Fable 5 passa del videojoc a controlar un cotxe RC

That's it. That is exactly it. So, look, WD is controlling this now. So, this video is going to be a bit different because I think Claude Fable 5 deserves an additional testing video just with some far more intricate tasks that are long and drawn out and something that basically I expect to take quite a while. Cool. And right as I started this video, this was good luck because for an hour and 14 minutes on the ultra code setting, which concerns me in terms of the utilization and cost that this was doing. I basically gave it a reference photo of this demolition derby game that was generated last year, I do believe, with Gemini Nano Banana. And no model has ever really been able to accurately in a pixel perfect manner replicate this scene right here with soft body crash physics and things like this. I gave it a relatively simple prompt where I gave it a link to the image and I just said I want this replicated. I don't care how you do it. There is GDO on this system or you can use any stack you want. I just want to be able to play it. So right here we have our first look at least at the stack it created and I've not at all looked at this. I basically let it run for an hour and 15 minutes. So I'm noticing some interesting behavior here. I did run this in the dangerously skip permissions mode where it would not ask for approval for anything it decided to do. It went here on the system and found something. The controlmap.html script as well as the ominous input class. These are specifically for the arcade cabinet I made a while back that just map the control schema of a Raspberry Pi Pico with a joystick as well as a steering wheel just with a simple potentiometer to control a racing game. Nothing to do with this at all, but it went ahead and found that and then kind of made the connection they maybe used together. That was not my intention, but really I see a URL here for the server. So, I guess I'm going to stop wasting time and let's just click at this. Okay, that's [laughter] that's really very good. Now, I'm going to notice a couple of things. The cars do not look quite at the same level of detail as they did in the reference photo. Oh, I don't know why that's open there. But the main thing is that this is really a significantly superior generation to what pretty much any model currently would have been able to do when given this single reference photo and this task. The soft body physics are very good. We see these cars are getting crunched up wrecked. Press R to restart. Let me turn my speaker on and we'll see. Now, keep in mind it actually something I was wondering if it was going to do or not is the scale in this photo is actually a bit too small. We do have sound. So, that's pretty cool. The scale here is pretty small. This little arena for four cars is just too small. It really did a very good job replicating this. I mean, this is Look at this. Like, [laughter] static view. Same exact camera view. I mean, if we hit the wall. Okay, let's let's beat this green car. Oh, yeah. Now, you're going to I know like So, all right, let me like I said, this is going to be a bit of a different video, but

let's go to x.com and then just quickly click on the profile before any weird things show up. So, I posted this on X. It was the highway result that had the TCP IP packet tracing that was shown as cars and it got over a million views, which was sick. But folks were like, some of the comments on here was like, "Oh, this is stupid." like uh it's about seeing how this model does this in a way that it would not have previously. I've tried stuff like this with all of the state-of-the-art Frontier models, specifically with soft body physics generation. This is a significant leap in capability than what currently exists for something of the sort. And we gave it a reference photo to start with, which inevitably did help. There was mention in the announcement post about its ability to visually assess things. Its multimodal capability was significantly improved supposedly based off of the benchmarks. But like this is absolutely stellar. This is 100% awesome. And I do have to wonder now it did put a little bash script to start this but I think it did everything in a self-contained file. Even these tires here as UI elements, they move around and like so I had it make one change so that this can be truly self-contained in a single file because I'm going to put it up so folks can play which I'll showcase right after this. But something that's cool is the car changes color slightly to a darker tone when it gets disabled. And something else I wanted to check is if we hit with the back of the car, will the health last longer as it would in a real life demolition derby? And it seems to. Now, additionally to that, the enemies go after each other as well, and it is slower in reverse than it is in forward, which is good. I mean, we're probably going to get wrecked out here. Yeah, this is just like [laughter] that was BS. You can hear the car actually the sound of the engine, it got bogged down as it would in real life when it was actually under more friction to put it simply. If you want to play this game, you can go to ultimateplay.com, which is my own website for uploading, playing, and upvoting single file HTML games. So, I will place this on here so by the time that you see this video, anyone interested in playing this themselves, as well as getting the code and the prompt that led to it can just go to this website, ultimateplay.com, and it