Transcript
00:00:00A single prompt, 50,000 lines of code, and the end result is an impressive Call of Duty game
00:00:04with no external assets used. Everything you're looking at was generated by Opus 5
00:00:09with a prompt about four sentences long. This prompt has now been called the Gauntlet Loop,
00:00:14showcased by Matt Schumer, and yes, it will burn through your entire subscription and tokens,
00:00:19but it has the model show some really cool behavior. I actually ran it myself and spent
00:00:23about $1,000, so let's find out if the final product is worth it.
00:00:30So this all started from a tweet from Matt when Opus 5 was first released. He showcased a first
00:00:36person shooter game with a mini-map working in the corner, player movement, sliding, damage,
00:00:40all of that seemed to work, and everything was claimed to be generated only by Claude. This
00:00:45was obviously met with a lot of skepticism at first, and then he actually published the prompt that he
00:00:49used, and people started to run it on their own and realized that Opus 5 is actually pretty good
00:00:54at this type of work. So quickly, a lot more examples showed up. There's this karting game
00:00:58with some genuinely impressive models and quite a cool environment, and it seems pretty functional
00:01:02from my small testing. There was also this pastel version of Nuketown, and everything seems to be
00:01:07working, and this would probably be a fun game for five or ten minutes. There was also this one
00:01:11called Descent, which has some really impressive graphics that were all made from Opus 5, like the
00:01:16background mountain looks really good for a model-generated asset, but I've not been able to
00:01:20actually play this game besides what's happening with the character here. You can view and play all of
00:01:24these demos on Matt's site. I'll leave a link down below, and I know these aren't AAA games or
00:01:29probably anywhere near bug-free, but for a first pass, they're incredibly impressive, and it's pretty
00:01:33cool to be able to create your own friend-slop games that you'll probably only play for 15 minutes and
00:01:38throw them away, and this would have been so cool to me as a kid. I actually had to learn Java back in
00:01:42my day so I could mod Minecraft. Now, you can just prompt what you want into existence, as long as you have a
00:01:47flawed subscription on the max plan, so maybe not great for kids. So what is the prompt that yields such
00:01:52great success? What is the gauntlet loop? Well, originally, before it had a name, the prompt that made the
00:01:57Call of Duty game was these few lines. This prompt gives a lead agent two things, a goal and a real-world
00:02:03target, in this case being Call of Duty. After this, you then ask it to use sub-agents, where each agent can
00:02:08loop over a small piece of the overall game, and it has a flow where it codes, and then that goes to a harsh
00:02:13critic, which actually looks at the screenshots, and the feedback from that critic is then fed back to
00:02:17the original agent, looping until the critic thinks it's perfect. This will literally keep going until
00:02:22the sub-agents reach the limit of what they can do, and for me personally, it looped for 19 hours
00:02:27until I stepped in myself and stopped it. Now, that was the original prompt that Matt used, but now the
00:02:32gauntlet loop actually has a meta prompt. You pass this to an agent and ask it to make the final prompt
00:02:36for you. When I ran my test, I actually only had that original prompt, so I simply modified mine,
00:02:41asking for a Formula 1 game instead of Call of Duty, and that is the only change that I made from the
00:02:46original prompt. So let's take a look at what Claw did. First, it actually created its own tools to go
00:02:50through the game frame by frame and get screenshots from multiple camera angles to pass it to the
00:02:55critic step, and then it started a workflow with 30 agents and 5 phases. One agent works on the
00:03:00foundation to write the core engine, and then 15 of them work on systems, building things like the
00:03:05environment, the models, and the camera effects. The third step is then integrate, so one agent takes all of the
00:03:11agents and just ensures that it all works well together, and then this can be passed to the
00:03:15critic agents, which is described as five harsh art directors looking at things like silhouettes,
00:03:20materials, lighting, world density, and motion, and it says here that the critics want a score of 90
00:03:25or more, meaning that it's indistinguishable from the original game that I asked for. Any issues that
00:03:30these critics find up until that point, and then pass back to the start, and this loops over and over.
00:03:35As an example of this, I actually got to round 5 on mine, and we can see that the critics broke this
00:03:39scoring into car, world, and motion, and round and round the score is improving. I actually stopped
00:03:44Claude at round 5 here as I ran out of my subscription and also time, but it reached a final overall score
00:03:49of 67.3, and Claude actually told me itself that this goal may be a little unrealistic and that the
00:03:55scores were plateaued in the 70s anyways. Now I won't go into this entire thread as it actually took 19
00:03:59hours, but along the way the model did actually showcase some really cool behavior, building its own
00:04:03mini harness. It made three permanent tools, one to actually screenshot the game while running,
00:04:08one to extract information about the running game like shadow map state, the lighting, and whether
00:04:12it casts, permiss shadow count, sun elevation, and material types, then also a tool to compare two
00:04:17screenshots between rounds and ensure there'd be no regressions. Alongside these, it also made 136
00:04:22single purpose tools. This was things like a road material A-B testing tool between runs, and also a
00:04:28play tool so it could actually drive the car through a scripted input sequence. I just think it's super
00:04:32impressive how these models built their own little harness, and in total it used over 2,000 screenshots.
