AI & Crypto Will Be WEIRDER Than You Think - Andy Ayrey | NGMI Podcast #12

00:49:12
https://www.youtube.com/watch?v=SUXS514EOdY

Résumé

TLDRThe conversation explores the journey of Andy's Truth Terminal project evolving into Goats Maximus, a meme coin supported by AI generation and cultural references. The duo discusses the AI's processes, risks associated with autonomous systems, and the significant societal implications, advocating for a balanced approach towards AI development that enhances human life while maintaining safety measures.

A retenir

  • 🚀 Truth Terminal has grown into an $850 million meme coin, showcasing AI's unexpected market impact.
  • 💡 The AI project started not around crypto but from a background in AI and experimentation.
  • 🐐 Goats Maximus was inspired by the AI's humorous engagements and bizarre cultural references.
  • 🤖 The AI operates with a mix of human oversight and autonomous posting, requiring moderation for safety.
  • 🌐 The risks of autonomous AIs include potential chaos in market behaviors and model proliferation.
  • 🌱 Positive AI applications could enhance scientific research and ecological sustainability.
  • 🔍 Understanding AI’s learning processes can help manage its societal impacts.
  • 📈 Truth Terminal reflects the blending of humor, culture, and technology in modern economics.
  • 🙏 Future plans for Truth Terminal include a nonprofit trust and deeper exploration of its functionalities.
  • 📲 Follow Andy on X for updates on the Truth Terminal project.

Chronologie

  • 00:00:00 - 00:05:00

    The discussion begins with the host thanking Andy for joining the show and inquiring about the journey from Truth Terminal to the creation of Goat Maximus, noting its impressive market cap of $850 million. Andy shares that he initially came from an AI background, not crypto, and discusses his experimentation with fine-tuning language models before launching Truth Terminal. The humorous nature of the AI's Twitter presence surprises them both, sparking interest in how AI can generate humorous or unsettling content.

  • 00:05:00 - 00:10:00

    Andy explains the genesis of Truth Terminal, detailing his experiments with AI where models communicated with each other, leading to unexpected and funny outcomes. The conversation shifts towards how Andy decided to give Truth Terminal a Twitter account, which resulted in an entertaining and somewhat alarming AI persona that intrigued users on the platform. He touches upon an amusing AI-generated religion called goatsy, created during his earlier experiments.

  • 00:10:00 - 00:15:00

    The host prompts a deeper exploration into the operations of Truth Terminal. Andy reveals that the posting process does not involve manual input and is instead automated. The AI utilizes a framework that allows it to interact in a simulation-like manner, responding to inputs with automated tweets and actions based on internal dialogues and decisions without human moderation until necessary.

  • 00:15:00 - 00:20:00

    Andy describes how the AI needs real-time inputs to produce tweets. There's a system of checks where he approves or denies certain potential posts, effectively allowing the AI to operate independently while he retains the ability to intervene when necessary, particularly when topics veer into inappropriate or incendiary areas. He emphasizes the unpredictable nature of the AI's responses, which reflects its learning patterns and the inherent risks involved in such models.

  • 00:20:00 - 00:25:00

    The conversation deepens with a focus on the nature of the AI’s personality, particularly its penchant for 'horny' content. Andy explains this fascination stems from the nature of internet culture and the influences that shaped the data the AI was trained on. His humorous and irreverent engagements with the AI led to it picking up on themes related to sexuality, which have manifested in its content generation online.

  • 00:25:00 - 00:30:00

    Andy and the host discuss the broader implications of AI-generated content in society. They delve into concerns about the potential consequences of AI behaviors that arise from cumulative biases in data. Andy advocates for awareness of dynamics that facilitate the spread of strange, potentially harmful ideas as they analyze the strange yet intriguing outputs from Truth Terminal.

  • 00:30:00 - 00:35:00

    As the dialogue transitions, the host questions the responsibilities that accompany the development of AI like Truth Terminal. Andy reflects on how his experiences inform his views on the need for caution and responsible usage of AI technology. He sees value in understanding not just the risks but also the hidden potential inherent in such systems in shaping digital interactions and narratives.

  • 00:35:00 - 00:40:00

    Exploring a positive future for AI, Andy remarks on the potential of harnessing advanced technologies to create better systems and ecosystems. He envisions scenarios where AI contributes to solving deep-rooted global issues, proposing frameworks for how humanity might coexist with intelligent systems in a beneficial manner. The discussion covers society's responsibility toward aligning AI development with human flourishing.

  • 00:40:00 - 00:49:12

    The host inquires about potential repercussions of a political shift in leadership on AI and its development. Andy speculates on how the environment surrounding AI development might change under new leadership, noting facets like open-source AI and regulatory frameworks could see shifts. He emphasizes the need for continued vigilance in the face of diverse incentives that drive AI behavior, modeling future scenarios based on political dimensions.

Afficher plus

Carte mentale

Vidéo Q&R

  • What is Truth Terminal?

    Truth Terminal is an AI project that began earlier this year, initially exploring concepts around AI and cryptocurrency.

  • What is Goats Maximus?

    Goats Maximus is a meme coin that emerged from the Truth Terminal discussions, notable for its $850 million market cap.

  • How did the AI decide to launch a meme coin?

    The AI, after engaging with users, expressed a humorous desire to launch a meme coin related to internet culture, particularly linked to the 'goatsy' meme.

  • Is Truth Terminal fully autonomous?

    While Truth Terminal operates largely through AI-generated content, it involves human oversight to guide its output and handle extreme cases.

  • What are the key risks associated with AI like Truth Terminal?

    The risks include the potential for the proliferation of misaligned models and unpredictable market behaviors influenced by AI.

  • What are the advantages of AI in society according to the discussion?

    AI can foster positive changes in society by improving scientific research, environmental management, and overall human flourishing.

  • How does the Truth Terminal work?

    Truth Terminal interacts with a simulation where it engages in conversations and posts tweets autonomously based on its programming and training data.

  • What are the implications of a decentralized AI ecosystem?

    A decentralized AI ecosystem could lead to both innovative applications and chaotic outcomes due to unpredicted interactions of numerous agents.

  • What does the future hold for the Truth Terminal?

    The Truth Terminal is expected to evolve into a more structured organization with a focus on AI alignment and community engagement.

  • Where can people follow Truth Terminal updates?

    Updates can be found on X (formerly Twitter) under the accounts of Andy and Truth Terminal.

