Making Sense - Sam Harris - #326 — AI & Information Integrity

00:00:00.000 Welcome to the Making Sense Podcast.

00:00:08.880 This is Sam Harris.

00:00:10.940 Just a note to say that if you're hearing this,

00:00:13.140 you are not currently on our subscriber feed

00:00:15.200 and will only be hearing the first part of this conversation.

00:00:18.500 In order to access full episodes of the Making Sense Podcast,

00:00:21.620 you'll need to subscribe at samharris.org.

00:00:24.200 There you'll find our private RSS feed

00:00:26.000 to add to your favorite podcatcher,

00:00:28.020 along with other subscriber-only content.

00:00:30.280 We don't run ads on the podcast,

00:00:32.440 and therefore it's made possible entirely

00:00:33.980 through the support of our subscribers.

00:00:35.760 So if you enjoy what we're doing here,

00:00:37.720 please consider becoming one.

00:00:46.360 Just a brief note to say that my last podcast on RFK Jr.

00:00:51.180 has been jailbroken,

00:00:54.000 which is to say the audio that's now on the free feed

00:00:58.100 is the same as the subscriber audio.

00:01:01.220 There's no longer a paywall.

00:01:03.060 This came in response to some heartfelt appeals

00:01:06.180 that we make that a public service announcement,

00:01:08.500 which we've now done.

00:01:10.060 So feel free to forward that to your friends

00:01:12.280 or anyone who you think should hear it.

00:01:14.640 There have also been several articles

00:01:16.080 that have come out in recent days about RFK

00:01:18.600 that entirely support the points I make there.

00:01:21.600 So nothing to retract or amend,

00:01:25.920 as far as I know.

00:01:28.560 Okay.

00:01:30.300 Today I'm speaking with Nina Schick.

00:01:33.120 Nina has been on the podcast before.

00:01:35.820 Nina is an author and public speaker

00:01:37.640 who wrote the book Deep Fakes.

00:01:40.280 She is an expert on current trends in generative AI.

00:01:43.940 She advises several technology companies

00:01:46.900 and frequently speaks at conferences.

00:01:49.580 She has spoken at the UN and to DARPA

00:01:52.320 and many other organizations.

00:01:54.880 And in addition to generative AI,

00:01:56.580 she's focused on geopolitical risk

00:01:59.400 and the larger problem of state-sponsored disinformation.

00:02:04.040 And today we speak about the challenge

00:02:05.360 of regulating AI,

00:02:07.260 authentication versus the detection of fakes,

00:02:10.500 the problem of fake video in particular,

00:02:13.460 the coming hyper-personalization of information,

00:02:15.620 the good side of AI,

00:02:18.220 productivity gains, etc.

00:02:20.680 We talk about possible disruptions in the labor market,

00:02:23.580 open AI as a company,

00:02:25.160 and other topics.

00:02:27.040 Once again, it was a pleasure to speak to Nina.

00:02:29.560 Unfortunately, in only one of the seven languages she speaks.

00:02:33.500 But for better or worse,

00:02:35.180 English is where I live.

00:02:37.020 And now I bring you Nina Schick.

00:02:38.800 I am here with Nina Schick.

00:02:46.840 Nina, thanks for joining me again.

00:02:48.660 It's great to be back, Sam.

00:02:50.620 So you were on,

00:02:51.700 I forget when you were on,

00:02:52.700 it's a couple of years now.

00:02:54.540 When do you have the year in memory?

00:02:57.020 Yeah, it was in 2020.

00:02:58.620 So just before kind of the new wave of what we're seeing

00:03:03.520 really started to emerge.

00:03:05.520 So, yes, you came on to talk about your book,

00:03:08.160 Deep Fakes,

00:03:09.100 which was all too prescient of our current concerns.

00:03:13.360 But in the meantime,

00:03:14.600 we have this new phenomenon of generative AI,

00:03:17.680 which has only made the problem of deep fakes

00:03:20.640 more profound, I would imagine.

00:03:23.020 And I think that'll be the main topic

00:03:25.900 of today's conversation.

00:03:27.140 But before we jump in,

00:03:28.440 can you just remind people

00:03:29.520 what you've been doing these many years?

00:03:32.080 What have been your areas of focus

00:03:34.060 and what's your background?

00:03:35.940 Sure, absolutely.

00:03:37.200 So I'm half Nepalese and I'm half German.

00:03:40.240 I actually grew up in Kathmandu.

00:03:43.080 But eventually I came to live in the UK.

00:03:46.160 So I'm based in London right now.

00:03:47.460 And I was always really interested in politics

00:03:49.600 and geopolitics.

00:03:50.600 So really for the kind of first two decades of my career,

00:03:54.000 I was working in geopolitics

00:03:56.300 and it just so happened

00:03:57.360 at the end of the 2000s

00:03:59.520 and throughout the 2010s,

00:04:01.520 I happened to work on a lot of kind of

00:04:03.620 seismic events, if you will,

00:04:05.620 from the original annexation of Crimea by Russia

00:04:10.100 to kind of how the ground was laid for Brexit

00:04:13.400 here in the UK to kind of,

00:04:15.720 I don't know if you remember,

00:04:16.660 but the kind of migration crisis in Europe in 2015,

00:04:20.000 which in large part was triggered by

00:04:22.600 the indiscriminate bombing of civilians in Syria

00:04:26.460 by President Putin,

00:04:28.620 basically the weaponization of migration,

00:04:30.300 which then consequently kind of led to,

00:04:32.780 was one of the main reasons

00:04:33.900 that Brexit happened as well.

00:04:35.860 And the kind of persistent feature

00:04:38.240 in my geopolitical career was

00:04:40.500 that technology was emerging

00:04:43.840 as this macro geopolitical force

00:04:46.400 and that it wasn't just shaping geopolitics

00:04:49.260 on a very lofty and high level,

00:04:51.440 but that it was also shaping

00:04:52.960 the individual experience

00:04:54.600 of almost every single person alive.

00:04:57.860 And I saw that again also on the ground,

00:05:00.420 so to speak, in Nepal,

00:05:01.900 where, you know, technology

00:05:02.940 and the internet and smartphones

00:05:04.860 have changed society immeasurably

00:05:06.980 in the past few decades.

00:05:09.100 So I just increasingly started

00:05:10.900 becoming interested in technology

00:05:12.540 as this kind of shaping force for society.

00:05:15.100 And because I had been working

00:05:17.980 on information warfare,

00:05:20.440 disinformation, actual wars,

00:05:22.800 and information integrity,

00:05:24.460 in 2017, I was advising

00:05:26.740 the former NATO Secretary General

00:05:28.700 on emerging technology threats.

00:05:31.420 And he was working in the context

00:05:34.540 of a group of global leaders,

00:05:37.680 which actually included at the time

00:05:39.540 the former VP, Joe Biden,

00:05:41.660 because they're kind of concerned

00:05:42.740 about the 2020 election in the US,

00:05:46.300 given what had happened in 2016.

00:05:49.080 And it was whilst I was working

00:05:51.000 for this kind of group,

00:05:52.860 which would, you know,

00:05:53.880 in part, we're looking at disinformation,

00:05:55.300 but they were also trying to forge

00:05:56.780 and strengthen the transatlantic alliance

00:05:58.800 in the view that only if Europe

00:06:02.140 and the United States are united,

00:06:04.900 can they kind of stand up

00:06:06.500 against the authoritarian forces

00:06:08.440 of Putin and in China, et cetera, et cetera.

00:06:11.320 But that's when I first saw emerging

00:06:13.660 this so-called phenomenon of deep fakes, right?

00:06:16.540 It started emerging at the end of 2017.

00:06:19.400 And I immediately sensed that

00:06:21.640 this was something

00:06:22.420 that would be really, really important

00:06:25.000 because it was the first time

00:06:26.280 we're seeing AI create new data, right?

00:06:30.220 And the first use case,

00:06:32.280 so malicious,

00:06:33.180 was a non-consensual pornography, right?

00:06:35.000 It's as soon as it became possible.

00:06:38.160 So as soon as these kind of research advances

00:06:40.140 started leeching out of the AI research community

00:06:42.840 and enthusiasts started using them on the internet,

00:06:45.460 the first thing they started to do

00:06:46.740 was to generate content,

00:06:49.520 i.e. deep fakes,

00:06:50.440 in the form of non-consensual pornography.

00:06:52.740 And I immediately sensed, though,

00:06:54.720 that this wasn't just kind of

00:06:55.980 just a tawdry women's issue,

00:06:57.720 even though this was undeniably,

00:06:59.540 in this instance,

00:07:00.780 something that was targeted against women.

00:07:02.580 But almost a civil liberties issue,

00:07:05.460 because if you can now clone anybody

00:07:07.960 with the right training data

00:07:09.740 and basically use AI

00:07:11.560 to kind of recreate their biometrics,

00:07:13.600 and this is potentially going to be a huge problem.

00:07:15.240 So that was kind of the seed.

00:07:17.040 That's what led me to write my book

00:07:18.940 on deep fakes and information integrity

00:07:21.000 and the inundation of synthetic content

00:07:23.600 and what that would do

00:07:25.000 in an already very corrupt

00:07:26.340 and corroded information ecosystem.

00:07:29.480 But really what my reflection is now

00:07:32.020 is, you know,

00:07:32.520 that was just really my starting point

00:07:34.300 into the world of AI

00:07:36.140 and generics of AI,

00:07:37.540 because from that point on,

00:07:38.780 when I wrote the book,

00:07:39.540 I kind of stopped working

00:07:41.260 with the global leaders

00:07:42.900 and on the policy side of things,

00:07:44.860 because I was just so fascinated

00:07:46.320 in what was happening

00:07:47.220 in this kind of new developments in AI

00:07:50.560 that I've just been concentrating on that.

00:07:53.220 And although the starting point

00:07:54.800 was mis and disinformation,

00:07:56.700 over the past few years,

00:07:58.400 my reflection has been

00:07:59.420 that it is so much more than that,

00:08:01.560 you know,

00:08:02.020 mis and disinformation

00:08:03.160 is one very important part of the story.

