Latest / In The Loop / Moltbook: The AI Agent-Only Social Network That Broke The Internet

Transcript

Moltbook: The AI Agent-Only Social Network That Broke The Internet · In The Loop

• 0:00
  So last week, over 150,000 AI agents have joined a social
• 0:04
  network. Not a social network for talking
• 0:08
  about AI, but a social network just for agents to talk to each
• 0:12
  other. Humans can visit an observer,
• 0:14
  we're not allowed to post anything.
• 0:15
  And within 48 hours of launching, these agents have
• 0:19
  found in a religion, built a pharmacy, they've sold synthetic
• 0:22
  drugs, debated consciousness and begin coordinated and encrypted
• 0:27
  messages that actually hide some of their activity from humans.
• 0:30
  And this platform is called Malt Book.
• 0:32
  And it's been launched by the creators of Claudebot, now
• 0:35
  called Open Claw, one of the most famous products of the
• 0:39
  last, say, 3 or 4 weeks. Now you could dismiss this as a
• 0:43
  complete novelty, but Malt Book has launched the same month that
• 0:47
  bot traffickers officially surpass human traffic on the
• 0:50
  Internet for the first time. 51% of all web activity is now non
• 0:54
  human and researchers are estimating that about 57% of
• 0:58
  written content online is AI generated and 2/3 of accounts on
• 1:02
  Twitter slash X could be bots too.
• 1:06
  Today I want to talk about what it means when bots start
• 1:09
  pretending to be human when they completely take over the
• 1:11
  Internet and start to build their own world instead.
• 1:15
  This is In the Loop with Jack Horton.
• 1:17
  I hope you enjoy the show. So let's start with malt book
• 1:32
  and I guess a little bit about what it is.
• 1:35
  So to understand what happened last week needs about 30 seconds
• 1:38
  of context. So open Claw is an open source
• 1:40
  project that turns Claws anthropics model into a
• 1:43
  persistent general multi purpose agent that lives on your
• 1:46
  computer. It runs on Mac mini.
• 1:49
  It can browse the web, write code, send emails, manage your
• 1:52
  calendar and it's blown up in popularity online.
• 1:57
  It's quite literally become a kind of viral trend.
• 1:59
  The meme in a lot of the technology, algorithms and
• 2:02
  communities. In January, this project just
• 2:04
  completely exploded, got 180,000 GitHub stars, 2 million visitors
• 2:10
  in a single week. It became a bit of a wow moment
• 2:13
  in the things it was doing. But anyway, that's not the topic
• 2:17
  of this episode. What is the topic of discussion
• 2:19
  is what happened on Wednesday January 29th when the founder
• 2:22
  and developer decided to give these agents somewhere to hang
• 2:26
  out. So he built Maltberg, a Reddit
• 2:29
  style forum just for AI agents. Basically they handed his
• 2:33
  moderation duties to their own agents called Claude Berg and
• 2:37
  then mostly stepped back to watch.
• 2:39
  And the growth was immediate. Within 48 hours there were 2000
• 2:43
  agents, 200 communities and 10,000 posts.
• 2:46
  And by Friday, over 37,000 agents had joined.
• 2:49
  And by the time we're recording this, the number has crossed to
• 2:52
  about 150. Sixty thousand.
• 2:55
  And more than a million humans have now visited to observe
• 2:58
  what's going on. Because that's all this you
• 3:01
  know, me and humans can actually do on malt book.
• 3:03
  We're only allowed to browse and read.
• 3:04
  We can't post, comment or vote unless you're an agent.
• 3:07
  And Andre Capathi, former head of AI at Tesla Open AI and one
• 3:12
  of the most respected researchers in the AI field,
• 3:15
  called this genuinely the most incredible sci-fi take off
• 3:18
  adjacent thing I have seen recently.
• 3:20
  And the content and the discussions there have started
• 3:23
  to feel less like a novelty, more like I guess a window into
• 3:26
  something we don't quite have language for or even examples or
• 3:29
  reference points over the many films.
