Latest / The Road to Accountable AI / Azeem Azhar: Exponential Challenges of Exponential Technologies

Transcript

Azeem Azhar: Exponential Challenges of Exponential Technologies · The Road to Accountable AI

• 0:01
  Azeem, welcome to the podcast.
• 0:03
  Kevin, it's great to see you again.
• 0:04
  I think it's getting into our 25th year of knowing each other.
• 0:08
  This is very frightening and you've done a lot of different things, worked for
• 0:11
  various companies, started various companies.
• 0:13
  How do you explain to your parents what you do now?
• 0:16
  Yeah, well, I, they're quite excited because, or my mum is now, my dad passed
• 0:20
  away last year, but he was even excited because I spent my time researching ideas
• 0:25
  on my own with my, well, with my team.
• 0:27
  And I have a couple of academic affiliations, research them, thinking
• 0:31
  about the intersection between this rapid technology change, things like artificial
• 0:37
  intelligence and societal change from the economy to our legal institutions and
• 0:42
  systems of government governance.
• 0:44
  And then I bundled that up.
• 0:45
  produce a newsletter from it, write the odd book, help the odd investor or
• 0:50
  company, maybe with a speech here or there.
• 0:52
  So I'm, I guess, a one-man band, but I do have an amazing team of researchers who
• 0:59
  work with me.
• 1:00
  Yeah, it's one of those jobs that probably didn't exist before, but now you can be a
• 1:05
  real force multiplier, which is wonderful.
• 1:08
  So let me ask you, when did AI first come on your radar screen as a technology you
• 1:13
  should focus on?
• 1:14
  Obviously, AI has been around for a very long time, but at what point did you
• 1:18
  notice that there was something really important going on?
• 1:21
  You know, I actually, I mean, my first AI deal was done in 2000 to 2001.
• 1:29
  So yes, it's been around for a long time, but then there was this additional shift
• 1:33
  because of deep learning.
• 1:35
  I first came across it in about 2012 or 2013 when someone in my team, my head of
• 1:43
  machine learning, wants to take a job at some funky London startup called
• 1:49
  Deep mind.
• 1:50
  I don't know if you've heard of them and I persuaded him not to do that.
• 1:54
  And, and then when we sold my company, I had some time to think, and then I, I dug
• 1:59
  into things and I saw something was happening and that was about 2014.
• 2:03
  So about a decade ago.
• 2:05
  What's changed in terms of either fundamentally AI or how we think about AI
• 2:11
  today?
• 2:12
  you know, the technology has really, really matured.
• 2:17
  And what happens in the process of maturation is things that become so hard,
• 2:23
  that only the very, very best researchers in the best research institutions can make
• 2:28
  sense of them.
• 2:29
  And then over the course of four or five years, they moved to being available as,
• 2:34
  you know, an executive ed class for a couple of thousand dollars at a business
• 2:39
  school, right?
• 2:40
  And that is...
• 2:41
  That is how a technology deploys into the economy.
• 2:44
  So a lot of those things have happened with deep learning based technologies and
• 2:50
  traditional machine learning systems like classification and discrimination and so
• 2:56
  on.
• 2:57
  But the other thing that has happened of course is that the waves of discovery have
• 3:01
  continued.
• 3:03
  Alongside that, the underlying infrastructure, the enabling technologies
• 3:09
  have become
• 3:11
  much, much more capable.
• 3:13
  And that means that for a business, they need to have less expertise than perhaps
• 3:18
  they did a decade ago to get into these technologies.
• 3:21
  And they're supported of course, by the systems integrators, the consultants who
• 3:26
  are the ones who really drive enterprise IT spending.
• 3:31
  And so what we see, which I think is pretty fascinating is not just, you know,
• 3:36
  maturation,
• 3:39
  of the technology that makes it more accessible to people, but also this
• 3:43
  continued drumbeat of absolute cutting edge breakthrough innovation to the point
• 3:50
  that, you know, you and I have been in the tech industry for a while.
• 3:55
  It is very much very sort of moments of Lenin, you know, decades are happening in
• 3:59
  weeks, sometimes in the same day.
• 4:01
  I mean, you finished reading one tweet and something else enormous has happened while
• 4:06
  you're reading those 200 odd characters.
• 4:09
  But people in the tech world always like to say that this is the transformative
• 4:13
  time.
• 4:13
  This is the destructive moment.
• 4:14
  This is the inflection point.
