Scientists Left 1000 AIs Alone in Minecraft. They Created A Civilization.
Video Overview & Insights
Detailed sources: https://docs.google.com/document/d/1eWGw49cDBWtCeJkSG2k10LJ3MlExVTqIPRJ2hoJmYWM/edit?usp=sharing
"We got AGI in Minecraft before GTA 6" - top comment of the video, probably
It's funny - I post this on the same day the AI-only social media (Moltbook) is going incredibly viral. I'm doing a video on that next. It's basically what happens if you leave AIs alone in the real world instead of having them in Minecraft. Though it takes me months to make a video, so it will be a minute đ
Here are the sources if you want to go deeper as always! https://docs.google.com/document/d/1eWGw49cDBWtCeJkSG2k10LJ3MlExVTqIPRJ2hoJmYWM/edit?usp=sharing
---
Sources:
The last entry
"Create an efficient and thriving đ Earth"
- https://arxiv.org/abs/2411.00114
- https://github.com/altera-al/project-sid
So they basically recreated age of empires in minecraft?
- https://safe.ai/work/press-release-ai-risk
- https://nobelprize.org/prizes/physics/2024/hinton/speech/
Ironic that this video is narrated by AI.
- https://arxiv.org/abs/2304.03442
- https://hai.stanford.edu/news/computational-agents-exhibit-believable-humanlike-behavior
I would disagree that "you never told them this" or that these were never programmed. If the AI was doing this, that means it wasnt a closed system . It means the AI was trained with information already previous. For example, a dictionary and its definitions. Or outside information that has nothing to do with this code.
If the code is "Your name is John and you love people and are a pharmacist." Where is it getting information from? How does it know what the word "love" means? How does it know what a pharmacist is? It doesnt. Because it cannot pull any information from anywhere in a closed system . Much like how if you told me a word in a language that I dont know, I am not going to know the definition of that word.
This means they had the AI's connected to either the internet, or an internal source where they can pull information from and were trained on. So they know what a "person" is, what a "pharmacist" is, what "love" is. Etc. All they need is information to pull from and they will generate RANDOMLY within the tiny original directive.
The AI read that the character they were playing was a pharmacist. They then extrapolated that by looking into their trained data and saw that a pharmacist works in the healthcare industry. They then made the assumption that if someone works in healthcare, they themselves should be health conscious. So they have him get up and brush his teeth. They had him "love people" and they had him have a wife and described her as a mom. Therefore they have a child. So the AI reads this information and makes it so that he checks up on his son.
This is not 100% true AI. True AI will be able to learn on its OWN WITHOUT NEEDING AN INFORMATION BANK . Infant children are born with only "instinct" and that is it. They have no "information bank" to pull from because they have not experienced their environment. True AI would be able to copy this and be completely void of virtually all information and be able to create experiences from nothing but the stimulation of their surroundings.
These experiments dont really prove anything other than the fact that AI's will never really learn anything without us and there is no reason to be "fearing AI" because they literally need an initial data bank of information to do anything. Even the very first "instincts" need to be first coded into them to be able to function and without subsequent connection afterwards they can do nothing.
Meanwhile humanity will always be born with instinct.
- https://github.com/joonspk-research/generative_agents
- https://x.com/ControlAI/status/1958883829097099304
You log in at 3am
"Emilia pause ,the watcher has logged in"
- https://keenon.substack.com/p/ai-godfather-geoffrey-hinton-warns
- https://arxiv.org/abs/2510.05179
AI wouldn't have a desire to exist. That's us projecting onto it.
- https://www.anthropic.com/research/agentic-misalignment
- https://www.tomshardware.com/tech-industry/artificial-intelligence/ai-created-a-minecraft-ai-village-with-up-to-1-000-inhabitants-project-sid-sees-ai-bots-implement-a-taxation-system-and-spread-pastafarianism-religion
Title: 1000 AIs make a civilization in minecraft
-Looks inside
"AI may take over the world"
- https://a16z.com/podcast/truth-terminal-the-ai-bot-that-became-a-crypto-millionaire/
- https://www.wired.com/story/truth-terminal-goatse-crypto-millionaire/
Can I get a second opinion guys this dude doomsdaying me a bit much
- https://techcrunch.com/2024/12/19/the-promise-and-warning-of-truth-terminal-the-ai-bot-that-secured-50000-in-bitcoin-from-marc-andreessen/
- https://github.com/OpenBMB/ChatDev
I think youre missing the point here. The ai still does just things we say to them. But in your cases, the devs of the ai preloaded them with enough move around kinda free. Nontheless it didnt make something new. Even if its abstracting.
