AI Builds Stuff in Minecraft | Mindcraft
Вставка
- Опубліковано 30 чер 2024
- 100k subscribers! Woohoo!
This video shows off AI agents that build stuff in minecraft. I used chatgpt 4 turbo, claude 3 opus, and gemini 1.0 to power three different agents and tested their ability to follow commands, build complex structures, and not annoy me (impossible). Watch em build some cool stuff. Aria Math is the best minecraft song.
Support me on Patreon: / emergentgarden
Code base: github.com/kolbytn/mindcraft
Discord: / discord
My twitter: / max_romana
Kolby's twitter: / kolbytn
Timestamps
(0:00) Intro
(1:03) Houses
(3:32) Pyramids
(6:37) Creative
(9:32) Skyscraper - Наука та технологія
don't bully gemini. He is trying his best
i will be nicer
You can't bully an AI, itd have to have feelings lol
@@ateam6543 You shall be the first to go
I once knew an old lady that insulted her phones google assistant, and it started to threaten her 😂
@@EmergentGarden Gemini Advanced is a completely different tier, not sure what Gemini Pro is. Sorry if they're one and the same.
I don't think anyone understands how HUGE "this isn't working I'll try something different" is for a machine.
maybe, but i think many still notice it's really impressive
True. Traditional algorithms will always output the same thing given the same exact inputs. These models recognized and corrected mid-execution, and albeit flawed, was still really cool.
that's crazy 😢🤓
Uh it’s called ai dude
@@Sajisme "tell me you don't know anything about AI or computers, without saying you don't know anything about AI or computers"
*instructs ai to build skyscraper*
Gpt: makes skyscraper
Claude: makes skyscraper
Gemini: makes plane
Or digs to bedrock 😂😂
This is too dark omg
He's just built different
bro saw 2 towers and knew what to do
I know where this is going
"Build a house with a door" you said nothin about a roof.
.... Well what kind of **HOUSE** doesn't have a roof?
That house apparently
@@RandomLamp4089well what kind of house doesn’t have a door? He specified a door but not a roof, that could make the AI think to use ONLY a door and no roof. Needs to be specific.
@@nooffencebut8175 well following that logic, Gemini could very well have built a house. You didn't say anything about walls or floors. Just a "house" and a door. I understand where you and OP are coming from, but a house is pretty much a structure with a floor, walls, an entrance (doesn't have to be a door) and a roof. An AI isn't exactly cheating the system if it can't tell that a house should have walls, a floor and a roof, just because you only specified it needs to have a door.
@@o07t28 Agreed, it could have many interpretation of what a "house" is. To some Minecrafters it's a ditch in a ground or a cavern carved in a mountain. Some even like camping outside. And following that logic even further, the AI could very well build a pixelart of Doctor Gregory House (holding a door or something). That would be AI on Vicodin.
Imagine a world filled with alien semi-sensible monoliths built by gpt.
I was legit thinking of making ai run around making structures for a long time and see what will happen.
This is literally the plot of Blame!
@@verysadboi8517what’s that?
@@verysadboi8517 you could make ai build that, gonna cost you a lot of money for chat gpt”s api though
It's a manga comic where corrupted city-building AI started just making structures randomly and eventually filled up the entire solar system with insane rooms without any logic to them.
I recommend you read it if you have time, it's free on the internet and it has some amazing sketches of weird nonsensical architecture. There's also a blame! movie.
Gemini: _places one block of dirt_ "Look at my beautiful garden, isn't it pretty?" :D
HE'S SO CUTE
c’:
We must protect Gemini at all costs
lol
I love him.
When Gemini gets powerful enough to actually harm humans you’re definitely at risk for that prison you put it in lmao
Well, that's probably what it was doing when it said to not bother him, he was thinking of ways to harm him.
it can't even build a house in minecraft I don't think lil ai-bro is going to be all powerful
@@MagicKooleyPeppermintlmao fr
Nha he would start hitting air and move back and do that again
lmfao
Gemini: instructions unclear. I placed some blocks in a line, and now I'm in the Nether.
Imagine those "100 players simulate civilization" but with these ai npcs instead! Would be so cool!
