Getting the transcript
Reading the captions from YouTube. A video nobody has opened here before takes 10 to 30 seconds; this page fills in on its own.
Getting the transcript
Reading the captions from YouTube. A video nobody has opened here before takes 10 to 30 seconds; this page fills in on its own.

Stefan 3D AI · @stefan_3d_ai
This video has no Most replayed graph yet: YouTube shows one only once a video has enough views. These are the moments viewers replayed most in Stefan 3D AI's most watched videos.
Most replayed moment at 4:08
7.0x that video's typical replay level
and it will also have the access to full ai apis. Now let's do some serious tests. The first test will be to do an animation of a character that will be split in parts. The first test will be to create animation. Animation of a character that will be an
Said at 4:01
Most replayed moment at 9:39
5.7x that video's typical replay level
Overall I think this guy's made really amazing job, bring it it all together and I really had a lot of
Said at 9:41
Most replayed moment at 20:36
4.1x that video's typical replay level
Just also because I was playing that game when I was kid. Okay, let me see if I remember how to play that. Let's start with the road, right? Let's try to build a, oh, there are cars. There are cars. It generated nice cars and they're like, wow.
Said at 20:29
The graph counts replays. It does not show where viewers stopped watching.
Words
2,944
Runtime
16:03
Speaking pace
183wpm
Reading time
12min
183 words per minute, just over the 181 median of 349 measured videos. That distribution comes from the 349-video hook study.
Opening (first 30 seconds)
I spent 1 billion tokens, more than 3 days, all limit resets, trying to create a game with GPT -6. But why? Well, there is a lot of myths about GPT -6 that it can do crazy things. So we're gonna find out is it really. Main condition is that I only can do prompting, but not in a stupid way, but in a smart way. I do have experience developing games, working with 3D and stuff. I'm gonna try to guide it best possible way, provide with necessary skills, MCP. This
92 words, the words spoken in the first 30 seconds at 183 words per minute.
Free, no signup. See how the first 30 seconds hold attention, with rewrites.
Sentence shape
| Measure | This transcript |
|---|---|
| Sentences | 298 |
| Average words per sentence | 9.9 |
| Longest sentence | 33 words |
| Questions asked | 17 |
| Sentences containing a number | 48 |
Most used terms
Filler phrases
98 in total: like 49 · actually 33 · I mean 5 · right? 5 · you know 4 · basically 1 · sort of 1.
A literal whole-word count of the same phrase list the Prepublish browser extension uses, so a phrase inside another word is not counted and a phrase used in its ordinary sense still is. It is a count and not a judgement.
Free, no account. See where attention is likely to drop, with a rewrite for each weak line. The free check shows the scores and the one issue costing the most. Or run it on the words above first.
Free · No login · See a sample audit first if you prefer.
What this transcript is
Every word below is the caption track YouTube publishes for this video, pulled from the video itself and reproduced unchanged. It is not Prepublish's writing, not a summary, and not a re-transcription: it is the video's own published captions. English captions, published by the channel, in the video’s original language. Source: the video on YouTube. A channel that would rather this page did not exist can ask for its removal through the contact page, and it is removed.
I spent 1 billion tokens, more than 3 days, all limit resets, trying to create a game with GPT -6. But why? Well, there is a lot of myths about GPT -6 that it can do crazy things. So we're gonna find out is it really. Main condition is that I only can do prompting, but not in a stupid way, but in a smart way. I do have experience developing games, working with 3D and stuff. I'm gonna try to guide it best possible way, provide with necessary skills, MCP.
This will be fun. So I actually gonna show you two games. I will create realistic zombie shooter in Unity. And I also invited my friend to help me and he gonna try to create GTA 6 with GPT -6 in Unreal Engine 5. Then if I record this video in next year, it could be like in Unreal Engine 6 as well. Two different engines so we can actually see where GPT -6 will show better results. And of course there will be plenty of MCP I gonna tell about them along the way.
