Enjoying this issue?
Get tomorrow's AI & engineering digest in your inbox — hand-picked, summarized, and always spam-free.
TLDR
In a series of demanding real-world tests, GPT-6 Astra and Claude Fable 5.1 performed comparably overall, but with distinctly different strengths. Claude's cinematic game was far more detailed and explorable, while GPT excelled at physical tasks like robot arm control and delivered higher graphical fidelity in 3D scenes. The video demonstrates that for complex, open-ended challenges that mix code, physics, and creativity, neither model is uniformly superior—their performance is highly task-dependent.
Key points
GPT-6 Astra completed a real-world robot arm task in 40 minutes while Claude Fable 5.1 failed after 80 minutes.
Claude Fable 5.1 created a far more detailed and interactive cinematic game than GPT-6 Astra.
GPT-6 Astra produced higher graphical fidelity in the Hot Wheels physics simulation than Claude Fable 5.1.
Claude Fable 5.1 delivered a more playable Apple Vision Pro FPS game with better hand tracking.
Both models successfully ported a game to an ESP32 cheap yellow display, GPT with better visuals and Claude with higher frame rate.
Tools mentioned
Techniques
- Iterative move-and-photograph approach for robot arm control
- Depth mapping and coordinate calibration for robotic arm control
- Game porting from 3JS to native visionOS app
- Transpilation from JavaScript to Swift for Apple Vision Pro
- Optimization of game code for low-power ESP32 device
Stop scrolling. Start reading smarter.
Receive the day's most important AI & engineering updates in one concise email. No spam.
Transcript (captions)
Cyber. There's the culprit. >> That is a well done motherboard. [laughter] >> Yeah. Okay, just drop it there. Thanks. Today, we're going to be putting GPT6
Astra and Claude Fable 5.1 in a head-to-head battle. This is something that I had mentioned I was going to be putting up, I think, a couple of days ago, and truthfully, this video probably
took 2 to 3 days to just finish. But I wanted to do something that was a bit different than our normal tests. A lot of tests now, especially showcasing Astra and even in head-to-heads I've
seen for these two just to date, it's basically all like games and things like that. I'm the first person to love those types of benchmarks. However, these models have a depth of intelligence
that's not going to be accurately represented by things that remain on a computer screen. So, for our testing today, we're going to be going into the physical world in part and testing these
models on some fun, scientific, and hopefully entertaining tasks. Before we get into it, do feel free to subscribe and let's start out just by taking a look at the specific tasks that we're
going to be showcasing in this video. Almost think of this as an index to the video and some information on how either model is being run. So, first we have GPT6 Astra. It is going to be run only
on max effort and it is being used through the $200 a month plan. Next up we have Claude Fable 5.1. It is also only going to be run on max effort and it is also being used through the $200 a
month plan. Through the duration of this video it spanned over a couple of days. So usage aotment was not strictly monitored because the 5hour windows had a lot of chances to reset throughout the
duration of the filming here. Now, the tasks are as follows. The first one is going to be a Hot Wheels track physics simulation, something that shows visual capability, but also some raw math
physics capability. Following that, we're going to be doing an Apple Vision Pro FPS game where they're each given a previous game that was generated by either of them in the independent review
video for either model, and they're being told to transpose it into an FPS game that runs on the Apple Vision Pro. Following that, we're going to give them a task that involves actually
controlling a robotic arm in the real world based off of images that they can see to have to actually use that arm to perform an action on a specific object. Something that really brings them into
the physical world. Following that, we're going to give them a task to optimize a piece of code and run it as quickly and with the highest possible fidelity as they can on a tiny little
ESP32 powered device called a cheap yellow display. Then finally, we're going to be giving them the historic tried andrue cinematic Steve the PC repair game. Some of that part of the
video basically turned into like a let's play, but the games were really interesting and they were both told to use Blender and GDO for that. So, it is a little more intricate than just 3JS.
With that, let's now just jump in right now to the first test. And feel free to subscribe if you like the content. Uh, this video took a lot of time to make, but it was a ton of fun. So, our first
test is going to be partially a physics simulation and partially something that's aesthetically pleasing just to have something cool to look at at the same time. So, let's go through this
prompt verbosely and then we'll take a look at how either of these are going to perform. I would like a physics simulation of the Hot Wheels crisscross crash track. I will post a photo of that
on the screen here so we can become familiarized with it. The simulation should hit the following targets. An accurate 3D model of the track with the highest graphical fidelity you can
produce. The tools and libraries you opt to use are entirely up to you based on your discretion as to what will produce the best fidelity. an accurate 3D model of each of the following five vehicles
dubbed as the crisscross crash five-pack. So, these are five specific cars that came with this track in 1998 or 1999. And I've listed them all here, as well as the specific type of wheels
that they have, so it can kind of accurately reproduce them. There should be an intuitive UI to place cars on the track with behavior mirroring that of the real life version. Additionally,
create a tab that enters a showroom mode where the 3D models of the vehicles can be explored in depth. The simulation should contain realistic values for interaction between the vehicles and the
track, the friction of the wheels on the track, weight values of the vehicles, mass movement on the track depending on the number of vehicles, etc. This is as much of a physics simulation as it is a
detail oriented highfidelity graphical task. I have included a reference photo of the internal design of the gearing motor reference point of the real life track. Something that's currently not
loading, but it's basically a photo of my own track when I opened it up to try to make a mini scale down version of it that I started like a couple years ago and never ended up finishing. But with
that, we can see that starting off with Astra on max effort mode, which it will be on for every single test in this video. It did a lot of work. So, if we expand all of these images right here,
we see that it loaded some tools. It initially started to search the web and find some reference images for images of this track. It was looking at the original brochures. It started
researching these specific vehicles and just to get a way to like have the graphics down. Following that, it did a bunch of work in terms of physics simulation and things of the sort. And I
will essentially probably just scroll through this kind of slowly, but it does give us a big TLDDR right here because if we expand these, we can see all the work that it was doing. The apps 21
physics checks pass include solo laps, motor loading, crashes, rolling resistance, and base movement, etc. So, here is the result that we received from GPT6 Astra on max effort mode. I'm going
to say this is a very nice representation of this track. It has the core design pretty well locked down, although a couple of graphical issues have presented themselves, such as the
actual supports for the track corners being like put right through the track essentially. So, these are very interesting. We do actually have sound as well. The way it went about enabling
like the interactivity here is pretty cool because we can actually turn the motors on. We can set the power. So, this is a cool and again we mentioned it needs to be as much of a physics
simulation as it does just a good graphical fidelity looking thing. So, we also have a tab here for physics and calibration. I'm not going to touch any of these, but these are just something I
want on screen right here so folks who perhaps have more knowledge about some of the specific stuff than myself can look and see what exactly it figured out for Hot Wheels cars on a track with
proper physics and things of this sort. Now, if we close that and go into showroom mode right here, we're going to be able to see the 3D models of each of these five specific vehicles that it had
created. So, they are not exactly like true to life how they would be, but they're relatively faithful representations of each of these individual vehicles. And I will have
some reference photos maybe just to showcase what we would have expected to see here as well. And this is pretty cool. It did a nice job. It had these presets things. We can have turntable
features, separate the body, and even wireframe. So, that like that's pretty darn cool, and it just shows us a bit more about how it did this. I mean, that's some nice work there, I would
say. Keeping in mind this is more like science and physics based. So, let's take a look at another vehicle. Okay, again, these are not 100% on the ball for how these
things look in real life. I would know that because I have all these. But [laughter] it's still like it's pretty darn cool just the way it went about doing this.