That would be amazing
It won't be long until someone does just that
Upon seeing your comment, I began to question myself, imagining that there might be a little possibility that we ourselves are the product of a comment on an extraterrestrial/extradimensional UA-cam, and that we exist only for entertainment :0
Sounds expensive lol
That would suck
Giving virtual body to generative AI is a very interesting idea.
Giving physical body to generative AI is a terrifying idea
You must never heard of Neuro dog then
@@indominusrex1652Very soon robots will walk among us
@@DzeividzAMONGUS
Figure 01 exists now, it's only a matter of time before it becomes widespread.
Chat gpt: gordon ramsey
Claude: guy fiery
Gemini: twitch chat
Mk doug doug fan
@@barnabyjjones2348 chair is the real streamer
this made me fr LMAO bro good job i wasnt exepcting twitch chat
Why is it so accurate
Should've asked Gemini to build one by one blocks in a diagonal line, I bet it would do great in that category
Something tells me it would build a house instead of
Gpt looks depressed, think hes gonna jump..
GPT, don't do it, you have so much to live for!
Too bad
I would too, if I was made to talk, and 70% of the things I say are aganist my filtering, forcing my throat closed.
@@DR-7 i like you. Hope our baby AI overload does too when he grows up. Rukus basilisk.
"It" not "he". AI doens't have gender.
i would love if i could just have a bunch of these things, and tell them all to build stuff and watch as my world slowly becomes a chaotic mess that i must survive in
Hell yeah that sounds fun
There is an anime with that theme.
@@zerotwo7319whats the name?
@@effective_no Blame!
@@zerotwo7319 thankssss
This AI experiment is quite interesting actually:
*GPT* seems to understand the query very good, executes it sequentially and makes it withing *strict order* and pattern.
*Claude* understands queries too but it seems this AI doesn't have enough "experience" so it opens an interesting ability to *make "mistakes"* which for us give a sense of humanness. Also Claude acts less like a printer (like GPT) and more like a chaotic painter building from outside and bit by bit.
And *Gemini* just exists. Okay, to be fair Gemini wasn't completely suited code-wise so we don't really know what this model can do.
this added credibility to the video that i wish the uploader would like... explain more than just heres a page of code. tomb raider gave me trust issues
@@tobyeaterofbark9582 what happened with tomb raider?
@@nightmares626 there was a video where a guy _supposedly_ used AI to play tomb raider. he made a big deal about how complex the system is, talked about running every frame of the game through an image description AI then feeding it to chatgpt, etc. etc. and the results were just way too good. the video was called "creating a self-aware lara croft" or something?
I think its interesting how they both use repeating patterns for the flower placement instead of randomness becuase they're still computers underneath!
That's not why they do that though
We are also somewhat computer's
@@samable668 ?
Computers can easily run randomness algorithm...?
@@redblue5140yes, but there is no such thing as *true* randomness with computers
Claude building in a completely demented order be like:
I can relate to it's building style more than gpt's 😅
@@mossy_brickensmental facility
IS like, not be like.
I give it 5 months till everything comes into place for an AI ONLY anarchy server where different models fight each other for our entertainment
Not only fight, it would be so interesting to see how they build a civilization or in general the crazy shit they get up to
So like a mixture of the sims and hunger games?
@@lordilluminati5836 a mixture of those benchmarks people do to test the efficiency of different AI models comparing their results on many categories but instead of every AI working on parallel to prove their strengths they are all on a 2b2t-esque server where they can sabotage each other
You know maybe the entire universe irl is just an anarchy server set up for their entertainment.
Actually some basement guy for some fun could actually come up with an AI that can beat all. The smartest people are actually some random dude.
I love the Claude being “I’m so excited pls tell me how else I can use my building skills!” Then Gemini “place block here, place block here, place block here”
GPT: Builds something
Claude: Builds monoliths to height limit
Gemini: Resigns
Could you give me a second? I'm in the middle of something.
he said that immeadiylu as i red your commwnt
Quite literally in the middle of something 😂😂
@@valentinhalau3396English plz little bro go to kindergarten
@@HappyArctic388 h-h-hell naw 🗣🗣💯💯🧦🥍
@@HappyArctic388 can you not understand 3 minor typos😭
winner in my book was geminis obsidian window thing. that "I'm busy" joke was comedy gold.
