Can AI Build a Race Track?
HTML-код
- Опубликовано: 14 окт 2024
- I was able to get ChatGPT to to fully create a race track in Zeepkist where all I needed to do was copy the pieces into a level file. Seen similar things done in Zeepkist and Trackmania before but not quite to this extent.
Thanks to Leiter and Owl:
/ chrisleiter
/ @owlplague
Check out these links:
linktr.ee/cala...
Go get Zeepkist by @Steelpan_Dev for yourself:
store.steampow...
chatgtp: "based on the complexity and legnth of the track we created, as well as the number of turns, checkpoints, and boosters, I would estimate that the racing time ofr this track could be around 2 to 3 minutes for an experienced player."
lol
I admire the time and dedication that went into making this amazing video, Great work Sandals!
At the end GPT understand humans more or less like a dog does, and has the same feel for colors: beige is king. 😄 Sandals, I want more AI track building training content!
I was missing this feeling like watching the Win95 screensaver with colored pipes, back in 1995. 🥺❤
Sandals - I love this, amazing concept and execution.
You should try the same thing with a human: find someone who doesn't know much about Zeepkist, and try to navigate them through building a map with text only, no pictures.
Having worked with Chat GPT I've got to say this is incredibly impressive to have even gotten it remotely this close to a real track. Had to be beyond frustrating at times
I think I kept telling people who would ask something like, "I had more fun validating JJ3" lmao
You have to remember that chatgpt doesn't use the whole chat history to generate next answer, you have token limit, so after a while knowledge about format is long gone and chatgpt uses previous responses as guide. I was thinking before about throwing many track to dedicated model, where lines would be converted to vectors, this would likely yield better results. Great video, for language model it's a pretty good result :)
🔥🔥🔥 vid! Nice job dude
Very neat concept! Super cool seeing such a thing possible for chatgpt to work with the tracks file format 😮
Dude. Other ZK tubers as you know have claimed "AI drawn tracks". This is what I actually wanted to see in those video, tho I was after zeepkist videos and enjoyed those races, I'd like to see HOW.
Thank you!
Even if the AI can pass the Turing Test, it still has to pass the Zeepkist Test. Haha. Some of the reactions during the building was funny though. Especially the "Is this a joke?" We could be near the point where AI could just be trolling us. 😂😂😂😂
@Sandals Love this. Can you post your file explanation prompt as well as the list of pieces? Basically the main prompts you feel would be useful to try this myself.
Ya of course, maybe post in CTR discord? trying to think of best place for that cause its a lot of text lol
@@Hi_Im_Sandals works for me
This was a great concept and very good execution. Awesome video man, very interesting, I´m surprised of the outcome. 😎👍👍
impressive that you managed to make it work through just messaging it with text, instead of doing something like training a model
I'm actually very surprised it was able to do it at all, very cool
smart I like the way you started structuring your "model" rule 1 make sure your give your assistant all your information before you have it do what you want 👌
"I think all the builders in Zeepkist should be worried" 🤣🤣🤣
I love how you told it not to overlap the tracks and chatgpt was just like fine ill just line em up in a row then.
Extremely logical but feels so sassy
Not a track making video I ever thought we needed but it’s amazing to see. Appreciate the time you spent with making this.
Is this a new series of videos that you’ll have longer tracks created by a A I ?
I feel like I could watch this thing go on and try AI designed tracks not until it gets good at designing but until I get good at racing.
And duuuude if I suck at racing...!
Chat gpt can only remember so many tokens, it essentially forgets your earlier messages. it might help to ask chat gpt to summarize as you go.
I noticed that, it seemed to remember more when I'd break it up into different sessions rather than one long session. Maybe was just something I was imagining and isnt accurate tho.
Does that actually work? I got frustrated with this as well. It felt like going in circles.
But it's still fun to tell ChatGPT: "dude, we already went over this" and get a friendly apologetic response :D
I would love to see how far this idea could go. Im sure it would be near impossible to have it create a cacky track, but i would think atleast verticality would be doable.
gsaghgahhaahg 🤣🤣🤣 4:30 it looks like me at primary schools. If I didn't want to do a task, I'll interpret given instruction in every legit form which DO NOT perform the task as intended. To piss you off until you gave up. 😂😂😂
You should re-do this with GPT 4 and its media upload functionality. So you could show it what the pieces look like in relation to the track file.
We need a longer ai track
Oh boy, ChatGPT became so much better nowadays compared to what you see in this video, especially with the latest 4o model
The initial tracks felt very much like some of the kacky ones I've had the displeasure of playing :D Still a long way for GPT to go, though.
Interesting that someone from CTR knows enough about computer science to have ChatGPT do this much.
I wonder if you could get with Yannic and have him download basically all the Zeepkist track files on his back end (or write a crawler to do it the hard way through Steam).
You can train your own GPT-3 model (not sure how expensive it is) with your own data on their servers or you can use something like GPT-2 or LLaMA (easier vs better) to have your own local instance that would have no understanding of things outside Zeepkist.
I'm guessing you would get things that look a lot more like real tracks with an arbitrarily definable complexity. Quality would kinda depend on how many garbage/broken tracks there are.
ChatGPT (or any AI for that matter) doesn't "understand" or "like" anything. It just returns the most likely text - slightly randomized - based on your input and the input history - up to a certain point. If you "talk" to it for to long, the older text is no longer considered and simply forgotten.
Nice! Will you continue on this topic? I wanna see more and how far you can take this while it's learning
So that's how JJ3 was made!
If I could like the video twice for all the effort. I would
It should be possible to make a loop where the AI adds one piece that then gets checked by a program and gives feedback if the piece is incorrect. And then repeat that loop for a while.
Human track builders are done for. Just look at what AI can do within 10h and a human coaching it all the way through.
Ai can make 19 semi-playable seconds of a track every 10 hours... Humans in trouble.
@@Hi_Im_Sandals AI doesn't have to sleep.
But the human overseeing it does.
Custom code with zeepkist specific rules would be probably more efficient. I'm actually somewhat interested to try coding a random track generator, but would need someone with more game-specific knowledge (as in available blocks and driveability)
Hard coding something would have certainly been easier, but wasnt what I was going for
Im so happy someone did this! I knew it was possible but i dont have zeepkist so i couldnt try it myself
having gone through this myself, I know your pain LOL
Give it 6 months, and we'll be trying to convince ChatGPT to stop putting risky finishes in all of its tracks.
Wondering what might think Yannic about this. 😆
Fun video!
I'm now wondering what ChatGPT4.0 would get to with connection to the internet
Amazing
Respect have been given
The real issue with AI is that it isnt gonna destroy humanity, not directly at least. The real problem is that humans will rely on AI as having the correct answers.
Hmm think on the larger scale (aka training AI) it would work better if you would just give it a big ammount of files and make it replicate the files instead of trying to teach it how it works.
Being able to give it a some sample maps could be useful but GPT didnt really like that very much.
Try a function collapse randomizer and in 2 hours you'll have infinite Maou maps.
I dont think any AI could ever understand MaouTech. Far to advanced.
For the effort. But your work is not yet finished until ChatGPT makes a decent track.
Oh wow, thanks I appreciate it :)
I think I am done with GPT for a bit at least. It really did take a long time to even get a 20 second reasonable track lol
Put it up against Google. Bard
i tried this with Clyde
didnt get as far as you and i think i modified the track.