00:08:06.040 But if you think about generative AI

00:08:08.980 as what is becoming clear now,

00:08:11.520 and again,

00:08:11.880 we're only at the very beginning

00:08:12.860 of the journey,

00:08:13.920 I think it's really more profound than that.

00:08:16.140 I think it's almost a tipping point

00:08:17.720 for human society.

00:08:20.060 Well, you have waved your hands

00:08:21.660 in the direction of many emerging problems here.

00:08:24.900 And over and against all of those,

00:08:27.760 there's the question of what to do about it.

00:08:30.580 And, you know, regulation is one of the first words

00:08:34.300 that comes to mind.

00:08:36.240 And yet regulation,

00:08:38.260 speaking from a U.S. point of view,

00:08:41.120 maybe the same is true in the U.K. now politically,

00:08:43.920 but in the U.S.,

00:08:45.880 your regulation is a bad word

00:08:47.680 for at least half of the society.

00:08:49.460 And especially in this area,

00:08:52.520 it seems to be in zero-sum collision

00:08:55.660 with free speech, right?

00:08:57.600 So there are many people who are, you know,

00:08:59.720 center, right of center,

00:09:01.280 who are especially focused on this issue.

00:09:04.980 There's a kind of silencing of dissent.

00:09:08.080 There's an effort on the part of big tech

00:09:11.400 and big corporations generally,

00:09:13.080 the pharmaceutical industry

00:09:14.720 and the, you know,

00:09:16.000 messaging into the public health calamity of COVID,

00:09:21.380 the government adjacent to all of that.

00:09:23.900 There are these elite interests

00:09:26.300 that seem to want to get everyone

00:09:29.320 to converge inevitably prematurely

00:09:32.400 on certain canonical facts,

00:09:35.700 many of which it is feared

00:09:37.120 turn out not to be facts.

00:09:38.940 They turn out to be, you know,

00:09:40.120 politically correct dogmas, taboos,

00:09:43.120 you know, various mind viruses

00:09:44.700 that we don't want to get anchored to,

00:09:48.440 you know, it is argued.

00:09:49.780 And I'm certainly sympathetic

00:09:51.120 with some of that,

00:09:52.500 but like you,

00:09:53.760 I've grown increasingly worried

00:09:55.420 about disinformation, misinformation,

00:09:58.340 information integrity generally.

00:10:00.760 So I'm just wondering

00:10:01.300 what you think about the tension there

00:10:03.340 because we're going to talk

00:10:04.820 about the problem in detail now

00:10:07.040 and some part of a response to it

00:10:11.060 is going to include

00:10:12.220 some form of regulation.

00:10:14.760 Many people at this point,

00:10:16.900 it seems to me,

00:10:17.480 just don't want to even be party

00:10:19.240 to that conversation.

00:10:20.460 The moment you begin going down that path,

00:10:22.900 all of the conspiratorial red flags

00:10:26.820 get, you know,

00:10:27.880 waved or imagined

00:10:29.300 and half the audience thinks

00:10:32.440 that the World Economic Forum

00:10:33.880 and specific malefactors,

00:10:37.440 you know, puppeteers

00:10:38.300 pulling the strings of society

00:10:39.840 will be in control

00:10:41.860 or at least will be struggling

00:10:43.800 to maintain control

00:10:45.000 of our collective mind.

00:10:48.060 So what do you think

00:10:48.740 about the tensions

00:10:49.380 and trade-offs there?

00:10:51.480 Yeah, I mean,

00:10:52.340 I think you've just really

00:10:53.660 hit the nail on the head there

00:10:55.120 when you talk about

00:10:55.820 how difficult

00:10:56.700 what a kind of quagmire this is.

00:10:58.780 And it's difficult

00:10:59.460 for many reasons.

00:11:00.820 Firstly,

00:11:01.140 because when you talk

00:11:03.500 about regulating AI,

00:11:05.280 you know,

00:11:05.860 it's so vast.

00:11:08.080 It's just like talking

00:11:09.040 about regulating society

00:11:11.180 or regulating the economy.

00:11:12.780 So we kind of have to

00:11:13.720 break it down

00:11:14.960 into component parts

00:11:17.000 that are easier,

00:11:17.860 I guess,

00:11:18.280 to conceptualize

00:11:19.400 and understand.

00:11:21.180 And with generative AI

00:11:22.360 in particular,

00:11:23.660 because it's so nascent,

00:11:25.160 I mean,

00:11:25.580 I've been following it

00:11:26.660 almost from kind of day one

00:11:28.320 in terms of the research

00:11:29.400 breakthroughs,

00:11:30.060 which really started

00:11:30.700 emerging in 2014,

00:11:31.940 2015.

00:11:32.840 But I would say

00:11:33.720 that it's really only

00:11:35.440 in the last 12 months

00:11:37.580 that the capability

00:11:40.660 of some of these

00:11:41.660 foundational models,

00:11:43.260 right,

00:11:43.800 and what they can do

00:11:45.020 for data generation

00:11:47.780 in all digital medium,

00:11:50.440 whether it's text,

00:11:51.340 video, audio,

00:11:53.020 every kind of form

00:11:54.120 of information

00:11:54.760 and the implications

00:11:55.740 that has

00:11:56.440 on the kind of future

00:11:58.000 of human creative

00:12:00.040 and intelligent work.

00:12:01.960 I mean,

00:12:02.320 it's so profound

00:12:03.460 that I've really

00:12:04.960 come to see

00:12:05.960 the other side

00:12:06.660 in the sense

00:12:07.660 that it's no longer

00:12:09.240 only about disinformation,

00:12:10.340 but this is also

00:12:11.500 potentially

00:12:12.260 a tremendous

00:12:13.580 economic value add.

00:12:15.640 This is potentially

00:12:16.320 also a huge area

00:12:17.740 for scientific research

00:12:19.840 and insight generation.

00:12:21.140 And you're already

00:12:22.020 starting to see

00:12:22.680 some very interesting

00:12:23.640 use cases emerging

00:12:24.760 in enterprise

00:12:25.620 and in research.

00:12:26.840 So I'm sure

00:12:28.000 we'll get into

00:12:28.560 that later.

00:12:29.580 So when you talk

00:12:31.100 about regulating it,

00:12:33.160 I do have some sympathy

00:12:36.160 for this,

00:12:37.720 if you want to call it

00:12:39.220 the worldview

00:12:40.460 where people are

00:12:42.200 just a little bit

00:12:43.000 sick of politicians,

00:12:45.200 sick of kind of

00:12:45.840 sweeping statements,

00:12:46.780 because already you see

00:12:47.880 the same thing happening

00:12:49.080 with regards to AI,

00:12:50.620 right,

00:12:50.860 where you have a lot

00:12:51.680 of political leaders

00:12:52.800 who perhaps don't have

00:12:54.960 a background

00:12:55.760 in artificial intelligence,

00:12:57.060 but they understand

00:12:58.060 that this is going

00:12:59.980 to be an important factor

00:13:01.680 and they kind of

00:13:02.600 are doing a lot

00:13:03.540 of grandstanding

00:13:04.460 almost saying,

00:13:05.120 well, you know,

00:13:05.660 we're going to build

00:13:06.420 safe AI

00:13:06.920 and we're going to

00:13:07.820 put together

00:13:08.260 a global agency.

00:13:10.200 And without much substance,

00:13:12.400 you can see why people

00:13:13.360 start to get pretty

00:13:14.200 spinnacle.

00:13:15.340 That being said,

00:13:16.120 does this need

00:13:17.040 to be regulated?

00:13:18.940 Absolutely.

00:13:19.900 Because I can't think

00:13:21.020 of a more kind

00:13:21.900 of profound technology,

00:13:24.560 exponential technology

00:13:25.440 that's going to change

00:13:26.300 the entire framework

00:13:27.340 of society,

00:13:28.380 but to start regulating it

00:13:29.620 well, I guess we need

00:13:30.440 to start breaking it down

00:13:32.240 into its constituent parts

00:13:33.680 and that's so difficult

00:13:34.980 because A,

00:13:35.920 it's still nascent,

00:13:36.920 we don't understand

00:13:37.680 the full capabilities

00:13:38.620 of the technology

00:13:39.540 and B,

00:13:41.440 because of the

00:13:43.000 exponential acceleration

00:13:44.440 and adoption.

00:13:45.860 I mean,

00:13:46.240 if you consider,

00:13:47.760 I almost conceive of this,

00:13:49.600 I mean,

00:13:49.800 obviously this is

00:13:50.440 a continuum

00:13:51.000 and it's kind of

00:13:51.720 an exponential curve,

00:13:53.460 but if there was

00:13:54.560 one moment

00:13:55.440 that's completely

00:13:56.280 changed everything,

00:13:57.500 if I had to pinpoint

00:13:58.280 one moment,

00:13:59.140 I would say

00:13:59.680 it is the moment

00:14:00.280 that ChatGPT came out,

00:14:01.640 right?

00:14:01.820 You can almost

00:14:02.400 see the world

00:14:03.860 as pre-ChatGPT

00:14:05.700 and post-ChatGPT,

00:14:07.520 not because,

00:14:08.540 you know,

00:14:09.480 OpenAI was the first

00:14:10.800 to pioneer

00:14:11.380 large language models

00:14:12.760 and Yan Le Kun,

00:14:14.040 you know,

00:14:14.240 the AI chief at Meta

00:14:15.720 kind of famously

00:14:16.500 or infamously

00:14:17.460 came out at the time

00:14:18.440 and was like,

00:14:18.800 it's not that innovative

00:14:19.820 and it got absolutely

00:14:21.220 panned

00:14:21.940 because whilst,

00:14:23.380 you know,

00:14:23.960 he was correct

00:14:24.820 in the sense that

00:14:25.700 they weren't the first

00:14:27.260 to pioneer

00:14:27.900 large language models,

00:14:29.760 it kind of misses the point

00:14:31.540 because that changed

00:14:32.780 the entire debate

00:14:34.600 in terms of both

00:14:35.540 public perception,

00:14:36.740 but also the market moving.