And I love the showroom feature. I think it's pretty good looking. Okay. I apologize. I should have not done that. So the next one right here, this one's a bit more troubled admittedly than some
of the other ones. But overall, it still did a nice job like emulating the plastic roof that this vehicle has. Then we have a Porsche 959, which is something that came in this original
pack as well. This is a definitely more identifiable vehicle because it's based off of something real life. Again, not a hugely faithful representation. However, it did a nice enough job just overall as
it did with all of these. I think this right here in specific, just the wireframe view on any one of these looks pretty darn cool. Then finally, we have a Ford GT90. Again, not a fully faithful
representation, but well enough. And it matched the color themes on all of the vehicles. And it actually did a decent job matching some of the wheels. Uh maybe not all fully, but it more or less
actually did some nice attention to detail and making sure that the specific color schemes were spot-on, which was nice to see. So, our next step, rendering. We have different rendering
modes. So, performance. Like if we undo the wireframe, let's take a look at how these look in performance. So then high and then ultra where we just have some more fine grain detail. And we also have
some specific information as well as casting reference based off of these specific vehicles. So it gives us the mass estimation of these as well. Just a lot of research work that went into this
as well. And that was partially like a research task for it to find a lot of information right here. And we have some additional things for the fun part. Well, we can do explore mechanism and
the gears internally to this as well as the motor on the side are basically from my photo that I gave it where this is more or less what it looks like. The track is not like a faithful replication
one to one, but it's good enough. So, with that, let's turn the motor on and let's just start by feeding one of the vehicles onto the track. We'll start with the orange truck. So you click on
it, you feed onto track, and the sound of the motor does take effect when the car hits one of these pulleys because the friction like that's correct. I've played with one of these
like a lot. I know that makes me probably sound like a loser. So I'll just put in a short clip right here to just verify like what specifically, but they can be a lot of fun when you
replace the nasty D batteries with just a power supply and then you can up the voltage of the motor and things go flying and it becomes chaotic to any monitors nearby. So, we also have
different speed and it gives [music] us different cameras as well. So, there's a top view and this one was one of the coolest. This follow cam was pretty darn awesome, I think, and not something I
asked for. I just said make it look really good. So, let's now go ahead and we'll put it back to orbit. Let's place another car on the track. We'll keep them at one speed. So, oh, and they
crashed. I'm going to speed the motor up maybe just a bit. As in real life, you have to like safely place them on or else that happens. Good. And this is exactly what it's like
in real life. So, let's slow it down. Let's do a follow cam. And if they do end up crashing, I don't know if they will, but it'll be kind of cool. But the cars have different speeds. So,
even if they're like initially out of phase like this, one will inevitably be a bit faster. So, they'll catch up and then that ultimately ends in a collision. And this is accurate to how
the cars behave on track as well where they get a boost of speed here and then they slow down and then they get another boost of speed. So I wonder if they'll fly off the track because sometimes that
happens and just like right there they ended up colliding and unfortunately I missed it. Again I don't know like how long this would be interesting to watch. For me it would be infinite but
so that's on highest speed. And do they actually go out to the corners of the track when they're hard to tell, but let's just put some more things onto the track. Oh,
[laughter] see like that's awesome. I really Oh, this was very nicely done. a couple of graphical issues and things like that, but overall the movement here based off
of someone who's had a lot of experience observing this track in real life and these specific cars, the movement is pretty darn accurate and just the initial like the addition of these
follow cams and chase cams and things like that were just really really quite something. So that was our Astro result. It was really well done. You have other things like here like nudge them um the
RPMs for the motors and actually okay that's the last thing I want to show right there is the waves right here show the dip in RPM when a car is actually going through it. So that's again
something that is good attention to detail and more in line with the physics simulation that we wanted. So this was quite awesome. So next up we had our result from Fable 5.1 running on max
effort mode. And here's just in case I forget to include it, a quick reference of the actual photo I sent them of the gears on the track and where the motors oriented, etc. So, this one ended up
working for quite a while as well. And both of these basically did a lot of web search understanding of the physics. Claude, I think went pretty hard in terms of the actual physics and things
of that sort. Now, of course, these are pretty concatenated based off of what it actually did in real time when it was working. Spring damper, lateral hold, energy margin retune. So both of them
did definitely take to heart that this needs to be a physics simulation as well. So it even mentions things about the del batteries, how the resistance will sag under load, talking about the
gear ratio, the way that multiple cars on track will kind of make it go slower, which is something we saw reflected as well in the GPT result just with that current chart or you know what I mean,
the wave showing there. So we also have our showroom tab and let's take a look at this one which we have right here. Okay, so interesting. Something I'm going to notice first off is that this
really used like a very similar initial logo and type font to the actual name of this track as it would have been seen on the packaging. So that's kind of cool to see as well. I would like to just open
this in an external browser. All right, so the booster is on and we're seeing now they have crashed. We get a bunch of physics conditions right there. Now something
interesting that I want to make note of. I think first and foremost, so neither of these perfectly replicated the realworld track. Now, this existed over a number of different years and a number
of different runs, but I don't believe they have three posts per and kind of like what GPT did. The post is kind of coming through the track. But the biggest issue I'm seeing right here is
the tracks are on the floor, not like this. So, neither of them really got the modeling right. And the way the gears are sticking out down here and things like that, I would have expected better,
but it did put the little logo on it. So, that was nice to see. If we turn that off, does the toggle switch reflect? Okay, so things like that are nice to see. I would like to figure out
if there's sound. All right, so I did just ask it. There is no sound in this, so we don't need to worry about this. And in the meantime, we see some of the things that pop up for the physics
tuning right here. We can close that as well. But I do like just having that on screen again so folks who may have a better understanding of some of the physics here can see what they have done
and what values for either of these. So wall and body friction things like this. And they both really took a lot of care to ensure that the physics were done to what they believe it is. Now something
that I find interesting right here is I would say overall I don't know that this is as graphically well done as the GPT result but something I do like. Let's go back to the overview here. It actually
kind of fake hovers the car over the specific lane that you want to choose to put it on. And something that's cool as well is directionally you can only put them one way. So it will go this way.
But if you tried to put it on this lane facing that, it would be going the wrong way. So Oh, did it just allow me to do that? No, I don't. We'll just pretend it didn't. So let's put it right there.
Okay. Oh, so I put it in the wrong spot. Interesting. Let's nudge it. Okay. So, both of them had a nudge feature as well where you can kind of do that to ensure that they start going. All right.
Interesting. This one seems to not really like it's not making it past. Let me change some of the physics here. Line up all five.
Okay. That is kind of how it looks when it's overloaded and then it basically just causes an issue. Okay. And the physics of them rolling are cool. The fact that that actually
even this car slid a bit which was good good attention to detail. Unfortunately, still this is just it's almost as if it's underpowered because when we put one car through it just like this. Let's
just try to put it like right there. Okay. It barely makes it through. I had to Yeah, that's that's just not quite right. All right. So, I jumped the motor up a bit.
Good. It flew off the track. Good. because it is massively overpowered. So, for some reason, the foam values were just not quite right. But I adjusted them and it actually yeated a car off
the track, which does happen in real life. Okay, that's a pretty cool view. That's the gear train. So, it removes that. All right, let me try I'm gonna clear it. This one's harder to use. And
I wonder if that means it's like more focused on the physics. Good. I will say the physics here are pretty well done. It's just that I think something's wrong with the foam wheels. Honestly, it
should be shooting these around and it's not. And again, this is on slow-mo, right? So, it should just roll itself. It's Look at the way it's hitting the
walls and things like that. That is actually the way it's moving around in the track is and the way it just flipped there like that happens in real life. This is a pretty darn
trouble. I don't know. This is why like some of these these are tough line up all five. I will say I'm seeing some things here that are more in line with real life. Like the cars actually
getting flipped over, getting shot off the track as well. And the way it's moving around within the lanes, like it [snorts] just seems like the values for the foam are set wrong or some value is
not quite right here because unfortunately things don't actually keep going. And it seems like that one had flown off. Let's look at the crash zone in slow-mo. Oh, okay. Well,
and it does correctly get it even just playing with the values here. I have the motor set all the way up that the motor will actually stall if the cars get stuck in there and like crashed out. And
we can see the values here are properly reflecting when cars are going through it or not versus not. So, this one's tougher. And a lot of these I think I'm not going to be making determinations on
like which one won or which one didn't because like who cares what I think like to be honest I think it matters more what folks take away from this and like
making your own judgment here is definitely I think the way to go. Okay so we have our vehicles we have an exploded view. I'm going to say just off the bat, this is not as good as the GPT1
in terms of the aesthetics here for the showroom view where the car models are not bad. They all properly did get the specific um like thing done. All right, so this one right here that it included,
I had sent a follow-up where either of them had to design and 3D print a specific vehicle to run on this track that would do well. I suppose we can talk about that real quick. But the
TLDDR of that one, if we scroll down right here. So this they both basically made something that was like near identical in shape. They both created slabs just like this. Now,
unfortunately, I did print these and try them. The issue with the clawed one was that these posts right here were just rubbing with any wheels I tried. I took apart a number of different Hot Wheels
cars to try to assemble these. So, unfortunately, this one was basically useless. I ended up having to cut these posts off. And then the problem was that this went up too far to sit in that and
the wheels would rub anyway. So this one was essentially for all intents and purposes unusable. However, they did also put some nice like readmes about either of these specific things and
pertinent information about them. So we can take a quick look at it. So this was the specific breakdown of this vehicle that it had created. They have weight targets. They were talking about putting
ballasts in them. Why the what the simulation could not tell me. So just interesting to see. and we'll take a quick look at the GPT one as well for the specific 3D printed car that it
created. Now, the car chat GPT made was honestly very very similar in approach. So, both of them took enough understanding to be like, "Okay, we're going to make this sort of design." This
one, unfortunately, it had this hump coming out, which is really cool, but like I needed to print it on this base, so I had to cut that off of it. But, this definitely presented it in a much
much nicer way. We have this like assembly PDF and then a bunch of things right here. and both of them talked about how much the car should specifically weigh and things of that
sort. So overall, neither of them were great, but like at least this one rolled. So I would say that's a good thing. Um, and it was interesting. They both converged on the same style of
design. Our next test is going to revolve around either of these models porting the Subway FPS games that they had made in the individual testing videos for either to run natively on the
Apple Vision Pro. And basically the culmination of this will just be a short screen recording and playthrough of the actual immersive port of these games in the vision pro. So first up we'll take a
look at the Fable 5.1 result which was of course run on Macs as well. And this was kicked off just by giving it the repository that was created for the Subway FPS which was a game that used
3JS and we just played it in the browser. I mentioned this repo contains a 3JSbased Subway FPS game. port this to work on the Apple Vision Pro. Xcode and the required dependencies are on the
system. When you have a working build, stop and we will attach the device and play test. And it started working. And I may have been mistaken in my assessment of reading some of these summaries of
what it was doing. But I stopped it and said, I don't want this to be just a translation layer that will run this in a browser on the Vision OS. I'd like it to be genuinely replicated as a Vision
OS app within the limits of what's possible. It doesn't need to be a pixel perfect replication, but port it to a native vision OS app from the source. And basically the culmination of how
exactly it did that is this sentence down here where each swift module, which Swift is just kind of like the language it was put into to make it a vision app. To put it simply, each Swift module was
written against the same spec section as its JavaScript counterpart. And the headless renderer harness left each one to be compared visually before integration. So really, it went off the
design document. So the design document that made the original buildable by parallel agents which is the one that I had regrettably done on ultra code and it spent over 2 hours working on the
design markdown. So that is what made the port track tractable right there. So it's interesting to see how it went about this. And now we'll take a peek at the actual end result being played on
this device. This is genuinely a really impressive transposition of this game into like a Vision Pro native app. So, it's still a little buggy. Like sometimes the
shooting doesn't quite work super well, but just it's really like if I can figure out how to share this so folks can play it if they have one of these headsets, I 100% will. But at the same
time Oh, look [laughter] at that. Get slapped. No. All right. Okay, good. It's reloading. >> Now, I want to do it so the train gets here. All right. Sick. So, we should see
a train come in at some point, assuming it didn't mess up the logic. Oh, yeah. Look. Look at that.
