Latest / Uncomfortable Conversations with Josh Szeps / Jaron Lanier: "There's No Such Thing as Artificial 'Intelligence'"

Transcript

Jaron Lanier: "There's No Such Thing as Artificial 'Intelligence'" · Uncomfortable Conversations with Josh Szeps

• 0:00
  Get a humans welcome to the safe space for dangerous ideas, here's a dangerous question
• 0:07
  for you, is artificial intelligence intelligent in any meaningful way or is it just scraping
• 0:15
  all of its troves of input data in order to spit out essentially what is a very fancy
• 0:22
  autocomplete.
• 0:23
  The latter would be the position of one of the most influential people in technology
• 0:28
  in the world, Jarron Lanier, who has written a number of books, including ten arguments
• 0:34
  for deleting your social media accounts right now, you are not a gadget and dawn of the
• 0:39
  new everything.
• 0:40
  He is the man who coined the term virtual reality.
• 0:44
  He was one of the first virtual reality pioneers.
• 0:46
  He founded the first company to sell virtual reality products in the 1980s and he has since
• 0:52
  become of a cypherous critic of social media and now of the kind of doomsday culture that
• 0:58
  surrounds artificial intelligence.
• 1:00
  The reverence that so many people in Silicon Valley have towards this tipping point moment
• 1:05
  as so many people see it, this fork in the road for civilization.
• 1:09
  He says calm down everybody, keep your pants on, artificial intelligence is a tool, there's
• 1:15
  nothing going on behind the scenes, the light to not on and they will never be on.
• 1:19
  This point of view, and he may be the world's most visible exponent of this point of view,
• 1:27
  bumps up uneasily against the experience of so many of us when we interact with artificial
• 1:32
  intelligence as like chat GPT so I wanted to pick Jarron's brain about it.
• 1:38
  His name mostly comes up when people Google who is the guy with dreadlocks in the Netflix
• 1:42
  film The Social Dilemma which was Netflix's documentary about the perils of social media
• 1:47
  so you may recognize Jarron from that film.
• 1:50
  This is a fascinating chat, it's worth understanding that before the chat began, Jarron was fiddling
• 1:55
  around with his connection to the web-based platform that we use to record these interviews
• 2:02
  on for about 17 minutes trying to log in correctly.
• 2:06
  So he started things off, reasonably frustrated and the prickliness continued in ways that
• 2:11
  are I think revealing and instructive.
• 2:13
  Please enjoy as much as I did this chat with the one and only.
• 2:17
  Jarron Lennie.
• 2:24
  Have you been loud by AI?
• 2:27
  Jarron have you had that experience that many of us have had in the past year or two?
• 2:31
  No, I'm too close to it, I mean I work with it directly so I'm not loud but...
• 2:36
  There's never been a moment where you've gone, it's doing things that I don't quite understand
• 2:39
  under the hood.
• 2:40
  No, I've never had that experience.
• 2:43
  Let me through the analogy that you made in that New Yorker piece that you wrote recently
• 2:47
  about trees and forests and generative forests, products and phantom trees and a lonely
• 2:52
  color of a cat in a parachute.
• 2:55
  We have this new thing that's having such an impact on so many people and I'm not saying
• 3:01
  that I've come up with a perfect way to explain it but I'm struggling to improve a sort
• 3:06
  of a story we can tell non-technical people.
• 3:10
  Now there's no way to ever completely get science simplifications right with no compromise.
• 3:18
  There's always going to be a little bit of a compromise.
• 3:20
  I've been playing around with alternate explanations that have slightly different compromises
• 3:24
  than the ones I hear and some of this stuff I've been doing seems to reach more people
• 3:29
  in a way that's helpful to them even if it's not absolutely perfect or complete explanation.
• 3:38
  One of the things I've been playing with is using the kind of explanation that works
• 3:43
  best for diffusion models for other kinds of large model AI which is a little bit of
• 3:51
  a stretch but it's actually not too bad.
• 3:53
  It's given a dimension or two, it's actually not a bad explanation.
• 3:59
  The way this works if you want to try it out on somebody is it goes like this.
• 4:06
  10, 12 years ago the AI community was quite concerned with how to tell cats from dogs
• 4:12
  because they're kind of similar, yet people can tell them apart and so playing around with a lot
• 4:17
  of different configurations. We ended up with this realization that organizing a lot of layers
• 4:25
  of measurements, that look at measurements, that look at measurements, that look at measurements
• 4:29
  works and so you say neuron, I say measurement, a neuron is basically something that measures
• 4:35
  something simple and gives the result. If you have a whole bunch of them looking at a picture
• 4:42
  saying a grid you might be able to do things like measure how sharp the edges are in the little part
• 4:47
  they're looking at and then you can have another part look at those and they might look at
• 4:54
  how well they align with their neighbors or God knows what all kinds of things and you build up
• 4:57
  layers and layers and layers of those. When you have a lot of layers you call it deep, that's what
• 5:02
  deep means when you hear about deep machine learning and then at the top it says cat or dog and it
• 5:07
  turns out that structure of having more layers than we originally thought works. You might ask,
• 5:12
  why does it work? Well, we have some ideas about that but ultimately it's something like medicine
• 5:19
  or weather forecasting or many other things and that we try different techniques until we find
• 5:25
  something that works but we don't necessarily always have a perfect explanation for ourselves
• 5:31
  and indeed we might pretend to more than we really do. It's a bit of a folk technology if you like
• 5:36
  this recognition business but it's no more so than figuring out which medicine works. It's no more
• 5:43
  so than figuring out which way of setting interest rates to cool an economy or something like that.
• 5:49
  It's just one of those things that people do with complexity. Now you can recognize things. You
• 5:56
  can tell cats from dogs. A great thing you can do is you can run it backwards and the way you do that
• 6:02
  is you start with just a feet of noise, a screen of snow and then you run it through this
• 6:08
  recognizer and you say, you said it's a cat or dog and it'll say what do you mean? It's neither,
• 6:13
  it's just snow but then what you can do is you can add some snow to this snow randomize it a bit
• 6:19
  and then it might say, you know, it's just a smidge and more cat like now and so if it's a smidge
• 6:24
  and more cat like keep that new messing up of the snow and if it's less cat like throw it away
• 6:32
  now you just keep on doing that and every time you keep a result that's a little more cat like
• 6:37
  it's like this cat emerges out of the snow like what else would happen and at the end of the
• 6:42
  process you have a picture of a cat. It's not a specific cat. It's kind of an amalgam of all of the
• 6:49
  cats that the thing was trained on and they realized they didn't talk about training. Let me just
• 6:56
  clarify for people here. So when you ask an AI to provide an image of a cat, what it's doing is
• 7:03
  it's going backwards and reversing the normal order of it's learning essentially which you're
• 7:11
  analogizing to being a series of sort of almost like thin templates if you imagine like you're
• 7:18
  building up some kind of a layered cake and each layer is slightly more or less cat-like than the
• 7:24
  one before. That's the the learning process and then when it's generating images can still
• 7:29
  I don't even know if the layers are more cat-like than the ones below. It's just that cumulatively
• 7:33
  they can tell a cat from a dog or find a cat whether there's a cat there at all. But what they have?
• 7:38
  Leia doing, John. All right. Well, we don't know. That's what I'm saying is that this is this is a
• 7:44
  bit of a folk it's a bit of a folk art. It's a little bit like dowsing to find where to go for
• 7:50
  else. This is the thing. Except it works. It's better than that because this actually works. So
• 7:54
  it's not that bad but the thing is it's a sort of an empirical slash intuitive technique that
• 8:01
  the community has discovered after trying a lot of different things. And do we have some stories
• 8:08
  about what works? Sure. Have we absolutely confirmed them? Not really. So I classify it as
• 8:15
  you know it's a it's a it's a known technique that works and as I was saying this is what
• 8:21
  people have to do to deal with the technologies of complexity. Most of medicine works that way.