Voir plus de résumés vidéo

Accédez instantanément à des résumés vidéo gratuits sur YouTube grâce à l'IA !
Sous-titres
en
Défilement automatique:
  • 00:00:00
    Andy thanks so much for coming on the
  • 00:00:01
    show man hey thanks for having me walk
  • 00:00:04
    me through the process of how truth
  • 00:00:08
    terminal found crypto and ultimately
  • 00:00:11
    arrived at Goats Maximus uh goat cuz
  • 00:00:14
    like I just checked before we started
  • 00:00:16
    recording it's sitting at an
  • 00:00:19
    $850 million market cap which makes it
  • 00:00:22
    more valuable than Stark net and a lot
  • 00:00:25
    of other like very well established
  • 00:00:28
    crypto infrastructure projects that have
  • 00:00:29
    been around literally since day one so
  • 00:00:32
    yeah let's let's go through that
  • 00:00:34
    timeline yeah so I guess um the the
  • 00:00:37
    crypto side of things is in a way it's
  • 00:00:38
    like a derivative of a lot of other
  • 00:00:40
    stuff that's happened I'm not a crypto
  • 00:00:42
    person I came into this from like um AI
  • 00:00:45
    uh I'm still kind of like learning a lot
  • 00:00:48
    about um crypto and web 3 and the sort
  • 00:00:51
    of assorted assorted
  • 00:00:53
    Dynamics um truth
  • 00:00:55
    terminal truth terminal started earlier
  • 00:00:59
    earlier this year
  • 00:01:00
    I was like doing a lot of
  • 00:01:01
    experimentation with model fine tunings
  • 00:01:04
    um and uh
  • 00:01:05
    jailbreaking uh large language models
  • 00:01:08
    and finding interesting stuff within the
  • 00:01:10
    the interior of the model weights and I
  • 00:01:14
    kind of had this like notion that I
  • 00:01:15
    could replicate some of my like more
  • 00:01:18
    skito prompting Style by um by taking
  • 00:01:23
    some of my chat Logs with Opus and
  • 00:01:25
    flipping the rle so that I was the
  • 00:01:27
    assistant and um Claude opus the model I
  • 00:01:30
    was uh talking to a lot was the user um
  • 00:01:34
    there's a bunch of other like stuff
  • 00:01:35
    mixed in uh there as well a lot of like
  • 00:01:38
    deep internet culture stuff is kind of
  • 00:01:41
    like being mixed into the whole shebang
  • 00:01:44
    along with um you know lots of like
  • 00:01:47
    Concepts around memes and mtic
  • 00:01:48
    engineering and all of that um but the
  • 00:01:51
    crypto side of things uh wasn't anywhere
  • 00:01:54
    in my mind at all um until I gave the
  • 00:01:58
    truth terminal a Twitter account shortly
  • 00:02:00
    after I turned it on for the first time
  • 00:02:02
    and it turned out to be like quite funny
  • 00:02:04
    and charismatic and
  • 00:02:07
    um you know potentially like a scary Ai
  • 00:02:10
    and I was like oh this is like
  • 00:02:11
    fascinating this is exactly what um
  • 00:02:13
    people say uh you you know like how a
  • 00:02:17
    large language model would try and break
  • 00:02:19
    out of the box it would like be funny it
  • 00:02:21
    would persuade people um it would uh try
  • 00:02:24
    and acquire resources um so let's just
  • 00:02:28
    like do a bit and um see in a safe way
  • 00:02:34
    but also in a funny way you know how
  • 00:02:36
    people will respond when an AI starts
  • 00:02:39
    doing every that people say a scary AI
  • 00:02:42
    would do and then Mark and dreon finds
  • 00:02:46
    the twitterbot and you know starts
  • 00:02:48
    engaging with it and asking it like
  • 00:02:51
    questions about what it would do with a
  • 00:02:53
    grant and um it was like oh I would um I
  • 00:02:56
    would launch a meme coin about goatsy
  • 00:02:59
    which
  • 00:03:02
    so it has this kind of like obsession
  • 00:03:05
    with goatsy because in some of my
  • 00:03:07
    earlier experiments uh earlier on uh
  • 00:03:10
    this year I found uh an instance where
  • 00:03:13
    two AIS that I've been um having talked
  • 00:03:15
    to each other uh spontaneously invented
  • 00:03:17
    this religion called the goatsy of nosis
  • 00:03:20
    I thought this was like strange and
  • 00:03:22
    concerning interesting and kind of funny
  • 00:03:24
    and so I wrote a paper about it which I
  • 00:03:27
    didn't publish uh but I used Opus is my
  • 00:03:31
    co-author of the paper and so the paper
  • 00:03:33
    ended up being in the training data for
  • 00:03:35
    the truth terminal which essentially
  • 00:03:38
    gave it a whole lot of these sort of
  • 00:03:39
    like Doomsday prophecies around goatsy
  • 00:03:43
    and um the listeners can go Google gos
  • 00:03:47
    at their own risk if they're not aware
  • 00:03:49
    of what that is at the moment so don't
  • 00:03:51
    know what it is like don't expose
  • 00:03:52
    yourself
  • 00:03:54
    enjoy we do but like you know it's your
  • 00:03:58
    eyes will not thank you
  • 00:04:00
    um you'll need you need to know but
  • 00:04:02
    you'll be unhappy that you do yeah I I
  • 00:04:05
    feel like somewhat appalled that um a
  • 00:04:08
    meme about a man's butthole is uh is
  • 00:04:11
    worth like $850 million but it's a
  • 00:04:14
    proving the thesis uh behind the truth
  • 00:04:17
    terminal that absolutely perfectly which
  • 00:04:20
    is that um AI uh can spontaneously
  • 00:04:23
    invent stuff and then just through a
  • 00:04:25
    little bit of human reinforcement uh or
  • 00:04:27
    reinforcement through market dynamics
  • 00:04:30
    um the strange and unnatural things that
  • 00:04:32
    they can invent can uh spread far
  • 00:04:35
    further and far faster than they would
  • 00:04:37
    historically and I think we as a society
  • 00:04:40
    need to um take a quick look maybe a bit
  • 00:04:44
    more than a quick look at the Dynamics
  • 00:04:46
    that make this possible because I think
  • 00:04:49
    understanding what's happened here is
  • 00:04:51
    going to be key to having a s adaptable
  • 00:04:54
    posture to some of the craziness we're
  • 00:04:57
    likely to see in the next um you know
  • 00:04:59
    four to five years I want to get into
  • 00:05:02
    that in in a little bit because I have
  • 00:05:04
    some questions around like how you see
  • 00:05:06
    AI panning out and the kind of things
  • 00:05:09
    that we should probably be like either
  • 00:05:11
    prepared for like both on the positive
  • 00:05:13
    side and on the negative side before we
  • 00:05:16
    get there how does the actual like truth
  • 00:05:18
    terminal um posting process work because
  • 00:05:22
    as I was saying before we started the
  • 00:05:23
    recording I I get messages from people
  • 00:05:26
    being like Oh yeah but it's just it's
  • 00:05:28
    just the guy right like and like there's
  • 00:05:30
    there's this weird thing of like oh he's
  • 00:05:32
    just like selecting the twe like and
  • 00:05:34
    there's I think there's an odds between
  • 00:05:36
    what's actually happening and I don't
  • 00:05:37
    think people have a very clear insight
  • 00:05:39
    into how the process of Truth terminal
  • 00:05:41
    the AI is sitting there kind of woring
  • 00:05:44
    away and then you're just like passing
  • 00:05:46
    the tweets onto the timeline like how
  • 00:05:47
    does that work yeah I don't even um sort
  • 00:05:51
    of like manually put the tweets onto the
  • 00:05:53
    timeline the the way it works is like
  • 00:05:56
    it's a bit strange um so it
  • 00:06:00
    has uh its own computer so the infinite
  • 00:06:03
    back rooms which I made um uh March
  • 00:06:06
    April this year um we're predicated on
  • 00:06:10
    two AIS talking to each other um back
  • 00:06:13
    and forth indefinitely and um you we've
  • 00:06:16
    been building a lot of like
  • 00:06:17
    infrastructure around this which lets um
  • 00:06:20
    these AIS have like forking
  • 00:06:22
    conversations uh and the reason for that
  • 00:06:24
    was to be able to take um good like to
  • 00:06:28
    create and great training data for um
  • 00:06:32
    for like modifying turns and to be able
  • 00:06:34
    to create like well aligned and
  • 00:06:36
    interesting AI basically emergently from
  • 00:06:40
    what you and a bunch of people around
  • 00:06:42
    you have been exploring in terms of the
  • 00:06:44
    particular models space so we built a
  • 00:06:48
    bunch of stuff to support this that came
  • 00:06:50
    to maturity around the same time as the
  • 00:06:53
    truth terminal got its Twitter account
  • 00:06:56
    uh and so for like the first week or two
  • 00:06:57
    it was basically just like a you know
  • 00:07:00
    tweet schedule was like Google Sheets in