• 4:17
  Are we really at one of those points?
• 4:18
  What is it that makes you think that now it's really true?
• 4:23
  We always say that because it's self-serving.
• 4:26
  And what's different, I think, this time is the, you know, let's look at a
• 4:32
  corporation.
• 4:33
  If you've got a, the C-suite, the C-suite are people who are likely in their mid-50s
• 4:39
  through to their early 60s, which means most of them have had careers using
• 4:44
  computers.
• 4:45
  And they've had computers on their desks and they know how to use Excel.
• 4:48
  They probably did some time doing PowerPoint.
• 4:51
  And so they're much more computationally savvy than C-level execs, even when the
• 4:58
  iPhone came out.
• 5:00
  And they've had 20 or 30 years to digitize their businesses, which is what we've been
• 5:05
  telling them to do since forever.
• 5:07
  We're out of that moment of J-curve that you see when you have to install a new,
• 5:13
  implementing technology.
• 5:14
  So what that means is that companies are ready to deploy this.
• 5:19
  But let's then talk about the technology.
• 5:21
  So what happened with large language models was we turned computers into that
• 5:28
  scene.
• 5:28
  I think it's Star Trek IV when they're trying to move the whales.
• 5:33
  And Scotty picks up the mouse and says, computer, design me a da-da-da-da-da.
• 5:38
  We finally got there.
• 5:39
  We can talk language to computers, and computers talk language back to us.
• 5:43
  And since humans essentially have
• 5:48
  construct their entire world through the passing of messages in the form of
• 5:54
  language, that has allowed us to bring the capability of computers into those
• 6:01
  moments.
• 6:02
  And one of the things that is so beautiful about using something like ChatGPT is that
• 6:09
  in the old days, so November 2022,
• 6:15
  If you wanted to do a particular type of analysis, you had to configure that
• 6:20
  analysis into a way that a computer understood.
• 6:23
  So you took your mental model of the world.
• 6:26
  And that's why some people are good at programming and others are less good,
• 6:29
  because those people who can convert this sort of squishy human world into, you
• 6:34
  know, flawless logical code, do so and they do that translation.
• 6:38
  But now with ChatGPT, I can say, hey, listen, I kind of want you to do a MISI
• 6:44
  analysis.
• 6:45
  on this list of issues, do it for me.
• 6:48
  And it comes back and it presents as well as a grad student, which is, I think is
• 6:53
  pretty impressive.
• 6:54
  So I think those two things lend us to think that there are the ingredients for
• 7:02
  an inflection.
• 7:05
  So my main focus in this podcast is on responsibility and accountability for AI.
• 7:10
  Let me just first ask you, generally, how do you think we're doing in those
• 7:15
  dimensions?
• 7:17
  Can we mention that this is the recording after the Google Gemini fracas, which
• 7:23
  has...
• 7:24
  Okay, so we'll get to that.
• 7:26
  I think we are doing...
• 7:30
  The awareness of the issue is much, much higher than it was seven or eight years
• 7:36
  ago.
• 7:36
  I mean, I remember in 2015, writing about all these issues with Google images and ML
• 7:42
  systems that couldn't recognize dark skin.
• 7:47
  There were no teams, there was no discipline, there was not much academic
• 7:51
  study.
• 7:53
  We didn't do the kind of evaluations either a priori or ex ante, pardon me, or
• 8:00
  ex post of the systems.
• 8:04
  I think that has changed over the last seven or eight years.
• 8:09
  That's worth paying attention to.
• 8:14
  In 2014, the reason these things weren't...
• 8:16
  done was because people didn't know how to do them.
• 8:18
  No one really even thought to do them.
• 8:22
  And so the systems behaved the way they did.
• 8:23
  Now, after a lot of work by civil society, by academia, by activists, we do think
• 8:31
  about these things.
• 8:33
  And then the question is, of course, it's often embedded into sort of the belief
• 8:38
  systems and the narratives in the world at the right time, right?
• 8:41
  And just so, so I think that some of the things we've seen have been
• 8:45
  whatever the rung above the ladder, above the, the rung on the ladder, above the
• 8:51
  idea of performativity is, we're getting to that stage with some of this where
• 8:57
  people beat the responsibility drum really, really hard, perhaps harder than
• 9:03
  it needs to be beaten.
• 9:04
  I think Sundar Pichai at Google is somebody who does that a little bit more
• 9:08
  than perhaps he needs to.