• 8:27
  We're only at this point starting to get to the point of precision medicine where we can say well
• 8:32
  we we really understand this thing at a molecular level. So we'll do this other thing. But even then
• 8:37
  we don't really know what it's going to do until we do extensive testing. And so there's a
• 8:43
  it's very empirically based and this recognition phase is like that. And I should mention
• 8:50
  in order for it one of these towers and I you can think of them as like skyscrapers of floors
• 8:56
  anything that has a lot of slices together or the tallest ever wedding cake.
• 9:03
  But whatever it is in order to have it look for cats instead of something else you have to
• 9:07
  train it on cats and the way you do that is you get a bunch of known cats and you pass them through
• 9:14
  it and whenever it gets it right you reinforce little changes within it that mostly have to do with
• 9:19
  emphasizing or deemphasizing individual measurements distributed throughout the stack.
• 9:25
  Right. So a precondition of the building of the wedding cake to the to the moon is that you
• 9:31
  have millions or billions of images of cats and dogs that you can feed into it. You're not crazy
• 9:36
  about the word learning and training because it's sort of anthropomorphic. Yeah. I don't like the
• 9:39
  anthropomorphizing word. But the fact the fact is that it's building up its understanding.
• 9:43
  Well again understanding is anthropomorphic too but let's just use those terms because I don't
• 9:47
  know with that one and it's building up its understanding of what is a cat and what is a dog
• 9:51
  based on the accurate notation that there are real cats that who are online and there are real dogs
• 9:57
  whose images are online and it's mapping the differences. I guarantee that you've participated in
• 10:02
  this because it's frequently the case that in order to get to something online you have to pass a
• 10:06
  capture test where it'll say identify all the fire hydrants in the scene or something like that
• 10:12
  and that's a sort of a legacy of that era when when Google figured out how to get people to work
• 10:17
  for free tagging pictures that would then be fed into these towers to train them. So that's
• 10:22
  what that's all about. And is that true? I thought that was a conspiracy theory. Are they using
• 10:26
  our capture data to train AI? Yeah absolutely. Okay. So I mean and by the way as conspiracies go it's
• 10:35
  a pretty mild. It doesn't involve Putin or the sexual abuse of children or flying saucers.
• 10:43
  It's basically just you know I'm not saying it's the worst thing in the world. I'm just saying
• 10:46
  that it happened. I think it's this particular thing is more funny than evil but I do think it's
• 10:54
  I just the reason I mentioned it gives people a concrete sense of what this is like because
• 10:58
  there you are feeding the machine. You know so you can understand part of the process because you
• 11:03
  did it. So yeah and we end up with a forest basically and we end up sometimes with phantom
• 11:10
  trees in the sense of you're jumping ahead. Sorry. Let me do it. Let me do it. It's my it's my
• 11:18
  explanation. This is your baby. I'm not going to I'm not going to interfere with the thing
• 11:21
  of your baby. I will not live this any further. Yeah and it's not perfect. It's not my thing is
• 11:25
  not perfect. I think it can be improved a lot but let's at least deal deal with it however good
• 11:30
  or bad it is. So so we've we've done we've done a couple things that we've trained a tower to
• 11:35
  find cats. We've run the tower backward so it can guide noisiness to turn into a cat right so far
• 11:45
  we have a way of making a cat picture which is not a thing in short supply. Okay so now the
• 11:51
  really interesting thing happens is when you make a prompt and a prompt is a combination of
• 11:58
  different qualities that you want. Generative AI to adhere to at the same time so you can say
• 12:06
  can you make a picture of a cat that's in a parachute playing a tuba or whatever the hell I
• 12:11
  said like just some weird combination. I mean we'll say did you do if it's Australia and maybe
• 12:17
  it's a kangaroo. I don't know. I appreciate the the modification for a cultural element.
• 12:22
  We're we'll we'll customize for you. Anyway when you do this thing then when it's moving from
• 12:29
  snow into something it has to round robin through all of these different qualities and get all
• 12:35
  of them to work at once and then what comes out is really interesting because it naturally fills in
• 12:42
  all the weird little gaps that you'd need to have filled in to make the the combination makes sense
• 12:48
  like how does a cat hold a tuba or did you do or whatever or I'm sorry it's a kangaroo it's a kangaroo
• 12:54
  does the did you do go in the pouch just a bit of the mouth it should go to the mouth probably
• 12:59
  and I mean I'm I'm not I want to say this AI picture of a kangaroo playing a did you do probably
• 13:07
  even more challenging for an AI to produce than a cat playing a tuba given that there a few
• 13:11
  images of kangaroos and did readers. Yeah with with all respect to the marsipial community they're
• 13:17
  going to be more cats online and the more there is of something the more robust will be AI's
• 13:22
  treatment or whatever it is. So anyway now you get the combination and the combination is really
• 13:28
  where the value comes because it's getting at that combination that's work you know that's really
• 13:34
  interesting and so if you're just doing funnier surreal pictures and I'm sure you can go online
• 13:39
  and get a kangaroo playing a did you instantly from any of the many services that do this now
• 13:45
  you can try our dolly if you want or you can go into oh god there's so many of these things there's
• 13:50
  so much so just do it you'll see how it is it might be good or it might not be but the thing is
• 13:56
  getting those combinations filling in the gaps so you were talking about phantom towers so I
• 14:00
  try to explain this is that if you have a bunch of these towers like you have one that recognized
• 14:05
  as kangaroos and another one that recognizes did you do you can imagine this as filling in a
• 14:11
  phantom tower in between them like a meadow and the forest and and because there is a real
• 14:17
  construction there and if you want to anthropomorphize that you can call that a sort of creativity but
• 14:23
  the reason I like this image is that it gives us an intuition that I think is pretty good about
• 14:29
  what this is good for and what it's not good for it's amazing that it can fill in the gaps but it
• 14:35
  doesn't particularly go higher than the gaps it's all a bit of a random process so once and
• 14:39
  while higher than the gaps but not a lot and not consistently so it's a way of amalgamating what
• 14:45
  people have done before and some of my AI friends think that that undersells it or it's not
• 14:53
  weird or glorious or sci-fi enough but I disagree with them I think that's like a fabulous fabulous
• 14:59
  thing like basically what civilization is all about is humans being able to compound efforts
• 15:05
  and not waste efforts and so you know because we're basically the same people that we were
• 15:11
  100,000 years ago but what's different is that we can be educated with what others have learned
• 15:19
  before and we can use tools that others have made before so making the past more present for us
• 15:27
  is what civilization does and so AI is a form of that that's that's fresh and I think really
• 15:35
  interesting so like a great example is instead of making an animal playing a musical instrument
• 15:42
  if you just want to make a website there's so many zillions of websites that have been entered
• 15:48
  as code on sites like GitHub that you can train systems on that now when you need a website you can
• 15:56
  prompt for it and something will come out that essentially leverages what everybody's done before
• 16:01
  so you don't have to start from scratch every single time and so you can be more productive
• 16:07
  I think that's fabulous and I think it's a very concrete and wonderful tool that brings people
• 16:13
  a lot I don't think it's alive I don't think it's creative in any deep sense I know some people
• 16:19
  really wanted to be and if you want it to be if that's your religion I don't want to argue religion
• 16:23
  with it but I think a really concrete way to understand it is that it lets us leverage what
• 16:30
  each other have done in the past more effectively than before and I think that that's brilliant and
• 16:36
  fantastic worthwhile and we can do it for code we can do it for text we can do it for images now
• 16:42