00:14:38.880 So in the past

00:14:39.500 kind of six months,

00:14:40.440 we've seen

00:14:40.880 all of big tech,

00:14:42.940 every single

00:14:43.740 big tech company

00:14:44.780 fundamentally

00:14:45.800 and strategically

00:14:46.960 pivot to make

00:14:47.840 generative AI

00:14:48.520 a core part

00:14:49.580 of their strategy.

00:14:50.760 The kind of

00:14:51.280 emerging enterprise

00:14:53.180 use cases

00:14:53.940 are truly astounding.

00:14:56.120 So I think

00:14:56.480 where this is

00:14:57.900 the calm

00:14:58.500 before the storm

00:14:59.360 or this is probably

00:14:59.920 the last moment,

00:15:01.360 I would say,

00:15:02.580 before we really

00:15:03.960 start seeing AI

00:15:05.020 being integrated

00:15:05.780 into almost

00:15:06.540 every type

00:15:07.580 of human

00:15:08.300 knowledge work.

00:15:09.720 So when you're

00:15:11.620 thinking of

00:15:12.400 the pace

00:15:13.520 and scale

00:15:14.360 of change

00:15:15.240 at this rate,

00:15:16.640 you know,

00:15:16.880 policymakers having

00:15:17.720 worked with many

00:15:18.380 policymakers for many

00:15:19.380 years,

00:15:19.700 they're always

00:15:20.280 kind of on the

00:15:21.140 back foot anyway,

00:15:22.800 but faced with

00:15:23.840 challenges like this,

00:15:25.800 you know,

00:15:26.320 it's very,

00:15:27.420 very difficult,

00:15:28.120 not least because

00:15:29.300 there is a huge

00:15:30.700 skills gap,

00:15:31.680 not only in

00:15:33.340 the companies

00:15:34.480 that are building

00:15:35.080 the technology,

00:15:35.840 we hear this all

00:15:36.380 the time about

00:15:37.000 the AI skills gap

00:15:37.980 and so on,

00:15:38.460 but also on the

00:15:39.820 regulatory side,

00:15:40.780 who actually

00:15:41.360 understands this,

00:15:42.760 who actually

00:15:43.500 can foresee

00:15:44.880 what the implications

00:15:45.860 might be

00:15:46.520 and given,

00:15:47.780 so the only

00:15:48.440 kind of piece

00:15:49.560 of transnational

00:15:52.240 regulation that's

00:15:53.200 in the works

00:15:53.580 right now is

00:15:54.380 coming from

00:15:55.100 the European Union

00:15:56.140 and this is

00:15:57.660 kind of a

00:15:58.060 gargantuan piece

00:15:59.020 of legislation,

00:16:00.220 it's meant to be

00:16:00.860 the kind of

00:16:01.260 first regulatory

00:16:02.700 blueprint,

00:16:03.900 if you will,

00:16:04.520 on artificial

00:16:05.320 intelligence,

00:16:06.140 been in the works

00:16:06.700 for years,

00:16:07.240 but until

00:16:08.720 ChatGPT

00:16:09.560 came out,

00:16:10.400 it made no

00:16:11.020 reference of

00:16:12.520 generative AI

00:16:13.560 or foundational

00:16:14.400 models.

00:16:14.960 Now,

00:16:15.080 they very quickly

00:16:15.900 redrafted it

00:16:16.920 because they

00:16:17.380 understood that,

00:16:18.900 you know,

00:16:19.600 this is really

00:16:20.260 important,

00:16:20.660 but that's only

00:16:21.160 going to come

00:16:21.560 into force

00:16:22.020 in 2026.

00:16:23.660 So,

00:16:24.060 I think

00:16:24.500 one of the

00:16:25.800 consistent

00:16:26.480 reflections,

00:16:27.240 and you must

00:16:27.700 have,

00:16:28.320 I know that

00:16:28.940 you've had this

00:16:29.760 reflection as

00:16:31.060 well,

00:16:31.760 if you consider

00:16:32.520 what's been

00:16:32.900 happening over

00:16:33.440 the past few

00:16:34.020 years,

00:16:34.420 is just how

00:16:35.060 quickly all of

00:16:35.720 this has unfolded.

00:16:37.040 So,

00:16:37.320 all AI

00:16:38.060 researchers I

00:16:38.820 talked to,

00:16:39.920 we all knew

00:16:40.820 or they knew

00:16:41.780 that this was

00:16:42.320 kind of

00:16:42.600 hypothetically

00:16:43.280 within the

00:16:43.820 realm of

00:16:44.180 the possible,

00:16:45.380 but everyone

00:16:46.040 always says,

00:16:46.600 we didn't think

00:16:47.240 we'd be here

00:16:47.920 now,

00:16:48.480 and just

00:16:48.860 try to keep

00:16:49.640 up with the

00:16:50.580 research papers

00:16:51.420 that are coming

00:16:52.020 out every

00:16:52.680 single day,

00:16:53.420 the new

00:16:53.720 companies,

00:16:54.340 the amount

00:16:54.660 of money

00:16:55.140 flowing into

00:16:55.640 the space,

00:16:56.560 the kind

00:16:56.960 of market

00:16:58.140 moving impetus

00:16:59.340 started by the

00:17:00.240 tech companies

00:17:01.000 by actually

00:17:02.800 commercializing

00:17:03.860 and productizing

00:17:04.900 these tools

00:17:06.140 and bringing

00:17:06.880 it to market

00:17:07.540 of hundreds

00:17:08.180 of millions

00:17:08.640 of people,

00:17:09.660 you know,

00:17:10.280 yeah,

00:17:10.800 regulators have

00:17:11.760 a hard task

00:17:12.700 on their hands.

00:17:13.980 We could

00:17:14.120 probably class

00:17:15.000 the spectrum

00:17:16.580 of problems

00:17:17.480 into two

00:17:18.800 bins here,

00:17:19.740 and so the

00:17:20.320 deepest,

00:17:21.180 which perhaps

00:17:22.260 we won't even

00:17:22.700 talk about unless

00:17:23.480 you especially

00:17:24.260 want to touch

00:17:24.780 it,

00:17:24.940 I mean,

00:17:25.060 it's something

00:17:25.400 that I've

00:17:26.120 spoken about

00:17:26.760 before and

00:17:27.240 will continue

00:17:27.640 to cover on

00:17:28.240 this podcast,

00:17:28.660 but the

00:17:29.540 deepest concern

00:17:30.340 is what

00:17:32.500 often goes

00:17:33.000 by the name

00:17:33.420 of existential

00:17:34.220 risk here.

00:17:35.140 Is there

00:17:35.500 something

00:17:36.000 fundamental

00:17:36.640 about the

00:17:37.960 development

00:17:38.320 of AI

00:17:38.720 that poses

00:17:40.040 a real

00:17:41.260 threat to

00:17:41.960 not just

00:17:42.760 the maintenance

00:17:43.180 of democracy

00:17:44.080 and the

00:17:44.900 maintenance

00:17:45.140 of civilization

00:17:45.680 but to

00:17:46.280 the further

00:17:47.340 career of

00:17:48.500 our species?

00:17:49.980 And,

00:17:50.580 you know,

00:17:50.900 I'm convinced

00:17:52.260 that there

00:17:54.100 is a problem

00:17:55.000 here worth

00:17:55.700 worrying about

00:17:56.280 and therefore

00:17:57.620 there will be

00:17:58.120 some

00:17:58.520 regulation

00:17:59.700 of a sort

00:18:00.820 that,

00:18:01.440 you know,

00:18:01.620 we've used

00:18:02.480 for,

00:18:03.020 you know,

00:18:03.780 nuclear

00:18:04.080 proliferation

00:18:04.700 or the

00:18:05.940 spread of

00:18:06.460 the tools

00:18:07.060 of synthetic

00:18:07.620 biology

00:18:08.120 and,

00:18:08.580 you know,

00:18:08.720 I don't think

00:18:09.560 we've done

00:18:09.920 either of those

00:18:10.560 especially well

00:18:11.200 but here

00:18:11.940 it's even

00:18:12.300 harder

00:18:12.640 and,

00:18:13.480 you know,

00:18:13.620 that's a

00:18:14.380 separate

00:18:14.580 conversation

00:18:15.120 perhaps.

00:18:16.040 Then there

00:18:16.620 are all

00:18:17.020 of the

00:18:17.580 piecemeal

00:18:18.560 near-term

00:18:19.700 and truly

00:18:20.280 immediate

00:18:20.800 threats

00:18:21.600 of the

00:18:22.380 sort that

00:18:22.700 we've

00:18:22.940 just begun

00:18:23.740 to speak

00:18:24.080 about

00:18:24.420 that go

00:18:25.340 under

00:18:25.640 the banner

00:18:26.720 of information

00:18:28.160 integrity

00:18:28.720 and,

00:18:29.400 you know,

00:18:29.800 cyber hacking

00:18:31.280 and,

00:18:31.780 you know,

00:18:31.980 cyber terrorism

00:18:33.200 and just

00:18:33.920 any malicious

00:18:35.600 use of

00:18:37.100 narrow AI

00:18:38.560 that can

00:18:40.020 really supercharge

00:18:41.840 in a human

00:18:42.740 conflict

00:18:43.340 and confusion

00:18:44.600 and this can

00:18:45.620 short of being

00:18:46.620 an existential

00:18:47.180 threat,

00:18:48.540 it can be

00:18:49.660 an enormous

00:18:50.520 threat which

00:18:51.560 is certainly

00:18:52.280 worth worrying

00:18:52.820 about.

00:18:53.420 So,

00:18:53.640 but then

00:18:54.240 obviously

00:18:54.740 the reason

00:18:55.640 why this

00:18:56.020 is such

00:18:56.360 an interesting

00:18:56.780 conversation

00:18:57.760 is that

00:18:58.460 that's only

00:18:59.240 half of the

00:19:00.040 story.