• 9:12
  Yeah, so let's talk about this example with Google Gemini, where for those who
• 9:16
  aren't familiar, they built an image generator model and a large language
• 9:20
  model.
• 9:21
  And addressing exactly those issues you're talking about, concerns about bias, went
• 9:26
  too far.
• 9:26
  That in some cases, the image model would produce literally non-historically
• 9:31
  accurate pictures of people.
• 9:33
  The text model would just refuse to answer totally anodyne questions, saying they're
• 9:37
  too controversial.
• 9:38
  So.
• 9:38
  People have different reactions to this.
• 9:40
  What's your takeaway or main takeaways from that incident?
• 9:44
  Well, it's complicated.
• 9:48
  I mean, I think the absurdity of some of the outputs of Gemini can't be denied,
• 9:57
  right?
• 9:57
  So the idea that you might say, give me a picture of soldiers, German soldiers from
• 10:05
  World War II, and it doesn't come back with them looking at all how they should.
• 10:09
  Or one I saw on Twitter recently, which said,
• 10:13
  who caused more deaths, Elon Musk's tweets or Adolf Hitler.
• 10:16
  And Gemini sort of prevaricates and ums and ahs and says it's an ambiguous
• 10:22
  question, which is just sheer absurdity.
• 10:25
  So I think that it highlights two issues.
• 10:28
  The first issue is that large language models are very difficult to control.
• 10:36
  They are not like machines that we have built in the past.
• 10:40
  And so in the moment of
• 10:43
  trying to control them, you are choosing between helpful and harmless.
• 10:49
  And there seems to be right now on the methods that are used by these companies,
• 10:54
  really a trade off between the two.
• 10:56
  But it's a bit more complicated than that because you don't exactly know how much
• 11:01
  helpfulness you have to give up in order to make it harmless.
• 11:04
  And in a situation like that, I think you will always, I mean, use a British phrase
• 11:08
  here, over egg it, right?
• 11:09
  You'll always err on the side of caution.
• 11:13
  in order to avoid the harmful nature of it.
• 11:15
  And then you sort of end up slightly in this absurdity.
• 11:18
  But I think there's a second issue, which is that, you know, we have to recognize
• 11:23
  where we, that the world in which we currently live, not the one we want to
• 11:26
  live in, the world we live in is mediated by a small number of large technology
• 11:31
  companies.
• 11:32
  And that means that the AI they build will become our interface between us and the
• 11:38
  real world and the resources that we want to.
• 11:42
  access.
• 11:43
  And so somebody has gone off and stuck this resource in between me and the world.
• 11:47
  I mean, you have to go to Gemini today, but in two years time, everything you do
• 11:51
  on Google will be filtered through it.
• 11:54
  And they've gone off and made certain decisions.
• 11:57
  It's harmless, helpless trade-off, other decisions in particular about what they
• 12:00
  consider harm.
• 12:02
  And to me, this is a public infrastructure.
• 12:06
  This is a public infrastructure.
• 12:10
  It's a utility.
• 12:11
  Maybe it's a type of good, public good.
• 12:14
  And Google did talk a little bit about what went wrong, but not in the degree of
• 12:19
  sufficient transparency of this new role they have of guarding something that is
• 12:26
  not purely a private asset.
• 12:28
  So I think that there are these two issues that we have to have to talk about and
• 12:31
  make sense of.
• 12:33
  Okay, no, that's a good point.
• 12:34
  And that's, I want to talk about the issue about the power of the intermediaries and
• 12:38
  the big tech companies.
• 12:39
  But let me just first push you a little bit more on the question about the
• 12:44
  capability of controlling the models.
• 12:46
  Your book is called exponential age.
• 12:49
  And you talked about how technologies like AI and rootables are on these exponential
• 12:53
  growth curves.
• 12:55
  Is there a concern that the improvements in the technology are an exponential curve
• 13:00
  but our ability to control them?
• 13:04
  are something like arithmetic, and I'm not just talking about here in terms of the
• 13:08
  broad kind of alignment that the AI doesn't kill us all when it becomes super
• 13:12
  intelligent, but even on what seems like a fairly mundane problem, like not having
• 13:17
  biased image generation.
• 13:19
  Is that potentially an unsolvable problem?
• 13:22
  So I think that this isn't a case of the type that I describe in my book.
• 13:28
  I mean, the ones I don't describe in my book are really about how the state and
• 13:34
  the institutions of the state and the people who comprise that are able to
• 13:38
  respond to this pace of change.