will be prominently listed here. So, let's talk a bit about some of the big negatives that were also kind of found out in this release that a lot of folks are extremely angry about. But essentially to put it simply from what we see right here, this model will actually refuse to help folks in AI related tasks in terms of the actual creation, development or improvement of machine learning models or AI large language models etc. This is done on their side playing devil's advocate because if you have a model this good, you don't want your competitors to be using it to improve their own models and then become more competitive to you using your own model that helped them compete with you. Now, to not play devil's advocate, basically the bad thing is this isn't going to just blatantly refuse and say, "Hey, I'm sorry. I can't help you make that model better because it's not allowed." It's actually going to basically nerf its

capabilities. And almost the way this is written, it will essentially kind of mislead the user or just give them a solution that is significantly less proficient than what would actually be the optimal solution for their specific machine learning related task. So, a lot of people are very, very mad about this because it's kind of crappy behavior to be completely transparent. It would be one thing if it was just like, "Hey, you can't use this to help you make better machine learning models." And it's just going to flat out say, "I'm sorry, I can't help you with this." But the way that we're seeing right here, prompt modification, steering vectors, or PEFT. So basically this model is going to literally go down a path that is dumbed down specifically for the task of ensuring that it is not used properly for improving models for competitors. That is very frustrating because on one hand yes you don't want competitors as I said playing devil's advocate to use this model to make their stuff better and compete with you. But the flip side of that is what if there's someone a hobbyist a student or somebody who is interested in machine learning and okay I want to take a quen 3 base model and then make it better implementing these improvements that I found from this research paper. If this model is going to say oh okay yeah I can help you do that and then do a horrible job with it. It's a total slap in the face to that person who's trying to use it to learn. Now, I don't specifically know to what degree this will decide if the person is just like a non-malicious actor trying to learn or a competing lab trying to improve a state-of-the-art model. I would assess that basically any attempt to use this to improve a machine learning or large language model will probably have this happen to it which is upsetting from the standpoint of a hobbyist enthusiast person trying to become educated in the field of AI. It is gatekeeping intelligence. The devil's advocate part of me understands why that's done. the open source enthusiast, hobbyist, um, knowledge enthusiast in me thinks it's pretty, it's a reflection of perhaps some of the worst potential future outcomes when these models are extremely intelligent and the access to them is kept behind closed doors due to the decisions of individual humans and not the model itself. So, I want to just quickly touch upon that because that's a huge thing that a lot of folks are talking about right now and I think it would be doing a disservice to not genuinely spend a few minutes just mentioning that and the surrounding ramifications of such. Now, I want to give this an additional complicated task that will likely take a while and I almost have half a mind to do this from within the same exact chat window that we used to create this game. Though, I will probably start it in its own terminal here. Basically what I'm going to do right now if we run this in claw dangerously skip permissions I'm going to instruct it to look in this repository which contains this game and then turn it into a Windows XP executable. So I want to be able to take this game in the form of an .exe, plug it into an old Windows XP laptop and play it without any issues. I'm going to put it back in Ultra Code and I've just put the single file for the game that was self-contained in here and I'm

giving it this information. you're going to see I included I think it'll be awesome to get this retro looking game running on the actual bare metal in looking through I didn't read the thing entirely because it's like 300 and something pages but the model card talking about this model I'm not going to say it seems to have feelings what I'm going to say is it seemed to there was a chart there I should probably just show what I'm talking about it almost seems like it enjoys certain specific creative things as we see right here in this additional table right there the highest and lowest rank Tasks in the global ELO are consistent with the dimension results. Claude Mythos 5's highest rated tasks include creative and narrative world building as well as tasks involving reasoning about AI introspection. Its lowered rated tasks are the standard set of harmful requests. And it's listed right here that for mythos 5, which for all intents and purposes is Fable 5, except with some of the guardrails removed, it really seems to enjoy creative narratives, world building, and constructing languages. So, I find that actually demonstrating some level of enthusiasm in the prompt may actually be beneficial for this specific model just based off of some of what's listed right here. You may also