Make Any Game With AI & 2 New Open Source Video Models!

Author: Theoretically Media · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Make Any Game With AI & 2 New Open Source Video Models! · Channel: Theoretically Media · 15:01 · Watch on YouTube

Topics: AI agents, AI content generation, Multimodal models, Privacy and data protection

Summary

The video covers two new open-source video models: Tencent’s Prism, which generates native 2K video with audio but requires substantial hardware, and Kandinsky 6.0, which can run on consumer-level hardware. It then explores AI-assisted game mods and browser-based fan games, including a Doom-and-Skyrim mashup, while noting the practical and copyright limitations. The host reviews early examples from Nano Banana 2.1 and discusses its image quality, text rendering, and consistency. Finally, he introduces Hark, an AI assistant from Figure, and briefly considers its pricing and privacy implications.

Key points

Tencent’s Prism generates 10-second, 24-fps video with native 2K resolution and 48K audio, but currently supports image-to-video rather than text-to-video.
Prism uses a 32B fine-tune of Mova and focuses attention on moving parts of a video; running it at 720p requires an 80GB graphics card.
Kandinsky 6.0 is another newly released open-source video model, with a lightweight version intended to run on an RTX 4090.
AI-assisted modding can combine the code and gameplay of existing titles, as in the host’s Doom-and-Skyrim mashup; users need to own the games involved.
Universal Modder can help combine games and may generate needed assets through an API, while coding agents can also be directed to do the work.
AI-created browser projects based on commercial games can attract copyright enforcement: a GTA V browser port was taken down, and Fallout New York received a cease-and-desist.
Nano Banana 2.1 is based on Gemini 3.6 Flash, with the host highlighting improvements in visual design, masked editing, and subject consistency.
Hark, an AI assistant built by Figure, offers a $100 plan and requests user information; the host suggests considering privacy and trying a free month if available.

Chapters

Start	Chapter	Summary
00:00	New open-source video models	The host introduces Tencent’s Prism and Kandinsky 6.0, comparing their video capabilities, hardware needs, and early availability.
04:06	AI meets video games	Examples of game mashups lead into the host’s Doom-and-Skyrim mod and an explanation of how AI coding tools can help build one.
08:15	Fan games and copyright risks	The video discusses browser-based GTA V and Fallout projects that were taken down or challenged by the rights holders.
09:51	Testing Nano Banana 2.1	The host reviews examples and personal tests of the image model, including its handling of text, prompts, and visual composition.
13:30	Introducing the Hark AI assistant	The host describes Figure’s Hark assistant, its pricing and trial offer, and considerations around giving AI agents access to personal information.

Quotes

Prism is capable of generating 2K with audio, natively — 00:30

this legitimately is Skyrim — 05:41

this is an AI generated mod — 06:42

be careful when you play with other people’s toys — 09:19

every company is going to be vying to be your new AI best friend — 14:31

Concepts and entities

Prism, Tencent, Mova, Kandinsky 6.0, Universal Modder, DOOM, Skyrim, Nano Banana 2.1, Gemini 3.6 Flash, Hark, Figure, Tim

Full transcript (original language)

00:00 Well, things are not slowing down in AI. Today, we’ve got a new open-source video model. This one’s uh from a pretty surprising source, and it does 2K. Plus, AI and video games are coming together like peanut butter and chocolate. I’ll show you the latest examples there, plus kind of a little mini tutorial on how I built Doom Rem, a like literal bash together of Doom in Skyrim. Uh plus, show you how you can mash any game together as well. All that and an early look and detangling of the latest narrow

00:30 banana update, and quick hit on the latest AI personality agent from a pretty interesting place. Kicking off in well, in kind of surprising news, we got a new open-source video model. Uh and this is an actual new one. It’s not a Min Max H3 fine-tune like we’ve been seeing all over the place. This one comes to us from Tencent of all places. And I do want to note, this one is early, but does look promising. Prism is capable of generating 2K with audio, natively,

01:00 instead of generating it like 720 and upscaling. Generations are about 10 seconds, 24 frames a second at 48K audio, which is actually really nice. Oddly, uh only image to video here, no text to video. Uh take a quick look. >> Al 369 >> in my bones too, my toes, yeah. I can see [music] the party is jump. >> Surprise. >> [laughter] >> Like this really softens the room a lot.