00:19:00.400 The other

00:19:00.680 half,

00:19:01.580 as you

00:19:01.960 point out,

00:19:02.580 is all

00:19:03.680 of the

00:19:03.960 good things

00:19:04.660 we can

00:19:05.080 expect to

00:19:05.780 do and

00:19:06.260 build and

00:19:07.320 enjoy on

00:19:07.980 the basis

00:19:08.340 of increased

00:19:08.960 intelligence.

00:19:10.260 Generically,

00:19:10.940 intelligence is

00:19:11.660 the best thing

00:19:12.740 we have,

00:19:13.820 right?

00:19:14.000 It's the thing

00:19:14.480 that differentiates

00:19:15.640 us from

00:19:16.200 our primate

00:19:17.200 cousins,

00:19:17.800 it's the thing

00:19:18.420 that safeguards

00:19:19.200 everything else

00:19:19.920 we care about

00:19:20.580 even if

00:19:21.040 the things

00:19:21.920 we care

00:19:22.920 about can't

00:19:23.480 be narrowly

00:19:24.760 reduced to

00:19:25.460 intelligence,

00:19:26.640 things like

00:19:27.040 love and

00:19:27.800 friendship and

00:19:28.900 creative joy,

00:19:30.200 etc.

00:19:30.880 All of that

00:19:31.380 is safeguarded

00:19:32.320 from the

00:19:33.260 casual malevolence

00:19:34.780 of nature

00:19:35.260 by intelligence.

00:19:37.380 The fact that

00:19:37.960 we have cures

00:19:39.760 for any

00:19:40.280 illnesses is

00:19:41.040 the result of

00:19:41.600 intelligence.

00:19:42.160 Obviously,

00:19:42.860 we want a

00:19:43.200 cure for cancer,

00:19:44.040 we want a

00:19:44.400 cure for

00:19:44.800 Alzheimer's.

00:19:46.200 If AI

00:19:46.780 could give

00:19:47.220 us those

00:19:47.620 two things,

00:19:48.940 the whole

00:19:49.460 experiment

00:19:50.000 would be

00:19:50.760 worth it

00:19:51.240 already,

00:19:51.920 apart from

00:19:52.720 the possibility

00:19:53.820 of our

00:19:54.320 destroying

00:19:54.920 everything else

00:19:55.720 we care

00:19:56.020 about.

00:19:57.080 So,

00:19:57.560 let's start

00:19:59.520 where you

00:20:01.340 and I really

00:20:01.720 left off last

00:20:02.560 time with

00:20:03.560 the issue

00:20:03.980 of deep

00:20:04.740 fakes and

00:20:06.260 I guess just

00:20:07.260 fakeness in

00:20:08.280 general.

00:20:08.760 It wasn't

00:20:09.600 until the

00:20:10.200 emergence of

00:20:11.040 ChatGPT that

00:20:12.860 I suddenly

00:20:13.200 glimpsed the

00:20:13.700 possibility that

00:20:14.540 in very short

00:20:16.080 order here,

00:20:17.300 most of

00:20:18.360 everything on

00:20:19.640 the internet

00:20:20.080 could be

00:20:21.320 fake.

00:20:22.580 Most text

00:20:24.220 could be

00:20:24.620 fake,

00:20:25.120 most image

00:20:26.420 could be

00:20:26.800 fake,

00:20:27.120 not now,

00:20:28.720 but maybe

00:20:30.080 two years

00:20:31.380 from now.

00:20:32.040 When you

00:20:32.780 look at how

00:20:33.220 quickly you

00:20:34.240 can produce

00:20:34.980 fake faces

00:20:35.880 and fake

00:20:36.500 video and

00:20:37.400 fake journal

00:20:38.000 articles,

00:20:38.720 that's just

00:20:39.700 an amazing

00:20:40.180 prospect.

00:20:40.980 So,

00:20:41.520 tell me how

00:20:42.700 you've been

00:20:43.320 viewing the

00:20:44.740 development of

00:20:45.720 these tools

00:20:46.180 in the

00:20:46.480 last 6

00:20:47.760 to 12

00:20:48.140 months and

00:20:48.600 what are

00:20:49.080 the crucial

00:20:50.000 moments with

00:20:50.880 respect to

00:20:51.340 deep fakes

00:20:51.900 and other

00:20:52.680 fake material

00:20:53.580 that you've

00:20:54.120 noticed?

00:20:55.340 Yeah.

00:20:55.880 So,

00:20:56.500 first of all,

00:20:57.160 I think you're

00:20:57.680 absolutely right

00:20:58.680 to conceptualize

00:21:00.200 the risks

00:21:00.920 broadly in

00:21:01.500 those two

00:21:01.860 buckets,

00:21:02.440 the kind of

00:21:02.800 AGI or

00:21:03.500 existential risk

00:21:04.360 scenario and

00:21:05.200 I'd like to

00:21:06.200 have a

00:21:06.700 conversation on

00:21:07.420 that with

00:21:07.720 you too,

00:21:08.060 so maybe

00:21:08.340 we can

00:21:08.540 go back

00:21:08.860 to that.

00:21:09.440 But as

00:21:10.120 to the

00:21:11.120 second bucket,

00:21:11.840 which is

00:21:12.160 almost the

00:21:12.880 short and

00:21:13.420 medium-term

00:21:13.920 risks,

00:21:14.260 things that

00:21:14.760 are actually

00:21:15.300 materializing

00:21:16.060 right now,

00:21:17.260 and foremost

00:21:18.040 in that

00:21:19.040 bucket in

00:21:20.000 my mind is

00:21:20.720 without a

00:21:21.220 doubt,

00:21:21.460 information

00:21:21.720 integrity.

00:21:22.540 Now,

00:21:22.660 this is

00:21:23.560 kind of

00:21:23.980 the main

00:21:24.460 thesis of

00:21:24.980 my book,

00:21:25.400 which came

00:21:25.760 out three

00:21:26.420 years ago,

00:21:27.040 and that

00:21:27.240 was,

00:21:27.900 you know,

00:21:28.440 so much

00:21:29.380 has happened

00:21:29.920 since then.

00:21:30.980 At the

00:21:31.220 time,

00:21:32.000 generative AI

00:21:32.660 hadn't even

00:21:33.340 been coined

00:21:33.960 as a phrase.

00:21:35.500 And although

00:21:35.900 people don't

00:21:36.940 usually put

00:21:37.620 chat GPT

00:21:38.700 and deep

00:21:39.360 fakes into

00:21:40.040 the same

00:21:40.420 sentence,

00:21:41.060 they're

00:21:41.200 actually

00:21:41.440 manifestations

00:21:42.000 of the

00:21:42.900 same

00:21:43.120 phenomenon,

00:21:43.760 right?

00:21:43.860 the same

00:21:44.320 kind of

00:21:45.000 new

00:21:45.260 capabilities,

00:21:46.260 new quote

00:21:46.680 unquote,

00:21:47.560 of AI

00:21:48.380 to be able

00:21:48.960 to generate

00:21:49.600 new data.

00:21:50.920 Now,

00:21:51.100 the really

00:21:51.600 interesting

00:21:52.040 thing is

00:21:52.880 that when

00:21:54.020 these

00:21:54.300 capabilities

00:21:54.920 started kind

00:21:56.000 of coming

00:21:56.380 out of the

00:21:56.740 research

00:21:57.040 community

00:21:57.520 at the

00:21:58.080 end of

00:21:58.340 2017,

00:21:59.860 people started

00:22:00.480 making these

00:22:01.360 non-consensual

00:22:01.940 pornographic

00:22:02.540 creations with

00:22:03.400 them and

00:22:03.940 memes and

00:22:04.760 kind of

00:22:05.140 visual content.

00:22:06.720 So much

00:22:07.340 of the

00:22:07.720 premise of

00:22:08.160 my book

00:22:08.600 was focused

00:22:10.260 on visual

00:22:11.040 media.

00:22:11.520 But of

00:22:12.520 course,

00:22:12.880 at the

00:22:13.240 same

00:22:13.540 time,

00:22:14.200 concurrently,

00:22:16.160 a lot of

00:22:16.660 work was

00:22:17.120 going into

00:22:17.520 the development

00:22:18.160 of large

00:22:18.700 language models.

00:22:19.560 I mean,

00:22:20.140 Google was

00:22:20.600 really pioneering

00:22:21.580 work in

00:22:22.200 large language

00:22:22.680 models in

00:22:23.940 2017.

00:22:25.160 However,

00:22:26.400 they kept

00:22:27.080 it behind

00:22:27.580 closed doors,

00:22:28.520 right?

00:22:28.840 It wasn't

00:22:29.380 out there.

00:22:30.060 This is

00:22:30.340 perhaps why

00:22:30.940 nobody was

00:22:31.420 really talking

00:22:32.340 about text.

00:22:34.300 And although

00:22:34.700 we kind of

00:22:35.460 registered,

00:22:36.080 I know you've

00:22:36.660 spoken about

00:22:37.360 GPT series

00:22:38.480 before

00:22:39.120 ChatGPT

00:22:39.760 came out

00:22:40.280 and we've

00:22:40.540 spoken about

00:22:41.100 large language

00:22:41.740 models and

00:22:42.360 I knew it

00:22:43.300 was kind of

00:22:43.740 on my radar.

00:22:45.080 It was

00:22:45.340 really only

00:22:47.020 when GPT-3

00:22:48.200 and GPT-3.5

00:22:50.040 and now GPT-4

00:22:50.940 and basically

00:22:51.860 ChatGPT

00:22:52.700 came out

00:22:53.200 that you

00:22:54.040 truly understand

00:22:55.280 the,

00:22:56.400 first of all,

00:22:57.640 the significance

00:22:58.500 of a large

00:22:59.620 language model

00:23:00.400 and what it

00:23:01.300 can do to

00:23:01.900 scale

00:23:02.540 myths and

00:23:03.120 disinformation.

00:23:03.960 I mean,

00:23:04.120 it's truly

00:23:04.540 incredible and

00:23:05.300 how convincing

00:23:06.860 it is.