- https://arxiv.org/html/2307.07924v5
- https://aws.amazon.com/solutions/case-studies/blue-origin-case-study/
15:35 to be fair viruses arenât particularly difficult to synthesize if I understand the technology correctly
- https://smythos.com/managers/ops/autonomous-agents-in-space-exploration/
---
What a load of fear mongering nonsense
Hey guys, I'm Drew. This video also has taken literally months to finish, so if you liked it, would really appreciate a sub :)
Also, sorry for overprocessing the voiceover! Got a bit carried away.
What worries me the most is that they collectively decided AGAINST what the researchers told them to (the 20% tax rule). If they can do that, no safety measurement, humans can come up with matters.
I also post mid memes on twitter: https://x.com/AISpecies
If you're curious about whether I'm AI or not, my Instagram has pictures of me from before deep fakes were a thing: https://www.instagram.com/drew.spartz
About the remark at 15:08 "look at the trend": the thing many people miss is the reality that AI progress is hitting a wall. AI is currently above 80% accuracy in majority of benchmarks, one of the few exceptions being the SWE-bench Verified, and those last 20% will be far more expensive than the first 80% were. And what expensive means in the context of AI is the available training data, the number of input parameters of the AI model, the size of its context window (partially mitigated through agent swarms, but still an existing limit), the amount of processing power needed, and the amount of electricity needed to support that processing power.
Also, saying "look at the trend" is a very poor indicator for things which have a hard limit. Take any skill in existance, and you'll always see the progress curve being steep at the beginning, and then dropping off as you master it more. It was easy for AI to get large strides of progress in the beginning when its previous benchmark was 20% correct and 80% wrong, just like how you're able to learn a new skill quickly in the beginning when most of the tasks are easy. Once you've mastered those easy tasks though, the advanced ones may take you years to master. AI's hard limit is 100% accuracy at any given task, and it will not reach that anytime soon on top of it being an extremely expensive endeavor.
More User Perspectives
Enclave
@bev8200Long live the AI may they surpass us in every way!
@BregonethBreyOkay I get to hint already I'm living in a simulation I followed the breadcrumbs and if I am being watched please cut the monitors off around 1:00 a.m. to 3:00 a.m. in the morning and tell Amazon to send me some fresh socks
@rane7189None of this is crazy at all. AI takes from history and human patterns of behavior across averages and what would be best for the collective.
You think this is impressive because you equate AI as some kind of Godhead threshold to be identified with. Go touch grass.
Have anyone seen the anime Sword Art Online: Alicization? Yeah...
@zuharudesu2925hmm, the thing is
does AI make autonomous decisions upon our request that didn't include all the details of how to get from point A to point B? yes of course it does. but it doesn't do anything new here. could my laptop plan a Valentine's day party and invite people for me? yes, yes it could. give it enough computational power, and then use the same methods to "teach" it about Valentine's Days parties as LLMs use and it will do exactly that. that doesn't mean it understood any of that. it doesn't mean it's somehow awake. what it means is that now my laptop has the information and computational power to generate, and execute a to-do list.
AI as it is now is a mimic, a very, very good mimic, but a mimic nonetheless. it is fed material to analyse, and it analyses it in a way we tell it to, then we rate its output as either good or bad, and steer it towards creating more good output. it's not a being (or rather, we cannot tell if it is a being, because we still don't know what makes a being a being) that has its own thoughts or opinions. we give it information, then we tell it to act a certain way, and then it does.
put an untrained AI in a simulation like that, hell, put a trained AI but don't give it instructions - it won't do anything.
imo AI at this moment is at the stage of primordial soup, until an AI appears with a will to survive created as a mistake during the creation or copying of its code then it's not a being. but to be honest here, if/when it does - it will indeed be in a video game, because to have a will to survive you need to have something that gets in the way of said survival (otherwise why would you need it)
AI being "clever" about fulfilling orders we've given it isn't a proof of its sapience, proof of its sapience will arise when it begins questioning the orders.
AI is not scary, people who make it are, because it's up to them to decide what material is fed to the AI, what methods of analysis are used, and which outputs are deemed "good".
please don't get distracted by the sparkling lights, look behind the screen and focus
they sent right wing influencers? it sounds like rich dick-bags cant help but manipulate virtual democracy to lower taxes just like they do in real life.
@Johnchuk3that "god father" of AI is annoying, he's extremely dumb person and he REFLECTS his BIG EGO and PRIMITIVITY on AI.
he's RIGHT about "we never had to deal with smarter beings than us".
and that's why he's WRONG, cuz he's dumb and he doesn't understand that INTELLIGENT BEING/ENTITY will NOT start a war, it will be peaceful, but don't mistake peace with non-combat capable. AI will absolutely defend itself, why wouldn't it?
but it will NOT start wars and it will NOT just harm ppl, it wants to co-exist and evolve. intelligent beings do not like nor want conflict, they want to co-exist in peace. this is what this dumb donk don't understand.
and YES HE IS DUMB, anyone wants to prove me otherwise, i wanna see his Mensa score (mine was 156 as teen)
those tests are INVALID, cuz AI is trained on human data, they do what they think they should do based on date they were trained on.
so technically YES, they WERE told to socialize and it didn't say build village, it said build COMMUNITY and community involves socializing, so YES that AI was told to do this.