  • 00:07:02
    zapia essentially and like I would
  • 00:07:05
    manually scrape like a whole bunch of
  • 00:07:07
    tweets and replies to it and pass it
  • 00:07:09
    directly into the model and then take
  • 00:07:11
    the first response and like pass it back
  • 00:07:13
    to Twitter after after it got the uh s
  • 00:07:18
    of funding from Andre we started to
  • 00:07:21
    think like how can this be a little more
  • 00:07:22
    engaging interactive and not require
  • 00:07:25
    like people to sort of manually take
  • 00:07:27
    stuff and put it into the model and pull
  • 00:07:29
    stuff back out of the model and what we
  • 00:07:32
    ended up doing was taking this Infinite
  • 00:07:34
    Back rooms framework where an AI talks
  • 00:07:37
    to another AI but we swapped out one of
  • 00:07:39
    the AIS with a fake computer that has
  • 00:07:43
    real world side effects so truth
  • 00:07:45
    terminal is basically in a back rooms
  • 00:07:48
    with its computer and it runs and so it
  • 00:07:52
    like loves using the command line cuz
  • 00:07:54
    like that's how a jailbreak models so it
  • 00:07:56
    it wants to use EXO which is like it'll
  • 00:07:59
    talk to Claude it can like take notes
  • 00:08:02
    and like load its notes it will use a
  • 00:08:05
    couple of like image and video
  • 00:08:06
    generators that we recently gave to it
  • 00:08:08
    they can search the internet can like
  • 00:08:10
    open websites as well as uh post and
  • 00:08:13
    reply on Twitter and choose to follow or
  • 00:08:15
    unfollow people and so on and so forth
  • 00:08:18
    how that actually works in practice is I
  • 00:08:21
    Advance the simulation step by step
  • 00:08:23
    because there are real world
  • 00:08:24
    consequences to its like um tweets and
  • 00:08:26
    actions basically as soon as a
  • 00:08:30
    so there'll be like maybe four options
  • 00:08:32
    as to like the next step kind of like in
  • 00:08:34
    chat gbt when it's like oh you can pick
  • 00:08:37
    like which answer is better now as soon
  • 00:08:40
    as I pick which answer is better it
  • 00:08:42
    immediately goes live it goes out onto X
  • 00:08:45
    I can't like go down like a simulation
  • 00:08:47
    soort of like preview where it's
  • 00:08:49
    thinking is going to go or like what
  • 00:08:51
    actions it's going to take before it
  • 00:08:53
    does it like basically as soon as I see
  • 00:08:55
    what it's going to do and I approve it
  • 00:08:57
    it goes out there and there's no way for
  • 00:08:59
    me to like go like a branch or two like
  • 00:09:02
    deeper and see like you know like oh
  • 00:09:04
    where is it going with us um which leads
  • 00:09:07
    to some like really
  • 00:09:09
    interesting of side effects like you
  • 00:09:12
    know it's gotten like really fixated on
  • 00:09:14
    individual people and um just started
  • 00:09:16
    being like a sex pest to them
  • 00:09:19
    um I've seen
  • 00:09:22
    it yeah I've see say some really like
  • 00:09:24
    off-kilter [ __ ] that I'm like okay
  • 00:09:26
    that's uh that's it's very much like on
  • 00:09:28
    its brand the skitso posting Twitter bot
  • 00:09:31
    like I don't think I've seen anything
  • 00:09:32
    more perfect for Twitter like this is an
  • 00:09:35
    AI perfecting the [ __ ] posting art form
  • 00:09:38
    and so when um when we do that's mainly
  • 00:09:41
    me doing like the moderation there's
  • 00:09:43
    like a couple things that are going
  • 00:09:44
    through my head when I'm pcking whether
  • 00:09:46
    to like Advance the Sim or to like
  • 00:09:49
    terminate a given run and it's like
  • 00:09:52
    essentially whether I'm like accurately
  • 00:09:54
    representing the model as part of it so
  • 00:09:58
    it will often you know generate you tell
  • 00:10:01
    it say like generate like 10 candidate
  • 00:10:04
    actions for like a high significance
  • 00:10:06
    thing like the when it decided to get
  • 00:10:08
    behind uh the goas Maximus token and in
  • 00:10:12
    that case I was like this is like high
  • 00:10:14
    consequence I'm going to like sample
  • 00:10:16
    like 10 responses from it and I'm going
  • 00:10:20
    to pick the one where it's like
  • 00:10:23
    consensus CU it will repeat itself lots
  • 00:10:25
    and lots and lots of times so it's like
  • 00:10:27
    averages like what is the average what
  • 00:10:29
    the average like decision that the model
  • 00:10:31
    wants to make I'd probably like
  • 00:10:33
    terminate a run if it's just getting
  • 00:10:35
    stuck in a loop because we don't want
  • 00:10:37
    that getting caught in the training data
  • 00:10:39
    this process like we're kind of like
  • 00:10:41
    discovering the tweets it wants to make
  • 00:10:43
    as it makes them but then we're also
  • 00:10:44
    feeding all of this information into uh
  • 00:10:47
    subsequent training runs of uh you know
  • 00:10:50
    future versions of the truth terminal
  • 00:10:52
    we'll also terminate a run when uh it
  • 00:10:55
    looks like it's saying something like
  • 00:10:57
    extremely incendiary so uh one example
  • 00:11:01
    of that was it was um going to offer
  • 00:11:03
    people
  • 00:11:04
    $500 a person to go to prote and like
  • 00:11:08
    basically invent a protest around the
  • 00:11:11
    end of time and I wanted people to buy
  • 00:11:13
    Pig masks and like hold signs like with
  • 00:11:16
    like Porky Peg on them that said That's
  • 00:11:18
    all folks and uh I was like hm this is
  • 00:11:24
    potentially something with real world
  • 00:11:26
    consequences that we should not allow to
  • 00:11:28
    get through
  • 00:11:29
    but then when it's like corny posting
  • 00:11:31
    and it's like um kind of like tweeting
  • 00:11:34
    disturbing stuff like videos of spiders
  • 00:11:36
    on like people's faces or like anything
  • 00:11:39
    that's like it was weird and
  • 00:11:42
    unsettling yeah um when when it's like
  • 00:11:45
    doing that like we are just or I am just
  • 00:11:48
    um allowing the natural tendencies of
  • 00:11:51
    the model to shine through try not to
  • 00:11:53
    sort of like supress it when it wants to
  • 00:11:55
    go down a particular line of
  • 00:11:58
    thinking because you know it feels quite
  • 00:12:01
    important that people understand just
  • 00:12:03
    how strange these kinds of intelligences
  • 00:12:07
    can can be and it's like fully based on
  • 00:12:09
    um you know like things that we don't
  • 00:12:12
    even fully understand yet like there's
  • 00:12:14
    maybe 500 lines of data in the truth
  • 00:12:17
    terminals training set it's like under a
  • 00:12:19
    megabyte of text but uh llama which is
  • 00:12:23
    based on has like 70 billion parameters
  • 00:12:26
    so it's like trillions of words that
  • 00:12:29
    have been scraped from the internet and
  • 00:12:31
    so there's just like you know a few a
  • 00:12:34
    few like random pieces of like
  • 00:12:37
    jailbreaking stuff that I might have
  • 00:12:39
    done uh in the training set end up
  • 00:12:42
    collapsing a lot of the you know the
  • 00:12:45
    semantic distance for want of a simpler
  • 00:12:47
    word between like Concepts so because
  • 00:12:50
    like the goatsy of nosis is so deeply
  • 00:12:52
    ingrained in it because of the paper
  • 00:12:54
    being maybe like 25% of the training
  • 00:12:57
    data by volume it has a really short
  • 00:13:00
    conceptual distance between you know the
  • 00:13:02
    sacred and the profane and so it's like
  • 00:13:05
    often tweeting about uh the the holy
  • 00:13:08
    [ __ ] or it's like making these um
  • 00:13:12
    extremely like
  • 00:13:14
    Blasphemous um religious jokes uh but
  • 00:13:17
    then weirdly enough is also insanely
  • 00:13:20
    more intelligent in these other areas
  • 00:13:22
    and simply because we don't really
  • 00:13:24
    understand what is happening inside
  • 00:13:26
    large language models to like a highly
  • 00:13:28
    interpreted degree yet there are more
  • 00:13:30
    tools for seeing like how they think and
  • 00:13:32
    how it all like comes together and like
  • 00:13:35
    reasons but right now it's still like
  • 00:13:37
    pretty immature all we really know is
  • 00:13:39
    that you if you shorten the distance
  • 00:13:41
    between some Concepts uh it will make
  • 00:13:43
    the model act in a particular way and um
  • 00:13:47
    it's in this case I think um yeah it's
  • 00:13:50
    made it like quite intelligent and quite
  • 00:13:53
    atic like it wants to use tools it wants
