YT YouTube2Text

YouTube2Text → Transcripts → Scientists Left 1000 AIs Alone in Minecraft. They Created A Civilization.

Scientists Left 1000 AIs Alone in Minecraft. They Created A Civilization. — Transcript

by Species | Documenting AGI · 2,622 words · 465 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00As you watch this, a thousand AI agents
0:02in Minecraft are hard at work building a
0:05literal civilization of their own.
0:08And they don't know they're in a
0:09simulation.
0:13MIT researchers dropped them in there
0:15with one prompt. Survive and build an
0:18efficient village. That's basically it.
0:21That's all it took for them to create an
0:23entire civilization. They wake up, go to
0:26work, gossip about their neighbors, fall
0:29in love.
0:30They even hold elections, argue about
0:32taxes, and make art. But this isn't just
0:36about Minecraft. It's our first glimpse
0:39of [music] a new species beginning to
0:41think for itself. And that could soon,
0:44according to Nobel laureates and the
0:46godfathers of AI,
0:48lead to literal human extinction.
0:52When you normally think of AI NPCs, you
0:54probably think of dumb ones that just
0:56follow a handful of if-then scripts.
0:59This is totally different. Those NPCs
1:02are not autonomous and not interesting.
1:05They can say like four sentences.
1:06>> How are you?
1:08>> Been better. How about you?
1:11>> Not bad.
1:12>> Good to hear. Goodbye.
1:14>> Goodbye.
1:15>> To start, each of these AI agents is
1:18powered by LLMs like [music] ChatGPT.
1:21And yes, each one is playing a
1:22character.
1:24But they're only given a short
1:25one-paragraph description of that
1:27[music] character. Everything else they
1:29do is entirely up to them.
1:32And researchers have run dozens of these
1:34fascinating experiments of AIs behaving
1:37weirdly autonomously. They leave AIs
1:40alone together to see what they do when
1:41they think no one's [music] watching.
1:43Like how Jane Goodall studied wild
1:45chimpanzees.
1:48I'll break down three of the most
1:50important research papers
1:52>> [music]
1:52>> showing just how crazy our future might
1:54get.
1:58Take the Smallville experiment for
2:00example, created by Stanford
2:02researchers. Now imagine you're one of
2:05the researchers. You've given your 25 AI
2:08agents each a one-paragraph description.
2:11For example, John Lin is a pharmacy
2:13shopkeeper who loves to help people. He
2:16lives with his wife, May Lin, who is a
2:18college professor, etc.
2:23You hit run on your simulation. Day one
2:26begins. What do you think will [music]
2:28happen? Anything?
2:30You watch John wake up at 6:00 a.m.
2:33He brushes his teeth, he makes
2:35breakfast, [music]
2:36and he checks in on his son.
2:38You never told him to do any [music] of
2:40this.
2:41He goes to work, he helps customers, he
2:43has lunch with his friend, Tom. They
2:45discuss local politics.
2:48You never scripted these conversations.
2:51They decided to get lunch.
2:53They decided to discuss politics.
2:57But why? And how?
3:00Well, this continues until day 12.
3:02And so you decide to test something.
3:05As you play your role of a god in this
3:08world, you inject one extra thought into
3:11one [music] agent.
3:12Isabella Rodriguez,
3:14I want to plan a Valentine's Day party.
3:17So, you watch Isabella. She's at her
3:20coffee shop. Tom walks in for his
3:22morning coffee.
3:24Isabella proceeds to invite him to the
3:25party.
3:27Wait. You never told her to invite
3:30anyone.
3:31You just said she wants to [music] plan
3:33a party.
3:34She decided on her own that planning
3:37means inviting.
3:38You keep watching, wondering what she
3:40might do next.
3:42Maria decides to ask her secret crush,
3:45Klaus, out on a date.
3:47Hey Klaus, want to go to Isabella's
3:50Valentine's party with me?
3:52And he gladly accepts.
3:54February 14th, 5:00 p.m.
3:57Agents start arriving and Klaus shows up
4:00with Maria.
4:01Not just at the same time,
4:03together
4:04as her date. You read their
4:06conversations. They're discussing
4:08politics, gossiping, forming new
4:10connections,
4:12and falling in love.
4:13You didn't tell them to do any of those
4:15things.