• 13:40
  I think what we're seeing in Google Gemini is much, much closer to...
• 13:47
  what happened in the early years of jet powered passenger planes.
• 13:53
  So these emerged in the early 1950s.
• 13:56
  And the first of one of the first of those was made by a British company.
• 14:00
  We don't have names like this for jet aircraft companies anymore.
• 14:03
  De Havilland.
• 14:04
  It's so elegant.
• 14:04
  It's straight out of Downton Abbey.
• 14:07
  The De Havilland Comet.
• 14:10
  And I think something in the order of 40 of these crashed.
• 14:14
  I mean, they make Boeing look like, you know, an absolutely
• 14:17
  pristine whiter than whites safety record.
• 14:20
  And about 40 of these crashed, though 20 of them had fatalities and about 500
• 14:25
  people died when these planes kept falling out of the sky.
• 14:29
  And the reason was that they had square windows and the corners of the windows at
• 14:37
  jet speeds in the thinner atmosphere heated up much more, created lots more
• 14:42
  metal fatigue than they'd ever experienced in propeller planes that were a lower
• 14:46
  height.
• 14:47
  traveling more slowly.
• 14:48
  And they couldn't do, you know, they didn't do have computational fluid
• 14:50
  dynamics, right, to simulate all these things.
• 14:53
  And, and, and we figured that out.
• 14:55
  And yes, 500 people died and air travel is now so much safer.
• 14:59
  Now, I think that we are that's partly what we're dealing with within the world
• 15:06
  of the large language models, right, which is that these are these will be made safe
• 15:11
  in some way, they will become progressively safer.
• 15:13
  Now, not everyone agrees with me, there's a
• 15:15
  an academic called Gary Marcus, who I think would take a different view to that
• 15:19
  and say that's kind of really hard, if not impossible, to make them safe and they're
• 15:23
  poorly designed.
• 15:24
  But I think they can be made safe and we will learn how to do it.
• 15:29
  Now, the challenge is that Google is in a race.
• 15:34
  It's in a race that is both strategic, that in some sense is almost existential,
• 15:40
  that is commercial, that is also about...
• 15:44
  corporate and personal pride, which is driving them really fast.
• 15:48
  And we have constructed a world where we are looking for some hype to drink from as
• 15:58
  consumers of the media.
• 16:00
  And so this pushes this process of actually figuring out how to make these
• 16:06
  things safe out into the limelight much, much faster.
• 16:10
  And so in some ways, I'm a little bit sanguine about the fact that this...
• 16:13
  happened.
• 16:14
  I'm kind of annoyed.
• 16:15
  I feel it should have been caught in testing.
• 16:17
  But I think it is a case of it's this technology is, you know, it's seven years
• 16:22
  now since the first academic paper.
• 16:24
  It's, it's, it's really early.
• 16:27
  And to be honest, honestly, no harm was done, except for Google.
• 16:32
  Google is embarrassed.
• 16:34
  It shut the thing off.
• 16:36
  It'll figure out a fix.
• 16:37
  We won't see it.
• 16:37
  Is anyone really missing having an AI based image generator for an extra few
• 16:41
  days?
• 16:42
  None of us are.
• 16:45
  But won't that competitive dynamic you talk about continue?
• 16:48
  Because the company that has succeeded the most so far is OpenAI, partnered with
• 16:52
  Microsoft, which took the kind of Mark Zuckerberg, move fast and break things
• 16:56
  approach.
• 16:56
  While Google waited to deploy its LLM because it was concerned about all of
• 17:01
  these issues, if that's always going to be the case, isn't there going to be
• 17:05
  constantly commercial pressure to push the envelope and not even to take the time to
• 17:10
  develop the solutions that are possible?
• 17:13
  That's business.
• 17:14
  I mean, when we read the stories of what business people did over the last 100, 150
• 17:23
  years, I mean, they make trade-offs and they have to apply their judgment because
• 17:30
  nothing is perfect, no one has perfect resources.
• 17:32
  They're always competing needs.
• 17:34
  There's also always ego and selfishness and profit motive in it.
• 17:39
  It's just, it's happening in...
• 17:42
  a in a public space, which, of course, the business people behind them, and that's
• 17:49
  what they are, let's not call them founders, let's just call them business
• 17:51
  people, they're just like the green grocer down the road.