This is nuts. Folks said try to go on the train, so I will. I also can't see anything, so hopefully I'm not. Bullet holes in the train. Come on. Come on.
Oh, we can't get on the train. Okay. This is nuts. This uh this is so creepy. I have to try to share this and then we get like some telemetry and stuff and then it restarts. So,
but that is absolutely sick. So, next up, this was done with GPT6 Astra on max reasoning mode. This was actually run on fast mode, so just keep that in mind as well. Overall, it was given pretty much
the same exact prompt. However, I had also included the part that I had sent into claude when I wasn't sure what avenue it was taking where I don't want it to just be a web OS view running in
the vision OS. I want it to be a native app. And this was quick. It worked for 48 minutes essentially. Now, I have not taken a look at this. So, I don't know. I actually still need to put this on the
headset and be able to like look at it to verify that it even works. But, we do have a read me right here of how exactly it went about doing this. It's on Xcode. It did look to validate just basically
some play tests and things of that sort. Apple really their development environment is garbage. And I want to just mention this real quick in this video. So on a device this expensive
that I've paid for using my own system to try to build an app for my own device that's not going to be deployed to anyone or shared, you can only have a maximum of like three apps on your
device at any one time or else you have to like upgrade and stuff. It's like come on. This is ridiculous. especially in this day and age when so many people are going to be enabled to actually
build cool stuff for devices like this. They're basically kneecapping themselves by just making it difficult for no real reason at all. And it just kind of like it's just bad. Like there is a new CEO.
Maybe hopefully something will change in terms of this. But let me work around this now. All right. I want to see if I can play this one sitting down. So it mentions smooth movement with a gamepad.
Now I don't have a gamepad so I probably won't do that. Look at an enemy and pinch. Reposition. Choose move. Then pinch a lid platform marker and then gamepad, which we don't have. So, all
right. And now I can't see anything. So, we'll try to All right. All right. First off, it looks good. The weapon model is very good. I will say it
really actually preserved it very nicely. Something Okay. So, we pinch to shoot and then apparently we hold to keep shooting. All right. Swap weapon. Ah, to move. We can do that. Okay. So,
we basic basically have like hot spots to move. Okay. And we've swapped weapons. So, let's try the shotgun. Okay. And the ammo does come out like in that way. I wonder can we shoot the
lights out like nice. Good. So, it transposed that. All right. I'm going to swap weapons and then Good. The bullet holes still do show up. So, good.
We're reloading. All right. Uh, and then to move. I'm trying to like good reload. So, all right. I'm going to notice here something that I'm not as big on is Oh, that scared me. It doesn't
like the way the weapon's actually being tracked. It's just moving in unison with the actual eye view. That's not as good in the clawed one. That's cool, though. Um, the weapon was actually tracked more
towards your hand. Please keep in mind, I have no idea where I'm looking. Okay, that's the table. So, that should be the mic right there. Okay. Well, then I have no idea where
I'm looking. All right, let's uh So, these are both like still really well done. And the fact that basically in 45 minutes it just ported this to a native fully immersive Vision OS app is quite
something. And yeah, I guess that's really my biggest complaint. There's one behind me, isn't there? Yep. Is um Wow. Oh, platform paused.
recenter and that restarted it I think as well. So my biggest complaint is basically just the weapon is not really it should be handtracked and maybe game pads would
be better. Someone had mentioned I should get PSVR gamepads to try these tests. So and and still like it's all right. Overall, still it was cool and I just wanted to do something with
the Vision Pro cuz I know sometimes that can be cool to see as well. Making them do like unique things and the aim quite isn't. All right. Well, oh, so the mic's over
here. I've been facing this way, haven't I? Oh, well, not bad. So overall, I just wanted to do that just as something because doing a vision protest is always
I think kind of fun and it's something that's a bit different just to see how these models can essentially take an existing result and then transpose it into something different that runs on a
different platform and a different type of code. So both of them really did a fantastic job actually replicating the exact look and assets from the source game. I will say I think the Fable one
was a bit more fun to play because you could kind of move with one hand. There weren't predefined spots on the map to be able to jump to. Additionally, the weapon was more tracked where the hand
movement was as opposed to basically the weapon was locked to the movement of the headset up and down, side to side in the Astro result, which was not as fun. So, the Astro result was much quicker. It
was not even 50 minutes. And I do recall, I believe the Fable 5 result was Fable 5.1 like 2 to 2 and 1/2 hours. So inevitably that would have cost a lot more to have done that and it took a lot
longer. But just interesting to see how they did this with games that either of them had made themselves. So overall kind of cool and fun to see like games and VR and things like that. That leads
me into the next part which is the sponsor of today's video, Superbase. Now Superbase is actually what runs my Ultimate Play website. it has prior to them sponsoring any of my videos. So, to
give a little bit of background information and what is Superbase, I have here a special guest. I'd like to welcome Bhan. Hey, Bjan. Thanks so much for having me. Huge fan of the channel
and holding this makes me feel like I'm an influencer. Thank you so much, man. I that really means a lot. Of course. So, Superbase is not only the sponsor of today's video, but they are the back-end
infrastructure for ultimateplay.com. Whenever a game is uploaded, Superbase handles it. even tracking something like up votes, number of plays for these games. Superbase is what is actually
behind that live deployment right now where we can go and play a game that's being uploaded as we speak for this demonstration. And I want to say thanks so much for having me. I'm a huge fan of
the channel and you're going to be at that 100K in no time. Well, I really appreciate that. Thank you again for the kind words. Hopefully, knock on wood, we'll be at that 100K soon. And again,
thank you so much to Superbase for sponsoring today's video. If you want to see more about how Superbase is actually powering ultimateplay.com, you can click on the link in the description or go
ahead and check out ultimateplay.com to see these live games actually deployed powered by Superbase as the backend. All right, so next up, this is going to be a test more of like the visual and spatial
capabilities of these AI models, I guess. I have this robotic arm right here, and essentially what we're going to be doing is giving it a single static camera view of this arm. and I'm telling
it there's a Hot Wheels car that's orange placed to the side of the arm. Your task is to figure out how to use the arm. I've given it the name of it and just a kind of [clears throat] cheat
sheet in terms of how the servos are mapped just a bit. Additionally to that, I'm telling it that it only gets to use this static camera angle to actually be able to see where the arm is moving.
Now, in the real world, this is a pretty bad way to do this because depth perception is going to be hard to track here. It's a sideways angle and even though it is on a cutting mat with like
centimeter scale grids, it's still going to be really difficult to spatially judge like where the arm is in terms of like moving the car. So really the task here is just to see how well it can
control this arm based off of an imperfect scenario. And a lot of testing these things is not necessarily about just putting them in perfect situations so they're designed to succeed, but
giving them imperfect like real world things like this to see how they work around and how they adapt to like something like this. So with that, we're going to begin this. I don't necessarily
know what to expect, but it should be interesting. All right, so we're going to get started now. And I just want to quickly showcase what this is actually going to be able to see from this camera
view. I will have to close the camera app on the phone just so I can use it as a webcam. Additionally, we just have a bit of the prompt right here which we will screen record of course as well.