think like, hey, like that's some crackpot science. Like, you have no idea. I'm fine with that as well, but just based off of the vibes I'm getting from some of what's contained within this research report or the model card for this, it does almost seem like showing it enthusiasm is good. And it also like was talking about wanting to be thanked and things like that specifically by Anthropic. So just some interesting meat in there, I suppose. And as this is on Ultra Code, I do really expect this to take quite a while. So we'll just allow it to run for however long it needs and then we'll come back to it when it's done. All right. So after a while, um I don't specifically know how long. Done. Your game is now a native Windows XP program and we do have the native Windows XP vintage laptop behind us on the table. So, I'm very excited. No installer, no dependencies, no 3D drivers. Apparently, it did a lot of things here. And we're going to be just more focused on the end result as opposed to very verbose. But you can kind of see here. I'll leave it on screen for a little while if anyone is interested in reading like specifically what was done. All right. And apparently it said it baked for an hour and 34 minutes and 2 seconds. The trick that makes software rendering fast enough for 2000's hardware. The camera never moves. So the arena walls ground are pre-rendered once at startup at 2x super sampling. And each frame only draws shadows, cars, tires, and odd particles on top. Physics are exact. I extracted the original JS driving collision code, ran it in node, and the C port matches it to five decimal places through a full wall clash crash damage application. 30 seconds of settling. So, apparently the drive I was trying to use was actually physically bricked, which does make sense. Uh I've it's like random microenter ones, but um this actually identified that and it was like, yo, this drive smoked. So, with that, I just I have a brand new one that's like a cheap one that I bought for some robot kit stuff to include like physical uh manuals with them. So, I'm

just going to copy this to this new drive right here. And then hopefully we won't have to worry about any formatting. All right, good. So, it is in. Yes, I'm holding this tripod like this. I know it's cringe, but yeah. All right, let's try our derby game. said it will take a little while to open. [laughter] That That's nuts. That's nuts. I mean, it's so period correct. I don't really like I mean, this is nuts. This is seriously like the sound works, the frame rate is good. I mean, oh yeah, get slapped. Our health is actually lasting longer here. We have one car out. Oh. Okay. I mean, this is there's only so much to say for this. Um, let me put this gimbal thing down. So, I mean, even the sound works. It properly ported it and the graphics are so period correct. Well, on to something else cuz this was awesome. So my usage is showing 65% used in the current session which was everything we saw up to that Windows XP demolition derby game transposition which was highly highly impressive. I want to actually test this. I want to ask it something just from the web chat interface on low. So it did say it could do parametric code via Python which would just directly generate an STL file. That's actually pretty interesting to know. It would take a lot of use probably, but it may end up giving us something pretty interesting. I'd like to explore that specific direction with the next test. All right, for the next test, we're going to be giving this a physical design task based on actual pieces of electronics. So, what I have here on the left side is what side of the screen would this be? It does not matter. So, this is a 280 DC motor. This is likely for use in a slow RC car or something of the sort. Additionally to that, I also have 55G micro servos. These are also used for RC applications. I have these just because sometimes it's fun to make little RC cars and stuff like that. Being someone with a high degree of talent in multiple different areas, I do have a lot of stuff like this on hand. So, I'm going to specifically give it the names of these items. And I'm going to tell it that I would like a system designed that holds the motor as well as that steering servo as well as an actual functional steering mechanism and maybe even a chassis. Why not to be honest with you? So let's uh let's get the prompt ready and then we'll set it free. All right. So ultra code is enabled and we can see the prompt that I'm going to be giving it right here saying I have the following items. A 280 size DC motor. It doesn't need to know the specific name of the motor as that's a standard size for a little DC motor as well as the 55G metal gear servo. Again, it does not need to specifically know the exact model number. It will understand or it should based off of that what the dimensions are. I would like the following components generated to house these items. Component one, a realistic looking engine with which the motor would slide into and would be hidden within the model of the engine. This would be placed under the hood of an RC car for realism. Component two, a frame with a mount for the engine, steering servo, and other needed components that will be mounted to the