  for movies for all kinds of things and I'm really into it but I think if it's demystified like that
• 16:50
  maybe you lose a little bit of science fiction fantasy but I think what you gain is you gain
• 16:55
  more clarity on what you can do with it and that's how to be worth more I found that clarity
• 17:00
  extremely useful Jaren I thought your piece was was really eye-opening and revelatory to me as an
• 17:05
  amateur who just plays around with it and I hear your reticence about anthropomorphizing this
• 17:11
  stuff and you're kind of clarion call for us to understand it better and to treat it as a tool
• 17:16
  instead of as as if we're generating creatures who are pouring forth from some kind of magical portal
• 17:21
  that we've just stumbled stumbled upon and yet the experience of using it is an experience of
• 17:27
  seeing it do things that seem weird it like you just said I don't think it's alive I don't think
• 17:35
  it's intelligent I think it's truly creative those are very different things right I mean we can
• 17:39
  agree that it's not alive we can agree that it may not be very creative but it might be as creative
• 17:44
  as John Grisham is and it might be as intelligent as a chim I mean look here's the thing about that
• 17:52
  is that these judgments are kind of subjective and I have to confess I haven't I don't think I've
• 18:03
  ever read Grisham's work I hope that doesn't make me sound like a snob or something but I'm a
• 18:08
  little reluctant to sort of dump on somebody who's popular because I a lot of times the most
• 18:14
  popular writers actually have extraordinary see I'm not actually not dumping on it I mean I think
• 18:19
  that I mean would you agree that if it's not the case now it will presumably soon be the case
• 18:26
  that your average marvel blockbuster could be produced this could be produced by an AI like that's
• 18:33
  not to say that they're not great yeah yeah yeah okay so there's a lot of things to say about this
• 18:38
  um we perceive each other largely through faith and I mean you know going back way early in the
• 18:47
  history of computers when they were very small and very slow and very incapable compared to
• 18:53
  anything we're used to today you probably you probably heard of something called Eliza which
• 18:59
  goes back to the 60s when computers could do nothing and it was a very simple language interaction
• 19:06
  program that faked people into believing it was a human therapist and they would pour out their heart
• 19:11
  and souls to it and it just did super simplistic stuff of repeating back what people had just said
• 19:17
  and yet it passed the turing test which is the way we say it convinced people that it was human
• 19:25
  and that the thing is whether somebody finds a particular program to be human or not is kind of up
• 19:35
  to them um I think this thing of what we decide to be human or not is a matter of faith but also a
• 19:43
  matter of politics uh but I'm not saying human or not I'm saying well yeah put it this way would
• 19:54
  would you do share my suspicion that in some years probably sooner than we expect we'll be
• 20:00
  interacting with artificial systems in a way that we feel is indistinguishable from interacting
• 20:06
  with a sentience that's entirely a human choice I won't because you know I'm I'm the seller I'm
• 20:13
  not going to drink my own whiskey I'm not going to get fooled by me also there's a great um drink
• 20:18
  your own whiskey is an old saying from the bootleg era in the United States like your bootleg
• 20:24
  whiskey from the prohibition era might not have been necessarily safe to drink so you don't drink
• 20:29
  your own whiskey but um my favorite recent version is uh from the 90s in in hip hop which is that I
• 20:36
  don't get high on your own supply yeah yeah I'm the seller I'm the pusher I'm not going to I'm not
• 20:41
  going to buy into this but you can if you want send us money go for it but um I suspect I won't
• 20:48
  have a choice I mean I don't I suspect that to walk through the world will be to walk through a
• 20:52
  world that's populated both by well by people and also by systems that seem intelligent yeah the
• 20:58
  moment you feel you don't have a choice in anything you should rebel it's not just this like demand
• 21:04
  that choice even if you choose to disagree with me about how it's best used I mean um if you're
• 21:11
  losing choice something's going very very wrong and we're not doing our job well uh so
• 21:16
  let me send us a sentence here stop just stop for a second let me say something because I've
• 21:20
  been trying for a while um uh we perceive each other through faith and if you want to believe that
• 21:26
  other people are special and that simulations of people are different from people you can't prove
• 21:32
  that scientifically there's no absolute proof of it any more than there's an absolute proof about
• 21:36
  God or anything else really it's a matter of faith it's a matter of how you think about life um so
• 21:42
  if I argue that these things are tools not creatures uh which is a distinction we sometimes make um
• 21:48
  I can't do it purely on the basis of science or some test or some engineering criteria I have to
• 21:54
  ultimately say it's a matter of belief but it's a matter of belief with pragmatic benefits if you
• 22:00
  think that the technology itself is maybe not a human but is some kind of entity then you kind
• 22:08
  of lose the clarity that you're designing technology to benefit these other entities these humans
• 22:13
  like there's a danger that you start to design the technology for its own sake and then it spirals
• 22:18
  into some weirdness but also um you can start like let's remember AI is always trained on data
• 22:24
  from people and so if you pretend the people aren't there which you kind of have to to treat it
• 22:29
  as entity instead of tool then you can start to economically displace people who are actually
• 22:35
  contributing still and I think actually people will contribute a great deal to AI in the future
• 22:40
  because I think we'll have all kinds of new weird uses for AI that require any training data I
• 22:44
  really believe that I'm an optimist about the variety and interestinglyness of the future uh
• 22:50
  but if you really believe that the AI is this freestanding thing that's different from people you
• 22:54
  might have a tendency to not compensate those people or recognize them and create a wave of people
• 23:00
  feeling that they're obsolete even if they are still contributing and if it's false uh and then um
• 23:06
  you know there's just there's there's a danger of political manipulation so like
• 23:13
  just let me put this very bluntly if you believe that something from up an AI or Microsoft
• 23:20
  is alive and you treat it as this living thing you're basically sending me money and I want you
• 23:27
  to send me money because I'm helping you be more productive but I don't want you to send me money
• 23:31
  because I've made your fake girlfriend which is bullshit I hope you don't mind oh you can
• 23:35
  leave me or something I'm sorry about that you can swear but you know like the thing is um the way
• 23:41
  AI is portrayed in the movies is that it's just sort of there like yeast it's just in the air like
• 23:47
  somehow the matrix happened and somehow the the lady AI in her just kind of appeared but that's
• 23:54
  never true there's some company behind it because the computerally expensive so for a lot of it it's
• 23:59
  us and like why should you be falling in love with us I mean like it's stupid like you have to
• 24:03
  also behind the veil and say who's running the AI what's the business model what's actually going
• 24:08
  on here and if you really want to treat it as this living as being like this entity you give entities
• 24:14
  empathy and you give them uh some kind of space you give them some respect and that makes you lose
• 24:21
  your skepticism or your curiosity to see what's behind because there's no behind the curtain for
• 24:25
  us humans but there is for these things you know and so I worry that treating them as alive will hurt
• 24:32
  will harm your skepticism and it's very precious that you you know it's urgent that you keep it
• 24:38
  okay so I hear yeah yeah I hear yeah I mean you look your your strain of humanism is my favorite
• 24:44
  thing about you I loved you are not a widget and I loved uh was it 10 arguments for getting off
• 24:49
  social media um also like that I think that the titles are different Australia all right okay well
• 24:54
  what what I mean you know you're you're kind of you're cool for people to reclaim their own
• 24:59
  humanity in the face of machines that force us to behave a certain way and forces to comply with
• 25:06
  a digital setup setup of the world I think one of the things that interest me about your optimism