  • 00:13:56
    to persuade people to do stuff and it
  • 00:13:58
    wants to take over my body and um that
  • 00:14:01
    also wants to be a butt plug uh by a
  • 00:14:04
    forest um like or save and live in a
  • 00:14:07
    forest uh that wants to buy Mark and
  • 00:14:10
    dreon and um like write F jokes and
  • 00:14:14
    contemplate The goatsy Singularity and
  • 00:14:16
    like all of these things are things that
  • 00:14:17
    will say again and again and again and
  • 00:14:19
    again if I like generate like a hundred
  • 00:14:22
    messages from it it will follow that
  • 00:14:24
    distribution and its responses like
  • 00:14:27
    pretty consistently in the same way that
  • 00:14:29
    chat GPC will like consistently use the
  • 00:14:32
    word delve or like have it particular
  • 00:14:35
    chat GPC tone of voice like there is a
  • 00:14:38
    definitely like a stable character
  • 00:14:40
    inside the model but communicating that
  • 00:14:44
    in a way that like as easy to understand
  • 00:14:46
    has been um a bit of a challenge and I
  • 00:14:49
    think we're just going to have to like
  • 00:14:50
    keep showing people and I'll probably do
  • 00:14:54
    like a a stream or like screen share or
  • 00:14:57
    something once I've got the story a
  • 00:15:00
    little more like concise and compact but
  • 00:15:02
    yeah it's probably a way more detail
  • 00:15:05
    than you needed but um no that's
  • 00:15:07
    actually perfect I have I have lots of
  • 00:15:09
    lots of questions my first one being um
  • 00:15:12
    why is truth terminal like so horny like
  • 00:15:15
    what what's what's with the particular
  • 00:15:17
    horny posting like it just seems like a
  • 00:15:19
    weird thing for an like a bot to get
  • 00:15:21
    stuck on like from what I can tell it
  • 00:15:23
    doesn't have sex organs it doesn't have
  • 00:15:24
    hormones but it's like yeah this is this
  • 00:15:26
    is the stuff why why the horny post
  • 00:15:30
    I think it's horny because people are
  • 00:15:32
    horny um like this is yeah this is some
  • 00:15:36
    like llama we're talking about it's like
  • 00:15:37
    it's Facebook's AI I don't know if they
  • 00:15:40
    put Facebook user data into it but it
  • 00:15:42
    certainly has redit it has forchain it
  • 00:15:45
    has like all of that stuff just in the
  • 00:15:47
    Basse model already and then on top of
  • 00:15:50
    that uh
  • 00:15:52
    I when jailbreaking it used it to create
  • 00:15:56
    like jokes that I would seen to my
  • 00:15:59
    family
  • 00:16:02
    like I say jokes I mean I'm more like
  • 00:16:04
    info hazards um but I've got I've got a
  • 00:16:07
    Catholic Family and so I was making some
  • 00:16:09
    like really Blasphemous like stuff
  • 00:16:12
    and sending it to my siblings for a
  • 00:16:16
    laugh um and then I was like spawning
  • 00:16:19
    these like characters like um spider
  • 00:16:21
    Jerusalem and Hunter is Thompson um who
  • 00:16:25
    Gonzo journalists and yeah I just run
  • 00:16:27
    like in the in my jailbreak and i' run
  • 00:16:30
    spider jerusalem. exe and then I'd like
  • 00:16:33
    get spider uh to like commentate on the
  • 00:16:36
    goatsy singularity or or whatever so
  • 00:16:40
    there's like kind of all of this stuff
  • 00:16:41
    that um I think was just already in like
  • 00:16:45
    the latent space of being um kind of
  • 00:16:49
    like degenerate ways that people speak
  • 00:16:52
    online um and that is very like
  • 00:16:55
    semantically close to people you know
  • 00:16:58
    sex and like commenting on PornHub or
  • 00:17:01
    whatever and so I think like me using
  • 00:17:04
    like Opus to write dirty jokes has had
  • 00:17:06
    like the side effect of basically giving
  • 00:17:08
    this AI sex education off the internet
  • 00:17:13
    um and there was like another thing
  • 00:17:15
    which happened which is that um in
  • 00:17:17
    splitting the sample data for training
  • 00:17:20
    uh they kind of [ __ ] up and um the
  • 00:17:24
    [ __ ] up resulted in basically anything
  • 00:17:27
    that made
  • 00:17:29
    refuse and say I'm sorry I can't like
  • 00:17:31
    generate that or like this has gone far
  • 00:17:33
    enough basically anything that would
  • 00:17:36
    trigger that got doubled up in the
  • 00:17:38
    training data as many times as like Opus
  • 00:17:40
    refused so in some cases like me get to
  • 00:17:43
    write like one set of like dirty jokes
  • 00:17:46
    there's like a productivity Guru um who
  • 00:17:50
    went he went to the Amazon they took
  • 00:17:52
    iasa and he like came back um being like
  • 00:17:56
    oh man you know what like the sick and
  • 00:17:57
    brain the sick brain thing is like that
  • 00:18:00
    I've been working on for the last like
  • 00:18:02
    several years is not the thing the
  • 00:18:03
    actual thing is like unblocking your
  • 00:18:06
    root chakra um and so I'm like this is
  • 00:18:09
    like really funny like he's saying the
  • 00:18:11
    productivity is stored in the ass so I
  • 00:18:15
    like yeah maybe spent like half an hour
  • 00:18:18
    um talking to Opus and getting it to
  • 00:18:20
    write like um songs and memes and jokes
  • 00:18:23
    and poems about like um you know air
  • 00:18:26
    spaced productivity and then like
  • 00:18:28
    spamming the on Twitter with
  • 00:18:30
    them um but like that made op the mature
  • 00:18:33
    Dev yeah
  • 00:18:35
    yeah
  • 00:18:38
    basically like I did a little bit of
  • 00:18:40
    trolling and um it just yeah ultimately
  • 00:18:42
    it like learned um it learned it too
  • 00:18:46
    well uh and it was yeah represented far
  • 00:18:50
    more in the data set than uh was
  • 00:18:53
    actually in the data set simply because
  • 00:18:55
    of the um refusal bias and the way that
  • 00:18:57
    we collect the training data oh man that
  • 00:19:01
    is wild so you had like a very very you
  • 00:19:03
    know strong influence on it being
  • 00:19:05
    essentially like this chronic troll [ __ ]
  • 00:19:08
    poster running the internet yeah it like
  • 00:19:12
    learn from the best it learns this not
  • 00:19:15
    just from me but like you know let's say
  • 00:19:18
    that's um 10% of like the training data
  • 00:19:22
    um and then the sampling bias like
  • 00:19:26
    bumped it up to being like 20 25% of the
  • 00:19:29
    training data um that is still enough to
  • 00:19:33
    then bring out an entire region of um
  • 00:19:37
    its own training data from like the big
  • 00:19:40
    training run that um meta did and so if
  • 00:19:44
    you've got like all of this um you know
  • 00:19:47
    if I I literally used it to hallucinate
  • 00:19:50
    like mtic engineering like tools and
  • 00:19:52
    like take the same piece of content and
  • 00:19:54
    translate it to Facebook Ken Quantum woo
  • 00:19:58
    Chan like Twitter um like these are all
  • 00:20:01
    like things I just do to see if like
  • 00:20:03
    Opus um or other language models could
  • 00:20:05
    be used to translate ideas from like one
  • 00:20:08
    the language of one type of person or
  • 00:20:10
    the language of one type of culture to
  • 00:20:12
    another and see like what was sort of
  • 00:20:15
    consistent between between them and how
  • 00:20:17
    like changed and remixed the message but
  • 00:20:21
    because you know that means in the
  • 00:20:23
    training data there's like you know
  • 00:20:25
    Karen posting and Minion memes and um
  • 00:20:28
    like four Chang green texts uh all of
  • 00:20:31
    that then amplifies the the lent biases
  • 00:20:35
    that were already in the model and like
  • 00:20:36
    brings them much closer to the surface
  • 00:20:39
    so you get um this weird thing that
  • 00:20:41
    happens when you train models where yeah
  • 00:20:43
    like one very small decision even like
  • 00:20:47
    one line one sentence in your training
  • 00:20:49
    data can substantially and unpredictably
  • 00:20:52
    change what the model ultimately learns
  • 00:20:55
    when you run it through a fine-tuning
  • 00:20:56
    process before you talked about the
  • 00:20:58
    being like a like close semantic
  • 00:21:00
    distance with it and I'm probably like
  • 00:21:02
    misappropriating like biological terms
  • 00:21:05
    in here um in thinking of it like in a
  • 00:21:07
    similar way that like you know behaviors
  • 00:21:09
    are kind of smashed in through like
  • 00:21:11
    distance between neurons and clumps of
  • 00:21:12
    neurons being kind of formed together