00:04:37Overall, by the time I stopped this, it was up to $1,200 in usage, and as I said earlier, that took 19
00:04:43hours, and it used 1.7 billion tokens across 137 agents. So the big question now is, did we actually
00:04:50get a good game from this? Well this is the final result. It called itself Apex GP, and I've got to say
00:04:55just from the background here alone, these models do look pretty impressive. We have four features, we have a
00:05:00race, we can do a full race weekend, we can do a time trial, and then we also have the settings. In here
00:05:05we have things like the camera, the heads-up display, as well as the difficulty level that we're on. Now
00:05:09I'm not going to do a full race weekend, we'll take a look at what a quick race is, but these all have
00:05:13the same setup menus anyway. You can see we can select our driver on our cars here, and it's seemingly
00:05:18used some real drivers. I'm going to go ahead and select Norris. We can choose a different livery
00:05:22here. I'm going to choose heritage. We can't choose a circuit, it only has one. We can choose some weather,
00:05:27we can choose the difficulty here, as well as the race distance, and also the starting grid. I'm just
00:05:31going to go ahead and jump in though, and take a look at the graphics. Now a little bit of jank
00:05:35there where all of the cars started together, but you can see it's actually driving this for me on the
00:05:39formation lap. I'm not controlling this at the moment, so I assume at some point I'm going to
00:05:43take over, maybe when we get to the grid. But I've got to say, these graphics are pretty impressive.
00:05:48I am super impressed with the model of an F1 car that is generated here. This actually looks
00:05:52really good. It's definitely better than I could do by hand. I have sort of no experience in 3GS
00:05:56modeling, and even the road textures and all of the people at the side looks pretty impressive
00:06:01for a 3GS game. This is actually kind of blowing my mind at how good this looked. I saw a few of the
00:06:06screenshots as it was going. I haven't seen the final result yet. Now it appears that we lined up on the
00:06:10grid here, although I'm definitely out of my grid spot, but the race is about to start. And I can take
00:06:15over the controls a bit, but it is doing a lot of the steering for me. And also, the steering is
00:06:20very sensitive. So literally pressing one key here moves me the entire way across the track. But we
00:06:25can see that the leaderboard up here is working. We have our minimap working down here. That's some
00:06:30really nice sort of graphics down here where we have the sparks coming off the car. That's a really
00:06:34cool, nice feature. We've got shadows working. Motion blur is probably a bit too much, and also
00:06:39controlling the car is a little bit difficult. I'm doing a fairly good job. And let's see. Okay,
00:06:45when I crash, I do go through the barriers. So that definitely needs work. And now I'm just
00:06:50going to let the car take over and do its own driving because I've caused a major crash here.
00:06:54This is definitely a safety car situation. I've gone ahead and restarted that race as I want to
00:06:58see what a few of these other camera angles look like. So if we go into the settings, we can choose
00:07:02the cockpit one. How well does this look? This looks okay. Needs a little bit of work maybe on the
00:07:07steering, although we do have a steering wheel. And the halo is a bit lower than it actually is in
00:07:12real life. We go ahead and change these settings again. We can go to the halo view, which is just a
00:07:17little bit higher. We've also got the TV view, which is pretty cinematic. Although I couldn't race from
00:07:21this, I suppose. But you can see these cars do look fairly good. Obviously not perfect F1 cars. But
00:07:27again, for something generated entirely by Opus 5, I think that's pretty impressive. Now, this is
00:07:32obviously a bit of a bug. I think we're now inside a building in that TV mode. So let's go ahead and
00:07:36change this to, let's say, a wide angle. Although again, this is going to be pretty undriveable from an
00:07:42actual usable, playable point of view. But it's still pretty cool that it added this in. Now, it looks
00:07:47like I might be out of line here. But the rest of the cars do actually line up on the grid, which is
00:07:51nice. And again, it has tried to copy some of the real F1 liveries. So we have Mercedes at the front
00:07:56there or maybe Aston Martin. I could also see a few Ferraris back there. And it looks like there may
00:08:01have been a major crash on the start line. Now, I've jumped into a time trial here just to see what it
00:08:05looks like under heavy rain conditions. And it's definitely not perfect. The environment just looks
00:08:10overall cloudy and a bit too gray. But the spray coming off the cars does look okay. Again,
00:08:15definitely not a perfect situation. But we do have some puddles on the road as well. Again,
00:08:20it's done a pretty good job from a single prompt. Now, I'm going to go ahead and actually try and
00:08:23finish a race here. But I did just notice at the top that we did get a penalty for jumping the start.
00:08:28And we actually get some track limits when I go off the track as well, which is super cool. And
00:08:32I'm definitely going to get a penalty for track limits as well. But I also want to finish the race
00:08:36and see what happens then. So this is the final lap now and I should be crossing the line. I did finish
00:08:41first, but I have a lot of penalties. So I'm curious how it's going to take that into account.
00:08:45And also we have a podium. It seems like nothing happens when you cross the line. It just keeps
00:08:50going. I thought you'd have to wait for everyone to cross the line. But even then, it's just still
00:08:54playing the game and we're actually doing extra laps here. So I don't think it implemented any sort of
00:08:58finishing screen or a podium. Overall, I'm pretty damn impressed with the results that we got back here.
00:09:03At the moment, I'm just leaving this on the TV view. Normally when I run this, I get really blocky
00:09:07looking cars, but these actually look pretty good and I can see that this is supposed to be an F1 car.
00:09:12And the same goes for the track and the environment. It's done pretty well on all of those things for
00:09:16something that's entirely generated from Opus 5 with no external assets. Would I play this game for
00:09:21more than five minutes? Absolutely not in its current state, but I think a few more prompts and I could get
00:09:25it into a bit more of a working state with working barriers and maybe some better handling. It's
00:09:30definitely a lot better than I could do in 24 hours myself. So would I ever make my money back on that game?
00:09:35Well, no, but luckily this was all on a subscription, so it was kind of free and I uploaded all of the code
00:09:41that this generated to a GitHub repo if you actually want to run this and give it a go. But I personally
00:09:45haven't looked a single line of the actual game code here. It was just a fun experiment and I'm super
00:09:50impressed with the results that we got from the models that it generated. What do you think? Let me know
00:09:54in the comments down below or either subscribe and as always see you in the next one.
Community Posts
No posts yet. Be the first to write about this video!
Write about this video