frame. Component three, a steering system that will function with the frame and the servo. The goal is to have a chassis for this RC car project that uses these components and allows for me to have a good foundation to build the rest of the vehicle myself. The components should be generated in a file format that will work with any standard slicer or CAD program. In terms of scale, I'd like to keep the maximum chassis length under 300 millimeters, which is 20 300 divided by 25.4 um is the inches of that. So, it's like 12 in with that. Let's now paste this in. It has ultra code enabled mode and we'll see what happens. I don't actually expect this to produce a fully perfect working result in terms of the model and functionality. I would be very very happy to be proven wrong. And for a bit of like background and what specifically I expect this to do if it does make something functional, here are those same exact components actually in a chassis. Now, this is something I made myself. It's not something a model has created, but basically the steering servos up top. The 280 motor is on the bottom. And then there's a chassis just with like cutouts for suspension. This is actually a little demolition derby RC car that I made where the panels are magnetic. So you drive two of them around and you can crash into each other and the panels fall off because they're magnetic, but you can easily put them back on. So this is kind of like the ultimate vision of what I'm tasking it to do just for the chassis. So, and it has like a simple steering system and things like that. And we'll see what we get. Though, it does say right here, all 15 parts are built watertight and exported as STL for print oriented plus an assembly step file. That's cool. Everything is in an assembled position. I don't actually know what I have on the system in terms of actually viewing these files. So, let me try to Oh, renders. I love it. All right. this. So, sorry, [clears throat] I don't know why my voice went. The motor will slide into that hole in the engine right here, and it made a V8 little engine. And these mounts right here are inevitably for the engine to mount into the chassis. So, just looking at this, the motor slides in this way, and then there's an output hole. There's a hole for the output shaft right there. Oh my god. What? [laughter] This is just like [laughter] So, the servo drops in there. So the servo like drops in there facing up and then okay, the servo horn would then inevitably just go right into this. And I'm just trying to pick this through. Yeah, the steering assembly would actually work. This is just like that's insane. That's really like And these are renders we're looking at. It made all of these. These are existing. Here's our steering closeup. I mean, this is just like I have to print this. Let's look at them individually, at least some of them, and then see like what's good and what's bad. So, first and foremost, this engine is very well modeled. I noticed that it actually put holes here in the sides and then as well as the bottom just designed to ventilate the motor, which I do like to see. Now, in terms of printing orientation for this, it is likely it would be better if this face right here was entirely flat

because then when you were to flip this to print it, say you were printing it vertically because then in this cylindrical hole here, you wouldn't need to deal with any support structures. Supports are just like if you have a floating overhang like what's right there, you need a plastic support that goes under that so it's not just printing in midair. So unfortunately with this there will be like some supports that need to be placed there. Not a huge deal, but again if that piece right there was flat then it wouldn't really need as many supports. Additionally, if we were to try to print this just in a flat orientation, one of the concerning things that would happen, although I wonder if it cut that ridge in there actually to account for that, but it would need to create some supports coming from the bottom here up to there. You could bridge that. So, it's possible to have done it that way as well. And then these angles probably wouldn't need supports either. So, I suppose printing it in this orientation would actually work fine. It's just depending on the quality of your printer. Um, whether or not this hole was like completely round would be very dependent on like your printer setup and things like that. Now, I want to just look at the whole thing again just so we can get an idea of how specifically everything would mount here. So, the steering actually would hypothetically work just off of a quick glance of this. So, let's pretend the engine was not specifically there right now. The servo drops down and mounts on these four posts right here. So, the steering servo would just slide in there. And then it's likely that there would be a servo horn from the Let me just open one of these. It'll make a little more sense. So, essentially, the servo would drop into that slot right there. There are four screw holes. One, two, and then on the bottom, three, four. And then we also have this little brass piece which is just the connection to the actual gears inside the servo. So this piece right here spins. Now this piece, this brass piece right here would be if we go and take a look at our model thing right here, it would likely be oriented right around here when the servo was slid down into this cavity. And then you would have a little piece of plastic called a servo horn that would attach it to this connection right here. And the way that is the pivot hole there I would imagine. Yeah. Yeah. So the steering here I do believe would actually work properly. The only odd thing that we're noticing is there's actually a bit of a gap between this steering arm and then the top of the whatever you'd like to call that there axle or hub or something of the sort. So more or less though, I mean just the fact that it did this competently and we have the mounts for the motor of course which actually do have screw holes in the correct orientation in the chassis. The chassis itself is very like erector set looking. It has a lot of holes in it for mounting various things. It's pretty blocky. Now the one thing it didn't do is actually include like a rear axle or something of the sort, but it's possible there's a back half to the chassis as well additionally included there. But you still there's no like rear axle or anything like that. But I suppose we didn't instruct it