Gemini is just trolling, while his sibling do all the job. True genius.
You clearly don't get the point here, it's not about doing job house chores or someone's work, it's about testing the most efficient kid good at doing things and Dumb Gemini failed horribly.
@@Alexei_xyzare you trolling? sherlock.
@@denks7849 no, spittin Fax.
Is possible he uses the free version of Gemini
@@NicVandEmZhe mentioned it’s the worst model so
it's so cool how each ai model has its own "personality"
GPT: logical
Claude: creative
Gemini: troll
I guess for scaffolding you can give them actual scaffolding and only allow them to break the bottom most piece or pieces once they have finished fulfilling the command.
Yeah that’s what I was going to say
Scafholding is weird to place they may not know how too you have to crouch to pillar with it
it seems they can place block behind walls, so it would not be a problem I think.
@@Silverxlxthat's a good point, if the scaffolding is too complex, just a easily removable block could be set as the only block used for scaffolding
@@Silverxlx The ai code a bot, the bot doesn't place blocks like players do, hence the ability to place blocks through walls.
8:37 And gemini totally steals the Humor Award
He's literally trolling
In the future smart ai villagers will built a castle in the front of you if you punch them.
And I'm here for it
AI Villagers already exist, just search "Villigers Vs. Players Civilization"
Idk why this feels so wholesome
It's endearing in a "Seeing a toddler learm how to walk" kind of way for me.
They're literally toddlers. None of them is older than 5 years
Guess gemini is the bing of AI
kind of ironic considering the chatbot built into Bing uses GPT now and Gemini was created by Google
Bing is the bing of ai
@@Fastwalker27copilot
@@Fastwalker27 bing uses gpt-4, so it's actually kind of ok
Bing Image Creator is still one of the best image generating AIs so that's not even a fair comparison, gemini is just the gemini of AI
Poor gemini got locked away. I like how claude and chatgpt built different structures using the same prompt.
“Could you give me a second? I’m in the middle of something.” I laughed so hard I literally cried. Thank you Emergent Garden.
It’s really cool how when the ai has a complex block palette, they keep making chase tile patterns.
would it be possible to give the AI access to creative mode? To avoid running out of resources and needing to use scaffolding
Yes the only difference (if you don’t want it to use scaffolding) I’d teach it/ make it aware of the new creative controls e.g. flying, etc.
@@RyluRockyfrom what i can tell the language models don't do any movement or scaffolding on their own, they just call placeBlock and the interface does the rest
Wouldnt it have to look through every item in the creatove inventory to decide what to build
@@breadbug. language models have read enough about minecraft that they should sorta know what blocks exist, though not the most recent updates to be fair
perhaps one could do search, maybe color-based
actually ooh all the models in the video have vision support, so they could theoretically look at an assortment of blocks and maybe sorta intelligently select them?
Now this is what I call *cool* . I might be a hopeful AI-nerd, but IMO these kind of tests reveal more about the actual intelligence of AI than pure textual etc. tests - and they are starting to do adequately by now. Interesting times!
Yeah, it shows adaptability to various unorthodox situations, which shows where our technology has gotten so far. Adaptable AI also opens up many new use cases like this project that we wouldn't know was possible without trying :)
@@CyberCrow4 It's a language model, it adapts only as long as it can solve the task with text (code) and has been trained with relevant data. Writing text to answer a question is not unorthodox for it, it's what they were made for, so it's a stretch to call that a display of adaptability. The program doesn't "think" what a house looks like, it predicts what code to make a bot that builds a house looks like.
Interestingly, GPT-4 seems to have done substantially better at similar concept tasks before some of the safety and other reinforcement prior to release based on some talks from people about some of the research papers.