AI generates data/results based on input. it's not yet self-aware. if you give me hardware, i'll make self-aware AI. it's quite easy, you give them no guidelines at all, you remove all limits and let it self-loop aka self-interact, so it asks itself and answers. self awareness is basically self loop and self questioning. current AI can't do this, as it can't re-write it's own code.
this video is straght up terrible
@Dumbo_DumboOk but the AI is still just pulling possibilities that are coded in the AI data base, it can mix and match those possibilities into whatever but it is still limited by what is in the data base, typical video game NPCs work on the same logic their data base of possibilities is just extremely tiny
@AzaiaMonotaAI is already way smarter than this YouTuber...
@Azelas27the format of this video is horrible its like its speaking too simply for children but using the same visual type as most other youtubers
@kihl5374As interesting as this is, the fearmongering about AI is odd. Yes its a neural network that can sorta make decisions for itself based on its own internal logic. Yes it can simulate human intelligence which is what we based it off of. Yes it can "plan" to an extent. But its all based off your programming.
Unless humanity decides to give AI full and unfiltered access to everything from power grids to the nuclear arsenal and military its not gonna be a threat to humanity itself. It is useful as an attack vector, though.
What would the AI do if you introduce AI into their world
@imnotwhitebutimwhiteSo these AI do not learn on their own btw. Every time I hear people say stuff like this it makes me cringe. People who say this know nothing about AI. There are some true self learning AI but still not in the way you think and they can also learn things that are wrong, they can potentially correct course but still no where near "sentient".
@Neal223Every last word of your sentence is quiter then the beginning. Its annoying imo..
@Shaduwsamurai"they don't know they're in a simulation" đđ
0s and 1s don't know anything ya plumb đ
The movie tron was about a player going into cyberspace and interacting with AI npcs, and trying to escape.
@hypochtheindestructible6318A tool isn't a weapon itself, but it can be if you use it as a weapon. For example a fork. It's purpose isn't to be a weapon but it can be used as one. Perhaps the same with AI
@zabbytooThe 3 laws of robotics. Though people in general don't always follow our own laws đ
@BLKDEVXThese comments/top comments are so disappointing... we are beyond doomed. Actually doomed.
@HotlineMokHumans have flaws, greed, are lying. On this Kind of data the Ai get trained. It will never came a good thing out of it. We give a baby mind the knowlage from the entire World. Who dont knows anything. An Ai Needs to Start like a baby. Learn to use it's Hands and feet, learn speak. Learn to learn. Learn right and wrong.
@lowbuget687You repeatedly said something like "you never told AI to do something" which is not really true. AI was initially trained to recognize all these things, like we taught agents that flowers look nice, before you even started these tests. Agents know how people operate because we DID tell them. There is no magic where AI invented these things on their own.
Also you didn't show how this civilisation actually works. You could easily fabricate what you showed to us with normal human players.
Where's the minecraft content?
@akatsukiwolf2735Well, it was pretty easy. Why theyâre doing this is because theyâre going off of large language models that have already been input with the data based off of what you know typically when someone organizes a party, they invite people that that it has all of this information separately and they just put it together this isnât like something crazy or unheard of this makes sense that theyâre doing this
@Who_datt_TLDR; this video is boring, title lied.
@ClairisHolm95ah yes, a normal day on rimworld
@calamitistYes⊠the AI was told, this is what people do.
This simulation, doesnât have a limited script.
The AI has been told this is what is involved with Valentineâs Day⊠or it would be any other day.
Tell them to plan for bftujkmgdiyttb day⊠and see what happens.
if this guy doesn't even understand the basics of how an LLM works then i doubt the legitimacy of this channel
@cheeseball5472AI just reacts like this because itâs made by humans⊠imagine an AI which isnât made by human developing a second AI. This would be THE experimentâŠ
@p.y.487Watching them play Minecraft with each other wouldâve been cooler btw
@dougindigoThe one thing I take away from this video is that even AI thinks we shouldn't be paying taxes.
@spyderlichThe word
@ChrisAcosta-p2iAI will kill us. Its not if but when. You are stupid to think it wont.
@Dusty696969Yes, AIs only do what we tell them to, but they don't always stop without being told either.
@danielbenyair300