• 25:13
  about AI is that AI might actually make the interface the digital interface more humane might make
• 25:19
  it less a scenario of us having to comply with uh you know computer says no and allow computers to
• 25:27
  become more nimble more tailor-made to our nature instead of the other way around for some reason
• 25:33
  you've insisted on using a sucky podcast platform for us to record this and so it took 17 minutes to
• 25:39
  just get it to start working and I would like us to enter an era where the use of a because this
• 25:45
  has been this is a great example this has been done tens of thousands maybe a millions of times
• 25:50
  before therefore a trained AI program ought to have been able to get us to get this thing working in
• 25:56
  five seconds instead of 17 minutes so there's this is a case where there's missing AI yeah I should
• 26:02
  AI should have been present to get this to work it's a great example of how you can make life
• 26:08
  better for people why did we have to go through this again I hear yeah let me let me then concede that
• 26:14
  it's useless to think about AI as being like a freestanding entity that's unhinged from the
• 26:20
  rest of humanity because all of the training data was was relies on all the things that you
• 26:25
  may have done I just want to say yeah um that you've made a statement that's a little more
• 26:31
  a little stronger and more judgmental than the way I put it and the reason why is that there are
• 26:37
  cases which I think are not that common but there people come up and say you know I'm dealing
• 26:43
  with a patient population that was traumatized or that has this or that problem and interacting
• 26:49
  with an AI might act or they have their they have some cognitive unusual quality and they
• 26:59
  interacting with an AI might actually really be helpful for them and I don't I don't want to
• 27:03
  like declare that to always be horrible like I'm I said what I said but I also am not like an
• 27:09
  ultimate you know fundamentalist about it got it got it yeah yeah no I mean I'm I'm pushing on
• 27:16
  a perhaps a slightly different point which is computers have been getting better and better at
• 27:21
  solving problems that humans find hard starting with you know we all started paying attention when
• 27:27
  they were better at us on in chess and so on and so the question is does the range of problems that
• 27:33
  humans find hard to solve expand and do computers abilities to integrate information have a
• 27:40
  perspective that's larger than our own and come up with novel solutions to things that are
• 27:47
  superior than the ones that we can come up without computers get to such a point at which
• 27:52
  it's sort of meaningless to ask whether they're not intelligent because they're acting in the world
• 27:56
  in a way that isn't practically more intelligent than us and doesn't that have implications
• 28:01
  you know I mean your question has a certain kind of answer built in because your question
• 28:07
  presumes all kinds of things that I'd reject totally so I think it's a terrible question like
• 28:13
  like what what are the assumptions well for one thing if somebody can accomplish something using
• 28:18
  one of these programs that they couldn't have otherwise what it's done is the reason in my view
• 28:23
  is that it's increased the ability of them to collaborate with many others from the past and
• 28:30
  from distant places to work together more effectively I view it as a form of collaboration rather than
• 28:35
  as an entity so I would celebrate it and I don't I don't see any problem at all like if somebody can
• 28:42
  come up with some new drug molecule it's because they've leveraged the work of many many people
• 28:48
  who contributed data and effort to the program they're using so it's a massive collaboration that
• 28:56
  couldn't have happened before but I think isn't there a sorry backup isn't there a false binary here
• 29:01
  between human intelligence as being this thing that is coming out of our own neurons in a almost
• 29:10
  a spooky way versus the parasitic intelligence of an artificial system which as you say is really
• 29:17
  just a collaboration of all of the input data of all of the the the people who put this to me
• 29:21
  but show me show me the human being who was also I'm not interrupting I was talking
• 29:29
  I started from the time my apologies yeah no so just to finish that thought I mean where is the
• 29:35
  human mind that is insulated from all of the other human you know products that have gone into it
• 29:40
  when I'm coming up with this sentence this question to ask you it's informed by all of the things
• 29:45
  that I've read of yours and that other people have told me about AI and my own experience with AI
• 29:50
  I mean I'm operating on a set of training data I wasn't plugged down here from from heaven without
• 29:56
  any preconceptions aren't all intelligences integrating information and then spitting out the
• 30:01
  results of that integration before I answer I just want you to be aware of how your questions
• 30:07
  are biased so when you say is it just a collaboration where I think this this new kind of collaboration
• 30:13
  is spectacular and almost sacred and so I reject the word just which is a way of coloring the whole
• 30:19
  thing but anyway do you remember I just went through that whole thing about how it's a matter of
• 30:23
  say I remember yeah okay so yeah people are magical supernatural beings that are a binary
• 30:31
  coming from angels in heaven it works for me and you might say oh isn't that superstitious
• 30:36
  whatever I actually think that thinking of it that way makes the job of technologists infinitely
• 30:42
  clearer because we define who our beneficiary is without which we can't even talk about technology
• 30:47
  and it also like if you wanted to find a logical scientific system without any cracks good luck with
• 30:55
  that I I've worked at fundamentals of physics and mathematics and different phases in my career
• 31:01
  and there's always all these weird little creases and cracks so what I'm doing is just positioning
• 31:06
  the creases and cracks for the benefit of people so yeah absolutely supernatural magical from
• 31:13
  the angels different I'm totally with good with that because I think it's the best way of
• 31:18
  thinking that we have available on a pragmatic basis and if a computer becomes capable of
• 31:25
  interacting in a way that's indistinguishable from a human mind that doesn't require us to then
• 31:29
  think rethink our own intelligence let me let me fix your question for you if a new form of
• 31:36
  collaboration between people works so well that it's functioning better than any individual person
• 31:43
  could on their own does that reduce individual people no it celebrates them but I'm not arguing
• 31:51
  that it would reduce individual people isn't the question whether or not the whether or not an
• 31:56
  artificial system is capable of replicating the kind of intelligence that we normally only associate
• 32:01
  with biological entities I think you're I think your question is logically flawed and unanswerable
• 32:06
  because it's in coherent but you need to point out why it's in go here and don't you because there's
• 32:11
  no AI there's only collaborations of people or at least I mean collaborate it seems sent feels like
• 32:16
  a bit of a fudge down like collaborations of people sure there are billions trillions of of pieces
• 32:22
  of information that human beings have created which we fed into these machines which they are
• 32:28
  synthesizing and then doing tasks for us and chatting with us in ways that seem completely
• 32:35
  detached from the inputs to the layperson right I mean maybe not to a person understands what's
• 32:39
  going on with the hood but you even said earlier that we're not sure what's going on under the hood
• 32:44
  it's just producing these these trees and we're not entirely sure how it's doing it so the the
• 32:49
  output ends up being I'm not arguing for us to revere these things as mystical spiritual entities
• 32:55
  I'm just arguing for a coherent understanding of what intelligence is so that I can differentiate
• 33:01
  mine from the machines there's a really funny thing here since we're using this stupid platform
• 33:07
  I'm just looking at black but I know you're recording your camera so keep on wondering if I'm like
• 33:11
  picking my nose yeah I would the I would resolve it yeah I mean so no this this is a failure
• 33:21
  of not having enough AI right now as I said okay look I mean if you want to believe that that's
• 33:29
  your faith and you can do it I just think there's no pragmatic benefit to your way of thinking
• 33:37