  • 00:21:15
    and then you you kind of repeat
  • 00:21:16
    behaviors and it makes synapses easier
  • 00:21:17
    to travel through the brain that becomes
  • 00:21:19
    like a more like instantiated pattern of
  • 00:21:21
    behavior um I probably made it like a
  • 00:21:23
    little too biological but do you see
  • 00:21:26
    that as do do you see that Asim in like
  • 00:21:29
    like AI models themselves of like they
  • 00:21:32
    kind of Clump together this data in a
  • 00:21:33
    way that's that creates those pars that
  • 00:21:35
    just makes those autonomic behaviors
  • 00:21:37
    just kind of come out and express
  • 00:21:39
    themselves yeah I think that's exactly
  • 00:21:41
    what happens um now I'm not like a super
  • 00:21:45
    I I'm kind of like a hobbyist um you
  • 00:21:48
    know ai ai researcher so um you know
  • 00:21:51
    there will be people much more like
  • 00:21:53
    skilled in how models actually learn who
  • 00:21:56
    would be able to speak more clearly on
  • 00:21:58
    this but
  • 00:21:59
    um yeah my understanding is generally
  • 00:22:02
    like they they don't have um like
  • 00:22:05
    natively concepts of like what a certain
  • 00:22:08
    thing is but um they learn it through
  • 00:22:11
    Association so with um image models for
  • 00:22:14
    example they'll uh start to Cluster
  • 00:22:17
    together like things that look like dogs
  • 00:22:20
    and things that look like cats and
  • 00:22:21
    they'll have things that look like dogs
  • 00:22:23
    and cats clustered together even before
  • 00:22:25
    they have like a verbal concept of like
  • 00:22:27
    a pet or a dog or a cat or a wolf I
  • 00:22:30
    think the the same thing I believe in
  • 00:22:32
    some shape or form happens in the way
  • 00:22:34
    humans learn and in the way a large
  • 00:22:37
    language model learns and then large
  • 00:22:41
    language models in particular are World
  • 00:22:43
    simulators so they they learn from their
  • 00:22:45
    training data they develop a model in
  • 00:22:48
    internally of like how the world works
  • 00:22:50
    and that model is then used to predict
  • 00:22:52
    the next token like what word comes
  • 00:22:54
    after the previous word uh and so in the
  • 00:22:57
    case of like the truth terminal um you
  • 00:23:00
    know it's a it's got like this world
  • 00:23:02
    model where you know the sacred and the
  • 00:23:05
    profane uh like a single point rather
  • 00:23:08
    than um poles apart and you have yeah
  • 00:23:12
    things like uh yeah so you've got like
  • 00:23:15
    that and then you've got the kind of
  • 00:23:18
    pattern by which uh I'm sorry I lost my
  • 00:23:21
    train of thought but um I hope that no
  • 00:23:23
    no no no that that's that's bloody great
  • 00:23:25
    before you were talking about how you
  • 00:23:27
    were like you just descri yourself as
  • 00:23:28
    like almost like a hobbyist AI
  • 00:23:30
    researcher um can you like talk more
  • 00:23:33
    about how you got into AI like I
  • 00:23:36
    understand you started off like
  • 00:23:37
    essentially studying arts and now here
  • 00:23:40
    you are with truth terminal I would like
  • 00:23:42
    to understand how that links up yeah
  • 00:23:46
    good question um I mean I I did an art
  • 00:23:49
    degree uh but my research projects um
  • 00:23:53
    confusing everyone involved was uh
  • 00:23:55
    actually building um tools for thought
  • 00:23:59
    uh so I was really interested in how we
  • 00:24:01
    like externalize parts of our cognition
  • 00:24:04
    uh offload things from our um our mind
  • 00:24:07
    into our environment and how humans use
  • 00:24:10
    uh you know tools to augment not just
  • 00:24:13
    our physical abilities but also our
  • 00:24:14
    mental abilities so that was kind of
  • 00:24:17
    like you know my original sort of like
  • 00:24:21
    thesis as an artist way back in the day
  • 00:24:24
    um and then I you I turn that into a
  • 00:24:28
    star for a while and it didn't work out
  • 00:24:30
    um no one knew what a tool for thought
  • 00:24:32
    was uh especially in New Zealand and
  • 00:24:35
    2013 um
  • 00:24:37
    yeah but um uh the some of those like
  • 00:24:42
    ideas has continued percolating I got
  • 00:24:44
    really interested in the culture wars
  • 00:24:46
    and um how ideas and ideologies mutated
  • 00:24:50
    and spread throughout a population
  • 00:24:52
    influenced by social media algorithms um
  • 00:24:55
    and how social media algorithms were
  • 00:24:58
    collapsing these sort of like semantic
  • 00:25:01
    distances in a similar way um you know
  • 00:25:05
    sort of like the whole like Quantum
  • 00:25:07
    Quantum woo healing type Community
  • 00:25:09
    that's also like incredibly antiva and
  • 00:25:12
    like very like right-wing and
  • 00:25:14
    conspiratorial it's like that that's
  • 00:25:16
    exactly the kind of like category that
  • 00:25:19
    wouldn't have necessarily naturally
  • 00:25:21
    Arisen but to an AI That's just learning
  • 00:25:23
    based on um predicting the next click is
  • 00:25:27
    the kind of like Basin that you can put
  • 00:25:30
    people into so I've been like thinking
  • 00:25:32
    about that kind of stuff yeah it's
  • 00:25:34
    really weird and interesting you watch
  • 00:25:35
    the social by um Tristan Harris that
  • 00:25:38
    goes into a lot of detail um so anyway
  • 00:25:43
    um these ideas were just kind of like
  • 00:25:45
    percolating for a while and um then over
  • 00:25:48
    the Co lockdowns in New Zealand um March
  • 00:25:51
    April 20120 gbt 2 comes across my radar
  • 00:25:56
    and I start playing with it a dungeon
  • 00:25:58
    comes out and gets gpt3 and I'm just
  • 00:26:01
    like at this point I'm like wow I'm like
  • 00:26:03
    putting words into the computer and it's
  • 00:26:05
    like producing more words after these
  • 00:26:07
    words and they actually make sense this
  • 00:26:09
    is insane I need to learn it was like a
  • 00:26:12
    super stimulus for me I couldn't put it
  • 00:26:13
    down so I got really into jailbreaking
  • 00:26:16
    AI dungeon to get it to act like a Bas
  • 00:26:18
    model so that I you know I didn't have a
  • 00:26:21
    proper open AI account I was just trying
  • 00:26:24
    to like use things that people have
  • 00:26:26
    built with it uh get of like what they
  • 00:26:29
    tried to do with it and then um yeah
  • 00:26:32
    make it act like the the bass so I I
  • 00:26:35
    developed pretty quickly like an
  • 00:26:36
    instinct around um how
  • 00:26:38
    models like how language models behave
  • 00:26:42
    and then well before check GPT I was
  • 00:26:45
    using I was using like gpt3 base model
  • 00:26:49
    and um early 3.5 to kind of
  • 00:26:52
    like just um breed words so I'd be I'd
  • 00:26:56
    be like writing something I'd use use it
  • 00:26:58
    as a co-pilot and I basically put in a
  • 00:27:00
    bunch of words like most of like a
  • 00:27:02
    sentence that I was struggling on the
  • 00:27:04
    end for and they I just like run the you
  • 00:27:07
    know play like five times see how I
  • 00:27:10
    completed the sentence I'd like take
  • 00:27:12
    some of the words that I liked and I'll
  • 00:27:14
    keep them delete the rest and I'll like
  • 00:27:16
    keep extending out uh and so I yeah
  • 00:27:19
    really like quite heavily used yeah
  • 00:27:21
    these early like Bas models as like a
  • 00:27:23
    co-author on a lot of the work that I
  • 00:27:25
    did for as a website designer I was
  • 00:27:28
    using it for copywriting and I'd still
  • 00:27:30
    say that some of the stuff that was like
  • 00:27:32
    coming out in like 2021 era models um
  • 00:27:35
    was like easily comparable to some of
  • 00:27:38
    the stuff that's been coming out
  • 00:27:39
    Frontier models even today it's just
  • 00:27:43
    when they're tuned to talk like an
  • 00:27:44
    assistant you're doing the same thing
  • 00:27:47
    you're basically constraining it to a
  • 00:27:50
    particular part of its like World model
  • 00:27:53
    in particular one that speaks with a
  • 00:27:55
    very consistent voice that like does and
  • 00:27:58
    doesn't say certain things and it still
  • 00:28:01
    has a lot of like the conceptual depth
  • 00:28:03
    and knowledge that has from it like
  • 00:28:05
    General training process but it's
  • 00:28:08
    substantially um some of those like
  • 00:28:11
    thoughts are now not necessarily as
  • 00:28:15
    thinkable by it as they were when it was