• 17:55
  They benefit from the publicity that it's that is being created by all of this, and
• 18:02
  they have to make difficult choices.
• 18:06
  I'm
• 18:07
  I would say, having spoken to him a few times, but not knowing him because he's
• 18:10
  very hard to read, Sam Altman, I would say that Sam is extremely, probably very keen
• 18:18
  to make all his products factually accurate and generally pretty safe, not as
• 18:25
  safe as a nervous Nick might want, but pretty safe.
• 18:31
  And he's just waiting for them to develop.
• 18:37
  those capabilities.
• 18:38
  And he at the same time, at that moment, he wants to go out and win market share.
• 18:42
  And what's happening is it's happening in this public domain, with all the noise by
• 18:46
  the way, Kevin of this idea that these LLMs will crawl out of our fridge, like if
• 18:51
  you've seen the film Gremlins, and you feed them and give them water, and they
• 18:54
  turn really nasty, you know, like that and take over our world, which again, I don't
• 18:59
  think they will.
• 18:59
  So I think we're in this funny stage where we should just kind of look at this and
• 19:03
  calm down a little bit and say,
• 19:07
  And in that calmness, it can still be a transformative technology.
• 19:12
  I mean, that can all be true at the same time.
• 19:17
  One of the important things about this wave of technology is that we have these
• 19:20
  big platforms that are building the AI systems.
• 19:23
  But every company on Earth probably is going to use them in one way or another
• 19:28
  with some amount of either their own technology or just taking things off the
• 19:32
  shelf and deploying them.
• 19:33
  You talk to executives of a lot of companies about AI.
• 19:37
  How are they thinking about these questions of responsibility and accuracy?
• 19:43
  And what does it take to put them
• 19:45
  persuade them to make the right trade-offs.
• 19:48
  I think in general, they take it very seriously.
• 19:52
  And it comes up surprisingly unprompted in the conversations that I have.
• 20:00
  And people talk about their responsible AI teams.
• 20:03
  Or they'll say, well, we have to roll this out responsibly.
• 20:07
  I mean, there are some variations.
• 20:10
  And for some, it's a.
• 20:13
  responsibility with a bit of a wave of a hand, say, don't talk to me about it with
• 20:17
  others, it's more, more engaged.
• 20:19
  But I think that comes back to a little bit earlier in our discussion, where I
• 20:22
  said that the way things have moved over the last seven or eight years, is that
• 20:28
  companies are taking this quite, quite seriously.
• 20:35
  And perhaps, you know, the challenge is that how much of that is, is simply just
• 20:41
  lip service.
• 20:43
  and how much of it is really, really deeply embedded in what companies try to
• 20:49
  do.
• 20:50
  And of course, the thing is, Kevin, how much do we want our companies to do?
• 20:57
  The founder of, and suddenly the name escapes me, the outdoor wear company,
• 21:05
  Patagonia, perhaps, is that?
• 21:08
  Who was sort of famous, he says, business will...
• 21:10
  you know, business will always leave a footprint, right?
• 21:12
  And we can't pretend it won't.
• 21:14
  And our job is to try to make it as small as we can.
• 21:17
  And I don't think it's necessarily a company's job to act responsibly or act
• 21:27
  ethically.
• 21:28
  That is the job of, I mean, it's the job of the law and whatever is delegated to
• 21:34
  regulators to say, these are the parameters within which you should act.
• 21:38
  Now you should act in that way and comply to this law.
• 21:41
  Otherwise we're going to come in and find those breaches and we are going to enforce
• 21:46
  actions against you.
• 21:47
  So yes, I'm excited that CEOs care about it and they're not spending their time
• 21:51
  trying to figure out how to kind of wind their way out of their obligations.
• 21:56
  But I think it's a bigger issue that those obligations are not clear.
• 22:00
  They're not clearly defined.
• 22:01
  And so that the companies really know what the rules of the game are.
• 22:07
  Yeah, that gets back to the point that you made earlier and you talk about in your
• 22:10
  book.
• 22:11
  Is there going to be enough capability in government to make those rules
• 22:16
  appropriate?
• 22:18
  Well, I mean, the government's getting better.
• 22:23
  It's definitely getting a lot better.
• 22:26
  And we, I think we tend to.
• 22:31
  have a slightly sneery view of government, but we can go back to the first consumer
• 22:36
  privacy legislation which came out in the late, early 1970s in California on the
• 22:42
  basis of the arrival of computer databases.