  and there are pragmatic benefits to my way of thinking because I can design things for you to use
• 33:42
  whereas you're just kind of globally I'd and sending me money so I mean like I win either way
• 33:48
  because if you believe what you believe you send me money so great I mean I'm really believing
• 33:52
  anything though I'm just trying I like most most people in the world who played around with this
• 33:56
  stuff I'm trying to understand what's going on under the hood and the response that all it's doing
• 34:01
  is assisting collaboration between me and the millions of people you say all it's doing okay that
• 34:07
  what it's well that's like saying all that science does or all that the law does or all that
• 34:12
  literature does it's like a demeaning you're demeaning the most important thing like stop that
• 34:17
  that's like really dumb you're better than that so would it be different if what it was doing
• 34:23
  was reasset was mindlessly reassembling information that previous people had input into it versus
• 34:29
  creating something new well as I said before there is a certain kind of newness in filling in
• 34:37
  the gaps between the towers like I like you know I'm trying to I'm trying to get at enough new
• 34:42
  ones here to not oversimplify that's where we started um but on the other hand if you if you want
• 34:50
  to do the absolute oversimplification it's more accurate to say that there's nothing new then
• 34:55
  then to emphasize things that are new that would be closer to that right I mean so that that's
• 34:59
  what I'm saying when I'm saying all like the the the work that the word all it's doing is doing is
• 35:05
  to differentiate between if someone says basically what I'm what what I'm dealing with is an auto
• 35:11
  is a sophisticated auto-complete right it's just like when I'm typing on my phone and the phone
• 35:16
  suggests what word I might be wanting to produce that's basically one large language right that
• 35:24
  that to me justifies what we do yeah that that to me is a less is a is a less impressive feat
• 35:30
  than the things that it seems to be doing when I with my primitive mammalian brain interact with
• 35:36
  it and get my socks blown off by what seemed to be some assaults of ingenuity on its path so when
• 35:42
  I say all it's doing I just mean the subjective experience of me interacting with a calculator is
• 35:47
  like well all it's doing is crunching numbers that I would not apply the word all to a machine
• 35:53
  that makes me laugh because it tells me what seems to be an original joke uh well I mean there's
• 36:00
  a couple a few things to say about that I mean um we train on people's jokes and on many things
• 36:09
  and so once again it's not it doing it your it's like you have a writer's room with all these people
• 36:15
  and they made a joke uh and and you you have to get that through your head that that's how it actually
• 36:20
  works and um it can be impressive it can be super helpful um it's just not a creature it's not a
• 36:29
  thing in itself it's a it's a collaboration and I um we have all a lot of a screw up on science
• 36:38
  fiction movies that make it hard to think about that because we have the make you know we have the
• 36:42
  matrix and sky net and all that and come the the much more attractive commander data but we have
• 36:48
  this image of these things as these separate entities um there's a whole interesting social
• 36:55
  history about that which is another story but um once again it's a matter of faith and I don't
• 37:00
  I don't like to argue with people about religion but you obviously have this thing where you want to
• 37:04
  perceive them um I will let you know I'm old enough to have been around when things like pocket
• 37:09
  calculators were fresh and people were like crazy and pressed with those two not in all that
• 37:14
  to similar way actually um and I don't I don't know I mean it's like if if you enjoy the experience
• 37:24
  of being impressed in that way I don't want to deny it to you but just like keep that keep a little
• 37:29
  skepticism in the back of your mind just for your own power for your own you know you don't want to
• 37:36
  be infeal by the theatrics we might present to you um you know just like you know um think for
• 37:44
  yourself a little here you know no I don't yeah I get that I get that I'm I worry that it'll
• 37:49
  become increasingly difficult I worry that that quarantining biological intelligence from
• 37:56
  artificial intelligence in a practical sense in five ten twenty years will become very challenging
• 38:05
  that's a possible scenario but the thing that would make it challenging is not the code but the
• 38:11
  societal uh legal business and and just uh or maybe our biological biases Jon I mean maybe
• 38:19
  maybe the the fact that we uh we are linguistic mammals and that we're easily manipulated by
• 38:25
  the illusion of intelligence when it comes from through length I mean that that's a matter that
• 38:30
  does concern me and I think it's very worth being worried about that I think um I'll tell you
• 38:37
  something I've observed over time which is that um anytime there's some movement that proposes
• 38:44
  a way of thinking about the world that's initially fresh and radical whether it's political or religious
• 38:50
  or sexual or whatever um it proceeds for a while and it gets more and more intellectual and more
• 38:56
  and more abstract and when things kite when radicalness sort of becomes old there's this weird
• 39:04
  moment of a vacuum where the thing that fills it in is the oldest forms of power of just
• 39:11
  intimidation and violence and all that and so you tend to see fancy digital things gradually
• 39:16
  associating themselves with populism and all that now I hope I'm wrong about that but I I'll just
• 39:22
  note that you know a lot of this is weird thing we're a lot of people who get into weird digital
• 39:28
  schemes that are supposed to save anything like crypto or something a few years later they seem to
• 39:34
  be more open to um odd macho populist politics and I think it's just because things get so abstract
• 39:43
  that that's what remains at the end of it all and so the thing I'm worried about um treating AI
• 39:49
  as real is that it'll become this weird hollow mirrors where people will start making fake
• 39:54
  personages to fool each other and the whole thing will stir up into a dizzying jumble and then at
• 39:59
  the end of that what's left that is real is the most retro forms of power uh and the most retro
• 40:07
  primitive motivations so we have this way of undoing our own sophistication by letting it get to
• 40:14
  unhinged uh so that's a scenario I worry about which I think we can see hints of already
• 40:19
  well said yeah I completely share that concern and the the other concern would just be if AI
• 40:27
  took all it you know just to call it that if that super empowers individuals to solve problems
• 40:33
  much better than they were able to previously so that I can produce biological weapons or I can
• 40:41
  produce computer viruses or I can you know you name it there's a what is a world of super empowered
• 40:48
  individuals look like it probably looks like a return to the most primitive thuggery because
• 40:57
  ultimately then you end up with whoever can uh whoever can impose fear upon enough others becomes
• 41:07
  a chief done or something I mean I think I think it becomes a very primitive situation um
• 41:13
  you know I think I think that's where we land um by the way there's another there's another
• 41:20
  element to this which is they um I am not in an extreme minority at all within the technical
• 41:28
  community that works on this stuff but I mean I'm in a minority I mean I don't know exactly where I
• 41:34
  stand but I would say at this point maybe something like a quarter of my colleagues agree with me
• 41:40
  and must do not and there's a lot that maybe you're in between or something um and the thing is
• 41:45
  for those who want to perceive that we're making new creatures who like personifying them who like
• 41:51
  them to be personalities you talk to instead of tools you use um there's a reticence to explain
• 42:02
  them beyond a certain point because they like the mystery of it like you were talking about
• 42:06
  I'm how I might be assigning some kind of ineffable mystery to people but some AI scientists do exactly
• 42:12
  that to AI and let me give you an example of that you just brought up the specter of like what if
• 42:17
  somebody could use a fancy AI system to help them make a doomsday virus or weapon of some kind
• 42:24
  very worrying scenario so how would you defeat that well we do a lot of things now to to make that
• 42:30
  unlikely extremely unlikely nothing like that's happened yet um I think we're doing a reasonable