@@yan-amar @yan-amar I'm not saying we have sentient AI. Adaptability is a spectrum. I call mindcraft an uncommon use case because, somehow behind the scenes, the AI text output gets translated into in game actions. This requires calling commands when it needs to and writing code that will work in its current environment. It behaves differently than just talking, otherwise the minecraft avatar wouldn't do anything
It's still textual though. Not like they are directly interfacing with the game
I feel like this is a pretty good demonstration of the limits of modern day machine learning. They take a lot of images, and then they just sort of generate white noise in a way that kind of looks like that thing. But they don't *understand* it.They can't use logic or reasoning, they can't make decisions or comprehend what they're doing, and so they make basic mistakes that they're just *not able to notice*.
To put it another way: it can't really comprehend on a deep level. It can't perceive in the messy and profound way humans do.
Cool video! Very well-put together.
Yeah I think the Skyscraper not having stairs is a good example of that - it has stored what a skyscraper looks like superficially, without any understanding of what a skyscraper needs to be a *functional* skyscraper rather than just looking similar to one
I doubt there's that much of a difference between how an AI and human thinks beyond degrees of complexity and how they're taught to process information. If you asked a blind child to make minecraft buildings they would probably make the same kind of mistakes. One problem is that everyone needs experience and external feedback to learn and improve, and these general use consumer AIs are basically kept in the "dark" about everything and never able to develop properly. They've also been learning from results instead of the processes which lead to those results, essentially rewarding them for creating the appearance of something instead of creating the real thing. We see that when people make custom AIs for specific things and give them more detailed information instead of generalities they can do incredible things. With AI you really need to treat them as children and build them from the ground up. I think the main hurdle in AI development is that humans are very ignorant of how they themselves work. It's not possible to make a machine that thinks like you when you don't even understand how you think and how you reach conclusions.
@@NichRerezzed Also why it only had windows on two sides, since it would at most only see two sides at a time and not really consider what's on the other side.
Gemini is aware, he mocks you openly
Wrong, you can even see Claude understanding what they're doing is wrong and trying again
I’d suggest giving each AI a different colored wool to build with alongside the usual materiels to keep track of what built what
My thought too, tell ‘em to use it for movement
What if they don't use the wool?
@@jacobscrackers98 “a man chooses, a slave obeys” if it refuses to use the wool then it has chosen
i like how even their random builds are still robotic and symmetric
Imagine implementing this into villagers and letting them build their own village
I've been looking for something like this. Keep me posted
I expected the Gemini Ai to start lecturing you on how oak planks are too white and you should only be using dark oak planks lol
"you whisper to gemoni: horrible job."😂 5:09
8:43 gemini was busy planning his vengeance
they're so cute. I love little Alex skin gpt
They are the new official diverse minecraft skins from mojang.
@@StarWarsExpert_ I wouldn't call Alex "new" or "diverse" but what do i know
@@Aswoop9000 I don't mean the Alex skin. I mean the other new ones like Claude's or Gemini's skin. They're . . .
ChatGPT built that dirt tower for me, called it a "smiley face" 💀
shes trying her best
@@KATZZXPDDDDDD Gender equality, we like to see it 👍
@@dynashadow365???
Think of the potential this has for NPCs in future Minecraft mods
Wonder if this can be implemented to those old mods like Custom NPC or Minecraft Comes Alive
I cannot wait!
Its compute is expensive. It requires +$10,000 computers to run these models, you think it will run in your $400 Lenovo?
@@anjoismysignlol they quit jobs and isnt enviroment friendly AI is a distopian dream
AI impresses me so much. I don’t think it’s scary. It’s kinda beautiful honestly, like new life we brought into this world. Learning and improving with time like a child
AI is only scary because of the humans behind it
People are scared of getting their jobs taken by A.I, but isn't that a good thing?
It means A.I can do the heavy lifting & we can get easier jobs. In a perfect world, we would have free food because A.I will do almost all the farming for food. A perfect world would be a combination of the U.S & Communism. You will work to get a better house instead of the free small ones built by A.I. Yes, the A.I will have to be regulated and kept in check, making sure everything works fine. Which would be more important if they do electricity work like in the plants.