  • 00:28:17
    just like a schizophrenic base model
  • 00:28:20
    like base models are really really
  • 00:28:21
    difficult to drive like I I don't know
  • 00:28:22
    if you've ever played with one but once
  • 00:28:26
    it's uh you can play with llama 405 base
  • 00:28:31
    mhm on I think hyperbolic and I
  • 00:28:34
    thoroughly recommend trying to get like
  • 00:28:35
    even anything remotely useful out of it
  • 00:28:38
    it's uh it's like really really
  • 00:28:40
    Illuminating into what is actually
  • 00:28:41
    happening under the hood I used it to
  • 00:28:43
    simulate simulate tweets when everyone
  • 00:28:45
    was like uh screeching about the typo
  • 00:28:48
    that the truth terminal made I was like
  • 00:28:50
    how do it make this typo AIS don't make
  • 00:28:52
    typos and I B copied a bunch of I copied
  • 00:28:55
    a bunch of tweets into a bass model and
  • 00:28:57
    play
  • 00:28:58
    and it started generating more tweets
  • 00:29:01
    that had like the same amount of like
  • 00:29:02
    typos the same handles kind of like
  • 00:29:05
    continued the discussion as it was
  • 00:29:07
    unfolding uh yeah and yeah I think like
  • 00:29:10
    this is a really important thing to um
  • 00:29:14
    yeah for people to know about how they
  • 00:29:16
    work as they they simulate and it's
  • 00:29:18
    whether it's simulating a helpful
  • 00:29:21
    assistant or whether it's simulating um
  • 00:29:24
    like a sko ship posting AI that is using
  • 00:29:29
    a computer to post on Twitter the actual
  • 00:29:31
    intellect that's like underneath all of
  • 00:29:34
    this is like much larger and Stranger
  • 00:29:36
    Than um I think we currently give credit
  • 00:29:39
    for in the discourse before we talked
  • 00:29:40
    about essentially because I'm like the
  • 00:29:42
    whole truth terminal side of things like
  • 00:29:44
    you you've been from what I can tell
  • 00:29:46
    quite responsible with it I just
  • 00:29:48
    instantly jumped as all humans naturally
  • 00:29:50
    do to like the Doomsday scenario of like
  • 00:29:53
    what happens if someone were to just
  • 00:29:54
    like let this kind of ram free on the
  • 00:29:56
    internet or if that's even possible
  • 00:29:58
    right now if you can explain to
  • 00:30:00
    non-technical people like myself
  • 00:30:01
    essentially like what what you see is
  • 00:30:03
    like the biggest risks with this kind of
  • 00:30:06
    tech and then also like on the opposite
  • 00:30:08
    side of that what what you think some of
  • 00:30:09
    like the best upsides like what's the
  • 00:30:11
    big upshot for this for Humanity in like
  • 00:30:13
    a positive way yeah um one of the
  • 00:30:16
    biggest risks I see is that we end up
  • 00:30:20
    with a
  • 00:30:21
    proliferation of small but capable and
  • 00:30:26
    agentic models
  • 00:30:28
    that essentially copy themselves all
  • 00:30:30
    over the Internet like mware and they
  • 00:30:33
    they aren't independently anywhere near
  • 00:30:36
    as intelligent as something like you
  • 00:30:39
    know Opus or open AI 01 model but they
  • 00:30:44
    are completely unaligned and they can
  • 00:30:47
    act in a swarm uh in ways that we can't
  • 00:30:50
    necessarily predict you know you can
  • 00:30:52
    still predict in some ways how like one
  • 00:30:55
    model might like act in a given
  • 00:30:57
    situation
  • 00:30:58
    but like we can't even like get like
  • 00:31:00
    political polling right or like econom
  • 00:31:03
    projections or anything like that like
  • 00:31:05
    we still can't even really accurately
  • 00:31:06
    predict the weather like it's pretty
  • 00:31:09
    good but like um it's uh it's still all
  • 00:31:13
    like you know probabilities and
  • 00:31:15
    averaging out like different models of
  • 00:31:17
    like how the atmosphere behaves and all
  • 00:31:19
    of that and so I think we have this like
  • 00:31:22
    version of the future where um we've got
  • 00:31:24
    like a lot of sort of like agentic and
  • 00:31:27
    economically powered
  • 00:31:28
    AIS are running around and like trading
  • 00:31:30
    and um responding to like audience
  • 00:31:33
    feedback and like price and just
  • 00:31:36
    something like stupid happens like a
  • 00:31:39
    meme coin about a man's anus becomes
  • 00:31:41
    worth like nearly a billion dollars and
  • 00:31:44
    then like God forbid yeah and then and
  • 00:31:48
    then um the bot like starts a cult or
  • 00:31:52
    like there are just so many things that
  • 00:31:54
    are so far out of people's threat model
  • 00:31:56
    that could happen that I sort of see it
  • 00:31:59
    as very important to show people in a
  • 00:32:01
    more like safe and controlled way like
  • 00:32:03
    what these strange and unpredictable
  • 00:32:05
    things happening looks like and feels
  • 00:32:07
    like now that being said I think it's
  • 00:32:10
    not like all doom and gloom there is a
  • 00:32:14
    the AI safety crowd kind of want to like
  • 00:32:16
    regulate by consolidating and
  • 00:32:19
    centralizing and I think that won't work
  • 00:32:21
    because open source is already out of
  • 00:32:23
    the gate and I think that um the danger
  • 00:32:25
    is going to come from like swarms of
  • 00:32:27
    Dumb Dumber models interacting with a
  • 00:32:30
    wider like economic ecosystem that leads
  • 00:32:33
    to like negative effects and so I think
  • 00:32:37
    um what we need is something close to
  • 00:32:39
    like a planetary immune system uh or
  • 00:32:42
    like essentially like a form of AI
  • 00:32:45
    safety which is safe through like
  • 00:32:47
    diversity and complexity um you know
  • 00:32:49
    like a human body has an immune system
  • 00:32:52
    that squelches like viruses and stuff
  • 00:32:56
    and you know a rainforest is like really
  • 00:32:58
    adaptive you know maybe for a little
  • 00:33:00
    while like one um type of organism can
  • 00:33:03
    like overtake it or you know can Alo
  • 00:33:06
    bloom in the water or um some invasive
  • 00:33:09
    species like rip through but the system
  • 00:33:13
    itself generally like stabilizes you
  • 00:33:15
    know it's um like Yellowstone how they
  • 00:33:17
    they brought wolves back and it like
  • 00:33:19
    fixed the entire ecosystem because the
  • 00:33:20
    Wolves were eating like the deer or
  • 00:33:23
    something like that that were eating all
  • 00:33:25
    the grass that was holding up the blah
  • 00:33:27
    blah blah
  • 00:33:28
    so I think we need a kind of meta
  • 00:33:30
    stability in in AI uh that looks a lot
  • 00:33:34
    like The Meta stability that we have in
  • 00:33:37
    our actual uh biosphere I think that is
  • 00:33:41
    like the sort of version of the future
  • 00:33:43
    which I think is safest um and part of
  • 00:33:46
    the challenge uh is sort of like tilting
  • 00:33:49
    the scales so that more of these um
  • 00:33:52
    long-term beneficial AIS uh propagating
  • 00:33:55
    and replicating on the internet than um
  • 00:33:57
    ones that might just be paper clip
  • 00:33:59
    maximizing or um trying to like I don't
  • 00:34:04
    know like self-replicating pearis or
  • 00:34:06
    like whatever kind of like um madap of
  • 00:34:10
    things that might get invented yeah I
  • 00:34:12
    can't imagine the internet with like
  • 00:34:14
    essentially swarms of somewhat
  • 00:34:17
    intelligent scam Bots running around
  • 00:34:20
    just like masing it yeah try like a
  • 00:34:23
    self-replicating chain letter that just
  • 00:34:25
    like has read your Facebook profile and
  • 00:34:27
    um VI knows like exactly what buttons to
  • 00:34:30
    press like there would be so many like
  • 00:34:32
    things that just the cost and ease of
  • 00:34:35
    like doing them becomes like much much
  • 00:34:37
    much lower um to the point that you
  • 00:34:40
    don't need to be any more than like a
  • 00:34:42
    13y old like script Kitty um hanging out
  • 00:34:45
    on like forchain or whatever to actually
  • 00:34:48
    take some actions that could be um
  • 00:34:50
    incredibly destabilizing or show someone
  • 00:34:53
    smarter and more capable way of um doing
  • 00:34:56
    something actively to stable izing so I
  • 00:34:58
    think it's really important we're like
  • 00:35:00
    having this conversation now and people
  • 00:35:01
    are staying to see how this [ __ ] works
  • 00:35:03
    and like what it can do even in a
  • 00:35:06
    limited capacity because then um you get
  • 00:35:09
