Playing Minecraft with ChatGPT
ฝัง
- เผยแพร่เมื่อ 22 พ.ค. 2024
- Meet Andy, a Minecraft bot that you can talk to and who can actually do stuff. He is built with ChatGPT and mineflayer. He is kind of an idiot, but occasionally smart. Watch as he mines, crafts, builds, fights, talks, and survives.
Mindcraft GitHub: github.com/kolbytn/mindcraft
Patreon: / emergentgarden
Discord: / discord
My twitter: / max_romana
Kolby's twitter: / kolbytn
Kolby's Paper: arxiv.org/abs/2301.12050
Voyager: voyager.minedojo.org/
Mineflayer: github.com/PrismarineJS/minef...
Music: • C418 - Haggstrom - Min...
Timestamps:
(0:00) Meet Andy
(1:11) What can Andy do?
(6:22) Surviving with Andy
(7:32) Mindcraft
(12:10) Andy is an idiot
(15:46) Dangerous Code (watch before running)
(16:34) Open End
☻/
/ ▌ copy me
/ \ - วิทยาศาสตร์และเทคโนโลยี
Andy so realistic he rage quitted when the Creeper exploded
I imagined it as him getting jumpscared
Lol
Erm, actually it crashed 🤓 ☝️
@@Yakqwerty u clearly didnt get the joke
@@learningwithaaronhe did get the joke l. He is just acting like what a nerd would say in this situation as a joke. You are clearly the one who can’t take a joke.
Imagine 100 or even a thousand of these working collaboratively. Imagine exploring in Minecraft and stumbling upon a civilization of AI bots all going through their daily routines to advance and automate processes.
this is exactly what i thought and would love to see someone set up an experiment of 100 ai in a minecraft world and watch to see if they collaborate
@@Mr_Tomato_Man we've already seen AI agents collaborating in virtual environments like in that hide and seek experiment Open AI did a while ago. I agree it would be super interesting to see agents collaborate in a more complex environment like Minecraft. The issue, at least from my understanding, is that it's difficult to get AI agents to work towards goals in an open-ended environment. Apparently creating intrinsic motivation is rather difficult.
it would just be hundreds of cobblestone pillars
old builder ai
Having the AI command other AIs would be incredible
The fact that he builds like a 3d printer is very funny for me
common robot ancestor xD
Exactly was I was thinking
6:57 ""Attacking a creeper"? What creeper? OH MY go-"
love how andy served as a warning but it failed anyway
Soon: Bro, your Minecraft house looks very AI generated
💀
That's honestly a fantastic way to roast a build.
:skull:
7:13 that's so cute after he crashed and rejoined he said sorry
The perfect base for a horrifying horror mod, which turns Andy into herobrine without your knowledge.
He slowly starts to creep around, and say darker and darker stuff.
That is just a good idea
Probably bit too expensive, either for paying for the GPT+ subscription, or for the resources you'd need to do it yourself.
@@Tomobata you can run a decent LLM on an RTX 3070. not easy to get but not extremely limited, especially if you already have a gaming setup
@@adora_was_taken A LLM takes up a lot of RAM, add that to running Minecraft?
The only way to make that work is to have 2 gaming PCs, one hosting Minecraft, and the other hosting the LLM, which is too much to ask for if you want to make a redistributable mod for a community.
@@Tomobata you'd be surprised. they usually only use VRAM on the GPU, which doesn't see a as much usage from minecraft. i just tested it with a 3070 and 32gb of ram and i'm able to run minecraft and gemma (google's local llm) simultaneously. you could probably get away with much less.
Think of a server.
A server where Andy is the admin.
No. Please no.
Yes
That would be very interesting
I find it so endearing honestly. ChatGPT is already so excessively polite and then when you make it into this assistant that's clearly working so hard, it's just so nice. The messages when it crashes or dies are so apologetic it makes my heart melt. I know it's only an AI but this thing is so endearing.
It's not really working hard it's an object
@@Arctic_roach objectifying ais smh
@@Arctic_roachbro is going to be turned into an object by Skynet
IT'S ONLY A MATTER OF TIME
@@adora_was_taken I mean, it kinda is. One day though...
It's a law of youtube that content creators will become more likely to upload a minecraft video with time
gaming + ai = views
I would absolutely love watching a little community of bots playing together. I feel like they’d get up to some interesting stuff and I’d love to find out more about their “thought processes” during. So cool
12:17 is some road runner and coyote scene 💀
its so much bs that the coyote will go through the effort of placing a creeper in a giant cave below his house, just for andy to somehow survive.
just like how i remember it
12:25 I love how Andy doesn't place a bed, and instead digs a hole to fight a creeper. 💀
imagine andy helping you find the zombie in your walls making you go insane
@@pokemoncraze13 yeeeeah, the one zombie who hides in your walls and blocks your sleep
5:40 “And he will defend himself from hostile mobs!” Proceeds to charge at the zombies and slaughter them without even giving them a chance to attack.