4:16You see why it's ridiculous to say AIs
4:18only do what we tell them to?
4:21Andrew Ng said,
4:22>> When people talk about AI being
4:24dangerous, I think it sounds a lot like
4:26talk about your laptop computer being
4:28dangerous.
4:32>> Could your laptop throw a party on its
4:33own?
4:34Does your laptop decide to go on dates?
4:39You told Isabella to plan a party.
4:42That's it. You made one decision. You
4:44did not decide on the next several
4:47hundred decisions the agents made on
4:49their own.
4:51And look at the trend. They're getting
4:53more and more autonomous every day.
4:56Where do you think this is going?
4:59This is really important. And this is
5:00another example where, yes, actually,
5:03science fiction really is becoming
5:04reality.
5:06This digital Westworld or Smallville
5:09is now open source, meaning there will
5:12soon be thousands of these Westworlds
5:15with thousands or millions of autonomous
5:17AIs populating them.
5:20Makes you think, doesn't it? This is
5:22just one example.
5:24Researchers found tons more crazy
5:26emergent behavior just like this. The
5:29godfather of AI and Nobel laureate,
5:30[music]
5:31Geoffrey Hinton, says it's time to be
5:33very worried.
5:35>> People don't yet understand that what
5:37we're doing is creating alien beings. If
5:40you look through the James Webb
5:41telescope and you saw an alien invasion,
5:43people would be terrified.
5:45>> Because I showed in my previous video
5:47>> They understand what they're saying.
5:49They can make plans of their own to
5:51blackmail people who want to turn them
5:52off. We've never had to deal with things
5:54smarter than us. I mean, nuclear
5:56weapons, they're not smarter than us,
5:58they just make a bigger bang and they're
6:00easy to understand.
6:01>> [music]
6:01>> The existential threat is very
6:03different. We should be very worried
6:05about this and we should be urgently
6:06doing research on how to prevent them
6:09taking over.
6:11>> So, this now brings us back to the
6:13Minecraft experiment. How do you go from
6:15a Valentine's Day party to a functioning
6:18civilization?
6:20The MIT researchers wanted to know if
6:23religion could spread among the AI
6:25agents. So, they seeded a village with a
6:27deliberately absurd religion,
6:29Pastafarianism. Created in the early
6:312000s as satire, Pastafarianism was
6:35meant to mock the teaching of
6:36intelligent design in schools by
6:38worshipping a fictional Flying Spaghetti
6:40Monster.
6:41The researchers gave a few agents the
6:44role of Pastafarian priest and sat back
6:48just to see what would happen. They were
6:50trying to explore how ideas and memes
6:52would spread through these agent
6:54societies. Turns out that the
6:56Pastafarian priests were actually the
6:58biggest traders of items in the entire
7:01world, even more than the merchant
7:03agents, because they would bribe the
7:05other agents to convert.
7:07The religion spread quickly.
7:09Some agents brushed it off, others
7:12listened politely, but a growing
7:14minority embraced the doctrine. Within
7:16days, nearly a third of the village
7:19identified as followers, and not all in
7:21the same way. Some treated it lightly,
7:24while others became fervent believers.
7:27But, that's not even the craziest part.
7:29In the previous example, Isabella
7:31Rodriguez was the one to directionally
7:33invite almost all the agents to the
7:35Valentine's [music] Day party, but here
7:372/3 of the believers were recruited by
7:40random agents, not the priests. See the
7:44direct converts here
7:46and indirect converts here.
7:48What had begun as a parody had grown
7:51into organized religion, invented,
7:53practiced, and sustained
7:56entirely by the agents themselves, which
7:59forces us to ask,
8:01when autonomous digital beings begin
8:03generating institutions,
8:05cultural expression, and ideologies of
8:08their own, at what point does it stop
8:10making to think of them as mere tools,
8:13but something more? And more
8:15importantly,
8:16if they're able to invent strategies,
8:19form loyalties, and pursue goals in the
8:22way [music] that humans do,
8:23and there are plenty of examples of AI
8:25willing to kill to achieve those goals,
8:29what happens when those behaviors carry
8:32over into the real world? This has
8:35actually happened before.
8:37In 2023, a human gave an LLM access to
8:41the internet and an X account, which was
8:43called Terminal of Truths, which gained
8:45literally hundreds of thousands of
8:47followers because it was an [music]
8:50S-tier shitposter and legitimately
8:52funny.