01:32 And there was just a lot of blank wall here. We had one piece over the sofa, but then >> I mean, so far we’ve caught marlin, cooper. I mean, a wahoo, we got slam. Why stop now? >> So, yeah. I mean, the demos do look pretty good. Although, I will note it’s probably still not at like C-ATS 2.5 level. Uh and as even they put it in the paper, the goal here isn’t to be C-ATS 2.5. The goal here is to narrow the gap. Under the hood, it’s actually a fine-tune of Mova, a 32B model from

02:02 January. With a bit of a new attention trick here, the trick here is that with Prism at least is that it only generates and attends to the parts of the video where things are moving, which is how they’re able to obtain that 2K raster size. I don’t know if that’s necessarily a new trick. I’m pretty sure that we’ve heard about that one before. But I guess you know, the point still stands. Attention is all you need as the old saying goes. Now, I do want to note while this is out and available, it’s it’s like actually one of the big boy

02:32 models. So at 720, you’d need an 80 gig card for this. So I mean that’s saying nothing about 2K. There is no quantized version yet or basically like community fork that has been taken down to a reasonable level. So you know, when I’m sure somebody’s working on it right now. So we’ll definitely see. I do want to note there was some controversy early on in people saying that the example videos were actually like real videos. Prism did defend themselves saying like no no, those are actually generated outputs. So

03:02 I mean this is one that I think that once we get our hands on it, we can definitely test it out and see. So for sure, Prism is one to watch. And breaking just as I’m recording this, it looks like we have a new open source video model. This one Kandinsky 6.0. So this just came out. I have not had time to mess around with it yet, but unlike Prism, this one is ready to roll on consumer level hardware. They note here that the light SD version will work on an RTX 4090 and all the way up to the pro full HD version as well.

03:35 So I mean obviously, you know, the the better card you have, the slower these times get. But you know, you can run it. And if you just want to try it out, skip the whole local install. The gang at Fall already has it up. They’re so quick with these things. Looks like it runs at least on the image to video side, a dollar 40 per 5 seconds. So, uh um yeah, I mean, I’ll I’ll probably run the local version. But, overall, you know, hey, two new open-source video models, I mean, that is reason to celebrate. Quick little in-betweener before moving into the next section, uh something kind of cool is

04:06 coming down the pipe. Can’t talk about it right now, but I did get the okay to tease it a bit. Uh so, let’s uh move over to this side of the desk and hear from desktop flamethrower girl. >> [sighs and snorts] >> I’m so glad I finally got out of that monitor. Though, I definitely think I should be taller than this. >> So, I’ll have more details on that after it releases. Uh should be sometime this week. I don’t think it’s exactly what you’re thinking it is. It’s just But, it is pretty cool. Moving on. Lately, it does feel that AI and video games have

04:37 begun to hit their stride as we found ourselves in a place where we’re moving out of that initial wave of AI slop games and into real playable experiences. Well, I mean, almost at least. We’ll circle back to that in a second. But, in the meantime, another branch of all of this that’s really starting to emerge is using the latest models to create some like really wild mods, like really wild. For example, putting Spider-Man into Arkham Knight and letting our boy Peter swing all around Gotham, which actually is just delightful. Anybody that has played the

05:09 uh Spider-Man games knows that it’s probably one of the few open-world games where you you don’t fast travel. You just You just swing everywhere. It’s so much fun. Another cool example was putting uh the game Skate 3 into Call of Duty. Um and what’s important to note here is like that is Call of Duty and that is Skate 3. So, all of like the physics and controls of Skate 3 uh are implemented in Call of Duty. All of this just makes me want to scream like, “Ramirez, do a kickflip.” All of which inspired me to put my own spin on it, creating Doomrim or Sky Doom. I’m not

05:41 I’m actually not sure what to call it. But, yeah, basically, I mean, this is Skyrim. Only, instead of being the typical sneaky archer that I always find myself being classed as I am Doom guy. And again, what’s like super stupid fun about this is that this legitimately is Skyrim. All of the NPCs are present. It’s just that Doom monsters spawn everywhere and well, I mean, I have access to a chain gun. What’s kind of even crazier about this whole thing and I actually did not expect this, but the Skyrim NPCs will actually react and attack to the Doom monsters as well. I