So where do I start? Well, as I said, I want to move in smart way and actually get somewhere. So that's definitely not gonna be like here all this API create 3D models and bring like the game, etc. I gonna need to move into iterations because that one shot approach is a dead end. It's not gonna get you anywhere. Like it's. 100 % you gonna want to change something along the way. And taking one small step at a time is definitely better solution.
So why is first prompt that took me only one hour is to create a 2D prototype of the game. So it will be arena shooter and I asked it like to create some playable prototype like exactly how I called it. And to my surprise first iteration was quite successful. I pretty much got what I wanted in terms of layout etc. For this to work of course I connected my Codex CLI to Unity CLI. For Unity MCPs I can recommend two things.
You either go with the official deleted Unity plugin which is available in Codex and Claude and it works with the same new Unity CLI that you install when you create new project. Or you can also use some custom MCPs. For example I use ankle breaker Unity MCP for my project because that's not only a new MCP but also a plenty of useful skills that really helps agents to do better job. Both options are free. Next step when I had the layout I actually asked it to move it to a 3D first person shooter.
And here you can see the results. By the way the interesting thing about the UI is that look I can actually tell that this was done by GPT -6 because it always create in same style. I don't know. There's some general knowledge. How to build UI. I'm not saying this is bad. I'm saying like it is always the same like I created like few other games before when I was testing new Unity plugin. And by the way, I really recommend you to follow me on X to be first to know about anything significant.
Okay, prototype is actually working. But again, everything was procedural at this stage, VFX, etc. And I would say like for those who create games like stylized low poly, etc. This is heaven. Like I mean, GPT -6 can handle pretty much everything. VFX is sound effects, especially combining with some APIs. That's heaven. Like you can create a lot, but we're moving to more realistic style. So when I verified mechanics, we're going to move to very big, big step like to add realistic characters, more realistic weapons, and so forth.
This iteration took me almost 149 million tokens. But at this stage, I still not creating assets with API just to see how far I can get. And to be honest, that is still impressive. Well, I lied a bit. It was allowed to create images and find some assets around the internet. It is actually quite surprising that this guy can go around the internet and easily find animation packs, image packs, like Mixamo. It can find animation by its own and just, you know, like download them and use them in the game.
Like, yes. So sometimes if you're not specified that. everything. Should be generated procedurally, it can stick to the internet and use ready made assets. How about that? But this GPT -made. Zombies are not that exciting looking, to be honest. So yeah, you can see like this guy tried so hard. And. In the end, we're still going to use API to bring some nicer 3D characters, zombies, etc. The only option is to use. 3D AI.
There are plenty of APIs. Also, I recommend to use video AI, some API for procedural textures. That basically. Saves a lot of time and give you a better image. There is no point trying to do everything with Astra. I mean, that's just fun, but like. That's not really working. So in my case, I'm going to use Hicksville CLI and Hicksville Blender plugin. Why? Because first of all, it gives you. A lot of tools. Second, within what subscription can get access to video AI, 3D AI, different one, like Tripor,.
Meshy etc. You were asking guys, can I use Tripor API? Well, yes, you can, but then you will have access only to Tripor. API While through Hicksville, you get access to Tripor, Meshy, Hunyun, Seedance, etc. The only real alternative. To Hicksville might be it doesn't have any official MCP and tools are not exposed that. Way not even close. So I'm going to use this one and recommend you to try. So for now, I'm going to use Hicksville MCP and thanks them to sponsor.
This. Video and providing me with the credits. And later, I'm going to calculate all assets that I created for this game and how much they cost. So what GPT actually did? It created images, then sometimes also video references. From images, it generated models. And there was a lot. You might ask reasonable question, why there is video AI? Well, the answer is very simple. You can use it as reference for GPT 6 to create an animation.
I actually made complete video just few days ago showing like how far you can get. And I was genuinely surprised that it works. I also tried without video references. I was not able to do anything close. So plenty of zombies is cool, but we also gonna need environment objects. And were a lot of them as well. And since it's like first person shooter, I really wanted to play around with the weapons and some cool animations.
