Transcript of GPT-6 Astra лучше Opus 5.5? Проверил на тех же тестах
Студия Игор
0:00Today we will test GPT6 Astra using the0:02 tests from the previous video. It will0:05 be interesting. Hello everyone and0:14 welcome to the Game Studio channel. My0:16 name is Igor, and today we are going to0:19 make a game. You really asked in the0:21 comments to also test GPT6 Astra to see0:24 how it can pass all the same tests I0:27 did in the last video. If you haven't0:30 watched the previous video, then pause0:31 and go watch it, because otherwise, it0:33 will likely be hard to understand what0:35 is being discussed. The link will be in0:37 the top right corner of the screen and0:39 in the description under the video. And0:43 the first test is making a car model.0:45 All prompts will be on the screen. You0:47 can pause if you want to look at them0:49 in more detail. I gave the exact same0:51 prompts to the Claude Opus 5.5 and GPT60:54 Sol models in the last video. Now on0:57 the screen, you can see the model0:59 building process, how GPT6 Astra does1:01 it. And now let's look at the result.1:05 And here is the result. And actually,1:07 the result turned out quite good,1:09 although I'm not sure if the1:11 proportions match the reference 100%.1:14 For comparison, you can look at the1:16 result from Opus 5.5 and compare it1:18 with the result from GPT6 Astra. You1:21 can see both results on the screen1:23 right now. Overall, it turned out1:25 pretty well. I can't say that GPT61:27 Astra was much worse in this test. And1:31 it seems the comparison will be1:32 interesting. So let's continue. By the1:35 way, time-wise, GPT6 Astra spent a1:37 little over 30 minutes on this test.1:39 That's about an hour less than it took1:43 Opus 5.5 to make the same model, and1:45 GPT6 Sol took about the same time. And1:50 the next test is one of my favorites.1:52 We will need to animate a humanoid1:54 character. I remind you that we made1:57 this character in Tripo AI, and it has1:59 no skeleton and no animations. I2:02 downloaded the animations themselves2:03 from the Mixamo site, and there, in2:05 principle, is everything except getting2:07 into a car, as well as turning the2:08 steering wheel left and right. This is2:10 the most interesting part. This is2:12 exactly what GPT6 Astra has to do now.2:15 So, the task is to animate the2:17 character model. We need to use the2:19 animation pack from Mixamo and add2:21 walking, idle, running, and jumping2:23 animations, as well as create the2:26 animation for getting into the car and2:28 steering wheel turns. And let's look at2:31 the result. This is the landing page2:33 with all the animations that GPT6 Astra2:35 prepared for us. I immediately see that2:38 the idle animation is actually not bad2:40 at all. We can also turn on the2:43 skeleton. And let's take a quick look2:45 at the other animations. This is what2:47 the walking looks like. Overall it's2:49 quite good, if you don't count the fact2:51 that the model leans back a little.2:53 This is what running looks like. The2:55 jump turned out a bit crooked and2:57 slightly resembles the result from GPT62:59 Sol. And now the most interesting part:3:02 getting into the car. And GPT6 Astra3:04 also used the real car model that will3:07 be in the game. Overall, I cannot say3:09 that the entry animation is done too3:11 badly. You can see that the character3:14 is trying to slide their legs. under3:17 the steering wheel and sit on the seat,3:19 not under or into the seat. But if you3:22 remember how the car entry animation3:24 was done in Opus 5.5, in my opinion,3:27 Opus did much better because it not3:29 only placed its foot on the step but3:31 also lowered its head when getting into3:34 the car. I mean the character. But with3:37 Astra, the character hangs in the air a3:40 bit when getting into the car, and the3:42 legs pass through the body. But on the3:45 whole, the result turned out, I think,3:47 average. And let's move on to the next3:50 test. To pay for subscriptions to3:52 various neural networks, you may need a3:54 bank card. That's why I want to tell3:57 you about the sponsor of this video,3:58 the company Zarub. It is a service4:00 where you can get a virtual card to pay4:02 for subscriptions to various services.4:05 And all this right in Telegram. No4:07 verification is required to issue a4:09 card. You can top up the balance with4:11 rubles via SBP. And if someone finds4:13 crypto more convenient, that option is4:15 also available here. After issuance,4:17 you get the card details and pay for4:19 the subscription yourself in your4:21 account. 3D Secure works for confirming4:24 payments with a one-time code. There is4:26 no monthly maintenance fee. You can see4:29 the issuance cost and commissions in4:31 the service itself. And if you need4:34 help with paying for a subscription,4:35 the bot has step-by-step instructions4:37 for various services. Get your card4:40 using the link in the description or4:41 scan the QR code on the screen, and use4:43 the promo code "Game Studio" to get a4:45 discount on opening the card. And the4:52 next task is to animate a dinosaur4:54 using video references. I have three4:57 videos in the project folder. These are4:59 walking, the dinosaur roaring, and the5:01 dinosaur attacking. Neural networks. It5:03 will be necessary to break all these5:06 videos into frames and animate the 3D5:09 model we made earlier in Tripa based on5:12 these frames. And let's look at the5:15 result. The result turned out to be5:17 quite decent. You can rotate the camera5:20, and I see that, on the whole, the5:22 idle animation is done quite well. The5:24 animations for when the dinosaur roars,5:26 attacks, and runs are also done quite5:28 well. Among the flaws, if one were to5:31 be really nitpicky, it can be noted5:33 that the rear part of the dinosaur5:35 turned out a bit static. This is5:37 especially visible in the roar5:38 animation. But overall, we consider5:40 that the neural network managed. GPT65:42 Astra had a good result in this test,5:45 especially if compared with