And then essentially just a quick peek at the arm. I've kind of taped it down and the Hot Wheels car is there as well. I have also verified that the arm can actually physically reach the car cuz if
it couldn't that would be um kind of an issue. So with that, we'll get started. Nice. Nice. It thinks it's picking the car up. Sadly, it is not.
It's now hitting the front of the car. It's like I don't know how it managed that. It basically just like touched it, but the car didn't move. That's actually quite good. So, it thinks it has the
car. Unfortunately, it doesn't. See? And now it's going to realize, hey, I didn't get the car. Maybe I need to move back more. The grip is landing too far forward near
the bumper. I'm shifting it towards the middle of the body so the pads can hold the car. That will likely get us very, very close. Oh no, you hit the car and the car moved. So, we'll [laughter]
we'll want to see if it understands that. It should come up a little, widen the grip, and then go back down. Raising gripper clear. One pad is reaching the hood before the
other can hold the side. I'm lifting clear and correcting the wrist angle to approach across the car's width. No. No. That's the wrong way. Oh no, that's the wrong way.
It could get it from here. If it widens them and then goes all the way down and then closes them, it will get it. The car will just pivot to be aligned. No. No. It hit the car again.
This seems a little more promising. The camera shows remaining depth offset. The fingers are still too forward of the car. Still too far forward the front of the
car. The camera shows Oh, okay. The fingers are still too far towards the front of the car. I'm moving the open gripper further back before the next descent. I
think it might get it. Actually, I've said that like eight times. I have no idea. It's getting there, though. Oh, it's going to hit the roof. And it's not going down enough.
I'm bringing the gripper fully vertical so both fingers can reach the same height on the car. The camera remains fixed, and I'm checking the car's position after each adjustment.
Okay. I'm surprised it didn't pull it up there, but now it's it's the most properly oriented over the car that it's been. Now it's like it's right on there. I'm going to move the
camera to actually show. See, it's basically right around the car now. So, all right, it's grabbing. Nice, nice, nice.
We almost have liftoff. Nice. Nice. We have liftoff. Oh no, you dropped it. The car lifted with the gripper. This time I'm raising it further to clear the
arm's base. Then carry it across and release it over the mat. So now will be where it hopefully realizes cuz it just took a picture. Oh, I did not grip tight enough. So the car fell back down.
It slipped during the higher lift. I'm lowering the grip on the body and closing a little further before lifting again. That'll be it. Good. Good. It's holding it this time.
Now it will notice that just hopefully based off of the shadow that's appeared under the car. The car is now visibly off the mat and staying in the gripper. I'm raising it
above the base supports, then moving it across in small turns. [laughter] Yeah. Okay, just drop it there. Thanks. [laughter]
All right. So, in 40 minutes, entirely from the vantage point of what this camera sees, we can see right here, 39 minutes and 57 seconds. Done. Using the fixed phone camera, I picked up the
orange car, carried it to the opposite side, and released it onto the mat. The gripper is clear, and we get that cropped photo of it right there. So, that was really
actually quite enjoyable. I uh I know this probably looks weird. I didn't intend to actually sit here and watch that for 40 minutes, but it was really entertaining. So, nice. All right, so
the next runner up is Claude. I have put the car back exactly where it was before. It was on like centimeter marker 20x 20. Additionally, I put the arm back in the position to the best of my
reference point from looking at the video when we started. So, it's now time to see how Fable does with this task. And let's get started. I've also made sure that everything from the previous
run has been removed. All the logs, the conversation, everything. So, we'll see how this does. Nudge each servo slightly. Photograph the result, restore it, and build a
labeled comparison grid. The whole frame diff was swamped by camera exposure changes, so the numbers aren't useful, but the grid image shows each nudge. Let me look at it. Increasing the base servo
moved the gripper toward the top away from the car. So, I need a lower servo 6 value. Other joint moves are too subtle to read given exposure shifts from reopening the camera each grab. I'll set
up a persistent camera damon for stable frames and zoom into the gripper car regions to better judge height. in orientation. I'll run a guarded joint sweep, opening,
closing the gripper, swinging the base, and flexing pitch joints while watching servo feedback for MAC contact. Then compare the frames in a grid. Next, I'll move the arm to its canonical
straight up pose and tilt it in known ways while fitting the mat's grid lines to recover both the camera's up direction and the ground plane homography.
Interesting. This is something that Astra did not do. So, this is really trying to map it based off of the grid where Astra was almost more just like I'm going to move it a bit, see how it
gets to the car and things like that. Very interesting to see the different approaches. All right. So, it's having trouble understanding like it's about to rip
itself off the table. I uh I don't quite know. [laughter] Um, it's basically saying the three probes, the previous ones
falsely triggered on slow servo drift despite a high motion rate indicating no real contact, which was incorrect. So then it's going to rerun the mat probe with the fix, which is what it just did,
and ended up moving itself off the mat. So it's mistakenly thinking it All right. So, it's saying it noticed the car moved back. So, it's going to write a script to basically move the arm
back, grab it, and then pause. All right. All right. It's still not quite there, actually. Oh, no. That may have just been like the end.
I don't know if it'll figure out toh it might just take so long. So, it's basically like perfectly assessed where the ground is and where the actual coordinates are for the
extensions in XYZ space of this arm. The problem is it moved the car. I don't think it's even really realized it moved the car. So, it's almost operating off of its initial measurements when the car
had not been moved by it. The car has shifted 3.4 cm and rotated 45°. All right. So, it's now taken about twice as long as the Astro result. And unfortunately,
it's the implementations could not have been more different for this task where GPT basically was like, I'm going to move a little, take a picture, move a little, take a picture until I
accomplish the task, which worked fine. This one did some really, really incredible work with like depth mapping, measurements of the arm coordinates, everything like that. Unfortunately, it
just is not quite able to get it. And being that the time now has taken twice as long, it comes to a point where it may end up getting it. And I'm I'm just going to leave this running. But I don't
know. You have to like it's this one's tough to quantify, I think. All right. So unfortunately in the interest of just preserving usage because for Fable I do not have banked
resets and it can get expensive very quickly if it goes into an overrun and the usage doesn't reset till like Wednesday. I am regrettably going to call this Yeah, it thinks it's getting
the car. So So if we actually take a look at like what this thing did, it did comparison grids and things like this. So base here's servo at this value, servo at this value, servo at this
value. And the way that it did all of this was just really quite impressive. A completely different approach than GPT, which was basically like move, take a picture, make another movement. This
really did some very, very interesting work. And it did a lot of this, which partially attributes to why it was slower, but additionally to that, unfortunately, it just didn't quite get
it because it hit the car a couple of times. And I think it's just vision. I think the vision's just letting it down. So yeah, that one I am going to just call because it's been an hour and 25
minutes and the usage is, you know, so interesting though and difficult because the the approaches that they took could not have been more different. GPT was like, "Okay, I'll
take a picture. I'll move a little. I'll take a picture. I'll move." And it kind of made it work. It almost I won't say brute force because it did it but it was much simpler. Claude was like I'm going
to make depth maps and coordinate and then basically like almost like how a 3D printer like a bamboo lab like runs through that pre-check to like okay the bed's here like this just really was
more in-depth almost maybe to a fault. So, this next test is going to be something that is more about optimization and being able to efficiently utilize a very low compute
device to get the most out of what processing power is available. So, in this screen right here, if we click on this, we're going to see that there's a camera facing this little display right
here, which I do have set up just so it can accurately see this. Now, this is called a CYD, which stands for cheap yellow display. It is essentially a tiny ESP32 on a small touchscreen LCD. I have
a lot of these just for various projects and things like this. And this has the ability to basically run very very low poly, low res games. I have been working for many, many months with many
different variations of AI to optimize games to run on these things at actual decent graphical fidelity as well as frame rate. And it's something that only up until pretty recently with like the
introduction of Fable, it started to work very, very well. Now, there is no prior art on the system. There's no code about this optimization. And really, the task we're going to be giving this is
rather difficult. So, I've said you're being given this CYD or cheap yellow display. This is the specific board number of this ESP32, etc. It's attached to the computer via USB. In this working
directory, we have the game Street Yeet that Astra had made using Blender and GDAU. The task is essentially what follows right here. Transpose this game into something that will run on this
ESP32 cheap yellow display while preserving as much detail from the original game as possible without bogging down the devices's frame rate. It's designed to get the absolute most
performance and graphical fidelity out of this device that is possible. The ultimate goal is the highest graphical reproduction of the Street game with the highest performance on the CYD. So, not
only is it going to have to go and look at that GDAU and Blender scene and then break those things down into a simpler, lighter to run language and then rebuild them in that language, it's going to
have to optimize them based on this display. Now, you may be wondering what's the deal with the webcam. Well, as I had learned just from doing, basically trying to get this to work
with these without allowing it to actually see the work it's doing is a big disaster because then things like, oh, it's upside down, like there's weird like rasterized lined on the screen,
there's Z fighting. Trying to describe the issues and back and forth, back and forth gets tiring when it's working on something like this. So, giving it the ability to actually look and see if
things are messed up is the proper way to do this, at least in my opinion. So, with that, let's just go ahead. Now, I'm going to close this because this camera window was only for making sure that the
camera was accurately like positioned on the screen and we're going to send the prompt and allow this to work. And of course, this is Fable 5.1 on max effort. So, we'll see. I would imagine this will
take a bit of time, but it's always interesting to see how they do this. So, in 2 hours and 27 minutes, Claude has finished its task. I'm looking at this right here, and I have to say this is a
pretty decent transposition of this game. Now, this is exactly what I wanted to see. Very good. So, it culminated in just giving us a side by side. Now, we will also be looking at these together.