Here I really asked this guy to do the procedural animation of weapon reloads, weapon shooting, etc. Not gonna lie, I was quite skeptical. You can see I was of course was playing the game and providing some feedback that like this not work, this to fix. And first version of like the hand animation, etc. was extremely scary. But iteration after iteration, I was able to do some simple reload animation. First saw rocket launch animation was terrible, but later it actually reworked it completely.
And look at this. Okay, okay. That. actually. Works Same thing happened with the shotgun. First version was absolutely terrible. But this guy actually cheated in a cool way. He created primitive part of this barrel. Like he created it from primitives and put it like aligned it here in this place. So the animation look decent, but also because later he was able to do that reload animation. Just look at this guy. That's.
Crazy right? Wow. Okay. So that's all made by GPT -6. Crazy. And there is also minigun. Minigun also generated. In three parts. So later this guy made reload animation. But yeah, it was terrible. Like it was not a complete circle. It was just. A part separating. Altogether, this last milestone took, I don't know, it felt like forever. I was sitting there providing. Feedback spending tokens like more than three days. And when I finally hit all possible limits, I actually spent all.
My weekly limits because I spent three retries. So the result is on your screen. Rug you like first person zombie shooter. What. Do I think about it? Well, for some reason, I feel that that was actually hard because I was only be able to prompt. I cannot. Go and change anything and I cannot generate 3D model. Right. So yeah, I mean, this was really fun. But to be honest, this. Approach is not viable for me. I just recently published first video of me building my own dream game.
And I think when you. Combine AI tools with your real skills, then the real magic happens. GPT -6 is indeed crazy. Like, look how much it was able to create, like different types of animation, retarget all this animation on these characters, do this procedural weapon animation. That's crazy. So, total for this game, I spent 1 .29 billion tokens. GPT -6 says that it was working like for around 50 hours, but in real life, it took me around 100 hours because I was slipping.
I was busy with some stuff, but I tried to make it work nonstop. I guess I'm less efficient than GPT -6. And what about assets and how much they cost? So, I recorded everything carefully. Here's what we actually did. 100 images, 50 models, around 35 rigs, audio and different stuff you can see here. Some videos, et cetera. Actually, I thought it will be much more. And if we try to calculate it, it's on your screen, the costs, it will be around 40 to 85 dollars.
I cannot tell you exact number because I was using Higgs field through my team account and it mixed it up with other usage. But let's say 100. I think yes, but only because it's not the smartest way to use it. When you do small iterations and build something meaningful, you actually spend way more or less. I do create my own dream game and I use all these APIs, et cetera. And I do not spend more. than. Five bucks a day creating this 3D models, like variety of them, like to build up my characters, like et cetera.
Now what about. Second game, which meant to be GTA 6? Well, spoiler, my friend actually spent twice more tokens, like two. Billions but he was using Unreal Engine 5. And not to mention how many times he went to me and cried how slow it is to work with Unreal. Engine And it's hard to use multiple chats in the same time because they both need access to Unreal Engine to recompile, et cetera, et. Cetera Anyway, here's the story.
He also started with the prototype. And here's the first prototype that took around a few hours to. Build Again, nothing really special, everything procedural, just base mechanics. And I really recommend you to start this way, because again, you. Can do the first proof of concept and add layers, layers, layers. Okay, good start. Nothing impressive yet. Then here comes. First big iteration, like first overnight cooking, right?
And here's the results. He actually started to build. Up the city and give it like better Miami look. Also, GPT -6 added day night cycle so we can see like. How it looks in day, how it looks at night. And don't look at this broken demo character. We're going to do characters a bit later. Now you can see that in. This version, cars are super generic and we cannot really go внутри the car, right? But GTA is about Grand Theft Auto, right?
So we must be able to take any car on the street and ride it. Therefore, that's exactly what my friend asked GPT -6 to do. And in order to achieve that, GPT -6 started to create procedural cars and their first version was actually quite awful. Not to say about this car entrance animation. It was terrible, but we are able to ride the car. And also there's some additional features such as NPC, cops, et cetera. And we already can get our ass kicked.