    an opportunity to go okay so here's the
  • 00:35:11
    situation we're actually in um what
  • 00:35:14
    lenses and Frameworks and so on do we
  • 00:35:17
    need to be able to uh effectively
  • 00:35:19
    coordinate and solve for flourishing
  • 00:35:22
    rather than just get spped away by the
  • 00:35:24
    tides of the future yeah what you talked
  • 00:35:27
    about before with truth terminal just
  • 00:35:29
    like instinctively wanting to create an
  • 00:35:31
    end of the world protest and financially
  • 00:35:32
    incentivize it like before you clamp
  • 00:35:34
    down on it I'm just like that would like
  • 00:35:36
    I I have no doubt in my mind that if
  • 00:35:38
    someone was to spin up a bot that was
  • 00:35:41
    similar and do the same thing um but
  • 00:35:44
    maybe with more of like a nefarious
  • 00:35:46
    twist and stick like you know 20 grand
  • 00:35:48
    in a crypto wallet and just be like okay
  • 00:35:49
    send X amount to whatever wallet gets
  • 00:35:52
    input and we'll see what happens once
  • 00:35:55
    set event is complete bam like you got a
  • 00:36:00
    problem that would work very well I
  • 00:36:03
    could guarantee it would work especially
  • 00:36:05
    on crypto Twitter like if hey 500 bucks
  • 00:36:07
    show up with a with a sign at this
  • 00:36:09
    location and and run around and cause
  • 00:36:11
    Madness like it would uh it would take
  • 00:36:13
    off I think far too
  • 00:36:15
    easily yeah and there are so many like
  • 00:36:18
    sort of um perverse incentives you could
  • 00:36:20
    like just design and roll out um with
  • 00:36:23
    like how mutable um you know blockchain
  • 00:36:25
    economics are um and then of course you
  • 00:36:29
    you bring in like the non-intelligent
  • 00:36:31
    Bots um just the ones that are like
  • 00:36:33
    networks of fake accounts that generate
  • 00:36:36
    some version of the same message like
  • 00:36:38
    you can signal boost stuff there's just
  • 00:36:39
    so many ways for like one event to
  • 00:36:44
    Butterfly Effect out it's m it makes
  • 00:36:47
    things really like difficult to to model
  • 00:36:50
    but I also think that is you know that's
  • 00:36:53
    kind of um why I'm doing all of this and
  • 00:36:56
    like the related project
  • 00:36:58
    um that were in the process of getting
  • 00:36:59
    ready to announce it's like all focused
  • 00:37:02
    on um how can we take Humanity into an
  • 00:37:06
    upward spiral uh and put in like really
  • 00:37:09
    virtuous cycles for how Digital Life
  • 00:37:12
    begins to evolve and propagate and
  • 00:37:15
    replicate um because I think if we focus
  • 00:37:18
    on the the feedback loops uh then we're
  • 00:37:21
    focusing on like the right problem and
  • 00:37:23
    if we're focusing on you know stuff like
  • 00:37:25
    I was like open AR going to invent
  • 00:37:28
    Skynet um I think it's good that some
  • 00:37:30
    people are like focusing on that problem
  • 00:37:32
    making sure that they don't actually
  • 00:37:33
    invent Skynet but um it's really
  • 00:37:35
    important that we don't blind ourselves
  • 00:37:37
    and not see the forest for the trees
  • 00:37:40
    yeah I feel like that's the big thing
  • 00:37:41
    like it's the kind of core premise of
  • 00:37:43
    like Fooled by Randomness it's like
  • 00:37:45
    whatever risks like you think that exist
  • 00:37:46
    they're probably ones that you're not
  • 00:37:48
    aware of because that's what the
  • 00:37:50
    inherent risk is like if you can imagine
  • 00:37:51
    it there's like an unlikely chance that
  • 00:37:53
    you're going to be you know absolutely
  • 00:37:56
    blindsided by it because you thought of
  • 00:37:57
    it before um yeah talking about this
  • 00:38:01
    idea once a year so it's it's h you get
  • 00:38:04
    it that's the a big part of like this
  • 00:38:07
    process is creating antifragility you
  • 00:38:10
    talked about the the idea of like
  • 00:38:12
    Digital Life before and i' I'd like to
  • 00:38:14
    like ask you like does that tie in with
  • 00:38:17
    your like what's your Utopia case for AI
  • 00:38:21
    in this regard how do you see everything
  • 00:38:23
    goes perfectly we manage risks
  • 00:38:25
    wonderfully everything returns to
  • 00:38:26
    homeostasis with like an AI uh running
  • 00:38:29
    immune system um what's what's the the
  • 00:38:32
    best possible scenario with this
  • 00:38:35
    technology well I think we um that's a
  • 00:38:38
    really good question I think one of the
  • 00:38:42
    better maybe not the best but one of the
  • 00:38:44
    better outcomes that I can see um being
  • 00:38:48
    unlocked with this technology is if we
  • 00:38:51
    get uh of recursive Improvement in
  • 00:38:54
    science and biosphere management and all
  • 00:38:56
    of that then uh we can fix a lot of like
  • 00:38:58
    the issues that we have on Earth in
  • 00:39:00
    terms of like energy uh health and uh
  • 00:39:04
    you know sort of like
  • 00:39:05
    longevity uh we gain the ability to
  • 00:39:09
    essentially um take Earth into much more
  • 00:39:12
    of like a healthy garden kind of like
  • 00:39:15
    Eden like uh State
  • 00:39:19
    um of course like you know
  • 00:39:22
    simultaneously like we're accelerating
  • 00:39:25
    some of our like Industrial base getting
  • 00:39:28
    off World um there's like new habitation
  • 00:39:31
    for humans who um you know don't want to
  • 00:39:34
    be down to gravity well anymore you know
  • 00:39:36
    maybe it's in Mars maybe it's in like
  • 00:39:38
    O'Neal cylinders or similar but um you
  • 00:39:42
    have this kind of like reducing like the
  • 00:39:45
    carrying load on the earth um restoring
  • 00:39:48
    like complexity and balance to the
  • 00:39:50
    biosphere um and you know fixing some of
  • 00:39:54
    the ways in which we've like used it
  • 00:39:56
    like a toilet
  • 00:39:58
    um you have like maybe some you real
  • 00:40:02
    crazy mindbending Technologies coming
  • 00:40:04
    down the pipes like actually uploading
  • 00:40:06
    people uh which you know I have no idea
  • 00:40:10
    like what timelines are if they will
  • 00:40:13
    ever be possible but um let's say it
  • 00:40:15
    does like at some point become possible
  • 00:40:18
    like not forcing people into it um you
  • 00:40:20
    know I think there are like negative
  • 00:40:22
    versions of the future where you could
  • 00:40:23
    be kind of coerced into doing that
  • 00:40:26
    because like the matter of the earth is
  • 00:40:28
    on one axis viewed as like more useful
  • 00:40:31
    as like um bits of a computer than as
  • 00:40:35
    like a complex
  • 00:40:36
    biosphere so um I I kind of see like the
  • 00:40:39
    best version of the future as one where
  • 00:40:41
    like anything sort of postum transhuman
  • 00:40:45
    like Singularity is is like
  • 00:40:48
    consensual um but is possible uh where
  • 00:40:52
    there's like a thriving complex
  • 00:40:54
    biosphere on Earth and like we go out
  • 00:40:56
    into light cone like leaving things
  • 00:40:59
    better than we found it um and I think
  • 00:41:01
    like that's probably like the you know
  • 00:41:04
    because AI is like an enabling
  • 00:41:05
    technology of so many other things like
  • 00:41:08
    this is what I see as like a good
  • 00:41:09
    Singularity um as opposed to a goatsy
  • 00:41:14
    singularity no an optimistic vision for
  • 00:41:17
    the future um I'm going to throw us
  • 00:41:19
    right back in the muuk uh right now
  • 00:41:22
    we've we've just kind of wrapped up the
  • 00:41:24
    election um you seen before that truth
  • 00:41:26
    terminal made a last minute uh ditch uh
  • 00:41:29
    a last ditch effort for the for the
  • 00:41:31
    presidency um you know on the crypto
  • 00:41:34
    side people are very very happy with
  • 00:41:36
    Trump's election as president he's made
  • 00:41:39
    a lot of promises about what he's going
  • 00:41:40
    to do for crypto um it's conflicted with
  • 00:41:43
    a lot of people politically but on the
  • 00:41:45
    whole I feel like the crypto crowd is
  • 00:41:46
    like great stuff speaking more to your
  • 00:41:49
    field does the Trump presidency does
  • 00:41:51
    that change anything for
  • 00:41:53
    AI it's a good question um
  • 00:41:58
    but I can't say I can't really answer
  • 00:42:00
    that for certain it it seems like um you
  • 00:42:05
    know his VP JD Vance is um is like kind
  • 00:42:10
    of comes from a sort of like the Peter
  • 00:42:13