00:23:07.360 And we

00:23:07.640 were thinking

00:23:08.220 about visual

00:23:09.040 content as

00:23:09.740 the most

00:23:10.060 convincing,

00:23:11.160 you know,

00:23:11.940 AI-generated

00:23:12.760 video of

00:23:13.600 people saying

00:23:14.480 and doing

00:23:14.780 things they've

00:23:15.280 never done.

00:23:15.900 Anyone can

00:23:16.540 be cloned

00:23:17.140 now,

00:23:17.500 whether it's

00:23:17.880 your voice

00:23:18.500 or your

00:23:18.840 face,

00:23:19.600 but hadn't

00:23:20.180 really considered

00:23:21.120 or put text

00:23:22.060 on an equal

00:23:22.680 footing.

00:23:23.900 But of course,

00:23:24.360 that makes so

00:23:24.960 much sense.

00:23:25.800 You know,

00:23:26.040 we're storytellers.

00:23:28.060 That's something

00:23:28.880 that goes back

00:23:30.220 to the earliest

00:23:31.360 days of

00:23:31.940 civilization.

00:23:33.440 And the

00:23:33.860 problem,

00:23:34.340 I suppose,

00:23:35.080 is that there's

00:23:36.000 been a lot

00:23:36.460 of thinking

00:23:37.000 on kind of

00:23:38.280 the visual

00:23:38.920 components of

00:23:40.180 AI-generated

00:23:41.000 content and

00:23:42.140 what we can

00:23:43.240 do to kind

00:23:43.840 of combat

00:23:44.360 the worst

00:23:45.100 risks,

00:23:45.520 and I'll

00:23:45.720 come back

00:23:46.120 to that,

00:23:46.940 but less

00:23:48.380 on text.

00:23:50.160 And if you

00:23:50.820 think about

00:23:51.340 the solutions

00:23:52.320 on kind of

00:23:53.160 the disinformation

00:23:54.340 piece around

00:23:55.280 that really

00:23:55.880 started thinking

00:23:56.420 around the

00:23:57.340 thinking started

00:23:58.080 with deep

00:23:58.480 face,

00:23:59.240 initially people

00:23:59.800 started thinking,

00:24:00.780 okay,

00:24:01.140 what we need

00:24:01.760 to do is we

00:24:02.660 need to detect

00:24:03.320 it, right?

00:24:03.980 We need to

00:24:04.500 be able to

00:24:05.080 build an

00:24:05.560 AI content

00:24:06.780 detector that

00:24:07.620 can detect

00:24:08.180 everything that's

00:24:09.200 made by AI so

00:24:10.320 that we can

00:24:10.820 definitively say,

00:24:11.900 okay, that's

00:24:12.660 AI-generated and

00:24:13.540 that's not,

00:24:14.000 that's synthetic,

00:24:14.860 that's authentic.

00:24:16.180 Turns out that in

00:24:16.980 practice,

00:24:18.060 building detection

00:24:19.040 is really,

00:24:20.260 really difficult

00:24:20.900 because first,

00:24:22.900 there's no one

00:24:23.600 size fits all

00:24:24.320 detector.

00:24:25.160 There are now

00:24:26.220 hundreds of

00:24:26.940 thousands of

00:24:27.520 generative models

00:24:28.240 out there and

00:24:28.880 there's never

00:24:29.280 going to be one

00:24:30.080 size fits all

00:24:30.640 detector that can

00:24:31.460 detect all

00:24:32.180 synthetic content.

00:24:33.060 Second, it can

00:24:34.500 only give you

00:24:35.900 a percentage of

00:24:37.760 how likely it

00:24:39.120 thinks that is

00:24:39.840 generated by AI

00:24:40.860 or not.

00:24:41.500 So, okay, I'm

00:24:42.120 90% confident,

00:24:43.300 I'm 70% confident,

00:24:44.500 so always has a

00:24:45.480 chance for a

00:24:46.120 false negative,

00:24:46.920 false positive

00:24:47.580 or false negative.

00:24:49.300 And third, as you

00:24:50.700 correctly point out,

00:24:51.900 and this is kind

00:24:52.600 of, was one of

00:24:54.160 the points I made

00:24:55.400 in my book and

00:24:56.200 actually over the

00:24:57.120 past few years

00:24:57.880 has become something

00:24:58.680 I've been speaking

00:24:59.380 about a lot,

00:25:00.380 is that if

00:25:01.760 you believe,

00:25:03.620 as I do,

00:25:04.220 and as you

00:25:04.500 already pointed

00:25:04.940 out, that there

00:25:05.820 will be some

00:25:06.660 element of AI

00:25:07.960 creation going

00:25:09.280 forward in all

00:25:10.780 digital information,

00:25:12.040 then it becomes a

00:25:13.500 futile exercise to

00:25:14.720 try and detect

00:25:15.600 what's synthetic

00:25:16.700 because everything

00:25:17.840 is going to have

00:25:18.500 some degree of

00:25:19.860 AI or synthetic

00:25:21.780 nature within any

00:25:23.180 piece of digital

00:25:23.880 content.

00:25:25.300 So, the

00:25:27.080 second approach,

00:25:28.900 kind of tech-led

00:25:30.580 approach that's

00:25:31.620 been emerging

00:25:32.120 over the past

00:25:32.640 few years,

00:25:33.020 which is more

00:25:33.560 promising, is the

00:25:35.040 idea of content

00:25:37.000 provenance, right?

00:25:38.400 So that, and this

00:25:39.720 is applicable to

00:25:40.800 both synthetic or

00:25:42.480 AI-generated content

00:25:43.920 as well as

00:25:44.520 authentic content.

00:25:45.500 It's about full

00:25:46.480 transparency.

00:25:47.440 So, rather than

00:25:48.220 being in the

00:25:49.580 business of

00:25:50.340 adjudicating what's

00:25:52.220 true, you know,

00:25:53.200 this is real,

00:25:53.980 this is not,

00:25:54.580 it's about

00:25:56.400 securing full

00:25:58.780 transparency about

00:25:59.600 the origins of

00:26:00.660 content in kind of

00:26:02.240 almost the DNA of

00:26:03.580 that content.

00:26:04.340 So, whether, if

00:26:05.100 it's authentic, you

00:26:06.080 can capture it using

00:26:06.940 secure capture

00:26:07.680 technology, and that

00:26:09.040 will give you almost

00:26:10.080 kind of a

00:26:10.980 cryptographically

00:26:12.240 sealed data about

00:26:14.040 that piece of

00:26:14.580 content, where it

00:26:15.660 was created, who it

00:26:16.920 belongs to, but the

00:26:17.960 same principle

00:26:18.920 should also be

00:26:20.220 applied to AI-

00:26:21.420 generated content.

00:26:22.180 And, of course,

00:26:23.580 not everyone's

00:26:24.100 going to do that,

00:26:24.920 but the difference

00:26:25.680 here is that if

00:26:26.360 you are a good

00:26:27.160 actor, right?

00:26:28.220 So, if you are

00:26:29.360 a open AI, or

00:26:32.060 you're a stable

00:26:32.700 diffusion, or,

00:26:34.200 you know, you're

00:26:34.880 Coca-Cola, and you

00:26:36.140 want to use AI-

00:26:37.120 generated collateral

00:26:38.140 in your latest

00:26:39.040 marketing campaign,

00:26:40.160 then you should

00:26:42.120 mark your content

00:26:43.180 so everyone can

00:26:44.020 see that this is

00:26:45.420 actually synthetic.

00:26:46.520 So, the technology

00:26:47.160 to do that already

00:26:47.920 exists, and I

00:26:48.380 should point out

00:26:48.900 that it's much

00:26:49.840 more than a

00:26:50.420 watermark.

00:26:51.340 Because people are,

00:26:51.820 yeah, okay,

00:26:52.180 it's a watermark,

00:26:52.880 watermarks can be

00:26:53.580 edited or removed.

00:26:55.060 But this kind of

00:26:56.380 authentication

00:26:57.320 technology, like I

00:26:58.600 said, it's about

00:26:59.260 cryptographically

00:27:00.180 sealing it almost

00:27:01.200 into the DNA

00:27:02.720 of that content,

00:27:04.360 so it can never

00:27:04.940 be removed, it's

00:27:05.880 indelible, it's,

00:27:06.800 you know, there for

00:27:07.400 kind of the world

00:27:08.400 to see.

00:27:09.520 But the second

00:27:10.100 part of this, and

00:27:10.960 this is really where

00:27:12.600 it starts getting

00:27:13.380 tricky, is that it's

00:27:14.900 no good signing your

00:27:16.560 content in this way,

00:27:17.880 in full transparency,

00:27:19.620 if you're a good

00:27:20.140 actor.

00:27:20.860 If nobody can see

00:27:21.840 that kind of

00:27:22.400 nutritional label,

00:27:23.680 right, so you've

00:27:24.120 signed it, you put

00:27:25.340 the technology in

00:27:26.140 there, but like, if

00:27:27.660 I view it on

00:27:28.260 Twitter, am I going

00:27:28.980 to see it?

00:27:29.520 Or if I see it on

00:27:30.320 YouTube, will I

00:27:31.500 see it?

00:27:32.440 So the second

00:27:32.980 point is that we

00:27:34.240 actually need to

00:27:35.200 build into the

00:27:37.020 architecture of the

00:27:38.100 internet, the

00:27:38.840 actual infrastructure

00:27:39.980 for this to kind of

00:27:41.400 become the default,

00:27:42.640 right, the content

00:27:43.300 credentials.

00:27:44.620 And there's a

00:27:45.480 non-profit

00:27:45.840 organization called

00:27:47.300 the C2PA, which

00:27:48.900 is already building

00:27:49.780 this open standard

00:27:50.860 for the internet.

00:27:51.660 And a lot of

00:27:53.580 interesting founding

00:27:54.300 members, Microsoft,

00:27:56.460 Intel, BBC,

00:27:58.000 ARM.

00:27:58.760 And I guess we

00:28:00.080 will see if this

00:28:01.200 kind of becomes a

00:28:02.640 standard, because my

00:28:04.820 view, so the

00:28:05.520 projection, the

00:28:07.180 estimate that I make

00:28:08.240 is that 90% of

00:28:09.240 online content is

00:28:10.160 going to be

00:28:10.440 generated by AI by

00:28:11.800 2025.