The best part is that he's not even that good at juggling, XD
Yeah, some terminator type stuff
Videos using GPT4 and a community of Andy bots would be cool to see.
also congrats when you hit 100k subs, it's coming up quick
now i want the "FUNCTIONS DESCRIBE THE WORLD" in every video
THE WORLD REVOLVING
FUNCTIONS THAT DESCRIBE THE WORLD
BIG SHOT
@@chickennuggetman2593 please don't tell andy that he can do anything
@@layla_gamesYT what
@@adamlolbonus4977 deltarune reference.
@@averageyoutubehandle497 okay
I cant wait to see the future of Minecraft chatbots. This fascinates me
You can use a multi model design where you have a potential fine tuned image to text generator so whenever you talk to Andy, it will take a screenshot, convert it to a text description of the world, and give it to GPT to assist with function calling since now it has more information of the world.
For example now “go to that mountain” would work
skip the screenshot step, and just give gpt a description. except, it would need to be more than a description and probably a list of coordinates with descriptions attached to each one.
@@steamkaptain9234 you need the screenshot to automatically generate a text description...
as for the coordinates, that data is already freely available in the mod, so that would not be difficult to implement, and it is likely already implemented.
OpenAI bills go brrrrrr
Finally! I can play with friends.
Coolest thing I've seen today, What a time to be alive!!
i think that a bot coding actions for itself is so cool but so scary
this has so much potential. imagine what it could be like in 10 years. acting like a real player, talking like one, maybe even having an ai voice connected to something like simoke voice chat
I think that's going to be even sooner
@@FXV56 It already kinda exists in real life! 😳 th-cam.com/video/Sq1QZB5baNw/w-d-xo.html This project called Figure is similar but it's an AI connected to a humanoid robot instead of a fake player. It would be pretty dang cool to get the same kind of responsiveness in-game though
bruh that sucks it better not take 10 years
@@uqnyChatgpt 4 and Chatgpt 4 turbo already have voice features for the AI of course you will have to provide it with voices to copy though.
Always thought Vedal and Neurosama would be able to get this done first.
But nice!
Holy hell, can you imagine neuro in control of this bot
Was searching for this exactly use of chatgpt in minecraft and found your video. Subbed and Hope to see more updates and similar projects.
Great work
I've read the Voyager paper too, and when I did I though that someone should do what you did! This has so much potential! Can't wait to see where it goes!
Wow this is actually incredible! Most other games or mods with ai npcs have you just talk to ChatGPT but here Andy can actually do things and even write it's own commands, woah. I think you're together with Kolby pushing the state of the art of incorporating ai in games
The limitations are unfortunate, but this is already so cool and I'm excited for what's waiting around the corner
all that stuff at the end about self-sufficient chatty npcs makes me think of Millénaire. but fr this project looks incredible, great work making an accessible video about it!
New here. This is fascinating. Can’t wait to see how much the bot progresses!
I have a half finished project like this myself, the agent has a local representation of the world of 5x5 in layers, represented in a table, and then a list of the top 20 interesting items nearby in a range of 20x20x20. The visuals have helped in some cases where the agent gets confused, but it also leads to confusion 😅
I also agree that Mineflayer is painful. The amount of times my agent gets stuck in tree leaves (I suspect there is a weird height issue and jumping doesn't work well).
I love this project and what you've produced this far, I got demotivated early but that was with early versions of GPT-3.5 and I suspect GPT-4 would be a lot more intelligent. It's also more affordable per token now.
Excited to see what you produce with this. btw, the newAction command is a great idea.
I've been looking forward to this specific application of AI bots. Awesome work, and channel.
I 10000% would pay good money for a very capable NPC "friend" especially if it can actually talk out loud rather than text. There are super low latency TTS models already too.
I love how Minecraft is the way TH-cam shows this channel to me for the first time yet the rest of the content on the channel is so interesting for me as well.
These videos are really fun to watch
Well, I think you have even surpassed the Voyager bot, although it has its flaws. There is a bright future with your experiment. This really has its value in gold. For my part, I am super grateful that we have people who launch themselves into these exciting projects.😁😁🤩🤩😍
This is so amazing! I absolutely love this project.