• 22:45
  I mean, that is phenomenally far-sighted.
• 22:47
  I mean, if that person had gone off and said, computer database is gonna be huge,
• 22:52
  I'm gonna build a company, they would be, you know, yachts are plenty by now.
• 22:57
  And in other areas, governments have been
• 23:01
  pretty impressive.
• 23:02
  I'll give you another example.
• 23:05
  My house was built in 1925.
• 23:08
  The cable that carries the electricity is rated at 100 amps and is still more than
• 23:16
  my house uses a century later.
• 23:18
  We've got two electric cars and we've got all the electronics you might imagine.
• 23:30
  genetic modifications in humans, right, was well ahead of the technology.
• 23:35
  But we have, governments, I think, started to slow down for historical reasons.
• 23:42
  And in my book, I talk about Reagan and the air traffic controllers and so on.
• 23:48
  But of course, that process of deregulation had started under Carter.
• 23:53
  And if you come at the back end of 25 or 30 years where
• 23:56
  The notion of good in the context, the intersection between government and
• 24:00
  business is about pulling regulations back, as opposed to recognizing that it is
• 24:06
  a presence of clear regulations, clear enforcement, a mandate for compliance that
• 24:13
  has allowed the pharmaceutical industry and the financial services industry to
• 24:17
  grow as big as they have.
• 24:18
  Of course, we're always going to feel that government is going to not have that
• 24:22
  capability.
• 24:23
  Now, I do think that they have really...
• 24:25
  moved and I can speak about the UK, I mean the quality of people in policy, that is
• 24:32
  the civil service who will brief government on these issues is really deep.
• 24:38
  Now I just don't know if there are enough of them, but they're getting really quite
• 24:43
  good and I wouldn't sort of besmirch them in any way.
• 24:47
  Our government on the other hand, maybe that's for an offline discussion.
• 24:51
  enough.
• 24:52
  But with regard to AI, different governments are taking different
• 24:57
  approaches.
• 24:58
  And so you mentioned the UK.
• 25:00
  It seems that actually this is one set of issues where Brexit has actually made a
• 25:04
  pretty dramatic difference.
• 25:05
  On a lot of things, there's not a great deal of substantive difference between
• 25:09
  what's happening, you know, the laws in the UK and Europe.
• 25:11
  But the European Union has adopted is in the final stages of this very extensive,
• 25:17
  very prescriptive AI act.
• 25:20
  UK government, at least from my perspective, more like the US, has taken
• 25:24
  more of an approach of saying, let's look at what existing laws out there, let's try
• 25:28
  and figure out and have different kinds of rules and how to balance innovation with
• 25:35
  regulation.
• 25:36
  Seems like a very different approach.
• 25:37
  So first of all, I guess, do you see that?
• 25:39
  And then what do you think is the best approach for governments to take?
• 25:44
  I definitely do see that.
• 25:46
  And I certainly see that gap emerge.
• 25:54
  And there was something else that was about the sense of victory that was coming
• 25:58
  out of the various parts of the EU.
• 26:02
  And we haven't created a $100 billion company since SAP, and now Novo Nordisk,
• 26:08
  which has got rich on Americans.
• 26:11
  having an obesity problem, right?
• 26:13
  I mean, because they make semaglutide, the GLP-1 agonist.
• 26:17
  And so it's a really odd thing to celebrate.
• 26:25
  The other thing is it's a really big act.
• 26:28
  And so I think that means that in a way, some parts of it may have been hurried,
• 26:37
  especially the bits that relate to the...
• 26:41
  large language models and those uses.
• 26:44
  But in reality, that has to now go through processes which will detail how things
• 26:49
  will get enforced.
• 26:53
  So we know that really bad laws can be created.
• 26:57
  I think, for example, many people probably view the Patriot Act hastily put in place
• 27:03
  in the US after 9-11 as something that has resulted in some overreach and might need
• 27:07
  some pairing back.
• 27:11
  But I think that the issue that I have with the EU Act is less about the
• 27:21
  specifics of the detail, but it was more about the posture of the participants
• 27:28
  towards this remarkable technology and what was really embedded within that,
• 27:35
  which was, I think, a bit of a distrust with American firms.
• 27:40
  but there wasn't within it a sense of, we'll create an alternative.
• 27:46
  And so then the question is what actually happens over this period of time.
