that it never turns off. And so I'm actually going to be talking about that setup. If you can fully and remotely control your Mac mini and it can run codecs.
And you can have a version of Codex running at all times 24 -7 that you can access. I think that's incredibly powerful. And so I'm going to be setting that up this week.
And I think that's going to be really fun. I've been using GBT work a lot. And it's really useful for automations with work stuff.
But a lot of my coding stuff is also related to the documents and presentations that I'm creating. And all of them, I want them all on one machine. And so I at least want...
one instance or version or computer of Codex running 24 -7, and I'm going to be setting that up. And that was my biggest realization after all of this, right? If AI can have full control over a computer, create 3D assets, can create video games, can vibe code any app you want on command, it can host them very easily, and then it can use everything on the computer to self -test so that it can improve until it reaches a...
A point where the AI is like, yep, that's good enough. We did it. It'll self -test and it's much better with Codex.
Why wouldn't I want a computer that runs Codex 24 -7? That's kind of my biggest realization after all of this. And yeah, I'm really excited about this new model.
I'm excited for these models to get better. I know that Anthropic coming up to their IPO, they're going to try and release a model that's better than GBT6. It's going to be very important to raise enough money to get to that $2 trillion valuation that they want.
And then OpenAI is going to quickly try and come back and release another model as well. We are in the era of exponentials. AI is going to get significantly better at controlling a computer.
It'll continue to get smarter. It'll get better at vibe coding. It'll get better at everything that you can do on a computer.
And so I'm feeling it now. I'm feeling the age of AGI. We are in a different era of AI.
I think the first threshold was Claude Code when it released early 2025.
We didn't feel it until mid 2025. I think we've crossed the era into kind of the AGI era. And I will keep you updated on everything.
I will be making a ton of videos going forward. Thank you guys so much for watching. I'll see you here in the next one.
The Hook
The bait, then the rug-pull.
A week of early access and $1,500 in burned credits later, one creator says GPT-6 Astra is the first model that has made him feel the agent era actually arrived.
Named ideas worth stealing.
00:27list
The Seven Realizations
- AI is better at 3D graphics than humans
- Astra creates high-quality video games
- Vibe coding is solved
- Computer use is insane
- Self-testing gets way better
- Much better with Codex
- Mac Mini setups remain supreme
The seven-part structure the whole video is organized around, recapped on a whiteboard slide between sections.
Steal forany 'first impressions of a new AI model' video that needs a repeatable structure instead of a rambling reaction
How they asked for the click.
VERBAL ASK
00:00link
“For Agent Native Resources: agentnative.inc/resources”
Soft CTA placed only in the video description, never spoken on camera.
MENTIONED ON CAMERA
Visual structure at a glance.