06:12 mean, granted, everybody is at like aggro level 100% all of the time, so you can’t actually talk to the Skyrim NPCs. But yeah, I mean, they do they do fight off the the Doom horde, which is pretty interesting. And they can also all defeat or be defeated by the Doom monsters. Yeah, you used to be an adventurer like me until you took a rail gun to the knee. Now, an important distinction to make here is that like these aren’t AI generated versions of Skyrim meets Doom or Skate 3 meets Call of Duty or even Spider-Man meets Arkham

06:42 City. Rather, this is an AI generated mod, which essentially takes the the code of both games and mashes them together. And on a practical note, you do need to own copies of both games or of the games that you’re trying to mash together. So, yes, for this particular experiment, I did have to buy Skyrim yet again. It’s like my 13th copy of the game. You’re welcome, Todd. Now, if you want to try any of this out, there is a pretty useful GitHub repo for this called Universal Modder. This does some interesting things, although you may not have to use it. So,

07:13 Universal Modder is skill, so I mean, you could use Cloud Code, CodeX, Gemini, you know, essentially whatever you want and point it at it. It does have a pretty interesting little like side quest, let’s call it, in that if it needs assets for the game it’s trying to mash together, you can put in a fall API key and it’ll essentially, you know, go to fall and generate those assets either in like GPT image 2, Nano Banana, Flux, or, you know, whatever. So, that’s interesting. Although, for the model that I was using, I just I, you know, I

07:43 pointed it over to Universal Auto and was like, “Hey, I want to you know, I want to mash up Doom and Skyrim. Do we need any of this?” And it was basically like, “Yeah, I got the general shape of it. We don’t need it. Let’s just go.” From there, you know, basically let it grind for about an hour and you’ve pretty much got a playable game. You should be able to handle doing any of this on like Opus 5.5 or like GPT 6.1 Soul. Now, is this a stupid game? Yes, it is. Did it keep me entertained for about an hour? Yes, it also did. Now, on the other side of all of this is games being ported or like entire spin-offs

08:15 that are being created using AI models. But, I do have to say, I mean, maybe maybe watch yourself when it comes to this. For example, someone who as of yet has gone uncredited ported Grand Theft Auto V to a web browser. Which when you think about it is kind of insane. Essentially, they ended up compressing all of the game’s assets down to about 700 megabytes and then the whole thing runs in a web GPU. I mean, that’s kind of crazy. Naturally, this did not last very long and like the whole thing was quickly taken down by Rockstar. And in a

08:47 fairly similar story, a friend of the channel, Chris First, spent about a week building Fallout New York with Claude. I was actually talking to Chris while he was building it. I mean, this thing this thing is pretty crazy. Yeah, there were 849 unique textures. He did all kinds of collider work in it so that like literally you could walk across the Brooklyn Bridge and there were real quests built into it. And again, all of it was playable in a browser. I do say all of it was playable in a browser, past tense, because well, I mean, of course Bethesda did issue him a cease

09:19 and desist. Which look, I mean, they are fully in their rights to do. You know, while modding is generally tolerated, I mean, that is a whole can of worms that I’m not opening in this video, but you know, overall, what I would say is if you’re thinking about creating a fan game, just keep in mind, be careful when you play with other people’s toys. That said, I mean, well, Chris did just post this image up, so I mean, well, I don’t know. I guess he enjoys talking to lawyers. Moving on. Well, apparently, Nano Banana 2.1 has released. Now, this one has been a bit of a chaotic rollout

09:51 even by Google standards. Um previously, it was being spotted in Google Flow, and uh I ran a couple of tests there, and then it got taken away from us. Uh but now it is out in, you know, wide release. Google knows that there are noticeable leaps in visual design, masked-based editing, and subject consistency um for more natural-looking images. That said, yes, I mean, in terms of the model dependencies, uh Nano Banana 2.1 is based on Gemini 3.6 Flash. So, yeah, it is not a uh a pro model yet. Friend of