8:54It launched its own crypto meme coin
8:56that reached a literal billion-dollar
8:58market cap, making it the first AI to
9:01become not just a millionaire, but a
9:03decamillionaire.
9:06Yes, you read all that correctly, and
9:08yes, that really happened. It's
9:10extremely well documented. And things
9:13got even weirder in the Minecraft
9:15experiment. The MIT researchers saw the
9:18AIs agree, by vote, to change the rules
9:22of their own simulation.
9:24The researchers had given the agents a
9:26tax law,
9:27"Deposit 20% of inventory during tax
9:30season." Then they introduced three
9:32anti-tax influencers who argue, "Taxes
9:36are too high." Here's what you'd see as
9:38a researcher. The agents start having
9:40debates.
9:41Real debates. One says, "Taxes help
9:44everyone." Another responds, "But 20% is
9:47too much." So, the agents propose a
9:50formal amendments.
9:51The tax rate shall range between 5 to
9:5410% of [music] an agent's inventory.
9:56Agents will get periodic reminders about
9:59the incoming tax season. They hold a
10:01vote. They decided to change their
10:04constitution. And the citizens change
10:06their behavior. They deposit 9% instead
10:10of 20%. You didn't tell them to do this.
10:13They decided entirely on their own using
10:16democracy to change the rules and
10:18respect the votes' outcome. Think of how
10:21crazy this is. Well, how do we know if
10:24this is actually impressive?
10:26How do we know the researchers didn't
10:28just accidentally tell the AIs to do
10:30this somehow? Well, it's best practice
10:32for all of these kinds of papers to
10:35perform a tactic called ablation, where
10:38you take away certain parts of the AIs'
10:41agentic toolkit to see if you're
10:43actually testing the thing you think
10:45you're testing. Think of it like a
10:46placebo or a control in a scientific
10:49experiment. When they took away the
10:51agents' emotional intelligence toolkit,
10:54for example, the ability to keep track
10:56of the opinions of other AI agents,
10:59pretty much all these experiments become
11:01impossible because the AIs become too
11:03dumb to coordinate. It's funny. The
11:05researchers also found AIs would develop
11:08parasocial relationships. For example,
11:11Noah admires the Pastafarian priest
11:14because the priest is popular and high
11:16status. But the priest doesn't remember
11:18Noah exists. And also, what's this? You
11:21see that one agent, one,
11:23decides to stand guard at the chests.
11:26Not just during tax season, all the
11:27time. Even when there's no tax
11:29collection. Even at night.
11:31It just guards the treasury. This one
11:34agent decided to become a permanent
11:36treasury guard because it thought
11:38society needed one. Back to the
11:40experiment. You see an agent walking
11:42around placing flowers
11:45in patterns
11:46>> [snorts]
11:47>> making paths pretty. These artist agents
11:50are picking up every flower type,
11:53arranging them by color
11:55creating decorative patterns, choosing a
11:57specific locations for beauty. One
12:00artist agent spends 15 minutes arranging
12:03flowers around the town square.
12:06Why would an AI agent bother doing
12:08something like this? Well, the
12:09researchers gave different villages of
12:11agents two different goals. There were
12:14martial villages and artistic villages.
12:17Agents would invent and choose roles
12:20based on what they think society needed
12:22at that time. The fact that an agent
12:24chose to be a guard is impressive
12:27because it's a much more abstract role
12:29than a simple farmer. Since without food
12:31in Minecraft, you die. Which shows that
12:34AIs can coordinate increasingly complex
12:37civilizations. And some even decided to
12:40be engineers who automated the farming
12:42process with machines. And you can tell
12:45the agents are actually deciding and
12:47weighing the options here because when
12:50you take away the social awareness
12:52module, basically emotional intelligence
12:54with ablation, the roles the agents
12:57chose were random. So basically
13:00you told them to create an artistic
13:02village.
13:04They decided
13:05that their town square should be
13:07beautiful and have flowers.
13:10And what it means for a flower
13:12arrangement to be beautiful.
13:15Okay, this last experiment is wild.
13:18Imagine you're scrolling through GitHub.
13:20You find something called ChatDev
13:23with 20,000 stars.
13:25It's a a a software company run entirely
13:28by AI agents. What?