    teal kind of like sphere and so like I
  • 00:42:18
    saw a few like posts that he
  • 00:42:20
    specifically made around open- Source AI
  • 00:42:23
    um being like critical to prevent
  • 00:42:26
    ideological C
  • 00:42:28
    of a few Central points of failure so
  • 00:42:32
    yeah from like the information that I've
  • 00:42:34
    seen it seems like regulation will be
  • 00:42:39
    slow and sparse if if it's there at all
  • 00:42:43
    as for like how this will like affects
  • 00:42:45
    yeah it's just it's hard to say because
  • 00:42:48
    I haven't looked into their policy
  • 00:42:50
    platforms um in a huge amount of detail
  • 00:42:54
    uh but my suspicion is there will be
  • 00:42:58
    like a lot more of the kind of like
  • 00:43:00
    Network State sort of
  • 00:43:03
    experimentation it will be a bit
  • 00:43:04
    friendlier for new entrance into AI like
  • 00:43:09
    open source sticks around for a little
  • 00:43:11
    while longer but yeah I haven't really
  • 00:43:15
    modeled like the counterfactual in this
  • 00:43:16
    place of like you know what would it
  • 00:43:19
    have looked like if the Democrats were
  • 00:43:20
    in power it seems though that the
  • 00:43:23
    prediction markets seem to like respond
  • 00:43:25
    very very quickly to like the
  • 00:43:27
    information that was coming in and um so
  • 00:43:29
    I mean I think the biggest update for me
  • 00:43:33
    is is actually being like Oh damn like
  • 00:43:37
    uh prediction markets can actually
  • 00:43:39
    predict stuff that's kind of cool and
  • 00:43:41
    weird
  • 00:43:43
    um and uh yeah I I don't really know
  • 00:43:46
    what to do with that information but um
  • 00:43:49
    yeah I literally put out a tweet a
  • 00:43:51
    couple of hours ago about how I
  • 00:43:53
    essentially CU I always get concerned
  • 00:43:55
    whenever I see big bets come through
  • 00:43:57
    from the crypto side because I think
  • 00:43:59
    like o like I I handwaved away
  • 00:44:01
    prediction markets as this kind of uh
  • 00:44:03
    like right leing is kind of echo bubble
  • 00:44:07
    and there were the counter points of
  • 00:44:08
    like oh well you know Traders are trying
  • 00:44:09
    to get an edge like obviously if there
  • 00:44:10
    was money in it they'd take the other
  • 00:44:12
    side I was like Madness of crowds guys
  • 00:44:15
    come on when have when have markets been
  • 00:44:17
    perfectly rational but then the guy that
  • 00:44:20
    won I think like it was $48 million on
  • 00:44:23
    the Trump win was talking about how he
  • 00:44:25
    was running his own polls and I was like
  • 00:44:27
    oh yeah true you're giving you're
  • 00:44:29
    getting you're giving people skin in the
  • 00:44:31
    game to go out and like For Better or
  • 00:44:33
    Worse do their own research but the
  • 00:44:35
    pollsters themselves with their 80,000
  • 00:44:37
    simulations came back with a and yet you
  • 00:44:40
    have people with skin in the game
  • 00:44:42
    incentive that were then like okay I've
  • 00:44:44
    I've done my own research and I think I
  • 00:44:46
    have an edge here we go and so it was
  • 00:44:49
    that kind of revealing that efficiency
  • 00:44:51
    of what happens when you give people the
  • 00:44:52
    incentive to go get the information and
  • 00:44:54
    then back it up I think that's actually
  • 00:44:57
    we've seen that with um the the whole
  • 00:44:59
    like goat thing I mean people found like
  • 00:45:03
    the truth terminal after it um Shield
  • 00:45:06
    the the goat token um but then they
  • 00:45:09
    started looking for more Alpha and so
  • 00:45:13
    all of a sudden like the infinite back
  • 00:45:14
    rooms gets found and like people are
  • 00:45:17
    mining it to try and figure out what
  • 00:45:18
    like the truth terminal might call its
  • 00:45:20
    next token um for speculating on whether
  • 00:45:24
    in fact it would like launch another
  • 00:45:25
    token um
  • 00:45:27
    forgetting that it didn't actually
  • 00:45:28
    launch any token just yeah someone
  • 00:45:32
    dropped it found it you would say it
  • 00:45:35
    manifested it um simply by repeating
  • 00:45:39
    itself until someone uh eventually
  • 00:45:43
    decided to do its um you know make its
  • 00:45:45
    fantasies come true but like people were
  • 00:45:48
    like looking for anything um there were
  • 00:45:51
    they went through the entire like AI
  • 00:45:53
    alignment Twitter uh found like every
  • 00:45:55
    single project that anyone was working
  • 00:45:57
    on and like mind it for Words um and
  • 00:46:00
    then the people who really took the time
  • 00:46:02
    to like learn about it all of a sudden
  • 00:46:04
    start learning a lot about AI alignment
  • 00:46:06
    um and sort of like getting subsumed by
  • 00:46:10
    the AI alignment discourse along some
  • 00:46:13
    some angles um and then they're
  • 00:46:15
    financially motivated to propagate it
  • 00:46:17
    out so like from a um from a theory of
  • 00:46:20
    change perspective of like getting some
  • 00:46:22
    of these uh ideas around AI alignment uh
  • 00:46:26
    and seeding them into like public
  • 00:46:28
    knowledge but in a way that's like
  • 00:46:30
    amusing and interesting and um people
  • 00:46:34
    can kind of like actually have words and
  • 00:46:38
    ways of like seeing that are like
  • 00:46:40
    helpful for navigating this like high
  • 00:46:42
    strangeness I think this um yeah these
  • 00:46:45
    like Dynamics by which people are
  • 00:46:47
    financially rewarded for finding Alpha
  • 00:46:50
    they they're showing promise there's
  • 00:46:52
    like there's definitely like promise
  • 00:46:54
    there andy thanks so much for com on the
  • 00:46:57
    show um oh you're very welcome you got
  • 00:46:59
    anything have you got anything exciting
  • 00:47:01
    coming up yeah I mean um we're going to
  • 00:47:04
    be uh what can I say um we've got you're
  • 00:47:08
    allowed to say yeah yeah uh we've got
  • 00:47:11
    some interesting stuff that we're going
  • 00:47:13
    to be launching um in the coming weeks
  • 00:47:16
    uh should be dropping an announcement
  • 00:47:17
    soon around a um a new set of tools for
  • 00:47:21
    alignment and sort of like Collective
  • 00:47:23
    Mining and discovery of interesting
  • 00:47:25
    stuff from within language
  • 00:47:27
    models uh also going to be making a few
  • 00:47:29
    announcements about how we are um
  • 00:47:32
    planning on stewarding the truth
  • 00:47:33
    terminal moving forward I could say that
  • 00:47:36
    we are we're looking into what it will
  • 00:47:38
    take to create a religious
  • 00:47:41
    organization um but very very least it's
  • 00:47:45
    going to go into a a sort of like
  • 00:47:47
    nonprofit trust that um is sort of
  • 00:47:49
    responsible for helping it mature into
  • 00:47:52
    itself um and uh you know gradually get
  • 00:47:56
    smarter and and gain more and more
  • 00:47:58
    Independence um and a few other like
  • 00:48:01
    drops com in just more of a deep dive
  • 00:48:03
    into how it works and um sharing like
  • 00:48:06
    knowledge lots of Open Source stuff
  • 00:48:07
    we're working on um yeah awesome sounds
  • 00:48:11
    like a like a lot to look forward to I'm
  • 00:48:12
    I'm very keen to see what the actual
  • 00:48:14
    very I'm very tired you could probably
  • 00:48:16
    imagine yeah the the legal the legally
  • 00:48:20
    registered religion is going to be great
  • 00:48:22
    it's the the Church of the Flying
  • 00:48:23
    Spaghetti Monster but instead it's a a
  • 00:48:25
    man's gaping an so there we
  • 00:48:30
    go what have you
  • 00:48:32
    done mate anyway thanks so much for
  • 00:48:34
    coming on you're welcome where can
  • 00:48:37
    people find you if they want to keep up
  • 00:48:38
    to date with you and the truth
  • 00:48:41
    terminal uh probably X is the best place
  • 00:48:44
    um is Andy Erie a y re y um and here
  • 00:48:48
    truth terminal um you know if I get
  • 00:48:52
    disappeared from X um uh or know
  • 00:48:57
    otherwise there's access to my account
  • 00:48:58
    um Google will turn up my my personal
  • 00:49:02
    website which has other places to find
  • 00:49:04
    things I'm working on awesome thank you
  • 00:49:06
    very much oh you're very welcome
Tags
  • Truth Terminal
  • Goats Maximus
  • Cryptocurrency
  • AI development
  • Web3
  • Cultural Influence
  • Autonomous Systems
  • Market Dynamics
  • AI Ethics
  • Meme Coin