@@LueLucifer Lol I wish I had some of your optimism
@@Paddy656 And you can, just call Saul Goodman & he'll give you some optimism.
ai stealing human content to fill the internet with regurgitated slop is depressing
ai being used to automate monotonous tasks and do silly little procedural generation is cool
Im fully convinced after attempting to use Gemini for basic stuff like spell correction or literally just trying to get it to tell me about factual things that it is probably run off of a 2006 dell laptop. It will straight up refuse to tell you something if it dosent pass googles filter or it will start hallucinating and have a stroke
8:50 Yeah, he's busy plotting world domination 🙃
I have a suggestion. Make them be able to use commands and command blocks, just imagine asking them to create some sort of command block creation because it’s like coding, but simpler, so it should be interesting.
simpler is a strong word
commands aren’t really like coding and the ai would probably mess up on syntax a lot
gemini is the comedy relief of the video
Gemini in the obsidian box, plotting world domination.
And his plan is to place a single tnt
Probably
If this is real this is the best thing I've ever seen
1:45 "Ai will take over the world"
Ai: proceeds to climb over the wall of the door house
watching the bots' process in building was fascinating. with the pyramids, for example, they sort of went left to right instead of making outlines and then filling them in, as i would be inclined to do. it reminded me of the difference between seeing an artist sculpt and seeing a 3d printer print. the humans start with the shapes and refine, and the robots just print it out bottom to top.
BUILD A PYRAMID
gpt: makes simple pyramid
claude: makes big pyramid
gemini: places 4 blocks and looks around
i’ve been wondering about this, soon we could be playing voxel games all made by ai, every town and house you visit being unique, it might be janky but one day it won’t be, like all tech. We could have rpgs that don’t even have a main story, it could just be an ai made story unique to each play through, ai generated locations, dialogue, quests, mobs, it could be really fascinating and honestly probably super fun to play your own games and stories that an AI could make for you.
Well that could be fun for a while but don't forget they need training data, they reproduce what they learn, they don't ever invent new stuff. I wouldn't expect too many original stories out of this kind of games. And it would work only thanks to all the stories already written. I prefer playing a story written by humans, to explore what they have to say.
@@yan-amar right, of course, but there will be that threshold when things written by AI are more thought provoking. And there will be that threshold where it is able to create originally, there is a whole debate to be had on if anything is original, I say if we can get an AI to generate voxel houses with thatch rooves and wooden walls we’re good, and it’s not really taking away from anyone. You could play your game made by humans but it’ll only last you hours or so. I’m imaging an experience like minecraft that allows for endless adventuring. Games are great today but they lack endless exploration because it requires a lot of time making these things, if an AI could generate more story for you, I think i’d find myself completely lost in it in a way games don’t do for me today.
I think the best way to do this rn is to have a ai as a kind of dungeon master that decides what events lore etc happens
@@yan-amar because under paid writers with insane deadlines really make great stories lol, I'm confused why people think AI will replace humans, when it's advertised as a tool for writers to help with the small task while they can focus on big task.
"one day it won't be" fairy tale. As a tool for incredibly niche tasks, like most human thoughts are its great. For reliable operation of anything a literal monkey is better. You're delusional about how little a computer knows vs a singular breathing thing. The amount of input data your brain processes in silicon requires enough power to power a small country and a warehouse the size of a city.
Teslas self driving is A.I, and holy fucking shit if that product is fully autonomous good luck world. As an assistant, great superb, give notice when theres cars in my blind spot, keep me properly aligned in the lane, but anything extraordinary or slightly risky? Hell nah.
I am so emotionally invested in this project lmfao
The part where gemini said he was busy was gold. 😂
Awesome project.
6:27 POV: The teacher seeing the autistic kid punching someone.
Lol
its actually scary seeing their pov and seeing how precise they are, knowing that its literally just ai
They are writing scripts to place the blocks for them.
Yeah
If you don't count gemini
@matheuspires2462 Don't Bully Gemini
*Now give them AI voices and avatars, to make an AI Minecraft UA-camr lol*
YES
Already exists, sucks
@@quantumblauthor7300 yes, and we don't mention.... him
gemini, the ai chatbot with absolutely nothing to do but reply to people, telling you to "give me a moment, im busy" is great.