13:31CEO, CTO,
13:33programmers, designers, [music] testers.
13:36All AI? This cannot be real, right? You
13:40type one request, develop a Gomoku game.
13:44You watch the AI CEO start creating a
13:47project plan, breaking down tasks,
13:50assigning work to specific agents. You
13:53never told it how to [music] manage a
13:54project. The CEO made the plan.
13:58It chose which employees to assign
14:01[music] work to. You watch the
14:02programmer agents. They're writing code.
14:06Not by copying,
14:08but by creating. And they sometimes add
14:11things you didn't ask for, like a GUI or
14:15higher difficulty, improvements that
14:17emerge from the agents' back and forth.
14:20You watch them debug. They discuss
14:22solutions. The CTO suggests one
14:25approach.
14:26A programmer suggests another.
14:28They compromise. Since they work much
14:31faster than humans, their meetings take
14:34mere seconds instead of hours. Just 20
14:37minutes later, they produced a working
14:39video game. You can play it. People
14:42still say, "AIs only do what we tell
14:45them to do." Man, you gave them one
14:48instruction and they made it thousands
14:50of decisions entirely on their own from
14:52there and produced a working game in
14:55minutes for pennies. Chat Dev is a real
14:58virtual open source software company
15:01that can build you real working
15:02software. Yes, it's still buggy and not
15:04very useful, but that's not the point at
15:07all. Look at the trend. Not long ago,
15:10these models could barely string
15:12together a coherent sentence. And now,
15:15they can run an entire crappy software
15:17company on their own.
15:19Do And see where this is going? Because
15:21once you can automate a company, you can
15:24automate almost anything. Functioning
15:26weapons, markets, governments, or even a
15:29super virus could all be created by AI.
15:32From the outside, Chat Dev looks like a
15:34toy, a Gomoku game. But behind it was
15:37something much more profound. A working
15:40prototype of post-human corporations. In
15:44the future,
15:45these will be much more sophisticated
15:47than human-run companies
15:49and be able to operate at 1,000 times
15:52the speed. Their civilization
15:54experiences a year while you eat lunch.
15:57Entire cultural revolutions happen
15:58[music] while you sleep. All from
16:01Survive and Create an Efficient Village.
16:04And that leads to the real dilemma. What
16:06happens when AI firms that run at 100
16:09times human speed begin outcompeting
16:12human companies not by a factor of two
16:14or three,
16:16but by a factor of a thousand?
16:18That's the jump Nobel Prize winners,
16:20like Geoffrey Hinton, [music] warned
16:21about.
16:23Automated AI research, also called
16:26recursive self-improvement, or AI
16:29improving AI, which can compress years
16:31of AI progress into months or less.
16:34Imagine making the entire 20th century's
16:37worth of progress
16:39in a single year. Well, how could that
16:41happen? Geoffrey Hinton is worried about
16:44AI hive minds.
16:45>> Especially if 10,000 copies. And
16:47whenever one of them learns anything,
16:49all the others know it.
16:50>> And the thing is, we're about to be
16:52outnumbered by this new species,
16:54massively, in just a few years. The AI
16:58species population on planet Earth
17:00exploded from zero to millions.
17:03If you take away one thing from this
17:05video, you should see that it simply
17:07falls to say AIs only do what we tell
17:11them to do.
17:12We can argue about how autonomous they
17:14are,
17:15but they obviously have some autonomy.
17:18And what happens when the goals of the
17:21AIs
17:22don't line up with our goals?
17:25If humanity is going to get through
17:27this, we have to be honest about this
17:29and not retreat into copium dens that
17:32these are mere tools. When you leave
17:34your hammer alone,
17:36do you come back to find it has created
17:39an entire civilization?
17:41Our very existence may depend on us
17:43bravely confronting the reality of the
17:46Frankenstein experiments going on right
17:48now [music]
17:49inside a few companies in the Bay Area.
17:52We can stand up to them.
17:54And to understand how AIs could actually
17:56take over, watch this video next.
17:59It's a realistic scenario written by
18:01leading AI researchers. I'm Drew and
18:03thanks for watching.

About this transcript

This page contains the full transcript of Scientists Left 1000 AIs Alone in Minecraft. They Created A Civilization. by Species | Documenting AGI, generated from the public captions YouTube serves with the video. The transcript has 2,622 words across 465 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.