GPT6 Sol.5:51 In total, we will need many 3D assets5:52 to assemble the game. I have already5:54 made some of them in Tripa, and some of5:56 them will need to be made by the neural5:58 network itself procedurally with code5:59 in Blender. We won't dwell on this task6:02 for long. Now you can see on the screen6:04 the result of the environment object6:06 generation that GPT6 Astra produced.6:09 The result is, actually, quite decent.6:12 Let's see how it will look in the game.6:15 And the next task is a sound library6:18 for our game. The point of this test is6:20 as follows. I prepared five videos of6:23 different lengths. In all these videos,6:25 a dinosaur is doing something.6:26 Sometimes it walks, sometimes it roars,6:29 runs, falls, and so on. The neural6:31 network needs to analyze the video,6:34 understand what is happening, and6:36 compile a sound library, which we will6:39 look at now. Later, we will use these6:41 sounds in the game. And now let's look6:44 at the result. The result turned out to6:48 be, actually, quite decent. All sounds6:50 are categorized. The neural network6:53 really gathered a fairly large sound6:55 library. And besides categorizing them,7:02 similar to Opus 5.5, it ran all the7:04 sounds through an analyzer. If you7:08 click the details button for any sound,7:11 you will see not only a description of7:14 what the sound is, but in some cases a7:17 note if the analyzer detected an7:19 anomaly. For example, footsteps mixed7:25 with a roar or something like that.7:28 That is, what to pay attention to if we7:30 want to check these sounds manually.7:32 The result turned out excellent. And7:34 let's move on to the next test. And the7:38 next task is to build a game level from7:41 primitives, or more accurately, a level7:43 prototype. In the future, the neural7:45 network will build the level in Unity7:47 based on this prototype. And let's look7:50 at the result immediately, without7:51 further ado. Overall, the result7:54 strongly resembles what GPT6 Soul had,7:56 especially from a presentation7:58 standpoint. Although there is already a8:01 certain cinematic quality that was also8:03 present in Opus 5.5. I specifically8:06 increased the playback speed a bit so8:09 we could evaluate the result faster. I8:12 immediately see that the road is8:14 winding, not straight like it was in8:16 GPT6 Soul, so it will most likely be8:18 more fun to play. And in many, many8:20 ways. Already now, specifically in the8:23 chase scene, I see similarities with8:25 the result of Opus 5.5. The level8:27 prototype turned out promising, and8:30 there are actually a lot of8:31 similarities in terms of presentation8:34 to how it was with GPT6 Soul. It is8:36 immediately clear that these are models8:38 from the same company. And now let's8:40 see how the neural network, based on8:42 this prototype, will assemble the final8:45 game in Unity using all the assets we8:47 prepared in the previous steps. If you8:51 like what I do, you can support the8:52 channel on Boosty, you can subscribe or8:55 make a one-time donation. It really8:57 helps the development of the channel.8:59 And we continue. And now let's take a9:02 look at the final game in Unity. And9:04 what would I like to say here first? It9:07 was not possible to make this game in9:09 one shot. The result, to put it mildly,9:11 was not very good. I didn't even bother9:13 recording it. And then I asked GPT-69:16 Astra to make the project quality even9:19 better. And after working for about9:22 another hour, Astra admitted that the9:25 project didn't work out. But overall,9:27 in her opinion, the quality of the game9:30 is quite decent now. And now let's play9:33 the game. I've already opened the9:34 project in Unity. This is what the9:36 scene looks like. Let's launch the game9:38. And I want to note right away that9:52 you cannot rotate the camera with the9:54 mouse. So, the game is quite linear in9:56 terms of movement right from the start.9:59 Next, there's this short cutscene where10:01 they introduce the dinosaur to us. I10:03 didn't see the dinosaur at first10:05 because it's sitting inside a rock.10:10 Then we need to quickly run to the car10:12 and start the engine. By the way, I10:14 intentionally didn't start the engine10:16 for a while, and you can see the10:17 dinosaur standing in the corner of the10:19 frame shaking its head. I want to10:28 mention the controls again. While the10:31 car is moving, you cannot steer it. It10:34 drives straight on its own, and we can10:36 only shift it left or right relative to10:38 the road. Well, and now I suggest you10:42 just watch the rest of this game to the10:44 end. And before we move on to the11:39 results of this comparison, a few words11:41 about the limits. In total, I spent11:44 about 17%of the weekly limit for the $11:46 200 subscription on the entire project.11:49 I remind you that for the same project,11:52 Opus 3.5...In the last video, it took11:54 me about 10%. And yes, the API cost for11:59 GPT-6 Astra is about 2.5 times higher12:02 than for Opus 3.5. If you want to know12:06 my subjective opinion, I believe that12:08 as of today, the model from Anthropic,12:11 Claude 3.5 Opus...It is the best model12:14 currently on the market. You saw the12:17 test results yourself. I will repeat12:19 once again that I am showing everything12:21 as it is and not exaggerating anything.12:23 And actually, in this video, GPT-612:25 Astra even had an advantage because it12:28 created the game in two attempts. Or12:30 rather, it created it first, and then12:32 improved it. And that's the comparison12:34 we ended up with. Give it a like if you12:36 enjoy these types of videos. Join the12:38 Discord, the link will be in the12:40 description. You can see the QR code on12:43 the screen. Leave comments on which12:45 game you liked the most. And also12:47 regarding the results of the two video12:49 comparisons. In your opinion, which12:51 model do you think is better today? And12:54 that's all for today. Thanks for12:56 watching, and bye everyone.
2,018 words · 289 lines