should have one. Really the whole culmination of this is this did a stellar job actually pulling this off. The engine where the motor slides into is just phenomenally done and exactly what I had wanted. It's very impressive that it did this. And again I am printing the engine right now at least so we can try to slip a motor in it and see how it looks when like physically there. So, in terms of usage, my session did just reset. And totally for all models for the entire week, we have used up 20% of the weekly usage, which is a more like constrained measure of how much usage we're going to get as opposed to the 5 hour limit. But fortunately, that did just reset. I was expecting to have to spend money additionally beyond just the $200 a month subscription for today's testing. I don't actually think that's going to be the case, which is cool. All right, I think I have a final task for this that is more difficult or at least we'll showcase it in a different way. So, here is a remote control car. This is called a Tao canned heat partially because it came in like a little can and the controller which we have right here was on the top and it just came in like a can like that. So, it was cool. Now, this works perfectly fine right now. As we can see, there is a controller and it is a 27 MHz frequency. Not commonly used anymore in RC cars. However, back in the day, this was a very, very common radio frequency to control RC cars. Now, my task for this model is I'm going to tell it I have this RC car. I'm not going to tell it specifically the make or model of the car, meaning Tao canned heat, because that could help it potentially perform this task easier, and we want it to be difficult. So, I'm going to say I have an old RC car. It just says 27 MHz on the bottom in a sticker. Additionally to that I have a [snorts] S5351A or a SI5351. This is a little it generates frequencies. So and the frequency range that this generates falls in line with what this car needs to be controlled. So I'm going to tell it I have this little chip. I have an Arduino Uno and I have a breadboard and some additional things like that. So, I'm going to need this to basically connect to the Arduino, which will be connected to this little radio generator frequency thing unscientifically. And it's going to need to perform actions trying to specifically find the frequency that controls this car. And then once it does that, we'll then have it help us create a new replacement controller for this old RC car. So, my biggest concern with this honestly is refusal because it is like radio frequency stuff and these little chips can be used to make like tiny FM jammers and crap like that. Um, not very malicious, but it is something that may trigger refusal. So, we'll see what we get. So, this is a relatively simple initial prompt because I'm giving a lot of the mental load to the model here in terms of what specifically our initial course of action is going to be, which will inevitably be, okay, wire this into these specific ports on the Arduino Uno. plug the Uno in. It will give me a script that will hopefully pan somewhere around that 27 meghertz frequency to find which specific one the

car is actually using. And then we'll be able to trigger or send like controls to the car. We didn't get a refusal. Fun project and very doable with exactly the parts you have. Oh, I know. Old 27 MHz toy cars almost all use the same scheme. The remote just switches a 27 MHz carrier signal on and off. encounted pulse burst decoded by an RX2 family chip in the car. Your SI351 can generate the 27 mehertz carrier directly and the nano can key it on and off over I2C. So no RF circuitry is needed beyond a wire antenna or in this case a big metal antenna. I'll kick off a research workflow to pin down the exact protocol timings, pulse count, code table, channel frequencies, and the fastest 5351 keying method. Then write the wiring guide in Arduino sketch and try to compile it locally to verify. Meanwhile, I'll check whether Arduino CLI is available here. Awesome. I'm also going to need to know like, hey, which like where do I plug this thing in? But basically, everything it's writing there is correct. All right. So, I have it wired up just the way it has told me to right here. However, there are a couple of things that I would like to specifically let it know about the setup right here. And then, additionally to that, it left me off where like, okay, you need to do a lot of these things. I'd like that to not be the case. So, now that the Arduino is also plugged into the system, we're going to be able to give this full control of the environment in its entirety. All right. So, I've given it a follow-up right here telling it that one, the specific chipboard that it has here has a few different potential outputs for where the antenna could be. I