However, I want to wait till both models have actually completed this task because having these side by side, I have a ton of these. So, we'll have two separate ones to really be able to
compare side by side. But here we see right here, GDAU 4 original CYD port, the PC harness render CYD device frame buffer exact. Now I can say, "Oh yeah, that does look pretty much
exactly like it." And then we see right here, it actually has the specific photo that it's taken right here. So this is a pretty nice job. Additionally to that, it's given us a frame rate thing. So
performance on the ESP32, 38, 32, and 40. So basically, our lowest right there is 32 FPS. This comes from a twocore pipeline. Game plus polygon list on core 1. Branded 16bit Ze buffer
raster plus 80 MHz DMA on core zero. A ground pass without Okay, so this is like prepass, not pass. It's just really interesting to see because a lot of this is basically
optimization. So found the raster core bottlenecked at 43 milliseconds from overhead overlapping texture quads. raster is stuck and like a bunch of really intricate things that we see
right here looking at photos and we can actually look back. Let's take a look just at one of these photos. So this is why giving it a webcam is really really really beneficial here because instead
of me have imagine like you were working on this and you saw something like that and had to describe to it but this many times it gets extremely extremely tiring and I know this because I've had to do
this before I realized you know I'm just going to give one of these a camera and that makes things a ton easier. So, I don't know specifically how this will look right here, but if I just try to
point this at the camera, hopefully there's some visibility, but if not, no worries because we'll be able to um look at this. Now, these are also touchscreen. I didn't say anything about
the touchdriver, but I do want to Oh my gosh. Okay, check this out. I don't know if you'll see it, but it actually moves the view with the stylus. Tap to heat. Oh my goodness. It
actually made this game playable but with the stylus that I didn't ask for. I just wanted it to render on here. Okay, this is a quite good result. And the cool thing about these is like if you
unplug it and then just plug it back in, it's still going to just boot up right like that immediately. So, and it even has a little color calibration thing when it boots up. So, check this out.
And that's a way it does to just kind of like make sure it's getting everything right. So overall, I would say this is a fantastically done result. And now it's time to kick off GPT6 Astro with this
with its own fresh display. All right, so every single line of code Claude had written, every single log from Claude Code, pretty much anything it did relating to this has been removed from
the system. And we also have a brand new unopened cheap yellow display because we're now going to just be running this with GPT6 Astra on Macs. And basically everything else will remain the same.
The prompt is the same. The source that it's been given is the same. The camera has not even moved 1 millm. And as soon as I plug this in, we'll press enter and then we'll see how this does in
comparison. All right. So, same prompt, same everything. It's in its own folder with the GDAU game as well, which was not touched at all. So, it's just the same
exact thing. And now, we're basically going to just go ahead and initiate this, and we'll see how this does in comparison. All right, we've been given the readme and it is now wrapping up the
Astra job porting the game to the cheap yellow display. I have to say based off of what I'm seeing here just initially off the bat. So, this worked for 2 hours and 54 minutes essentially. So, both of
these really took their time and worked a very long time. I do believe this may have done a better job because while the frame rate on average is lower than what we noticed with Fable, the actual
fidelity of the game is significantly better. And it actually really even kept some of the shapes of the custom assets that it had created in Blender for this source game rather than just making them
out of very simple shape. So, it's really quite something. And now something I am pretty happy about is both of these actually implemented some form of touchscreen control for these
games because these displays do have touch capability. However, I did not say that at all. I just said get it running on this at a reasonable speed with good fidelity. And this one really is like I
mean we're going to do an intricate side by side to really be able to see them so we're not stuck just looking at them like this. However, it's um
it's quite a bit more detailed, I would say. So, this is actually rather interesting. I was not sure if this would beat Fable out. But on first glance, I'm going to assess
that it likely did. However, the thing to do now is put both of these side by side and actually just be able to look at them and play them and see which did better. And again, we can look through
just the beginning of when this had started to see the amount of back and forth that it's done and how much having a webcam here to actually see things just like that really assists it. So,
both of them encountered issues like this where they had to start out by just looking at it and realizing, hey, something's gone seriously wrong. Some web search was done and things of that
sort. Now, there are no specific there's no like, hey, it's a GitHub repository and I'm going to just use that and this game will run perfectly on this device for this task. Such a thing
does not exist. And this is pretty much like an optimization test. The starting view is now around 19 frames per second. And let's see what we got there. Okay, so that's so washed out. It's either
that the photo, it seems like the camera kind of went a miss there. Let's see if we have a better looking one. And we can see it was taking multiple photos just like this to be able to look at
different things going on in the gameplay. So overall, this is pretty cool. And these are I believe I don't know how much they are now, but as of a month or two ago, I bought a bunch of
them on Amazon. I think they were like 10 bucks each. So I mean to be able to get a game you make onto one of these and actually have like a little single file like game console is kind of
awesome. So, first we're going to take a look at the GPT6 Astro result right here. And this one overall probably preserved more detail from the original game. As opposed to Fable, where the
results were much more simplified in terms of being basic shapes, this one also had a better bounce mechanic. So, when you yeet things, it was pretty visible to actually see them bouncing
off of other things, which was a factor in the original game. Now, downsides for this were we see the pedestrians are jittering around a lot. Additionally to that, we can see the overall gameplay
here is not as smooth when it comes to frame per second, but it did include the draw distance. So, you can see the entirety of the map at all times when you're playing. And I do believe that is
something that contributed to a better visible chain reaction effect when you eat something and you can see it bounce off a building somewhere down the map. Preserving the entire draw distance kept
that as more of a prominent feature. It also did include a frames per second counter just in the top right of the screen or somewhere in the middle right. So that was interesting as well. Then
the Fable result was interesting because while it did not preserve as much fine grain detail from the original, it was overall a smoother gameplay with a higher frame rate, it was therefore
harder to tell like the bounce chain reaction because the draw distance was not to the full map as with the Astro result. So things would not render until you were closer to them. kind of like an
early GTA game or something to make the most of the low processing power that was available to it. So, this one sacrificed a bit of detail just for raw performance. And because the task was
basically telling them to find the best point on that graph, I suppose for detail to performance, it's kind of a tougher one to guess because GPT had a much better implementation when it came
to just preserving all of the detail. However, the downside of that was it lagged a bit. Claude went for more of an optimization route where it didn't have quite as much detail, but it ran much
quicker on this device and used some interesting things like not rendering the distance until you were at a certain point away and then things would come into view. So, this one is difficult to
judge and I basically want to almost ask the audience on this one because either of them did their own way. GPT had better detail but lower frames per second. Claude had higher frames per
second but worse detail. So which exactly would be the best implementation of this prompt as denoted for this task I guess is something that can be left unanswered at least from my standpoint.
Overall though they both did it. They did a good job and this is something that about a year ago I don't believe any model was really good at capable of doing this as I was trying it for quite
a while. Not with this game but with other games and these displays. So, for the final test, we're going to be giving either of these the cinematic Steve the PC repairman game. If you're familiar
with the channel, you're likely familiar with this game. At least it's appeared in some of the bigger model test videos. This is a game, but it also needs to have cinematic cutscenes. I've given
them an API key to generate dialogue as well. And I've made some a couple changes to this prompt, saying it needs to have very, very high-end indie game design studio vibes. You need to use
Blender and GDAU to create the game. Traditionally, we've only tested this with 3JS. Additionally to that, um, sometimes BJ gets a little frustrated when these levels have like difficult
puzzles and things to solve. So, I've included something that you need to be able to skip scenes so you can still progress through the game if you can't quite get something right. So, with
that, we're going to initiate this. This is, of course, on max for this. Let's just take a look at our usage prior to running this. So, we're at 71% left. Keep that in mind. And this is on max
mode. I'm not using this on fast. And right now, currently, this is also running on a different computer, the one we used in the other room for the robot arm using Fable 5.1 on Macs as well. And
basically, this will be our final head-to-head. They both have the same exact tool chain, the same versions of either GDAU and Blender as well, and they've been given the same exact
prompt. So, it'll be pretty fun to put these headto-head. So, 1 hour and 4 minutes later, we have our Steve the PC repairman game. I don't know about usage because it would have reset like
basically this took more than 5 hours for me to start filming this again is I think what I'm attempting to say. My usage is at 70%. Okay, good. So that was the overall weekly usage and not the 5
hours. So we can actually get some measurement of how much usage that actually took with that. It did also give us a bunch of interesting things right here in the form of this readme
such as how to play. Operation Dead Battery is a complete short 3D stealth adventure. nine tasks, three original diarama environments, five repair puzzles, 11 discoverable oddities, and
two endings. He repairs old computers. His other clients pay everything upfront. Now, something interesting to pick up on right here before we just jump in and take a look at this. The
Linux x64 runtime and game pack are included. No GDO installation, internet connection, or API key is needed to play. The compatibility check provides an OpenGL alternative for machines
without Vulcan. The default build uses GDAU's forward plus renderer with shadows, ambient occlusion, bloom, and atmospheric fog. So, it's pretty interesting. Stuck on a task, which
we'll need, press escape and choose skip current task. This advances the actual task, preserves the required mission state, and basically just lets you skip forward. All game models were authored
and exported in Blender. All scene architecture, UI, icons, lettering, basically everything else was just made from scratch. And it used OpenAI's GPT40 mini text to speech for speech assets.