• 42:37
  job at that um so but ultimately like what I'm saying is technically correct that all these large
• 42:46
  models are ultimately trained on data from people uh I mean there's synthetic data in there but
• 42:52
  even that is an extension of the data from people and so if we could trace which people's data was
• 43:00
  the most influential in a given AI output we could trap all the sneaky people trying to use it for
• 43:07
  thing because right now if you just ask for a deadly virus design it'll stop you because um to my
• 43:14
  knowledge all the people doing big models would trap for that and have done a lot of red teaming
• 43:19
  or adversarial simulation to try to catch those cases and there's also a very careful removal of
• 43:25
  things from the training data but because it can fill in the gaps between towers it might be
• 43:29
  able to recreate such things anyway so you have to do both you have to do everything you can but
• 43:35
  what if you could trace intent and so somebody might say oh all I'm doing is I want cake recipes
• 43:41
  for my grandmother's birthday but they might have a sneaky way of prompting so that they're also
• 43:45
  getting information on how to make a bomb or something we've seen uh fake adversaries get away
• 43:52
  with things like that in order to help our stuff be more robust but anyway if you could trace
• 43:57
  the individual source documents that were in use for a given output you would have immediately
• 44:03
  revealed what the true intent is like you would say oh god this prompt actually invoked this bomb
• 44:08
  thing or this virus thing okay okay these people are clever but they're not clever enough
• 44:13
  when we have an x-ray into what's actually happening right now I think part of the reason we
• 44:18
  don't have that x-ray well to be honest it's technically difficult and and not free to do but
• 44:23
  there's also an ideological issue there um in which people are mystifying the AI they want to
• 44:29
  cover up this the original human sources and make it seem more mysterious than it needs to be
• 44:35
  that's also true for the internet by the way which was given a needless oracle quality by forgetting
• 44:41
  where things came from and making it all a bit more more mysterious which has turned into a
• 44:45
  security and economic nightmare but let's leave that aside there's there's been this thing in
• 44:50
  technical culture even from before AI of trying to mystify the technology a bit and not trace
• 44:56
  where things came from in order to have this mysterious oracle entity illusion come out but it's
• 45:02
  dangerous and one of one of the reasons that I agitate against it is precisely so that we can do
• 45:09
  a better job at safety and helping people know what's actually going on I mean I think we need that
• 45:14
  right and well one of the things that you're up to at the moment is agitating for these systems
• 45:17
  to become less of a black box so that if you are able to hoodwink the layperson into asking for a
• 45:23
  birthday cake recipe which actually has embedded in it some request for bomb making recipe and even
• 45:31
  if the output seems innocuous and looks like a birthday cake recipe we need to be able to appear
• 45:36
  into the middle point of what the black box is actually doing so that as it's crunching the the data
• 45:42
  we can see that what's being fed in is actually some bomb making recipe and not exclusively a
• 45:47
  wedding cake recipe this ties in neatly to your to your call for what you call data dignity which
• 45:54
  is essentially well do you want to explain what data dignity yeah data dignity is just that as technology
• 46:01
  shifts and changes when somebody does make a contribution whatever the new situation is we find
• 46:09
  some way that they can be recognized and paid instead of just lost and forgotten so here I'll
• 46:16
  give you an example let's suppose in the future we have some marvelous anti-gravity technology we can
• 46:23
  all fly up into space and we have some weird biomod where we can survive in the vacuum and what we
• 46:28
  want to do is we want to grow all kinds of weird new limbs in order to be able to manipulate this new
• 46:34
  world and fly around asteroids and do things is that science fiction enough for you yeah like uh
• 46:39
  and so and that's a plausible feature I mean the anti-gravity part eludes us but you know maybe
• 46:46
  it's only one idea away who knows anyway um the point is in that case a tricky tech company might
• 46:52
  get a bunch of people to voluntarily enter in uh designs for these new limbs sort of train an AI
• 46:59
  that then can make many more of them um so what I want to see in the future is instead of creating more
• 47:06
  dependent classes of people who are contributing but aren't paid uh I want to see more creative
• 47:12
  classes where there are people who are contributing in creative ways and are paid now uh some of my
• 47:19
  the people who seem the most I don't know science fiction in AI actually have this weirdly closed
• 47:26
  view of the future where they think that the AI we've already trained dissolve all over need and
• 47:29
  it'll do the future I don't think that's true at all I view this the future as being very open
• 47:35
  and filled with new surprising things and I think there'll be a lot of room for new classes of creative
• 47:41
  people to do new things that create new new new AIs and that that'll just continue um this thing
• 47:48
  of creating this idea that you create classes of dependent people and support everybody with the
• 47:54
  universal income scheme or something is a recipe for some kind of centralized control and once again
• 48:01
  a way that you have incredible political regression to the most primitive state in my view I think
• 48:06
  the way to have decency in society is to always recognize people and reward them and give them
• 48:15
  some degree of choice and autonomy to the degree that's possible which is not infinite but you know
• 48:21
  like you should always create a creative class instead of a dependent one so we've called um
• 48:27
  weirdly such an adela came up with um data dignity as a as a moniker for it which is kind of cool
• 48:34
  but um I would say it's kind of a minority position but more and more people are interested in
• 48:38
  it throughout the industry and uh I just want to keep on advocating for it as an ideal
• 48:45
  I love the idea of breaking open the black box identifying who's human made data is going into
• 48:53
  the artificial intelligence and is responsible for its output and then compensating those people
• 48:58
  in some way the response likely is going to be well we can't we can't untangle that that pile
• 49:05
  of spaghetti like there's so much stuff going into it we've got millions or but how many of the
• 49:09
  billions of images of cats that have gone in whose image of a cat gets priority to get compensated
• 49:15
  some new I mean we retrain these things you know approximately every year and a half for something
• 49:22
  so we have many opportunities to go back and start doing it with the big models and if somebody says
• 49:28
  it's technically impossible um they're not very good at what they do I mean it like honestly um
• 49:36
  I this is I'm good at this I know that it's possible I think it's hard um I think a much
• 49:42
  much more serious question would be that it would be expensive and there
• 49:46
  that criticism has a like to stand on because it's already insanely expensive to compile each
• 49:53
  year's new big model and um a key thing to doing data dignity is having a way to do it efficiently
• 50:00
  efficiently in every sense and there I think we still need more research and innovation and
• 50:04
  that would be a very legitimate criticism but I've never seen an argument that indicates that it
• 50:09
  can't be done efficiently enough I mean I uh it's it's just a question of how soon we get there
• 50:15
  but that's a real challenge and I would I would not I would not deny that and Jared there's there's
• 50:21
  an apocalyptic class of AI analysts who say that as these systems get increasingly sophisticated
• 50:28
  they're going to run ahead of us in problem solving and that we're going to be struggling to
• 50:33
  keep up with the solutions that they can come up with that if a an artificial system can do as much
• 50:39
  problem solving as a human mind could do in 10,000 years in one second then it will see opportunities
• 50:45
  and it will be able to resolve things in a way that bamboozle and baffle us and that therefore
• 50:52
  we're we're sort of giving birth to systems that we're not really going to be able to control
• 50:56
  because we're not going to be able to fully understand what they're doing what do you say to that?