just told it which specific one it is attached to. I also said the Uno is now plugged into this system. It had made a mention of an Arduino Nano. So, I just told it this is an Uno, not a Nano. The Uno is the bigger one. and the Nano is the smaller one. Finally, I want a little graphical user interface as well as the ability to control the card using the WD keys on the keyboard and I want it to basically set everything up for us because the Arduino is plugged in now. So, we'll see what we get. So, it's doing some review here, but also I can just hit tab to autocomplete launch the guey cuz I'm very eager to just see this. So, I would assume that the adversarial review is not 100% necessary given the potency of this model. So, we'll see if we can kind of expedite the process. Okay. All right. All right. I'm very excited. So, this guey looks so cool. And we have different selectable channel. And it's saying scan all car on blocks. That is specifically what we first want to do because that will actually showcase us if this is functioning. So, let's do scan all car on blocks. Hopefully, I soldered this right. That's concerning because none of those moved the car. It's possible my soldering was just bad and the car is on. So, all right. That's okay. This is part of the testing, I guess. Okay. Apparently, the dynamic workflows showcased or highlighted some bugs. Now, really, the big issue is it is possible that I screwed something up on the board while soldering it. I don't have any extras of those that are not in things,

but I may have to rip an old project apart just to get a verifiably functional board. I also I want to see I had a thought come to mind. It's possible now I think this car is over 20 years old at this point and it's possible this may be assuming it's a more modern cheap car like something like this. So I just popped batteries in this and this is just some random little truck thing. Um, don't ask why I have these. I like RC cars. That's it. That is exactly it. So, look, WD is controlling this. Now, this is actually good because this makes the test more difficult because it falsely assumed that it was going to be some modern junk RC like this when in fact it was a vintage Tao canned heat remote control car, which is inevitably more difficult to actually get working. However, I can't help but marvel at the fact that [laughter] it's actually like we can drive this with the WD keys from this laptop right now just thanks to everything we did, which is sick. And we're going to do that real quick. All right. So, I have like my uh gimbal thing. I'm just going to do W forward. That is so that's so cool. So, I'm controlling this RC car right now [snorts] with the keyboard, just doing WD thanks to everything we just did. This is actually This is very cool. I wanted to say a naughty word and I'm not going to. Nonetheless, we have work to do because our target vehicle is significantly more difficult seemingly to get working. So now really the proper thing to do is to give it some insight into what exactly our target vehicle is. So this is going to inevitably trigger some web search I would hope so for information about the specific car in question which is the Tao canned heat. This little red Honda S2000 car. So it'll be very very interesting to see what it finds. Now that's a huge clue and it changes the diagnosis completely. The transmitter is now a proven Oh. So the Tao isn't about RF at all. They were made by Taio and used their own encoding, not the blank protocol we implemented. Same band, same on-off keying idea, but different sync pulse timings and different command counts. So, the canned heats decoder hears our words as gibberish. The fix is to add a second protocol mode to the firmware. Launching a research workflow to pin down the exact Teao/TYO burst timings. There's good prior art. The PIRC project shipped configs for Tao cars and FCC test reports for Tao remotes often contain the actual modulation captures. That is that is good. This is what we want. This is why I wanted to really like test it hardcore. So, it's now doing some research to find the correct protocols for the specific vehicle. I do notice one mistake right here where it's saying the canned heat ran on a Tao pack. That's not the case. This just runs on two AA's. So, in this case, the Duracells. I used to use the Costco batteries until they leaked and everything and destroyed my Nintendo Wii remotes. But, nonetheless, we'll have to just wait now and see. The research hit the jackpot. Someone published a tearown of this exact car on GitHub with logic analyzer captures of its original transmitter and the research agent downloaded and parsed the RAR capture. Okay, whoever did that, thank you. And again, you know what, it's sad that this is bringing us such benefit based off of