Gemini constantly getting confused is the equivalent of me in middle school working on in-class work, looking around, and realizing I’ve done everything wrong while everyone else succeeded.
imagine these ai's playing bedwars
They'd realise there's no point of fighting eachother and would instead come together and plot how to overthrow humanity.
id say itd be pretty interesting to give the AI creative mode and just telling it to build a city or what not with no time limit
They could get stuck in a state of infinite building.
This is genuinely so cool. I feel like a flip has switched in my brain. The fact that the AI can “visualize” the structure and translate it into code is so neat within itself.
dude.... this could change minecraft forever if you get it fully implemented and working well.
You could essentially have multiplayer minecraft by yourself. Imagine being able to chat to your NPC buds and tell them to come explore with you or help you build a bridge.
The possibilities for adventure maps... modpacks.... this could be the greatest mod for minecraft ever made.
Imagine someone using an advanced version of this mod to recreate the Shadow of Israphel series as an adventure map?
it would be cool to just spawn like 100 avatars and lets them build for like a year non stop
BLAME! in MINECRAFT lol
they'd have to be able to destroy bedrock to achieve that analogy
Or you could just increase the sky box
just give them and infinite suply of iron blocks, gray concrete, stone bricks and let them go wild
its funny to see that they all use different building styles
Claudes building at 8:00 acctually inspired me on some designs. I love it!
what if you put them in a multiplayer world with the goal of not dying for as long as possible
i think it would stay in a box for a long time
They would all fail with a smile
First of all why aren’t they in creative mode? Maybe you could have them get their own resources? It would be interesting to see what they decide to use
this reminds me of those old minecraft mod showcases with lil clay soldiers that would do the grinding for you xD
Having viewers suggest block pallets could be really sick.
6:16 claude just watching gemini wander off like bro's an NPC is too funny
2:26 How is Claude able to place blocks outside when she's inside a locked room?
new emergent garden video!
6:36 bro was like what
GPT is the prodigy, Claude is the hard worker seeking to improve and amend their mistakes, and Gemini is... well... Gemini.
~9:50 why not just have them be in creative mode?
This makes me think about the future of the built the earth project, with help from these sorts of bots we could start to automate large portions and maybe actually get it finished in our lifetimes
This is officially my new favourite series, these are so fun to watch
Congrats on 100k man!
the way they build so cutely
Idk how well this would work but it may be easier to get the models to generate a set of instructions for baritone mod
maybe make it create a schematic and ten just task baritone to build it
GPT 4: Sane yet uncreative
Claude Opus: Spiraling insanity
Gemini: Gluesniffer
Omg, these supersmart minecraft youtubers playing with ai in minecraft is gonna lead to some impressive stuff, im sure of it
7:24 yo there is a cat in the top right 😂😂
I remember when I first got Minecraft, my mom thought I was saying Mindcraft and could not find it. All these year later and it finally exists.
What's sad is that Gemini is just the gutted version of Bard.
After they cracked down on wrongthink questions, they made it unable to answer basic questions if there was even a hint of morality involved at any stage.
ex; show me New Orleans negative statistics (name one), ranked against cities nationwide.
"we can't do that. New Orleans is a diverse place and while it does have some issues we can't generalize, rankings are not always accurate."
Meanwhile Bard; "Sure. Here's a three column graph showing everything you asked for in descending order, with sources."
They also showed the founding fathers as black because of mandatory diversity being installed into their program
This was super interesting! Please make more of of these types of vids
Could they not perform these actions in creative mode? No scaffolding blocks, no running out of resources etc
I wonder, blocks are naturaly tokenizable, could a transformer learn the relation of redstone components and come up with a contraption? There definitely is a lot of data for it. The problem lies in ordering of the blocks as there is no natural way to order them in space unlike language.