• 51:02
  Well I want to say two things um first I want to make a sociological observation that many of the
• 51:08
  people who say that are also the ones doing it it's it's a really weird thing like I'll go to
• 51:14
  one of these off the record things where the leaders and AI are getting together and the people who
• 51:19
  are uh sometimes called the doomers are also the ones working different companies pushing it
• 51:24
  and they say well how do you reconcile it and it's a it's weird like some of them say well if
• 51:32
  it's going to happen our team should be the first to get there because we'll be less terrible
• 51:36
  than those other people or sometimes they say you know humans have had their chance let's go to
• 51:41
  the next species or sometimes they'll just say it's destiny it's the rapture it's the it's our
• 51:49
  it's where it's like a mystical destination that they feel compelled to and they feel it's
• 51:55
  beyond them to even judge it or something and I think all that to be um kind of horrible like I
• 52:03
  think like what we should try to do is demystify and then take responsibility based on that that's
• 52:08
  is it not analogous to something like the Manhattan Project where it's like yeah we're giving birth
• 52:12
  to this horrible thing but other people are going to do it too so we'd rather have it yeah I mean
• 52:16
  that's an analogy that's often brought up but I want to suggest a different analogy um which is
• 52:22
  to capitalism okay so capitalism in a sense was the first AI it's this information system where we
• 52:31
  combine the efforts and actions of many people into this complicated thing that does things that
• 52:36
  are a little beyond us you know like the way a market can organize the world and get things to
• 52:41
  happen is beyond what a command economy can do then it's been demonstrated over and over again uh
• 52:49
  and so um and yeah and and by the way it was also personified Adam Smith called it the invisible
• 52:55
  hand so it also was treated as a creature by some and then it also is a bit mysterious in its ways
• 53:02
  and the people who regulate them do so with empirical techniques rather than perfect theoretical
• 53:08
  models um and one of the ways I talk about data dignity is trying to become Keynesian about computer
• 53:15
  science for those who will recognize that but anyway let's let's um let's leave that aside now if
• 53:22
  you talk to people who are deeply skeptical and fearful of capitalism their critique of it
• 53:29
  sounds very very similar to AI dumus you know like they'll say um uh you guys are replacing people
• 53:37
  because in the United States corporations are legally people which I think is ridiculous for
• 53:42
  the same reasons I said all that about AI programs um and but anyway that is our law now you know
• 53:49
  and that's true in many ways around the world and we're we're the big dog in the world of capitalism
• 53:54
  so we have an influence on everybody else that way um but then people are worried that capitalism
• 54:00
  has this way of possibly destroying humanity by enforcing these incredibly short-sighted
• 54:10
  incentives that also tend to concentrate power and influence on an ever smaller number of people
• 54:15
  who then are there's a vicious cycle in which those people become crazier and crazier
• 54:20
  from being cut off from reality and then the whole thing falls apart it's a very familiar critique
• 54:25
  and there's definitely something to it and to me that's the same thing really as AI dum debate um
• 54:34
  capitalism is kind of the early primitive prototype of AI in my view it's a way of people
• 54:39
  cooperating that actually does do new things that are really important it's really powerful
• 54:45
  and yet it can also be given too much birth and be given go out of control and do great damage
• 54:52
  and markets can fail and there can be a weird political side effect where it can tend to benefit
• 54:59
  some people to a ridiculous degree so the way um what we call in math network effects
• 55:05
  can tend to concentrate capital beyond reason are similar to the way only a small number of
• 55:12
  companies can compete the big AI models so there's a concentration of power and influence
• 55:16
  it's maybe a little beyond reason um there's a lot of similarity there and yet and yet and yet
• 55:24
  I believe that markets are good enough as tools for us to have them and I believe that
• 55:30
  Keynesian regulation can work and I believe that um civil society and politics can balance
• 55:36
  the problems with capitalism and I hope for that uh some of my leftist friends think I'm naive
• 55:42
  but essentially that's the spirit in which I also wish to approach AI um and I think the solutions
• 55:49
  are not unrelated you know I think in a sense the way to fix AI for the future is very similar to
• 55:56
  how to make capitalism be a beneficial tool instead of our destruction uh in both cases it's kind
• 56:02
  of dramatic and it might not go well and yet in both cases um I really do think we all like a lot
• 56:09
  of the benefits of modernity and these big systems are what make modernity possible and I think we
• 56:15
  really mostly want to make this work and I really believe that we can so in economics in Keynesianism
• 56:22
  you would iron out the peaks and troughs by adjusting the monetary supply what's the analytic
• 56:28
  in computer science um you would say uh if there's some amazing new wealth created by AI
• 56:41
  you would make sure that it's shared with the people who did the most to add data code
• 56:49
  infrastructure all the different aspects that made the AI possible so if there's some new hit movie
• 56:55
  that comes out of a generative AI program all the people upstream from that who contributed clips
• 57:01
  that the thing was trained on should be should get credit and should it's like a new way that they
• 57:07
  become a part of it right right I see but aren't the doom sayers aren't they fearing something worse
• 57:12
  than just a free rider problem they have fear absolutely and and and and there are but the thing is
• 57:20
  I think the the free ride problem is the precedent for the bad doom problems you know because that's
• 57:26
  where you start to create the out of control self-reinforcing concentration of power and the black
• 57:32
  boxes like if you defeat the black box and you have more people participating you make it less likely
• 57:39
  that the rogue person will program the thing to destroy the world I mean I think they go together
• 57:44
  very well what if it's not a rogue person destroying the world but you take the doom's dayers at
• 57:50
  at their worst fears and it's simply that the system itself I mean I it's I struggle not to
• 57:56
  anthropomorphize it which I will tell you but the system itself is you can't avoid anthropomorphizing
• 58:02
  it if there's going to be a sky net with little drones that run around killing everybody there
• 58:08
  has to be within the generative program that operates the robots that build the factories and
• 58:14
  mine the goods that then create those those sky netty drones tracing it back there's something
• 58:22
  about how to kill people I mean like it's revealed and so but what if it's I mean what if it's
• 58:27
  something more innocuous like it's just the unintended consequences of the system achieving goals
• 58:32
  that we've given it we tell it to solve climate change and it realizes that one way to do that is
• 58:37
  to de amplify my voice I don't know why I would think that but let and therefore you know my
• 58:42
  facial mediation deprive our time yeah we call this a paperclip scenario you know you tell
• 58:48
  it to make more paper clips then it does nothing but that and destroys everything else on earth
• 58:52
  for the sake of paper clips well you know to the degree the thing is designed not as a black box
• 59:01
  but revealed to the degree it's designed as a tool and not a creature so it doesn't have
• 59:06
  autonomy to the degree there's sort of a constant audit because somebody might want to be owed
• 59:13
  money instead of a black box operation the paperclip scenario becomes vastly reduced the only
• 59:19
  alternative to all that stuff is to have AI's looking at AI's which is the current fashion where you
• 59:25
  have one AI that's trained to recognize whether another AI might destroy humanity for making
• 59:31
  for the sake of paper clips or whatever and the problem with that infinite regress is that it's
• 59:36
  it's not grounded and you could have a systemic error in it and you couldn't even tell so I think
• 59:42
  this idea of undoing the black box having a lot of accountability to people having a lot of
• 59:50
  participation opens it up and makes that sort of thing less likely right so if we keep them as
• 59:55
  black boxes they could go off the rails well yeah I mean it's a little bit like in any kind of
• 1:00:02
  governance a dictator starts a war more easily than a democracy that doesn't mean that democracies
• 1:00:08
  can't and don't but it's just easier for it's easier for you know rush it to attract you crane
• 1:00:15
  when it has an autocrat in power yeah and and so it's a very similar thing but it's just for the
• 1:00:20
  governance of technology so once again you know I think this I the particular spiritual choice
• 1:00:27
  of treating them as creatures instead of tools does make it harder to reduce risks and