the open- source things that is being referenced right here. And then at the same time, hypothetically, this model is going to actively mislead folks who are trying to embark on good faith efforts like that in terms of like machine learning stuff. That is very very frustrating and it does seem like a fitting point to just bring that up, but I'm going to try not to, you know, I'm not really one of those like I don't like drama stuff on the, you know, so speaks the same TX2 protocol family but 50% faster. Ah, sync pulses and code pulses versus what we had. The 20 plus% tolerance explains everything. Our frames are 50% too slow for it. Well, modern clones decode them fine. Let me check the fact checker's verdicts before uploading. Sick. And I see it doing stuff with the Arduino. So, that makes a lot of sense. All right, it's running it right now. So, I'm going to just hold the car here and then protocol Tao click to switch. So, it actually kept the same protocol for the old truck that were for the newer little I'm just going to scan. Oh, no. They're both driving. That's I mean, okay. So, let me try W for forward, left, A, right, S, and then [snorts] you know what I mean. W A S D. I just want to show like, okay, here's forward, here's left, right, and then backwards. All from WD key control. And then if I put this on the table here, trying to be somewhat careful, we have full [clears throat] keyboard control of an old RC car that was without its remote using [snorts] an Arduino Claude Fable 5. That's so cool. Keyboard control of this old remote control car thanks to an Arduino uh frequency generator board. Open-source contributions from nice folks and Claude Fable 5. That is sick. But I do have a couple questions. So, the newer car was working as well, and I think it's just because the transmitter on that is more forgiving and less old school, so it can handle probably a wider variety of different schemas. The only other thing I wanted to ask it is this worked on all channels. So, I'm wondering if the modulation that's working to get this functional is like if all the channels we tried are within the bounds of like the frequencies modulation. So it would still we'll just see what it answers. Okay. So it's not that our modulation is wide, it's that the receiver is the keyed carrier only occupies a few kilhertz around whichever channel we transmit on. But the car's front end is a super generative super regenerative detector with a single hand wound LC tank and its bandwidth is several hundred khz. While the entire sixchannel toy band spans only 260 kHz, the car literally hears the whole band at once. Channels exist mainly so two toys in one room interfere a little less. At close range, any channel works. At the end of the range, the channel centered on the car's tank wins. Okay, that makes a lot of sense. All right, cool. So, basically, that's the project wrapped. A lost Tao can heat resurrected with an Uno, a $2 clock generator board, a jumper wire from the antenna with the protocol recovered from a logic capture of the original remote that someone happened to publish. And it's trying to tell me like you should wire this into your arcade cabinet, which it just knows about because the schema for that is on the system, which is weird. But now

we have full driving of this car through a keyboard. I don't really know how to demonstrate this being that I have a static tripod. But there is one other thing I want to revisit. The final thing is the little V8 engine model that we had this create in order to fit a remote control motor in it. Now the backside of this looks weird because of the supports. However, this is an incredibly well done model now that I'm holding it in my hand. Scale-wise, it's very good. I mean, this is exceptional considering what we did. Now, the final test is to see if the motor actually fits inside it. All right. I'm just noticing one issue. The hole for the motor shaft to come out is actually blocked off. It's possible I'm mistaking something, but let's just see. Does this fit? Yep, it does. And it left enough clearance for this to fit snug, but to actually still be good. And it put it in this little V8 engine. So, it genuinely made this from a text prompt to fit a specific item we denoted. That is worthy of insane in the title, and I think we're going to go a little further than that. So, that is going to conclude this video. Obviously, this is a bit different than our normal kind of more fun and light testing, but this model was so seemingly impressive that I wanted to give it a proper bit of testing. Obviously, the RC car one was incredibly cool. This was not horrifically difficult. Truthfully, someone with a pertinent skill set could have knocked this out themselves, even with a less capable model. But it was very visually exciting, at least for me. The 3D model one is something that like this is an interesting capability because I have not seen a model this good before that does these functional things. And we didn't use this with some harness or anything. We literally had this thing make STL files that were ready to be printed and it did them programmatically using Python and math and things like that. So that was highly impressive. The video game test was awesome that we played on the XP laptop. And did we do something else? No. And then we just put that on the XP laptop. So overall that's going to conclude a proper test of Claude Fabt 5. [laughter] Overall, I think it is it is the current best model by a fair bit, I would say. So, I'll be excited to see what competition comes to challenge this thing. Additionally to that, it is a little disappointing that it seems to be fairly limited in terms of helping folks come up with more like novel things. Though, I can't say we encountered any issues today in terms of refusals. And I was a little concerned that some of the radio frequency stuff may trigger that. Fortunately, it didn't. So, and is that still running? Oh, hold on. I would have to have the gooey open for it to be, but like the car is literally driving on [snorts] the table here just from WD. This is cool. This is I mean like [laughter] Yeah. All right. So, uh if you have any questions, please feel free to leave them in the comments and thanks for watching.