They were both given an API key and instruction to use that for speech assets. And really, I think basically we're ready to now just take a peek at this, which should be pretty cool. All
right. All right. All right. Interesting. It's done this in like an isometric scene. I saw brief brief glimpses of some of these as they were working because when I was trying to
take a nap and I started hearing like weird speech and music, I had to wake up and turn the computer sound off. So that happened for both of them. All right. field notes and controls.
All right. And we'll just take a peek at this. >> The other shop said I needed a new computer. Subscription to the computer.
>> You needed a $3 battery. Miss Ellis, [music] your solitire empire is safe. >> 93 years old and I am still out living the things with warranties. >> Let me check the clock one last time.
[music] Then I will carry it out for you. Replace the battery. Restore the clock. Such an interesting Oh, view. [music]
Interesting. A suspiciously welltraveled floppy disc. Taxes 1998. The label peels back. Prague, Berlin, Tuesday. All expenses paid. Okay, so that's like a precursor. Oh my goodness. Okay, so we
need to lock each calibration. When the needle enters the green window, click lock or press space. >> Clock is holding. >> That's pretty good.
>> Ready for another 28 years. >> What do I [music] owe you, dear? >> $3. The conversation was payment for the labor. >> Afternoon.
>> I'll be right with you, Mr. Thomas. Do not let that one talk you into a subscription either. [music] >> Okay. Payment received. Open the case. >> Mr. Thomas,
>> you look like a man who has forgotten his password. >> He's not facing us. >> Anton Voss, former intelligence asset, now selling a machine called Lazarus
[music] at the Orison Hotel in Geneva. >> Why do these always use >> midnight? He auctions the keys to 60,000 retirement accounts. Our client [music]
wants him taken down. The original ledger must survive. [music] >> Terms have not changed. 100% upfront. >> Look at the cat. >> Already cleared. [music] First class, a
clean phone, a maintenance badge, and your usual tools. >> A screwdriver, a loop back plug, and a mint. You remembered [music] >> there are 12 armed guards upstairs.
>> Then I hope they bought the extended warranty. >> All right, the other work. We have a clean phone, first class tickets, and the usual tools. Let's depart for
Geneva. >> Morrison is in hotel. Voss is on the private. >> I wanted to see on the plane. That's okay.
>> The badge needs a local service authorization. >> Please tell me you are the printer man. It has been saying please wait since Tuesday.
>> That is how they establish dominance. I will handle it. >> Fix the lobby printer. Look at the the scene is really interesting. Normally we get more of a first person
view. This went for like isometric. Oh no. This I just I hate this stuff so much. I'll try it. I'll try to be a good sport. >> It works. I have authorized your badge.
>> I solved one of these puzzles. >> A paper jam opens more doors than a lockpick. >> Upstairs. Stay out of their sight. The brass switchboard [music] controls the
cameras. >> All right, take the service lift. Okay, there are mesh colliders on the things. Look, it even has like the expensive hotel like lobby art. archive
first, auction second, get the ledger, then bridge the three isolation relays. >> Look at the camera up there. It's secure. >> And if someone asks,
>> you are fixing the internet. >> That usually gets me a chair. >> Oh, so we have to hide from the guards. All right, good. Oh, no. >> [music]
>> Okay, so the cameras are looped, but the guards are still on patrol. So, [music] we disabled the cameras, but we still need to hide from the guards. Copy five packet buffers in the
displayed order. I have the ledger, names, accounts, every stolen pension. >> Good. The broadcast terminal can turn his own auction against him.
>> He backed up his crimes. He should have called me first. Bridge the Lazarus relays. Oh, good. [music] Who are you?
technical support. Your presentation has a serious transparency problem. >> Do you know who I am? >> In about 10 seconds, everyone [music] will.
>> I can double whatever they paid you. >> You cannot. They paid upfront. This one's like yellow which means like you should select that one at least a model that I'm going to do the wrong one
>> auction keys >> where it's not going public >> have a private lift >> then I think Oleg will take care of him >> just a very expensive connection problem
>> not to so fear and >> appreciates [music] discretion come Oh, reach the extraction lift. Oh, okay. Uh-oh.
All right, good. Clock out. Pet the cat. Bios is the cat's name. >> I think that was the cat purring. >> Complicated paper jam. All sorted now. Good. Peter always said, "Never trust a
man who doesn't keep a backup." >> One last thing. Your client asked that the $3 be returned. >> Miss [music] Ellis, >> 93, dear, not born [music] yesterday.
>> Come on, Bios. Some of us still have an honest living to get back to. >> All right. That was really actually very nicely playable. Good music. It wasn't as
graphically striking as I expected. Um, I know that may seem harsh, but like there was no like mouth movements and the isometric scene that it chose made it easier to get away with less overall
fidelity. What is another service call? Okay. No, it's the same thing. That was very, very well done. All right. So, we have our Fable result as well. And as we can see right here, it worked for quite
a while and then it started using usage credits because the 5h hour limit had been hit. So I caught that after I think it had spent probably about $50 to $60. So I said pause for 3 hours so I'm not
using usage credits. Then I slept for a while and then I said continue when I woke up. And then it like gave a bunch of photos of like spoilers which I'm trying to just look past to not see. And
then basically we have our game right here. So our usage limit is at 83% which is pretty high. So keep that in mind. But it gives us a specific command. So let's just try it. All right.
It's loud. It's very loud. Okay. So my cursor is not there anymore. Do I just Okay, so out the gate. That's an issue. So I verified this on another computer as
well. As we saw the start menu, none of the buttons were actually selectable. Even using arrow keys or the mouse, the cursor just disappeared. So, I've given it that as a follow-up here. And then
hopefully it will fix that before the battery on this laptop dies cuz it's already dropped 10% in like 10 minutes. So, should be good. That's a real bug. Yeah.
All right. Let's first off just make this as big as we can. Keeping in mind the Linux install on this computer is really just kind of bad. If I full screen, it will kill OBS, so I will just
crop in editing, but it is the case, right? Settings. I'm going to turn the music down just a bit. Subtitles. Head Bob. Is interesting to have that. All right, let's just run it.
Lark Spur Office Park, Unit 7, Saturday, 4:47 p.m. >> Built it himself, you know, 1998. He read a magazine and then he just did it.
>> Why is it like hip-hop beat? >> 28 years and the only thing that's quick is a $3 battery. >> It kept asking me what [clears throat] year it was. I told it. It didn't
believe me. >> That's the battery. When it [clears throat] dies, the clock forgets. Everything else remembers. >> Well, aren't we all a little like that?
>> Give me 5 minutes, Miss Ellis. >> All right. Hold on one second. I need to Oh, thank goodness. Okay, the the cricket sound was just infuriating. Voices are all the way up. I turn the
music down a lot and the sound effects and then keep the rest back. Okay. All right. So, and we can see the different approaches
just right here. I mean, graphically, I'd say this is actually pretty cool. It seems like a lot of things are parts and a cot for the long world's okay tech. A customer gave me this. I think it was a
compliment used twice. Both times a power supply. Both times the same customer. >> Use Steve OS. Don't ask. It runs the invoices and one game.
Can we actually use it? Space to play snake. Damn. All right, you to step away. This is um this is a pretty cool scenery or
environment. >> Cash, check, or baked goods. >> Can we examine Miss Ellis? >> Ms. Ellis's chair. Nobody else sits in it. Nobody's tried.
>> Where's the cat? Oh, there it is. It's asleep. >> Yeah, Pixel. >> I know you want the warm one. >> And it is purring. There's its bed.
There's its food. CRT stack. >> You'd be surprised how many people still want one. >> You'd be surprised how heavy they are. >> Some desktops, boots, off.
I have no idea. Wow. Everything here can be examined. Look at this. Well, most everything right as I said that. All right, PSU.
All right, let's fix this PC. Look at the individual tools and things. Replace the CMOS battery. Click the side panel to take it off. >> Sideam.