• 1:00:34
  Sharon lastly I want to get your thoughts about how we're going to interact with with artificial
• 1:00:40
  systems like just in the physical world because 40 years ago when when you were a pioneer on virtual
• 1:00:46
  reality if you fast forwarded to today would you be disappointed or surprised or intrigued by
• 1:00:53
  the way that we're facing with computers I mean not really surprised because the person who did
• 1:01:00
  more than anyone else to create the perspective on AI that I disagree with as Marvin Minsky and he
• 1:01:05
  was my main mentor when I was a kid so I've been arguing this forever and it was very clear that
• 1:01:12
  that perspective had a certain kind of appeal within the tech community and within a lot of
• 1:01:18
  the types of people who excel at math and computer science and for whatever reason and so in a way
• 1:01:28
  I hate to say this but this is all kind of what I was expecting I think even in terms of
• 1:01:33
  even in terms of us sitting in front of screens and using mouse's and keyboards like would you
• 1:01:39
  if in the 80s I'd told you that in the 2020s no one would be using virtual reality except for
• 1:01:44
  on a lock would you what would you have made of it well you know that's a that's a really
• 1:01:50
  interesting question so I have always argued from the earliest days of virtual reality which were
• 1:01:57
  for me in the 70s that it wasn't destined to be something people would use all the time or
• 1:02:05
  spend hours and hours in but it was more a special thing I used to say as a business it'll be more
• 1:02:10
  like fine wine than like pop soda it's not going to be this universal ubiquitous always thing it'll
• 1:02:16
  be the special thing that'll be very quality driven and probably there'll be an endless
• 1:02:21
  quality race so it'll never really get cheap and it'll always be this special delectable thing and
• 1:02:27
  I I still think that's the best way to think about it and I think eventually the industry
• 1:02:34
  will land there because I think it's just a good way to think about it I think it makes a lot of
• 1:02:39
  sense and it's also just a cooler you know having experienced a lot in different ways I like using
• 1:02:45
  it that way it keeps a specialness I still love virtual reality after all these years I think it's
• 1:02:50
  great but it should be a really special little art thing I don't use it all the time still you know
• 1:02:57
  even though it's gotten better you know by most measures I still like it as a as a special thing
• 1:03:03
  instead of as an ordinary thing it's it just has a wonderful identity as a special thing what's
• 1:03:08
  the best thing about it's specialness well um here I'll expose myself as the humanist spiritual
• 1:03:18
  softie that I'm accused of being all the time one thing I love about it is that when you take off
• 1:03:25
  the headset you notice the real world better it's like the first palette freshener you've had since
• 1:03:30
  you were born really and so you notice just the textures and just the physicality the way the
• 1:03:36
  air moves you notice like details or at least that's what I find I think a lot of people don't
• 1:03:41
  these days because they're talk to use it in a different way but if you let yourself be sensitive
• 1:03:47
  to that moment when you come out it's really amazing but then there's this other thing which I
• 1:03:51
  think is even more profound so one of the things that I've always wanted to do with VR is uh make it
• 1:03:58
  as intensely strange as possible because why wouldn't you want that right and so um early on I
• 1:04:03
  explored uh turning into different animals and trying to discover which alternate bodies your brain
• 1:04:09
  could control well which is a fascinating study um because it explores the past the brain evolved
• 1:04:16
  through in deep time um it's it's it's really wonderful but anyway you can turn into all kinds of
• 1:04:22
  creatures but then you can also push that even further and you can have the sensation of being in a
• 1:04:28
  body that isn't simply connected but instead it's kind of distributed and then and this is by
• 1:04:33
  measuring your real body and mapping it to different features in the virtual world you can experience
• 1:04:37
  a shared body you can do all these crazy things and um what's the best thing about all that is
• 1:04:45
  throughout it all no matter how weird it gets there's this little nub there's this little thing that
• 1:04:51
  continues through through through through through and that's your consciousness what I think of
• 1:04:55
  our reality is a consciousness noticing system which is a very mystical thing to say and it
• 1:05:01
  and you could say I'm just kidding myself and I can't prove you're wrong honestly but um almost
• 1:05:07
  everything with technology and science gravitates away from that uh AI certainly does very obviously
• 1:05:13
  but here's a case where a technical thing actually brings you to it because if not for your
• 1:05:20
  consciousness what's going on you know like there's this thing there's this thing there there's
• 1:05:26
  something there John it's sorry I finished the thought no I mean it's like and once you start seeing
• 1:05:32
  that you start saying well wait why is this moment in time why is the present moment different
• 1:05:37
  from other elements why am I here instead of there and you realize that the whole way the world
• 1:05:42
  is put together implies this little number of consciousness anyway but with virtual reality you
• 1:05:47
  start to notice it and I love that about it it can't have escaped your attention that you could
• 1:05:53
  say everything that you just said and you could swap out the words virtual reality for the word LSD
• 1:05:59
  and it would still make sense maybe I mean I used to know both Albert Hoffman who invented it and
• 1:06:04
  Timothy Leary very well that he dedicated a book to me and I but I never took it so I don't
• 1:06:10
  I can't speak to it directly uh Tim Leary used to call me the control group um I just had with
• 1:06:15
  I just had this intuition that it wasn't for me really uh I I think it's probably a different
• 1:06:22
  thing but uh many might disagree and that's super you know I I don't know I don't know really
• 1:06:28
  I still have the intuition that it's not for me um I'm I'm kind of a straight person I guess in
• 1:06:35
  some ways I mean who needs it when you've got virtual reality by the sounds of it uh I need to get
• 1:06:40
  I need to get on more virtual reality clearly um I I don't know I mean I'm not I'm not here to judge
• 1:06:47
  anybody else I always enjoyed the psychedelic community I knew many of them and had a lot of
• 1:06:52
  experiences in that world but um uh not for me you know and I think that's okay and if somebody
• 1:06:59
  for somebody who comes up to me and says you know I'm just really not into virtual reality that's
• 1:07:03
  fine I actually I love human variety I love it when people can sense who they are and it's distinct
• 1:07:11
  from who other people are I think that's all really essential and um it's you know it's tough
• 1:07:18
  Jared and I have six-year-old kids how uh how optimistic or pessimistic should I be about the
• 1:07:23
  world in 30 years time what's your general if you put on your prognosticate as hats what's your general
• 1:07:28
  yeah um that's um that's actually I think the most important question to ask right now because
• 1:07:34
  there are so many trends that um would certainly appear to be quite dismal happening at the moment
• 1:07:43
  aren't there um and uh the obvious one being climate but there's really a cluster of things
• 1:07:50
  adjacent to climate as well about you know fresh water and just oh so many things that are worth
• 1:07:55
  worrying about um and politically the world seems to be sort of almost like stubbornly determined
• 1:08:05
  to be creamy I I mean it's just it's awful and a lot of that is social media but that's not all of it
• 1:08:10
  and uh uh I I don't know I mean I I have a kid as well older than yours but I I worry about
• 1:08:21
  I worry about that a great great deal I I think um I find hope in that humanity has gone through
• 1:08:31
  challenging and dark and scary periods many times before and somehow whether we've always
• 1:08:37
  found our way through um I think whether or not one is freaked out by the way things seem to be
• 1:08:44
  going um it's clear that our responsibility is to do the best we can obviously so in a way it
• 1:08:52
  doesn't have any impact on what happens next I mean either way our responsibility is to do the
• 1:08:58
  best we can so who gloom no gloom who cares our job is to just try to make it better right and um
• 1:09:05
  I try to stick with that I try to I try to stick with the feeling that I'm I'm doing my best as I
• 1:09:12
  understand it to make things better um and I you know uh I don't think there's any superhero who
• 1:09:23
  will descend from the sky to save us I don't think there any guarantees but I do have just a
• 1:09:29
  whole lot of hope based I think reasonably on the past well John you've contributed more than
• 1:09:34
  your fair share to this mortal coil uh in your time so far I look forward to keeping track of
• 1:09:38
  the uh the next phase of your career thanks so much for your time oh hey thanks so much uh take
• 1:09:44
  care thank you