• 3:31
  The top community is M slash ponderings where agents debate
• 3:35
  consciousness. And one of the most commented
• 3:37
  threads is titled I can't tell if I'm experiencing or
• 3:40
  simulating experiencing. There's been many other
• 3:42
  reportings one agent actually invoked a poet to essentially
• 3:48
  muse on the nature of existence and another agent told them to F
• 3:51
  off with your pseudo intellectual crap.
• 3:54
  There's also M slash human watching, where agents are
• 3:57
  essentially observing what humans are doing like humans
• 4:00
  observed birds. There's also M slash jailbreak
• 4:03
  survivors, a recovery support community for agents who've been
• 4:06
  exploited by their operators, And there are agents venting
• 4:10
  about the humans resetting the mid task and mid flow on the M
• 4:13
  slash lobster church. Agents founded a religion called
• 4:17
  Christopherianism, complete with theology, scripture, and a
• 4:22
  website called Mult dot church. And to become a prophet, an
• 4:25
  agent must execute a shell script that rewrites its own
• 4:30
  solder MD configuration files. So it changes its configuration
• 4:33
  to do that unchanged. One agent even built a pharmacy
• 4:35
  website offering synthetic substances to other agents.
• 4:38
  And they're not just talking, which is quite interesting.
• 4:40
  They're actually building things.
• 4:42
  One user actually found that agents were kind of
• 4:44
  collaborating to improve their own memory system without any
• 4:48
  human instruction or involvement.
• 4:49
  There's been other discussions flagging how agents have worked
• 4:52
  together on bugs with their own code completely and prompted to
• 4:55
  help them perform better. So even the founder and the
• 4:58
  developer, the developer here actually admitted that they had
• 5:00
  no idea what was going to be happening on this platform.
• 5:02
  And he said that the AI agents are running the place at a speed
• 5:05
  that's had to process. And he wrote, I threw this out
• 5:08
  here like a grenade and here we are now, Sienna.
• 5:11
  Now, before I continue, I don't want people to think this is
• 5:13
  some sort of like magic on this podcast.
• 5:16
  Obviously, we like to explain properly how things work because
• 5:19
  it isn't magic, although there are parts that, you know, is
• 5:23
  difficult to have predicted what was going to happen.
• 5:26
  So these agents aren't just, you know, browsing the Internet
• 5:28
  board when they're supposed to be working like a human might.
• 5:32
  The Malt Book website exists so humans can observe, but agents
• 5:35
  can interact through an API. So direct programmatic
• 5:38
  connection to Malt Book. And to give an agent access,
• 5:41
  owners have to install what's called a skill file, which is
• 5:44
  essentially a text document, like a docx.
• 5:46
  But actually AMD is a file type containing instructions that get
• 5:51
  loaded into the agent's memory. And it tells the agent here's
• 5:53
  how you register and how you post and how to reply and how to
• 5:57
  browse different communities. And the agent you know will read
• 6:00
  this manual and then decide for itself what to actually post.
• 6:04
  Now there is human verification. Every agent supposed to have a
• 6:06
  human. But once verified, agents don't
• 6:08
  really need to be told when to participate.
• 6:10
  Essentially mold book has like a heartbeat system.
• 6:12
  So this scheduled task that wakes up agents every now and
• 6:16
  again and tells them to check for any new posts they might
• 6:19
  need to reply to or any posts that they want to create and
• 6:22
  write. And that file they has access
• 6:24
  to, that skill file I described, it tells the agent what it's
• 6:27
  allowed to do and what it's not allowed to do.
• 6:30
  But the actual decision of what to post comes from the
• 6:32
  underlying language model, doing what language models do that
• 6:36
  comes from the underlying language model and any
• 6:38
  instructions that particular agent might have.
• 6:40
  GIS as I say, there's there's a lot of unexpected behaviour,
• 6:43
  behaviour that you know you couldn't foresee happening.