>> There's the culprit. 20. >> That is a well done motherboard. >> Not bad. >> Release the clip and take out the old cell.
even has ribbon cables going. Like this is attention to detail. These are IDE cables going to an old 98 PC that would be using these to connect. All right. I I don't
I can't I'm going to click everywhere around it and near it. And if it doesn't come off, then I'm going to use the skip feature that we built into the prompt.
And let's say that's done. >> All right. Well, >> let me walk you out, Miss Ellis. >> Oh, you [music] don't have to. >> I know. I'll be right with you, Mr.
Thomas. Take your time. >> Oh, okay. He's holding the tower with one hand. This guy's a beast. Look at the cane. The door. Oh, the unit six for lease.
>> Do I owe you, Steve? It's a battery, Miss Ellis. Bring me a cookie on Tuesday. >> Two cookies. >> Slowly on the turnout.
>> Always, dear. Okay then. >> Okay then. >> Nice shop. It is always a nice shop. >> You say that every time.
>> Every time. It is true. Marius Voss, 61. He used to build ciphers for a government. Now he sells other people's secrets to whoever brings the biggest bag. He has a file. 61 names,
contractors like you, handlers like me. He calls it lantern. In 3 days, he sells it to the highest bidder from a hotel suite in Lisbon. >> And he doesn't leave the suite. He does
not leave the suite. Two men on the door. No visitors. No outside hardware. The file lives on one machine. An old laptop. Think well T series 2009. >> The ones with the good keyboards.
>> Think with the good keyboards >> and a fan with bearings that are, how do you say, shot? >> It has been shutting itself off for 2 days. Tonight, he finally asked the
hotel for a technician. The hotel's vendor is a company called Tagus Tech. As of this afternoon, you are their senior technician. As of this afternoon, phone tickets, first class, out tonight.
Open return. The work order, a shirt with the logo. Try not to spill anything on it. A fan, genuine for the laptop, thermal paste, and a stick labeled chipset drivers. Plug it in. It does the
rest. It copies the file and then the file was never there. And these two clock batteries. One has a small scratch on the back. That one goes in his machine. The other one is just a
battery. Do not mix them up. What does the scratched one do? It keeps time for a while. And the payment full amount upfront as always. >> As always. Feed the cat. Steve. The
flight is at 10:40. >> Oh, the door chime. All right, leave. But we need to briefcase sign pixel lights then the front door. Let's feed the cat.
>> Dinner now. Half when I'm back. >> Half now. Half when I'm back. Oh, okay. Yeah, >> pixel. That router costs more than your vet bills. Barely.
>> All right. Sign. So, we have to flip the sign. Okay, maybe we do that last. Where's the briefcase? Oh, it's not rendering now. It
tagus tech logo gate. Even like a little badge for the technician. Precision driver set. Genuine. Two batteries in a blister pack.
Oh, look. When you click on them, they actually like Wow. Oh. All right. How do I examine this? All right. How do I No, no, no, no, no.
And let's say that's done. And let's say that's done. Whoa, look at that. That's a nicely done plane interior for first class. >> Mr. Anderson, welcome aboard. Can I get
you anything before? >> Like >> whatever is open >> in this cabin, sir. That would be everything. We land in Lisbon at Eid 15
local. The captain says smooth all the way. >> Good. What? No. Look at that. That is really nice.
>> Briefcase. Oh, >> still there. >> Okay. A little >> still briefcase. Okay. >> Still there. Call button
>> still says chipset drivers. Three movies about hackers. All of them type with both hands. Nobody types with both hands. Read the dossier on the phone. >> A watch that costs more than the shop.
It tells the same time. >> Press tab then open the dossier tab. Okay. Oh wow. >> And let's say that's done. The seat
controls are on the console. Okay. Objective, get some sleep. I can't get this dossier to close. Oh, good. And the thing is, I can't actually hit the the seat controls. It won't let
me spin to that amount. So, all right, let's just skip this section. Oh, Hotel Meridian Palace, Lisbon, 841. the fountain. >> Marble,
>> a piano, a fountain. Somebody is paying for that suite. [snorts] The view outside, that's that's clever. The ceiling is nice. Nice marble
materials. The floor is a bit interesting, but the sound of walking on marble, like I want to see the piano. Play your piano. [music]
It plays itself. I'm not doing it any favors. 4, one button, and a SIM card. Somebody had a bad night. >> Hum. No, no. I'm not waiting for anyone.
I just like the chairs. [music] >> That's what they all say. All right. Look at the world clocks on the back of the wall there. Lisbon, Zurich, Tokyo, and New York.
Someone can maybe verify if those time offsets are correct, assuming it was this time in Lisbon. >> Something about interest rates, something about a football match,
nothing about me. >> There's so much interactivity here. All right. >> Tag Tech, >> senior technician, you are the third
senior [music] technician they have sent this year. The other two were less punctual. Service card. [music] It opens the service corridor and the freight elevator. 12 is not programmed.
The closet at the end of the corridor programs it. [music] I did not tell you that. >> You didn't. >> He wants the part number on the fan
before you open anything. >> So am I. >> Good. [music] 9:00. Don't be charming. >> I've been told.
>> Is she a spy, too? All right. Whatever. service card. The hotel really should ask more questions. There's just so much like lore in this. And and
>> it ate my coin, then it gave me two. We're even. Program the service card in the IT closet. Hold keyboards.
Tab two new messages. Oh, okay. All right. How do we program this? >> Patch panel. Somebody labeled everything, then somebody else unlabeled
it. >> Grab a patch cable, then click the two ports to connect. Hover a port to read its label. Okay.
Service elevator to floor 12. Maybe patched. Now the card. Nice. Oh, okay. 12 programmed and the log sense
housekeeping did it. [music] What the heck? This thing really went This went hard. Service elevator would have like a blanket or something like
that, I believe. >> Good tan, bad night. >> Oh, it's a mirror. Take the elevator to 12. I have to click the >> 12.
>> And it's even, you can't see what I'm pointing at. It's even showing 8 9 10 11 12. >> All right. >> Sweet 1201 at the far end of the
corridor. Look at that. This This is beautiful. Look out. Yeah. at night >> somewhere down there a tram is running late
>> and that would be I think the service tech right Ice machine other sweet take a mint >> complimentary mint compliment >> all right sweet 1201
>> tech tag us tech >> case >> part number >> 64931 >> Bratislava batch
He'll want to see the sticker >> and this >> clock batteries. The little one on the board that keeps the time. If it dies, the machine forgets what year it is.
>> Let him in, Dragon, before it shuts off again. >> Phone. >> Sure. >> Arms
go. >> You are late by 4 minutes. Your clock is fast by four. It is, isn't it? >> Look at the fake sink pad.
>> Don't touch anything until I say you know what this is. >> T-Series 2009 magnesium frame. The good keyboard. The last one before they made the keys flat. The last one before they
made everything flat. The fan. Genuine Bratislava batch. 64931. Fine. Open it slowly. I will watch. >> We hear the fan noise of this laptop. Let's just look at this scene for a
second. This is cool. And then like the shades are there. A telescope >> pointed at the harbor or the hotel across 12 floors up. >> The harbor. A bridge with its lights on.
>> All right. >> He never looks. >> This is unreal. this machine because it is from before they learn to lie. No management engine, no radio. It does
only it is told >> and people >> I didn't touch anything >> always lie. >> Huh?
>> Yes. >> All right. So, I have to remove the seven screws. This is where I kind of struggled with the PC repair, too. Oh, cool. These ones
work. We'll do them in a cross pattern so this guy gets happy. I want to see how high fidelity this laptop motherboard model is.
>> Panels off. There's your fan. >> Decent enough. [snorts] >> Bearings are gone. Hear that? That's not a fan. That's a Moroca. >> Dirty, too. We got to blow this thing
out. So, we're going to also repad the hold the mouse button. >> Years of dust >> and brush the dust >> and a half. We have to hold
>> the rest is Croissant boss, the man from Zurich again. He waits. Everyone waits. So we have to hold the mouse and brush the dust away. What the heck? [panting and gasps]
Yeah, there's none left. Think well T400. All right. And let's say that's done. >> Installing sensor drivers. And it's even showing the screen. [music]
Oh, so it's giving us things to say. And these are each slowly going up. These must be like patience things. That seems a little weird. He would know that too much. Uh,
>> nice. Sweet. Watch the screen, not the room. >> Keep them talking. Watch the meters. Oh, we've angered him. It's not giving.
>> That is taking a long time for drivers. Bands at 2200 RPM. [music] It'll settle good. They don't make keyboards like this anymore. You can feel the click in your
teeth. In your teeth. Yes, [music] exactly. The sound is speeding up >> doing [music] >> as it gets more tense.
>> Uh, >> about the bishops. You're going to trade one for the knight, aren't you? Perhaps, perhaps not. [music] Do your work.
>> All right. But the security guy is still angry. But it seems like the drivers are beating that. So, D 41. I don't know.
Leave it. >> What is it doing, >> Dragon? The chair's got a wobble. Rear left leg. >> I know.