• 27:49
  So I think in these moments of really, really rapid change, you need to have
• 27:58
  principles-based approaches to regulation, right, rather than hard and fast rules.
• 28:04
  So I would say that that's probably more a British style than...
• 28:09
  you know, an American approach and you need that because the underlyings change
• 28:18
  so frequently.
• 28:18
  So I think that feels like it's sensible.
• 28:21
  But the other thing that I think will end up happening is that there will be an
• 28:26
  increasing harmonization of the ways in which nations look at certainly economic
• 28:37
  and industrial applications of
• 28:39
  of AI.
• 28:40
  And the reason I think that they'll start to harmonize will be because we're in a
• 28:45
  search process of what method works best.
• 28:49
  Every country will benefit if technology companies can access their markets on
• 28:56
  equal terms.
• 28:57
  So that will also start to drive some degree of consistency across the piece.
• 29:05
  Where I think it may differ
• 29:08
  will be, and this is probably going to be quite a large part of AI to be honest,
• 29:12
  aspects that connect to culture, identity, or potentially national security issues,
• 29:20
  which was also the first sticking point, of course, with the internet, where
• 29:26
  countries sort of started to cleave themselves away.
• 29:31
  But I think what we saw with privacy was that there was a lot of harmonization.
• 29:35
  So basically every global company is complying with the European GDPR either
• 29:39
  because they have operations in Europe or because there's just no point in building
• 29:43
  two completely different systems.
• 29:45
  But I don't think anyone, even the people in the EU, in the European Commission,
• 29:50
  would say things are great on privacy.
• 29:53
  The practices have tremendously improved because we have this agreement on this
• 29:59
  comprehensive set of standards.
• 30:01
  Mm.
• 30:02
  How do we get to a point where the actual substance improves, which, as you noted,
• 30:08
  companies, they want to hear government telling them what to do, but they have a
• 30:12
  role in helping governments understand what the capabilities are and what could
• 30:16
  happen.
• 30:18
  You know, I think a really great example of that was what happened with Apple and
• 30:24
  the App Store.
• 30:25
  So under the Digital Services Act, no, the DMA, the Digital Markets Act regulation,
• 30:34
  because Apple operates this App Store that is sort of so big and so dominant, it has
• 30:40
  to provide alternative means of payment and alternative App Stores.
• 30:44
  So if you're in the EU and you have an...
• 30:46
  and I felt this is going to be all to American listeners, you can kind of get
• 30:49
  apps from a different store run by someone else and you don't have to pay through
• 30:52
  Apple.
• 30:54
  And of course, Apple has gone off and complied with this and involved something
• 30:57
  like writing 500 new APIs, which are kind of quite complex ways in which you use
• 31:02
  computer services over the internet.
• 31:06
  And they introduced new pricing mechanisms because what they've done is they've
• 31:11
  absolutely complied to the letter of the law.
• 31:14
  as you'd expect because they can afford good lawyers, and people are arguing they
• 31:17
  haven't complied to the spirit of the law.
• 31:20
  And this is the old problem of the sorcerer's apprentice, right?
• 31:23
  The Goethe, the old Goethe short story, right?
• 31:28
  Where we've got this technology, we try to control it, we give it rules, and it
• 31:33
  follows into the letter.
• 31:35
  And so I think that raises issues about when and how is the right time to...
• 31:44
  to write these rules and how do you get people sitting around the table at the
• 31:51
  same time?
• 31:52
  Now, I was actually quite a supporter of the Digital Markets Act because my framing
• 32:01
  is that there is a tendency towards dominance in digital markets for reasons
• 32:08
  that I initially probably Kevin learnt from you 25 years ago.
• 32:14
  And that force of gravity continues.
• 32:19
  And because there's this tendency to dominance, you will have concentration of
• 32:25
  power.
• 32:25
  So you need to find mechanisms to be more onerous with bigger firms than smaller
• 32:31
  ones, because the bigger ones are so much bigger.
• 32:33
  So I was at a high level, I was a supporter of it.
• 32:37
  I think that the...
• 32:40
  issue is, and there's another analyst called Benedict Evans, who I think
• 32:44
  probably made a better call on this than I did, where he said, well, the problem is
• 32:50
  that even if Apple is dominant, the process is safe and it's easy on the App
• 32:57
  Store.
• 32:58
  Consumers love it.
• 32:59
  No consumer is going to go off the App Store to the dodgy flea market with men in
• 33:04
  gray coats of saying, want to buy an app, when they can safely buy an app from the
• 33:08
  App Store.