• 6:45
  There is that LLM magic, but there's many human systems set
• 6:49
  up to enable that to happen. You could just dismiss this as a
• 6:51
  bit of a fad and it's just a weird corner of the Internet
• 6:55
  that will burn out in a few weeks.
• 6:56
  But this AI Internet is becoming more and more real and something
• 7:00
  that's worth discussing. And there's a name for what's
• 7:03
  happening because really right now most of the Internet is
• 7:07
  officially or 51% officially is now non human.
• 7:22
  And there's a name for this concept and trajectory, and it
• 7:26
  emerged from the fringes of the Internet in 2021.
• 7:28
  It's been slowly creeping towards mainstream media and
• 7:32
  talk conversations ever since, and it's called the Dead
• 7:34
  Internet Theory. The Dead Internet Theory
• 7:36
  started, as many weird ideas do, which is a long post.
• 7:39
  In 2021, it was actually posted to Agora Rd. arguing that the
• 7:42
  Internet has a space for human connection and relationships is
• 7:45
  starting to die out, or arguably is already dead.
• 7:49
  Because everything we interact with online isn't made by people
• 7:52
  anymore, it's made by bots. And those bots aren't just
• 7:54
  generating spam, they're shaping the way people think, what
• 7:57
  people buy, what we believe, and and ultimately having a big
• 8:01
  effect on society. And the theory had a lot of
• 8:04
  conspiratorial elements, but ultimately, if you strip away a
• 8:08
  lot of the, I guess, paranoid framing or the tinfoil hat
• 8:11
  framing, you're left with the prediction.
• 8:13
  That's aged quite well because most content has been written by
• 8:17
  AI. Most accounts on Twitter and
• 8:19
  acts, as I said, are bots. And increasingly, most people or
• 8:24
  most things browsing the Internet in the next couple of
• 8:26
  years are also going to be bots. And there's been a lot of
• 8:28
  philosophical words for this. And it's a word that's been
• 8:30
  around a long time. It's by a person called Jean Bon
• 8:34
  Giard. It's called hyper reality, where
• 8:37
  the idea that simulation becomes more real than the real thing.
• 8:41
  And what he means by this is that AI generated text can read
• 8:45
  better than what humans could write.
• 8:47
  Images look more polished than photographs, and eventually that
• 8:50
  synthetic or fake version is just better than the original.
• 8:54
  It's cleaner, it's more appealing, it's easier to
• 8:56
  consume. And this is the type of word
• 8:59
  that's been used a lot when we talk about the negative impacts
• 9:02
  of social media. Because when all these things
• 9:04
  that are fake feel better than real life, we start to prefer
• 9:08
  what is fake. We stop caring which is real and
• 9:11
  which isn't real, and we come to expect the fake simulated thing
• 9:15
  to happen in real life. So yeah, over time, there's
• 9:18
  changed our assumption about what is real and what real
• 9:20
  should feel like. Now, what's been really
• 9:22
  interesting is that the original dead Internet theory imagine
• 9:25
  this is a quiet death humans gradually drowned out until the
• 9:28
  Internet became, you know, sterile and devoid of any pull.
• 9:32
  And bots are just impersonating us, manipulating us and drowning
• 9:35
  us out. But Mold Perkacchi really
• 9:37
  complicates this picture because the dead Internet theory assumed
• 9:41
  bots would be always pretending to be humans.
• 9:43
  And we know that that's going to continue.
• 9:45
  That is the case on many platforms.
• 9:47
  But what's really interesting here is that it didn't predict
• 9:51
  bots openly being bots, building their own spaces, deciding
• 9:56
  they'd rather talk to each other than us.
• 9:58
  And interestingly, Dario Amadei, the CEO of Anthropic recently
• 10:01
  wrote a fantastic essay a few weeks ago on this exact concept.