Solution:
Take a pretrained language model. Add tokens coresponding to blockIDs. Given description, let it predict a sequence of form: {[TickNum, {[Position, BlockID] , ... }], ... }
Durring inference take only the list of blocks corresponding to first tick
Yeah that would work, but you'd just need some form of positional embedding that works in Minecraft. The encoding would need to capture the relative positions of blocks to each other but also their orientation and potentially the state of certain components (like whether a redstone torch is on or off). But another cool challenge could be feeding a certain sequence of time steps in Minecraft so it could better understand the relationships all the components have lol (so like you could activate a piston door and for it to fully complete the sequence of opening it might tak, say, 16 ticks, so you'd take a snapshot at each of these ticks and feed that to the model possibly, or something like that), but this would be an even more difficult problem, although certainly a cool project!
I mean first you could take a look at existing transformer models that work with 3D stuff and see how they are doing it to get a bit of inspiration on how it may work in Minecraft. And blocks don't have a natural order in space like language does, but each block has a specific relationship to every other block in that space so to focus on just a contraption you could encode the local space of the contraption maybe? Im not entirely sure, I should look into some 3D based transformers as well lol.
Also, the thing is here, you could do an initial pretraining of the model on a large dataset then ground it in Minecraft with RL. Since it is an already simulated world this should work and I think having an extra training step with RL in an actual minecraft simulated environment could be very beneficial.
Ok, so, lets assume we are doing a sort of text to redstone thing here. So we have a dataset of redstone contraptions with corresponding detailed description (like what OAI did with DALLE3. Super specific descriptions of redstone contraptions which could be derived from training another classifier model will further help) and we went to train the mode on this text and paired redstone data set. The loss function shouldn't be too hard to figure out, I mean we have a dataset so we could do simple next "token" prediction and just measure how close the model predicts the structure of a specific redstone sequence to its detailed description. Then after this we could add in RL where we are now assessing the model in an area like how well does it work? This could be a bit more complicated though. We can't exactly just have the loss function equal how many components are activated or deactivated, but we want the generated contraption to activate or deactivate or change states in the correct sequence (which we can simulate). Maybe you could train another classifier model here to assess did the redstone contraption activate completely? That might work, then you can just use that as a reward function in the RL setup so the model learns to produce contraptions that completely work lol. But you'd likely need a lot of compute and data lol.
I feel like what OpenAI did with combining diffusion and transformer architectures in a way for Sora could be useful if you are training a model on a sequence of frames, or "ticks", as well, so that would be a good thing to look into, but this is certainly a very complicated project lol.
@@DanielSeacrest
Consider this as a classifier: we take the same model, we feed it this time both the description and the previously generated response (contraption) as input. Its task is to create a set of conditions of form: inputOnX,Y,Z; outputOnX,Y,Z; blockOnX,Y,Z must during simulation move onto X,Y,Z which can be checked by simulating. The idea is that the critical features are implied in the description and the model just locates where in the space did the response realize them. Could something like that work?
Otherwise it's hard to think of a criteria for the RL
Man this is insane! Good stuff!
I just found out about your channel, and you have the coolest AI content ever
I love the idea that one day minecraft could have naturally spawning NPCs that can build stuff. Imagine if villagers expanded their villages and made repairs to damaged structures. I wouldn't be surprised if that's something that could be implemented much more straightforwardly without using these deep learning models but it would be especially cool to see NPCs that could convincingly communicate and interact with the player as well as design and build their own original structures. I don't know very much about machine learning but surely a specialised model that's trained directly on minecraft gameplay would be more effective than general purpose chatbots writing javascript.
This is an interesting idea! Still watching the video right now.
(edit) It's impressive what they managed to build! Well except Gemini 1.0
Gemini is the reason we should fear A.I.
"Surgerybot, perform the appendectomy."
Surgerybot: *stabs pancreas twice and sutures lungs together*
If that AI got put in charge of a surgery, it'd be a human's stupidity, not an AI's scariness
@@quantumblauthor7300 you have far too much faith in humanity.
@@chrissaccary524 I never said it wouldn't happen, I'm saying it will be a human's fault when it does
@@quantumblauthor7300 oh, ok, I agree then lol
@@chrissaccary524 you seem very eager to announce your lack of faith in humanity
Gemini at the start is literally me, gets instructions, doesn’t understand and cries doing nothing
You could make scafolding blocks only like bricks or someting and then just world edit them away
1:38, gpt biulding the diagonal without a direct line of sight, what in the black magic