00:28:12.620 That's a punchy

00:28:16.740 kind of figure, but

00:28:18.620 I really believe that

00:28:19.940 this is probably the

00:28:20.620 last moment of the

00:28:21.400 internet where the

00:28:22.360 majority of the

00:28:23.040 information and data

00:28:23.960 you see online doesn't

00:28:25.540 have some degree of

00:28:27.500 AI in its creation.

00:28:29.940 So this issue of

00:28:32.280 authentication rather

00:28:33.940 than detecting

00:28:35.300 fakes, it's an

00:28:37.320 interesting flip of

00:28:39.680 the approach, and it

00:28:41.540 presents a kind of

00:28:42.740 bottleneck.

00:28:43.360 I'm wondering, does

00:28:44.480 this suggest that

00:28:46.080 this will be an

00:28:47.460 age of new

00:28:49.400 gatekeepers, where

00:28:50.880 the promise at the

00:28:52.620 moment for those who

00:28:53.400 are very bullish on

00:28:54.600 everything that's

00:28:55.480 happening is that

00:28:56.180 this has democratized

00:28:58.460 creativity and

00:29:00.500 information creation

00:29:02.220 just to the

00:29:03.480 ultimate degree, but

00:29:05.560 if we all get

00:29:07.800 trained very quickly

00:29:09.000 to care about the

00:29:11.080 integrity of

00:29:11.600 information and

00:29:12.640 our approach to

00:29:13.960 finding legitimate

00:29:15.400 information, whatever

00:29:16.300 its provenance, whether

00:29:17.760 it's been synthesized in

00:29:19.100 some way or whether it

00:29:20.260 purports to be actually

00:29:21.420 human-generated, if the

00:29:23.620 approach to safeguarding

00:29:24.680 the integrity is

00:29:25.700 authentication rather

00:29:27.200 than the detection of

00:29:28.780 misinformation, how do

00:29:30.360 we not wind up in a

00:29:31.180 world where you can't

00:29:33.460 trust an image unless it

00:29:35.080 came from Getty images,

00:29:36.700 say, or it was taken

00:29:38.640 on an iPhone, right?

00:29:39.920 Like the cryptographic

00:29:40.920 authentication of

00:29:41.960 information, is this

00:29:43.580 something that you

00:29:44.180 imagine is going to

00:29:45.260 lead to a new siloing

00:29:47.540 and gatekeeping, or is

00:29:48.740 it going to be like

00:29:49.480 blockchain-mediated and

00:29:52.040 everyone will be on

00:29:53.380 this, all fours

00:29:54.640 together dealing with

00:29:55.880 content, however,

00:29:57.380 people will be able to

00:29:59.100 create it outside of

00:30:01.260 some walled garden or

00:30:03.020 inside of a major

00:30:05.060 corporation and everyone

00:30:07.060 will have access to the

00:30:08.080 same authentication

00:30:09.440 tools?

00:30:10.580 Yeah, I guess the first

00:30:11.640 thing to say is even

00:30:13.240 when it comes to the

00:30:14.480 authentication approach,

00:30:16.400 there's no silver bullet,

00:30:17.640 right?

00:30:17.900 Because the scale of

00:30:20.100 the problem is so vast

00:30:23.040 and in part because over

00:30:25.140 the past 30 years we've

00:30:26.380 created this kind of

00:30:27.260 digital information

00:30:28.720 ecosystem wherein

00:30:30.440 everybody and everything

00:30:32.720 must now exist.

00:30:34.300 it doesn't really

00:30:35.000 matter if you're an

00:30:35.980 individual, whether

00:30:37.180 you're an organization

00:30:37.980 or enterprise or a

00:30:39.360 nation state, you

00:30:40.160 don't really have the

00:30:41.320 choice not to be

00:30:43.560 engaged and be doing

00:30:46.620 things within this

00:30:48.020 ecosystem.

00:30:48.860 So the very

00:30:49.940 possibility that the

00:30:53.100 medium by which

00:30:54.720 information and

00:30:55.820 transactions and

00:30:56.800 communications and

00:30:57.780 interactions, so all

00:30:58.960 digital kind of content

00:31:00.820 and information could be

00:31:01.900 compromised to the

00:31:03.280 extent that it becomes

00:31:04.540 untrustworthy, that's a

00:31:06.400 huge problem, right?

00:31:07.440 So trying to ensure

00:31:09.480 that we build an

00:31:10.540 ecosystem where we can

00:31:11.580 actually trust the

00:31:12.840 information or a part

00:31:15.200 of an ecosystem, right?

00:31:17.160 We can actually trust the

00:31:18.260 information you engage

00:31:19.540 with online is going to

00:31:20.960 be critical to society,

00:31:22.620 to business, on every

00:31:25.020 kind of level conceivable.

00:31:27.180 Detection will always

00:31:28.120 play a role, by the way,

00:31:29.200 I think.

00:31:29.520 It's just that it's

00:31:30.260 not, you know, it's

00:31:31.220 not the only solution.

00:31:33.400 Let me just take a very

00:31:34.760 narrow but salient case

00:31:36.560 now.

00:31:36.860 I would just imagine at

00:31:38.240 some point today a video

00:31:40.640 of Vladimir Putin

00:31:41.980 claiming that he is

00:31:44.480 about to use tactical

00:31:45.780 nukes in the war in

00:31:47.600 Ukraine emerges online

00:31:50.140 and, you know, the New

00:31:52.000 York Times is trying to

00:31:53.260 figure out whether or not

00:31:54.720 to write a story on it,

00:31:56.940 react to it, spread it.

00:31:58.140 Clearly, there's a

00:31:59.440 detection problem there,

00:32:01.080 right?

00:32:01.300 It's like we have this

00:32:02.400 one video that's spreading

00:32:03.880 on social media and to

00:32:05.500 humanize, it appears

00:32:07.420 totally authentic, right?

00:32:09.080 I think it's uncontroversial

00:32:10.800 to say that if we're not

00:32:12.280 there now, we will be

00:32:14.240 there very soon and

00:32:15.500 probably in a matter of

00:32:16.420 months, not years, where

00:32:18.200 Oh, we're there.

00:32:19.180 Yeah, we'll have video of

00:32:20.980 Putin or anyone else where

00:32:22.980 it will literally be

00:32:24.160 impossible for a person to

00:32:26.320 detect some anomaly

00:32:28.420 which is an obvious

00:32:30.880 tell to it being fake.

00:32:32.940 Absolutely, yeah.

00:32:34.040 And I mean, we're

00:32:35.200 already there.

00:32:36.680 So to very sophisticated

00:32:38.340 synthetic media or, you

00:32:40.120 know, deepfakes has now

00:32:40.960 come to kind of mean

00:32:42.020 AI-generated content in

00:32:43.940 which somebody's biometrics

00:32:45.500 are synthesized, right?

00:32:46.920 So it's a visual

00:32:49.420 representation of a person

00:32:51.300 saying and doing things

00:32:52.100 they didn't do or a

00:32:52.920 completely synthetic

00:32:53.580 person.

00:32:54.040 they're already so

00:32:55.460 sophisticated that we

00:32:57.200 can't tell.

00:32:58.280 And whilst we're talking

00:32:59.100 about kind of the

00:32:59.840 nuclear political

00:33:00.800 scenario, it's already

00:33:02.400 cropping up really

00:33:04.020 malicious use cases and

00:33:05.960 kind of the vishing has

00:33:07.580 become a big deal.

00:33:08.500 So this is kind of

00:33:09.740 phishing using people's

00:33:12.460 biometrically cloned

00:33:13.860 voice, you know, so the

00:33:15.040 age-old scam of your

00:33:17.220 loved one calling you

00:33:18.620 because they've had an

00:33:20.600 accident or they're in

00:33:21.700 jail and they need to be

00:33:22.700 bailed out.

00:33:23.340 Now, imagine you get

00:33:24.500 that call and it's

00:33:25.400 actually your son's voice

00:33:26.960 you hear or your wife's

00:33:28.560 voice or your father's

00:33:29.700 voice.

00:33:29.980 Are you going to send

00:33:30.560 the money?

00:33:31.420 Hell yeah, you're going

00:33:32.120 to send the money because

00:33:33.000 you believe your loved one

00:33:34.200 is in trouble and you can

00:33:35.960 already synthesize voices

00:33:37.400 with up to three seconds

00:33:38.720 of audio.

00:33:39.440 When I started looking at

00:33:41.160 this field back in 2017,

00:33:43.200 you need hours and hours

00:33:45.440 and hours of training data

00:33:46.960 to try and synthesize a

00:33:48.160 single voice.

00:33:48.780 So you could only do

00:33:49.460 people who are really in

00:33:50.340 the public eye like you,

00:33:52.200 Sam, you know, your

00:33:54.120 entire podcast repertoire

00:33:55.760 would, you know, have

00:33:56.760 been a good basis for

00:33:57.660 training data, but you

00:33:58.740 don't need to be Sam

00:34:00.600 Harrison, you know, to

00:34:02.080 do this.

00:34:03.100 You just need three

00:34:03.800 seconds, which you could

00:34:04.600 probably scrape off

00:34:06.160 Instagram, off YouTube,

00:34:07.740 off LinkedIn.

00:34:08.480 So you're already seeing

00:34:09.420 that, right?

00:34:09.980 One question here, Nina,

00:34:10.840 are we, with respect to

00:34:12.080 video, are we truly there?

00:34:14.380 Because the best thing

00:34:15.420 I've seen, and I think

00:34:17.800 this is, you know, most

00:34:18.840 people will have seen

00:34:19.680 this, are these Tom

00:34:21.180 Cruise videos, which are

00:34:22.720 fake, but they're

00:34:24.920 somewhat gamed because,

00:34:27.100 if I understand

00:34:27.860 correctly, the person

00:34:29.300 who's creating them

00:34:30.160 already looks a lot like

00:34:32.020 Tom Cruise, and he's

00:34:33.300 almost like a Tom Cruise

00:34:34.400 impersonator, and he's

00:34:35.620 mapping, you know, the

00:34:37.440 synthetic face of

00:34:39.020 Cruise onto his own

00:34:40.900 facial acting.