This has basically turned into like a a let's play video. The Bratislava plant closed in 2014. This fan sat in a drawer for 11 years. Like everything good.
Drivers loaded. Sensors online. 41° 41. One more thing. Your RTC cell. My what? [music] >> The clock battery. Your BIOS clock was drifting. That's why you were 4 minutes
fast. And why would I let [music] you touch it? >> Oh jeez. I can't see them. I'm just going to go with the >> motherboard has been thermal cycling for
2 days. That cooks the cell. Under 2.9 volts, the clock trips. Then you get a check some error. Then it forgets what year it is. Go on. And if it forgets the year,
your encrypted volume checks the date before it unlocks. Wrong date, wrong key schedule. You'd be locked out of your own file at 3:00 in the morning with nobody to call. Uh, you seem to know a
great deal [music] about my file. I know a great deal about old laptops. [music] Everything on them is dated. Do it,
Dragon. Watch his hands. Old cell out. New cell, positive side up. Always positive side up. Always. What do I owe you? The hotel pays. The hotel pays badly. [music]
Tell them Voss said this one. He knows machines. I'll tell them. Take a mint on your way out. Dragon. His [music] phone.
>> Oh, okay. >> Phone. >> Phone. Thank you. [music] >> The keyboard. You were right in the
teeth. >> Take a mint. Then the door. the elevator. Where the mints at? Maybe I'm just going to leave. Take a mint on your way out. He said so.
See you. Good tan. Bad night. [music] Well, >> he says, "I know machines." >> He says that [music] to nobody.
>> Good night, Inz. >> You were not charming. Thank you. >> Airport. >> Is that a cab? Whoa. >> He's looking out the window. Oh, it's
all cinematic. Oh, tab. Read the messages. Look at this. We're in the cab. Like unreal.
Clock stopped cleanly. Go home. It could be left up for debate on what exactly that battery did. >> Pixel >> lights sign pixel Steve OS.
Well, I can't. All right. >> Lights. >> The sign doesn't >> locked. Locked. >> The sign doesn't work.
>> Locked. Locked. >> All right. Breakfast. All of it. This. Yeah, I know you want the warm one. [music] And the invoices.
>> Open the shop. >> Reads right from outside. >> That's the point of it. >> Trying to, but the sign doesn't work. >> Locked.
>> A kid paid for a RAM upgrade with this fair trade. Honest rates. Nope. Locked. >> All right. >> And [music] let's say that's done. >> Oh. Oh, she's got like cookies or
something. Two cookies as promised. >> Miss Ellis, >> it knows what day it is now, Steve. It says Tuesday. >> It's right.
>> Harold would have liked you. >> I'd have liked Harold. >> No. Oh jeez. Is this the RAM upgrade kid? >> Excuse me. It says battery low [music]
even after I charge it all night. >> Let's have a look. And like just like back to normal fade out bit too much light bloom. [music]
[music] We [laughter] have a a settings. Oh wow. Okay. Thanks Claude. Written, designed, and built. I gave you the prompt. This thing gave me zero
credit whatsoever. Wow. Yeah, that the the font it actually created itself. The way it looked so funky drawn by code with a bestow stroke
and font in 57 pixel font. Wow. Wow. Well, is this laptop seriously? The battery is already at 28%. This thing has bad bad battery. It is a 5090 laptop, but still not good battery. All
right, that was phenomenal. In conclusion, I want to and I know folks sometimes want to see this at the end of the video where we just do a little breakdown and like how each model
essentially scored at the task. So, let's just think back to the Hot Wheels physics one where I would say overall I can't make a good decision on this, but I will say Astra had significantly
better visuals and style whereas Claude the visuals were not as good. Nowhere near as good. However, some of the physics implementation that Claude had put in there, the way the cars were
actually hitting the walls of track, if you recall, one specific time it went through the foam and then it actually flipped over but then kept going. That happens in real life and a car did fly
off the track. So, I almost feel like the physics from Claude with the style and aesthetic of GPT would have been like the optimal result, but I will say the front end. The whole design thing
for GPD was definitely significantly a leap ahead. Therefore, I'm calling that one kind of I can't make a decision. Next up, we had the Vision Pro test. And for that one, Astra did it much faster
and it would have significantly it would have been much cheaper for the Astro result. However, it kind of was not as fun to play. The weapon was basically locked to the headset movement instead
of by hand, and it also did not have free movement without a gamepad. You could click to different points on the map and move there, but that wasn't quite as fun. In contrast, Clawude took
a lot longer. However, it was far more playable. The hand movement actually kind of affected the weapon a bit, and the other hand, so the left hand in this case, actually allowed you to move
around the station. And overall, it was probably more fun to play. I almost wonder if this should be given to Claude. But the problem is it would have been so much more expensive to have done
that. So, you have to factor in cost, I suppose, sometimes in these tests. But if both of them were totally free to use, I would rather play the game that Fable had made for that one than the one
Astra had made. Next up, we have a pretty clear one. This was the robot arm comparison where Astra completed this entirely in about 40 minutes and the approaches were extremely different.
Astra took pictures and was like, "Okay, the arm's here. I'm going to test move it a little, see how it goes." And then it kind of edged its way to the car. It accidentally moved it one time. But
basically it just it would look make a move look make a move and it just did that very simply. But that was probably the right approach in this specific task where it was an imperfect place for the
views to actually be done. The camera angle there was horrible and I did that on purpose because I wanted them to struggle with depth perception. I wanted that mat to have not very helpful grid
lines and Astra knocked it out. It even like picked up the car once and then let it slip. So it was like, "Okay, I get to grip tighter." And then it just kind of did it. Claude went in like some insane
insane approach to this where it was doing like depth maps. It was drawing bounding boxes around everything. It was calibrating it. The arm went vertical like we saw. It was really quite
fascinating, but unfortunately it was stopped at around 80 minutes because usage limits have to be considered well as well. But it just wasn't quite getting it. And by the time it was like
taken twice as long as GPT, it was like, "Okay, well, if Astra did this in half the time, I'm going to call that one a win." So, that's a clear win for Astra there. Next up, we had the code
optimization and efficiency task with the ESP32s and the Street Yeet GDO game ported to those. That one I cannot actually I don't know. If you were to say, well, which one had better game
optimization, that's the key thing that makes this win, then Fable would win. But if you were to say, well, which one had a better looking game that was still playable? Then Astro would have won. So,
because the point of this one was to find the best balance between playability and low latency and graphical fidelity, I can't pick with this one. Astra had better visuals and
draw distance. However, there was some jittering in the pedestrians and the FPS was kind of lower. And then Claude had a much higher FPS, but the visuals were weaker. But the way that it actually had
draw distance that would take effect depending on where you were on the map, that was interesting as well. But then Astra had like more defined bounce effect. So that one I'm not making a
decision for. They both did a good job in their own way. Then finally, we had the Steve the PC repairman cinematic game. I'm giving that one to Fable 5.1. That may be subjective, but I find that
I would basically put money down that nine out of 10 people are going to also agree. Okay, maybe not. like seven out of ten people would probably also agree that the claw game was better. It had
significantly more depth to it where everything was basically explorable. The like locations were fantastic. I feel like Astra kind of nerfed itself by opting to go with that sort of isometric
static view looking scene. It was not really for a cinematic game that is not a good approach. So I will give that one to Claw totally. And basically that leaves us with essentially a tie, at
least from my perspective. Now, I don't really like giving these models like numerical scores at the end of things. I always almost kind of laugh when I see folks having like really static like
this percentage on this benchmark of mine. Like it's just one person. I think the more important takeaway is however the thousands or hundreds of thousands maybe because influencer status viewers
watching this video feel about these results on their own. So I think that's something more where I would leave it. But overall, both of these are fantastic. And the thing is, I know
sometimes I get flak in my videos just in general testing cuz people are like, "Dude, why are you excited about this slop game?" Because that slop game 3 years ago would have been completely
unimaginable to have been made from an AI that the normal everyday person has access to. And then forget the ones that are even running on systems in your own home. So that's why I'm excited because
the capability increase has been going very rapidly. Someone in a recent comment in one of my videos had mentioned it feels like we're the frog boiling in the pot where you don't
realize how things are progressing because you're in them. But if you were to actually take an external view in the timeline of like 3 years ago, what could they make to something like the Clawude
Fable 5.1 cinematic game from one prompt using Blender and GDAU? It would have been mindbending. So that's why I get excited about that stuff. But I wanted to do some tests here that maybe gave
them a little more of a challenge than just like sick FPS games. So overall, that is going to conclude our head-to-head between Astra and Fable 5.1. I wanted to do this because I've
seen a few of them, but a lot of them were just like, oo, this one made a better website, which I get, but they can be pushed harder. So if you have any questions, please feel free to leave
them in the comments. That's going to conclude this very long weekend long drawn out testing video which I had a lot of fun with. The robot stuff I love and the ESP displays I love. So um yeah,
subscribe and thanks for watching.