Transcript of AI와 함께 젤다풍 월드모델 만들기 #shorts
홍정모
0:00Oh, what is this? The enemies appearing0:01 are scarier. It's scarier than the0:03 original game. It has a bit of a horror0:05 vibe, and it doesn't respond very well.0:07 Sometimes the input doesn't register0:09 when I press a key. Well, it turns out0:11 okay, I guess. Yes. So, I haven't0:14 trained it all that much. If I use a0:17 more sophisticated model and train it0:20 for a longer time, I think the results0:22 on the right would come out much better0:24. So, confirming that it works like0:27 this is meaningful when studying. Oh,0:30 the sword, it reacts. When I swing the0:32 sword, an enemy appears. The reason for0:34 this is that it doesn't understand0:35 cause and effect. We swing the sword0:38 when an enemy is there. But to this0:41 model, based on the training data, the0:43 enemy and the sword are always together0:45. Since it's common to swing a sword0:47 only when an enemy is present, if0:49 there's a sword, it draws the enemy0:50 along with it. Oh, that's interesting.0:53 Oh, two characters appeared on the left0:54. I think things like this are just a0:56 limitation of the training. Well,0:58 seeing the limitations first and then1:00 improving upon them is the standard1:02 path. For example, if it's a Zelda1:04 style, we can quickly whip up a1:05 Zelda-style game. Extract data from1:07 there, train the neural network, render1:10 the results, and walk around in this1:12 world model—yes, that's what I plan1:15 to do today. Selecting, making the game1:18, data collection, and verifying the1:19 training; it organized everything1:21 neatly. Well, right away, teacher, I am1:24 starting to make the adventure game.1:26>> Adventure game creation is complete.1:28>> Since the game is made, let's take a1:29 look. Shall we look at the report for a1:31 second? I have launched the game window1:32. You can try it out right away. Ah,1:34 the game is working.1:35>> Teacher, remote data generation and1:37 training are starting.1:38>> Data generation and training are1:40 running on the PC with a 4090. 16 bots1:43 collected 200,000 frames from different1:46 worlds in 40 seconds. It says it1:48 defeated 3,950 enemies. Diffusion model1:51 training is running at 14 steps per1:53 second. It's already running. Oh, the1:55 GPU is running at 100%. This right here1:57 is the PC with the 4090 GPU. Starting2:00 preparations for the diffusion model2:01 test.2:02>> Looks like the diffusion training is2:03 finished. Oh, what is this? The enemies2:05 appearing are scarier. It's scarier2:06 than the original game. It has a bit of2:08 a horror vibe. It doesn't respond very2:11 well. Oh, it says the training took 262:13 minutes. It's honestly hard to get good2:15 results with just this much. Now,2:16 explain the world model. I've gone2:19 through everything from creating a new2:21 Zelda-style game, training it on a 40902:22 PC, running the demo on a laptop, to an2:24 easy explanation of world models. Live2:27 mode, Tiny World, Zelda review, similar2:30 project research, Ninja Adventure2:31 assets, game creation, increasing2:33 enemies, connecting bots, training data2:35 generation and training, reviewing2:37 training data samples. Wow, it...it2:40 actually remembers everything I asked2:42 it to do. Prepared for diffusion model2:44 tests, analyzed easy-to-understand2:46 documentation on diffusion models, and2:48 built a world model. Personally, I find2:50 doing these kinds of things quite fun.2:52 I really hope you all try doing this2:54 too.
554 words · 86 lines
![[선데이 라이브] 당신도 AI 시대에 살아남을 수 있다!](https://i.ytimg.com/vi/D95UC9U5ShA/hqdefault.jpg)