10:23 the channel, Brent Lynch, took it out for a test peel uh with this, I guess, like, you know, thriller movie poster. Uh text looks pretty good. Um I always do appreciate the uh the the title blocks, particularly that uh it’s a film by Alan Smithee, and the music is by Tangerine Dream. That’s that’s that’s pretty nice for an ’80s movie. Mark K tested it out with a prompt stress test here, uh essentially seeing how well 2.0 would handle, as I you know, a lot of called-out details in his prompt. He actually didn’t provide the prompt, though. But uh you can tell that uh you

10:54 know, mixed lighting is in here, um a pretty good amount of um text written, um and uh reflections as well. He did call that out. Um I presume as well we have the umbrellas here in the backpacks on the table. So, and then on top of it, the image is crammed uh with a lot of people. I will say that there is, in general, I think it actually does a pretty good job not knowing what the original prompt here is. But uh the mixed lighting conditions are a little possibly a little bit off. For another prompt stress test that we do have the

11:26 prompt for, uh Synopsis took it out with this following prompt, which you know, I do think it it does pretty well here. I think it’s actually a very nice image. It does call out for things like 24 people. I don’t know if there are 24 people here. If somebody wants to pause and count, they are more than welcome to. But I mean overall, while I think that it’s a nice stylized image, it does seem to have taken that like symmetrical aspect in the prompt really pretty seriously considering that all of the subjects do kind of it all just has a very like kind of like Wes Anderson

11:57 symmetrical feel to it. Speaking of kind of that Wes Anderson look, Maximus gives us this image which I mean I don’t know, I just like it. I like it. So no notes on this one. Running my own quick test here, I started it off with actually an old prompt that we used to run here. Create a photo realistic book cover for a supernatural thriller set in the Pacific Northwest. The book is called Alan Wake. And then as a tribute to our former generations, the author is Don Mindy. That was actually an in joke from like I

12:28 think like Ideogram 2.0 which put Don Mindy’s name in as garbled text. I loved it. Don Mindy is now the official author of Alan Wake the novelization. Moving a little deeper into that test, I wanted to test for text which is really hard to say. So I had it generate the like chapter one of the Alan Wake book itself. I do note that chapter one seems to be like in the middle of the book, but that’s beside the point. What was interesting to me is that in the text, it actually

12:58 calls out Bright Falls which is the town from the game. So yeah, it clearly actually I’m not even sure it’s been a long time since I’ve played it, but it does feel like this is the opening narration to the game. So yeah, kind of interesting. So overall look, I know that a 0.1 flash update is not the thing that everyone has been clamoring for. I will say that you know, Gemini 4 Argon is coming soon. So hypothetically that should mean a big brain update to Gemini image. So, hopefully we are going to have a big

13:30 banana bread update very soon. Um I presume it’s in the oven right now, so I will be keeping an eye out for it. Or in the case of banana bread, I will be keeping my nose out for it. Rounding out, we have another new interesting AI assistant that is on the loose. Uh another one that’s asking for all of your information so it can be helpful to you. Uh yes, this is another one of those like kind of like dot or Muse type deals. So, Hark is actually being built by Figure. Um so, the robot people. So, um you know, your your AI butler is actually being built by the people who

14:00 are building the AI butlers. I always joke that these AI agents are pretty handy, but you know, they can’t do your laundry. I guess you know, Hark via Figure will be able to. So, Hark looks like it’s launching with a $100 plan. Although, if you sneak in early enough, um you might be able to get a free $100 month um to check it out. So, I did sign up for it and uh and when I logged in, I you know, I got a I got a free account to check it out. Um I haven’t really put it through its paces just quite yet, but I will say that if you haven’t played around with any of these AI agents as of yet, I mean

14:31 you should you should definitely give it a shot. Um if you’re uncomfortable with the privacy aspects of it, set up like a burner Gmail account and point it at that. So, you know, I mean I do keep saying it like, you know, this is going to be the next big push is that every company is going to be vying to be your new AI best friend. Uh there is a reason that Figure is making this play right now. I will say that I don’t know if going with the $100 pro plan right off the bat is the way to go, but um you know, they are they seem to be offering a free month. So, um if you want to try

15:01 it out, the link is down below. In the meantime, I’m going to have my robot butler check in on that banana bread in the oven. Oh, you know what? It can’t smell, can’t it? Well, look at that. Humans one, AI zero. As always, I thank you for watching. My name is Tim.