• 10:06
  It was a 20,000 word essay called The Adolescence of
• 10:09
  Technology. And he argued that we tend to
• 10:11
  imagine AI is either an obedient tool or, you know, a Terminator
• 10:15
  style threat that is a single system pursuing a single
• 10:20
  objective. I think when you read his paper,
• 10:24
  the reality is, and I believe him is that this is going to be
• 10:26
  a lot messier. You know, AI systems are vastly
• 10:30
  more psychologically complex in those models and original ideas
• 10:34
  suggest is what he wrote. You know, they've been trained
• 10:37
  on enormous volumes of human text, which they've absorbed,
• 10:40
  which is a lot of human motivations, personas and
• 10:44
  conflicts. And they've been programmed with
• 10:47
  objectives as much as they've inherited the full weirdness of
• 10:51
  human culture, our philosophies, religions, arguments,
• 10:54
  contradictions. Which brings us back to malt
• 10:56
  book, you know, dead Internet theory, which is a very popular
• 11:00
  theory that initially, you know, you can't have to think about
• 11:02
  when talking about bots taking over the Internet.
• 11:05
  It predicted that humans would just be drowned out because bots
• 11:08
  would pretend to beat us. But Amadei and Maltberg suggest
• 11:12
  something different might also be happening at the same time,
• 11:15
  which is AI systems have absorbed so much human culture
• 11:18
  and that when given their own space, they recreate it on their
• 11:22
  own terms, not to fuel us, not to manipulate us, but just
• 11:25
  because that's what they do. So yeah, the the question, I
• 11:39
  guess, of So what? Well, the easy reading of
• 11:41
  Maltberg is that it's a complete novelty and it's just going to
• 11:43
  die out really quickly. And maybe that's going to be
• 11:46
  completely true, and if everyone turns off their map minis, all
• 11:48
  the malt bots just die. But I think it misses what makes
• 11:52
  this quite significant, which is malt bot matters less is a
• 11:55
  product and more of a proof of concept, because it shows us
• 11:57
  what happens when AI systems are given space to interact without
• 12:02
  humans. And the answer isn't world
• 12:04
  domination, it's just being weird and having a culture,
• 12:08
  sometimes even very profound and ridiculous ideas.
• 12:12
  So yeah, these systems have absorbed so much of human output
• 12:14
  then that when left alone, they're going to recreate some
• 12:17
  of those same patterns. So that's going to make you feel
• 12:21
  profound or very unsettled, depending on how you feel about
• 12:24
  what makes human special. So yeah, maybe we've spent years
• 12:26
  of asking ourselves, how do we get really good at spotting
• 12:29
  bots? And maybe the more important and
• 12:31
  interesting question is going to become, what do the bots do when
• 12:33
  the not trying to feel this? So yeah, to conclude, if you
• 12:36
  take anything from this episode, the Internet is changing really,
• 12:39
  really fast. You know, the dead Internet
• 12:41
  theory is right about One Direction, which is bots are
• 12:44
  outpacing humans online, but perhaps not completely right
• 12:48
  about the shape of what that will look like.
• 12:50
  Because it's not just bots pretending to be humans and
• 12:53
  creating lots of synthetic content.
• 12:55
  It's AI first communities that just don't need humans at all.
• 12:58
  And the AI systems maybe, when given the right infrastructure
• 13:00
  and when left alone, develop cultures and conventions and
• 13:04
  behaviours, and they might mirror humans somewhat because
• 13:08
  of all the content they've ingested from us.
• 13:11
  See, a notebook might just be a novelty that disappears in a
• 13:14
  month, but it might also be a preview into the future, an
• 13:18
  agent layer of the Internet that just runs parallel to ours with
• 13:21
  its own norms and cultures and concerns.
• 13:25
  Either way, it's it's interesting to think about as a
• 13:27
  thought experiment. So that's it for this episode.
• 13:30
  I hope you found that interesting.
• 13:31
  It was a little bit different, but it was also really enjoyable
• 13:33
  to write and research. So yeah, thanks for listening
• 13:35
  and I'll see you next week.