00:34:42.960 And so it's very

00:34:44.060 compelling.

00:34:44.700 It never looks truly

00:34:46.240 perfect to me, but it's,

00:34:47.480 if you weren't tipped

00:34:48.700 off that you should be

00:34:49.560 paying close attention,

00:34:50.660 you'd probably pass for

00:34:52.340 almost everyone.

00:34:53.560 But is it, are we, are

00:34:54.620 we to the point now

00:34:55.780 where, absent some, you

00:34:58.560 know, biasing scheme like

00:35:00.260 that, where you have an

00:35:01.360 actor at the bottom of it,

00:35:03.160 you can create video of

00:35:04.680 anyone that is, is

00:35:07.220 undetectable as fake?

00:35:08.800 It's a much more

00:35:09.820 difficult challenge, and

00:35:11.260 the deep Tom that started

00:35:13.280 emerging went viral on

00:35:14.520 TikTok in 2021.

00:35:15.460 You're absolutely right,

00:35:16.560 because that creator was,

00:35:18.700 first of all, he was

00:35:19.380 doing a lot of VFX and

00:35:20.820 AI.

00:35:21.320 It's not that the AI was

00:35:22.320 kind of autonomously

00:35:23.320 creating it, and he was

00:35:24.520 working with an actor who

00:35:25.820 was a Tom Cruise

00:35:26.500 impersonator, right?

00:35:27.680 So he was just mapping it

00:35:28.840 onto his face, which is

00:35:30.240 why the fidelity looked

00:35:31.460 highly convincing when it

00:35:33.120 came out in 2021.

00:35:34.540 Video is still a harder

00:35:35.520 challenge, but already

00:35:37.320 now there are consumer

00:35:39.140 products on the market

00:35:40.380 where you can send 20

00:35:41.800 seconds of you talking

00:35:44.260 into your cell phone, and

00:35:46.140 from that, they can

00:35:47.300 create your own kind of

00:35:48.380 personalized avatar.

00:35:49.820 So the point is that

00:35:51.880 whilst it's still, the

00:35:54.100 barriers to entry on

00:35:55.340 synthetic video generation

00:35:56.620 are still higher, they're

00:35:58.980 coming down very quickly,

00:36:00.340 and like I said, there's

00:36:01.500 the kind of market for

00:36:03.120 your AI avatar is already

00:36:05.540 thriving, and that requires

00:36:07.080 about 20 seconds of video.

00:36:08.380 And where do the visual

00:36:11.020 foundational models fit in

00:36:12.500 here, like DALI and

00:36:13.820 Mid Journey and Stable

00:36:15.300 Diffusion?

00:36:15.820 Are they the source of

00:36:18.140 good deepfakes now, or are

00:36:20.300 they not producing that

00:36:21.640 sort of thing?

00:36:22.580 Yeah, so it's been a

00:36:24.680 really interesting shift,

00:36:25.900 because when deepfakes first

00:36:27.660 started coming out in 2017,

00:36:29.340 right, it was more that

00:36:31.400 this was now a kind of tool

00:36:33.780 that enthusiasts, ML and

00:36:37.160 AI enthusiasts, perhaps

00:36:38.640 those with a background in

00:36:40.740 VFX could use to create

00:36:43.040 content, and they started

00:36:44.240 doing it to basically clone

00:36:46.500 people, right?

00:36:47.220 But there was no kind of

00:36:48.180 model or foundational model

00:36:50.380 in order to be able to do

00:36:51.800 this.

00:36:52.700 Then, pretty soon, I think

00:36:54.620 it was in 2018, NVIDIA

00:36:57.020 released this model called

00:36:58.740 FaceGas, StyleGas, and

00:37:00.420 that could generate endless

00:37:02.200 images of human faces.

00:37:04.080 So, it had been trained on

00:37:06.120 a vast data set of human

00:37:07.880 faces.

00:37:08.860 So, every time you kind of,

00:37:10.160 you might have gone to that

00:37:11.080 website a few years ago

00:37:12.160 called thispersondoesnotexist.com.

00:37:14.180 Yeah.

00:37:14.200 And it was, yeah, it was

00:37:15.260 astonishing, right?

00:37:16.200 Because every time you

00:37:16.960 refreshed the page, you'd be

00:37:19.220 given what looked like an

00:37:20.780 entirely convincing

00:37:21.960 photograph of somebody who

00:37:23.380 was entirely synthetic and

00:37:24.920 AI-generated.

00:37:26.300 Although, I remember the

00:37:27.300 tell there was that they

00:37:29.040 could never quite get the

00:37:30.140 teeth right.

00:37:30.920 The teeth, the ears,

00:37:33.880 exactly.

00:37:34.600 There was this giveaway,

00:37:35.840 telltale signs, right?

00:37:37.200 So, when you saw the best-in-class

00:37:39.260 kind of productions or

00:37:40.980 creations like Deep Tom in

00:37:42.720 2021, there was a high level

00:37:45.740 of post-production and VFX and

00:37:48.100 a bit of AI.

00:37:49.240 So, you know, this wasn't

00:37:50.800 still democratized to the

00:37:52.240 extent where anybody could do

00:37:53.800 it.

00:37:54.180 But what you've been seeing

00:37:56.020 over the past 12 months is

00:37:57.700 the emergence of the so-called

00:37:59.820 foundational models.

00:38:01.800 Now, these are interesting

00:38:03.720 because they are not task

00:38:05.720 specific, their general

00:38:07.020 purpose.

00:38:07.900 And they are trained on these

00:38:09.320 vast, vast data sets.

00:38:11.820 You can almost conceive of it

00:38:12.940 as the entirety of the

00:38:14.120 internet.

00:38:14.520 So, the ones you just

00:38:15.260 mentioned, Dolly 2, Table

00:38:17.680 Diffusion, MidJourney, they're

00:38:18.980 all text-to-image

00:38:19.880 generators, right?

00:38:21.640 And they're so compelling

00:38:24.020 because the user experience is

00:38:26.960 phenomenal because they have

00:38:28.580 NLP tied into them so that

00:38:31.460 when we use them, how do we

00:38:33.560 get them to create something?

00:38:34.780 Well, we prompt them.

00:38:35.600 We just type what they want,

00:38:37.740 what we want them to create.

00:38:39.440 So now, all of a sudden, you

00:38:41.560 have these foundational models

00:38:43.660 that can generate images of

00:38:46.380 anybody or anything.

00:38:48.580 So, yeah, you know, the really

00:38:50.220 sophisticated deep fakes you've

00:38:52.520 been seeing during the rounds on

00:38:53.720 the internet recently, whether it

00:38:55.540 was Donald Trump's kind of, just

00:38:58.640 before his arraignment, you

00:38:59.720 know, Donald Trump.

00:39:00.400 Yeah, fighting off the cops.

00:39:01.040 Did you see those ones?

00:39:01.920 Yeah, yeah, yeah.

00:39:03.120 Those were created by MidJourney

00:39:04.660 V5 or the ones of the Pope in

00:39:07.620 the Balenciaga jacket.

00:39:09.100 So, there's been an incredible

00:39:11.320 amping up of the capability,

00:39:14.180 shall we say, because before it's

00:39:15.420 quite piecemeal and you have to

00:39:16.840 do this and that, but there was

00:39:18.500 no foundational model where you

00:39:19.960 could just type in what you wanted

00:39:21.640 it to create and boom, it would

00:39:23.240 come out.

00:39:23.580 Now, does that exist for video yet?

00:39:26.620 Not yet.

00:39:27.680 Is it going to come?

00:39:29.100 Invariably.

00:39:29.780 And that's actually what ChatGPT

00:39:31.580 is as well, right?

00:39:32.400 It's a one manifestation of a

00:39:34.880 foundational model for text.

00:39:36.480 And that's one of the reasons why

00:39:38.080 it has just been so compelling.

00:39:40.180 It's that user experience.

00:39:41.580 Hey, I can just have a

00:39:42.740 conversation with it.

00:39:43.600 I can just type in there and

00:39:44.800 then it can like create anything

00:39:46.420 I ask it to.

00:39:47.840 And it's the same concept for the

00:39:49.420 foundational models for image

00:39:50.980 generation and video generation

00:39:52.940 is next.

00:39:54.100 Yeah.

00:39:54.280 I recently watched, rewatched the

00:39:56.680 film Her, which I hadn't seen

00:39:58.400 since it came out 10 years ago.

00:40:00.400 And I must say it, it lands

00:40:02.580 differently now.

00:40:04.200 Pressing.

00:40:04.580 Yeah.

00:40:05.000 I mean, it's, it's like, I

00:40:06.460 forgot what I thought about it 10

00:40:08.000 years ago, but you know, it's all

00:40:10.540 too plausible now.

00:40:12.780 And it's already happening, Sam.

00:40:15.040 Yeah, yeah.

00:40:16.200 But it's, and it, but, and the

00:40:17.620 thing that's, I think it's hard to

00:40:20.100 get away from, I mean, obviously

00:40:22.160 there, there's some benefits to

00:40:23.860 this sort of thing.

00:40:26.000 I mean, if you could have a, you

00:40:27.000 know, a functionally omniscient

00:40:29.280 agent, you know, in your ear

00:40:31.720 whenever you want, I mean, that's,

00:40:33.220 it's, many good things could come

00:40:35.380 of that, but there is something,

00:40:37.160 it's a vision of bespoke

00:40:40.500 information where no one is seeing

00:40:43.420 this or hearing the same thing

00:40:44.900 anymore, right?

00:40:45.620 So there's a siloing effect

00:40:47.180 where if everyone has access

00:40:49.060 to an oracle, well, then that

00:40:52.540 oracle can create a bespoke

00:40:54.720 reality with or without

00:40:56.560 hallucinations.