I'm thinking more about how to get it to see the world around it. Coming from the text adventure approach, it would be really cool if we could somehow deconstruct the terrain generation algorithm to be able to give a prompt like "You are in a forest, with many treas around you. There is a pond west of you. exposed in the pond is a vein of coal. Far south of you there is a mountain, with a cave entrence in it" - this output itself can be generated by giving GPT a list of points of interest and asking it for a human-like output, with points related to eachother and prioritised. It might be possible to integrate raycasting into this to only feed in visible surroundings
Of course any non-natural structures would have to be somehow described as well. I think the best way to do that is just to always remember structures somewhere whenever an agent builds them. Then there could appear "a small hole dug by agent_12", "a big tower constructed by yourself".
I also just thought, asking GPT to describe the terrain/world as a D&D style DM would be kind of perfect
This is a pretty cool! seems fun to have a bot play along and help you with things :)
Really interesting video as always!
Masterclass, thank you, keep up the project, dude!!
Definitely going to check out your project. Recently, I've been playing with a fork of Voyager (Voyager-Ollama) that runs using local models. After making some changes and getting it running, many hours of runtime in "learning" mode against 10 or so models, I've found a few things. First, it's been very unsuccessful without GPT-4's massive amount of training. For instance, the llama2 model isn't aware of Wooden Pickaxe when I ask about it. Several models insist that you need a tool to mine wood, so coming up with the first tools fails. Sometimes my prompting to counteract this works, other times it doesn't. JavaScript code generation also frequently fails to produce something runnable. Overall, in my opinion this just shows how much the "skill" (training) of the language model on the backend contributes to the success of automation projects like Voyager. Choose a model that doesn't know a lot about Minecraft (or possibly worse, believes some incorrect things about it), and it's going to have a hard time.
That's super interesting, i hope it gets further development
love the names you have for andy, mindcraft, mineflayer.
it might sound weird but andy extremely reminds me of when i first started playing minecraft, down to his building and such. it's giving me very weird nostalgia, but it goes to show how genuinely cool usage of AI can be in this certain context.
Great video and amazing work !
I'm fascinated with this, it has to be one of the most interesting applications of AI I've seen
Found you through your minecraft stuff but the fractal and life videos you make are cool as balls too
just like a young child learning minecraft
This is and incredible project! Keep going.
Thank you very much for the great video.
I'm really looking forward to seeing what we can do once we "log in" to the real world.
As robots, a fusion of hardware and software, come into play, and with the energy issue about to be resolved with the upcoming operation of fusion reactors,
it seems we humans are about to start creating something new, driven by our creativity and curiosity.
Emergent garden video! Buckle up!
I'm biased, but my vote goes for implementing evolutionary survival.
Also, have you tried using chain of thoughts and agents to make chatGPT come up with the orders that Andy should follow?
For example I just tried with the prompt "I just started a game of Minecraft, can you explain me step by step what should I do first to survive?" and then "I want you to assume the roles of a strategist, an engineer and a javascript coder. Simulate a chain of command on how they are going to execute step 1 of your plan. Strategist should define the goal, engineer the way to proceed and coder create javascript code to actually execute the actions. All three roles are part of the same minecraft player." and the result looks like something that could become usable.
You know this is Box of Pandora. You create Life
"yes you can, try again"
I find it both funny and frightening that that actually works
This is absolutely mind-blowing
On hyvä Robotti. Todella hauska! 👍
Now can you imagine this kind of improved AI on an actual real life robot?
Wow, incredible work! I have to ask, is there a reason why you didn't use Baritone?
Have either of you looked into baritone? it may have a few inspirations to take for this project if any. I'm aware how primitive baritone is compared to this, still thought worth a mention.
Will definitely look into this
We need this to speedrun the game
But to be honest this would be very fun to play with in the future where these bots can easily be used in worlds and act almost human and this can lead to a lot of cool stuff to see and do!
Would it be possible to add a build option to use LLaMa instead of GPT, so as to make it completely free?
Yes, but you'd still need to run llama on some pricey hardware
@@EmergentGarden How about quantized models or would the performance just be too slow? I'm working on making an local inference server instead of using OpenAI's API
@@EmergentGardenNow that llama 3 8b is released, i could try
It would be really cool to see the AI complete the entire game and get every achievement!! Or even more complicated, like you mentioned, a community of bots in their own SMP with a giant modpack they could play to make things even more complicated :)
With gpt4, you could potentially make the bot take like 6 screenshots and then merge them together to make like a 360 degree image which it could use to see what you're talking about. "Mine that tree" "where should we build our house"
To improve his overall performance you could maybe use a combination of image to text and functions that allow to check for specific blocks (like nearest block in this direction, highest/lowest exposed block in the near vicinity, a cave finder or an output of a small box of blocks) chat got is actually pretty good at interpreting tables.