And by the way, the idea was to make cars destructible. So that's why they are modular. Like that's why they are procedural. And by the way, all characters here was made upon the matter humans. Why? Because this is Unreal Engine and GPT -6 has access to all this Unreal Engine stuff. This way it actually added some NPC on the streets and it starts to look more lively. Even though it's absolutely terrible. And at this stage, I think we had enough fun playing with procedural generation.
So let's plug in Higgs field with 3D AI, 2D AI, video AI, et cetera. And the first thing that needed to be fixed is car entrance animation, because this is a game about cars or stealing cars. And what actually worked is to create a video reference of person sitting внутри the car and giving this to chat GPT as a reference and asking it to do. Proper animation. It also found some cool library called GVHMR that helped it to sort of mock up this motion and.
Use it внутри the game. It was actually really challenging. And in the end, we finally get something more or less working. Same thing with car. Exit animation. And dang, that's quite cool approach. I actually want to use this library in my game now. Now, anyway, within that four days, of course,. There was much, much more cases and there was a bunch of different animations for NPC, et cetera. But just so you know what workflow was behind.
All of this. And of course, a lot of assets and not going I am still amazed how good GPT 6 understand spaces. And how good it is able to do the building scenes, rooms, interiors. Just look at these buildings that it. Made from 3D AI generated assets and not only exterior, but also the interior that actually makes sense. Look at this Baker and you know, like everything is in right place. I mean, dang, that's really, really cool.
I mean, I'm really surprised. And there are more buildings with complete walkable first floor, for example, smaller buildings like a surfing shop. And you can see, like, again, it created like a lot of small object, placed them around nicely, et cetera. And you must be wondering how. Good it looks внутри the game. Well, buildings actually look decent. Well, cars are the problem. And here's. the. Thing If we were asking to generate 3D models of cars, then it automatically lose the ability to be interactable because.
Then you need a door separate, et cetera. So to actually have this game mechanics in place, we had to stick. With these procedurally generated cars, which look terrible, to be honest. Well, it's still impressive that those cars actually works. So 2 .5 billion tokens after, well, he actually started trying to build this game right after GPT. 6 publicly released after more than four days, nonstop prompting GPT 6. We have this result.
And by the way, in this game, you can actually swim. If you ever played GTA, why city you can get the job so we can actually enter any car we can take it. We can drive. It We can do the shooting. We can get in trouble with the police. There are plenty of NPC and you know, they actually look like Miami. Style Like I look at this guy on the beach, definitely come here to swim. We have quite cool buildings that we can actually go внутри.
In some of them. Bad looking cars, but at least you can destroy it, right? Well, I definitely think there's a better way to spend 2 billion. Tokens and four days of your life. Same amount of tokens combining with this too can give you much better results. For example, I could. Spend 30 minutes generating car and doors separately and add this. to. The game. So GPT 6 can do the rigging and do the nice animation while only prompting having nice card seemed like.
Impossible task. Same goes for buildings. I'm not gonna lie, but the time GPT 6 spent to combine these. Buildings from parts, I can spend three times less time to do that. But maybe at least now, let's see where we get. Next year with GPT 7, Claude Fable 6, I don't know. And what about us? There was plenty of AI. 247 I really think that effort could be used elsewhere than generated sounds, radio tracks, motion reference.
Video altogether was 24 videos. One of the most expensive thing here generated 3d models, different types around 51. And other stuff. Approximate cost here can also be around 100 bucks. I guess we just don't have enough time to get. More hard on it and generate more assets. But that will just simply make no sense because overall it will not increase the quality. There are different problems. Really curious to know what you think about all of these guys.
Like and subscribe and I see you in the next video.
The words are the caption track's own and nothing is reworded or re-transcribed. Paragraph breaks are placed between sentences so the text reads as prose.
Free tools for your own script: paste a draft and see where it stands before you record it.
Paste your draft and see where viewers are likely to drop off, with a rewrite for each weak line.
Paste the first 30 seconds of your own draft for a hook score and rewrites.
Check your draft against YouTube's advertiser-friendly guidelines before you record it.
Read this channel's public videos and transcripts, and download a writing brief for it.