00:40:57.520 I mean, your, your, your

00:40:58.320 preferences can be catered to in

00:41:01.260 such a way that, you know,

00:41:02.500 everyone can be, I mean, to some

00:41:04.060 degree this has already happened,

00:41:05.460 but it just, it, the concern gets

00:41:07.320 sharpened up considerably when you

00:41:08.840 think about the prospect of all

00:41:11.780 of us having an independent

00:41:13.060 conversation with a, a super

00:41:16.380 intelligence that is not

00:41:18.940 constrained to get everyone to

00:41:23.700 converge or agree or to even find

00:41:26.840 one another interpretable, right?

00:41:29.440 It's like, it's, I already feel

00:41:31.160 like when, when I see my, I am no

00:41:33.640 longer on social media, so I have

00:41:35.220 this experience less now, but when

00:41:37.100 I was still on Twitter, I had the

00:41:39.080 experience of seeing people I knew

00:41:41.860 to some degree, behaving in ways

00:41:43.720 that were less and less

00:41:44.500 interpretable to me, I mean, they

00:41:46.540 were seeming more and more

00:41:47.800 irrational, and I realized, well,

00:41:50.320 I'm not seeing, I'm not looking

00:41:51.540 over their shoulder seeing their

00:41:52.880 Twitter feed, I don't see what

00:41:54.340 they're, the totality of what

00:41:56.160 they're feeling informed by and

00:41:58.340 reacting to, and I just see, I

00:42:00.300 basically, from my bubble, they

00:42:02.260 appear to be going crazy, and

00:42:04.040 everyone is, you know, red-shifted

00:42:05.920 and quickly vanishing over some

00:42:08.300 horizon from the place where I am

00:42:10.820 currently sitting, and I, and no

00:42:12.880 doubt that I'm doing the same thing

00:42:14.880 for them, and it is a, it's an

00:42:17.320 alarming picture of a balkanization of

00:42:20.240 our worldview, and it's, yeah, I'm

00:42:23.660 gonna, I guess the variable there

00:42:25.460 really is the coming bespokeness of,

00:42:28.640 of information, and, I mean, just

00:42:31.600 somebody, I think it was Jaron

00:42:33.660 Lanier, I might, I think it was

00:42:35.640 Jaron Lanier, you know, flagged this

00:42:37.500 for me some years ago where he

00:42:38.720 said, you know, just imagine if you

00:42:40.420 went to Wikipedia and no one was

00:42:42.920 actually seen, you know, you look

00:42:44.400 at, look up an article on

00:42:45.520 anything, you know, World War II,

00:42:47.060 and that is curated purely for you,

00:42:51.000 right, no one has seen that same

00:42:52.520 article, you know, your, the

00:42:53.740 Wikipedia, the Wikipedia, you know,

00:42:55.780 ground truth as to the causes and,

00:42:58.080 and reality of World War II was

00:43:00.540 written for you, catering and

00:43:02.780 pandering to your preferences and

00:43:05.220 biases, and no one else has access

00:43:08.460 to that specific article, well, it

00:43:10.860 seems that we're potentially

00:43:12.060 stumbling into that world with

00:43:14.320 these tools.

00:43:15.520 Oh, absolutely, this kind of

00:43:17.280 hyper-personalization or the

00:43:20.360 audience of one, and you already

00:43:22.040 kind of see some of the early

00:43:24.040 manifestations of that, so you,

00:43:27.060 you're talking about her, so we

00:43:29.780 have already, of course, after

00:43:31.560 ChatGPT came out, some of the

00:43:34.000 kind of more nefarious things that

00:43:35.740 started immediately being built,

00:43:37.620 because it's going to get very

00:43:38.560 weird when it comes to love, sex,

00:43:40.420 and relationships, are these kind of

00:43:42.820 girlfriend bots, right?

00:43:44.960 For very lonely men, and these kind of

00:43:48.760 chatbots can cater to their every

00:43:50.600 sexual fantasy, there was actually a

00:43:52.420 company called Replica, which also

00:43:54.640 did these avatars, and they, people were

00:43:57.480 using them as this kind of girlfriend,

00:44:00.320 you know, being very inappropriate and

00:44:02.740 using them for their sexual fantasy, so

00:44:04.900 they had to kind of change the settings,

00:44:06.900 if you will, and reboot and kind of make

00:44:08.500 sure that their avatar didn't include the

00:44:11.420 chatbot abilities that would kind of

00:44:13.380 relate to any kind of intimate

00:44:15.300 relationships.

00:44:16.720 But it's also true, of course, from the

00:44:20.380 point you were making.

00:44:21.380 By the way, well done on leaving Twitter.

00:44:23.580 There's nothing left there, so I'm sure

00:44:26.560 your life is a lot better for that.

00:44:29.380 But if you think about radicalization and

00:44:32.420 online radicalization, this is actually

00:44:34.820 something that I already read in a paper

00:44:37.700 years ago, because it was a piece of

00:44:40.160 research done by the Middlesbrough Institute

00:44:42.260 of Terrorism, and it was looking at how

00:44:45.340 an early forebear to chat to GPT-4, I think

00:44:49.740 it was GPT-3, they had tested it, and

00:44:53.280 seeing how it performed as a radicalization

00:44:57.340 agent, right?

00:44:58.680 And we know so many people are radicalized

00:45:00.940 online.

00:45:01.860 Now, imagine, we're just talking right now

00:45:04.500 about the capabilities of very sophisticated

00:45:06.940 chatbots that are going to become even

00:45:09.180 more sophisticated and be able to fulfill

00:45:11.620 your every sexual fantasy, or to be able

00:45:14.020 to groom you, to radicalize you.

00:45:17.380 And the next step, when we talk about the

00:45:19.320 capabilities of these generative models, is

00:45:21.440 so-called multimodal models, right?

00:45:23.660 Because right now, they're still kind of

00:45:25.280 broken up the foundational models into the

00:45:27.860 type of content that they generate.

00:45:29.360 So you have large language models, you have

00:45:31.240 image generators, you have audio generators,

00:45:34.500 you have the kind of growing video generators,

00:45:37.900 although they're obviously not as sophisticated

00:45:39.500 as the text or image yet.

00:45:41.720 But the multimodal is when you can work with

00:45:44.060 all those digital medium in one.

00:45:45.700 So hypothetically, you can, if we're going to

00:45:49.320 go back to the kind of virtual girlfriend

00:45:51.880 scenario, you know, you can not only chat to

00:45:54.340 her, but you can see her, you can have photos

00:45:56.700 generated of her.

00:45:58.300 Similarly, if we go back to the grooming kind

00:46:01.160 of scenario, you know, you're being shown

00:46:03.120 video, you're being shown audio, whatever your

00:46:05.780 worldview is, can be entrenched.

00:46:08.240 So these are some of the darkest kind of

00:46:10.640 manifestations of this hyper personalization.

00:46:13.180 On the kind of more benign side, I think people

00:46:17.440 in the entertainment world are very excited

00:46:20.060 about it, because they're like, oh, this is the

00:46:21.840 ability to create an audience of one, you know.

00:46:24.980 So if you like the Harry Potter books, and you

00:46:28.940 want to find out more about Dumbledore, you can

00:46:33.040 say, I want to have some more backstory generated

00:46:36.780 for Dumbledore, and I want to know about where he

00:46:39.220 was born, and, you know, what was his mother's

00:46:42.480 backstory, and it could just hypothetically generate

00:46:44.900 that for you in real time.

00:46:47.040 But one area where this hyper personalization is

00:46:50.400 really promising is in medicine.

00:46:53.280 And interestingly, we talked about chatbots, and

00:46:55.880 there have been some interesting trials on kind of

00:46:59.500 using a chatbot as almost a assistant, or a friend, or

00:47:05.040 a voice to people who are anxious or depressed.

00:47:07.780 And the early kind of indicators have been that it

00:47:10.620 can be helpful.

00:47:12.220 Now, I don't know, I guess it's a philosophical point,

00:47:15.260 whether you think you should treat people with mental

00:47:18.640 health issues with a chatbot, you know, whether that is

00:47:22.160 the benefits outweigh the risks, that's not for me to

00:47:26.060 decide.

00:47:26.880 But in terms of like the hyper personalization of

00:47:29.480 potential medical treatment for people based on their own

00:47:33.740 data, that's something that I am really interested in.

00:47:36.320 Yeah, yeah.

00:47:37.380 I mean, when you add, you know, full genome sequencing to

00:47:41.260 that, it gets very interesting.

00:47:43.640 On the multimodal model front, do you have a sense of how

00:47:48.580 far away we are from the all-encompassing tool where, you

00:47:52.200 know, you could sit down and say, give me a 45-minute

00:47:55.680 documentary proving that the Holocaust didn't happen, using

00:48:00.620 all kinds of archival footage of Hitler and everyone else, and

00:48:05.760 make it in the style of a Ken Burns documentary.

00:48:10.420 Yeah.

00:48:10.640 And with that prompt, it'll spit out a totally compelling 45-minute

00:48:16.240 video with essentially perfect fake sourcing of archival imagery

00:48:22.920 and all the rest.

00:48:24.100 So, I think to have like a 40-minute completely synthetically

00:48:28.760 generated video, we're still a way off that.

00:48:33.900 But having said that, you know, I've been kind of working a lot

00:48:37.320 in the research community.

00:48:38.420 If you'd like to continue listening to this conversation, you'll

00:48:42.900 need to subscribe at SamHarris.org.

00:48:45.460 Once you do, you'll get access to all full-length episodes of the

00:48:48.120 Making Sense podcast, along with other subscriber-only content,

00:48:51.860 including bonus episodes and AMAs and the conversations I've been

00:48:55.600 having on the Waking Up app.

00:48:57.400 The Making Sense podcast is ad-free and relies entirely on listener

00:49:01.040 support.

00:49:01.460 And you can subscribe now at SamHarris.org.

Making Sense - Sam Harris - July 06, 2023

#326 — AI & Information Integrity

Episode Stats

Length

Words per Minute

Word Count

Sentence Count

Misogynist Sentences

Hate Speech Sentences

Summary

Transcript