I would love to see these bots able to communicate in some way, such as asking one another for resources and they go get it, or have one bot as a sort of ‘admin?’ They can help clear commands and fix the ones who are stuck as well as direct and ask other bots to progress the game, gather resources. That sort of thing.
It would be cool to make whole server with this bots. Theoretically ChatGPT can plan tasks for each bot-player for achieving some goal. Also it’s possible to block all commands from users, leaving only ChatGPT to decide what to do next. Optionally every bot-player can use chat when it has time and good situation for it. So we will have server with 5-10 bot-players, that run on some virtual machines and have their own goals and plans, that can discuss them with players or between each other
Oh yes, you mentioned at 17:25 the same idea👀
The revolution of Minecraft is here 😭 I can't even imagine this
Now AI is going to take Minecraft content creators jobs 💀
Cant wait for chat gpt to tell me the rarest block in minecraft
Will the ai also do the mandatory child groomer when they hit 100k subs?
@@Fieldperson imagine getting groomed by an ai bruh 💀
If you want to implement voice commands into the bot maybe try using the Simple Voice Chat mod for Minecraft - this mod enables you to directly speak to the bot through the game like you do with the chat.
Maybe you could write a "look" command, that describes his surroundings in different levels of detail via letters layer by layer (kinda like Dwarf fortress portrays land).
In any case I'm very interested in this project and curious how much better GPT4 would do. Keep us updated!
1 year from now: Andy beat Minecraft
That MC Skin is fancy
It's perfect
Definitely perfect and not causing me to get annoyed at any particular pixels
13:27 Would it not be possible to use an image recognition ai, to feed chatgpt with a world view? And give it a better general understanding of what is happening around it?
I suspect a "world view" would not be possible to translate into text for chatgpt to use. But giving it a better understanding of it surroundings should be possible. With info like who is standing infront of it, What is written on signs, and stuff like what direction something is in. I'm not into ai that much, but what is your opinion on this?
LOl "apparently im a minecraft youtuber now." got me dieing
Cool stuff. I wonder what would happen if you put the bot in a chat loop. Something like.
If Status Idel then
SelfQuestion: ChatGPT what can I do in my Minecraft World?
ChatGPT to Bot: >>example build a House
just you wait...
this is so cool. I think i might try making my own with Baritone and pyautogui so it can just type Baritone commands in chat
This is so good now imagine a minecraft rpg server with npcs like these
i would love to see this attempted with local models, something like llama. would be really cool, especially if you fine tuned it specifically for minecraft.
It might be interesting to allow it to take screenshots and send them to some computer vision/image description model so that you could actually tell it to do things like walk to the mountain over there. Granted, I would not expect these models to work particularly well on Minecraft images without some finetuning.
yay, i can finally have friends to play minecraft with me, thank you, internet stranger
next week's video: andy builds a lossless redstone creeper farm
How I wish I could participate. This is very endearing, and I already have some ideas for its' possible uses.
You're dreaming of the Stargate Replicators in Minecraft. Be *VERY* careful what you wish for.
It just a matter of time before it starts to beat minecraft
sounds familiar? same voice actor / tts as upper echelon? either way, would this work with a locally hosted llm? in theory there are openai compatible api's build in into oobabooga.
Welcome to minecraft 😊
Andy is cool, be like andy.
Wow this is cool, i got to try get this working
This is awesome
You goddamn wizard. This is super cool.
amazing idea
thumbnail is so intriguing.
I think it would be interesting to store all successful actions and use that data to fine tune a model like Mistral or maybe Gemma 2b to run on a local machine and save money.
Sooner than you think Andy bot will be strong enough to defeat the Ender dragon.
Another idea for how it could get better is that you could make a training model that you play with and tell it if something it does is either good or bad, thats how i use stuff like this
This is really really interesting 🤔
it'd be cool if a minecraft challenge youtuber tried to get andy to beat the game
This would be a bit of a nightmare but using some kind of trained image recognition so it could identify a mountain if you ask it to climb a mountain. Idk how the implementation would work but i think thay would be the turning point for me. Once its got that it could play the game entirely on its own.
awesome project! I don't know javascript but to me it looks like you are not using the function calling feature in the openai api. If that's correct, could you tell me why?
I believe they are. But they are writing allowing the model to pass code as a parameter (I think thats how they do the newSkill thingy)
edit: apparently they are looking for !thisIsAFuction(parameters) in the output manually
One thing I can imagine that would be really cool is one on one teaching an advanced AI how to play Minecraft, with the AI no prior knowledge of Minecraft or internet access
This finna blow up