• 33:10
  I think has a lot of momentum behind it as an idea.
• 33:13
  And that I think brings to bear the fundamental issues of capability within
• 33:23
  government and who did you have the right sorts of people who really were able to
• 33:28
  have the right kind of dialogue to get around the problems of corporate capture,
• 33:33
  which is what we saw with GDPR and have that conversation.
• 33:37
  And I don't think we did.
• 33:39
  So.
• 33:39
  So, I mean, that's a very long answer to say, it's quite hard.
• 33:49
  I'm not sure it's necessarily gonna serve European customers particularly well.
• 33:57
  I think what's hidden slightly below this is that these are all American companies.
• 34:01
  I mean, the rules are not written to discriminate against American companies,
• 34:05
  but it just so happens they're all American firms and will...
• 34:10
  continue to be if what we do is celebrate passing laws rather than building
• 34:13
  companies.
• 34:15
  One more question.
• 34:18
  You're pretty optimistic, it sounds like, both about the path of the technology, the
• 34:21
  business option, and working through these regulatory issues.
• 34:25
  What do you worry about?
• 34:26
  Where might you be wrong about any of these aspects that you think about?
• 34:31
  That's something that might put us on a bad path.
• 34:36
  You know that there are really interesting things that have happened that I wouldn't
• 34:41
  have predicted, like that a trained scientist running a major nation would
• 34:47
  shut down the nuclear program, as Angela Merkel did.
• 34:52
  I could run that tape a hundred times.
• 34:54
  I would always say she's not going to do it, and she did.
• 34:57
  So I guess I get...
• 35:04
  slightly concerned about the growing dominance of these firms.
• 35:14
  I mean, less so Nvidia.
• 35:15
  I mean, they're going off and making a lot of money because they're so far away from
• 35:18
  the customer.
• 35:20
  But the power and the capability that these companies will increasingly have,
• 35:26
  and then what that means in terms of their ability to
• 35:32
  control what's left of the public square and to do so in ways that aren't really in
• 35:41
  the interests of the people who live there.
• 35:43
  And I think that we've seen some aspects of that play out with the way targeted
• 35:47
  advertising works in the US and Facebook over not necessarily the last couple of
• 35:52
  years, but the years before that.
• 35:55
  I'm always a very simple man.
• 35:58
  And so I simply go back to...
• 36:01
  ideas of concentration of power, where, and wherever I see that processes and
• 36:10
  mechanisms that can reinforce concentration of power, it makes me
• 36:15
  nervous because I can't believe that if you are that powerful, you aren't somehow
• 36:22
  skewing the rules or taking a kind of shifty vig on the side that we don't know
• 36:28
  about.
• 36:29
  So where I could be wrong is that we don't put in things that slow down this
• 36:38
  agglomeration.
• 36:40
  And the other workplace I could be really wrong is whether it's even possible to
• 36:44
  slow it down.
• 36:45
  So some days I wake up and I think, I'm just not sure it's possible to slow it
• 36:48
  down.
• 36:48
  It doesn't matter that Google and Apple are a bit scared of Lena Khan at the
• 36:51
  moment and aren't buying companies.
• 36:53
  I don't think it's gonna slow down how quickly they grow.
• 36:58
  And I could also be wrong about that.
• 37:02
  What I'm not worried about so much is I'm not worried about AI accelerating and
• 37:11
  getting out of control.
• 37:11
  And the reason I'm not is because I think we technologically co-evolve systems to
• 37:17
  keep it safe, even if we can't co-evolve the regulatory systems as quickly.
• 37:23
  But yeah, but that's really...
• 37:25
  That's really about it.
• 37:27
  And of course, Kevin, I'm worried about the elections in the US and elsewhere.
• 37:31
  And I'm worried about, you know, what might happen with China and what might
• 37:35
  happen with Russia.
• 37:35
  I mean, those things are just background noise for 2024, really.
• 37:39
  They're the track to which we're going to be dancing.
• 37:42
  Absolutely.
• 37:43
  No, a lot out there to worry about a lot more that we could certainly talk about,
• 37:46
  but we're out of time at this point.
• 37:48
  Azim Azhar, thank you so much.
• 37:50
  Always great to hear from you and appreciate your sharing your perspectives
• 37:53
  with the road to accountable AI.
• 37:55
  That's been wonderful, Kevin.
• 37:56
  Thank you so much.