WEBVTT
00:00:00.000 --> 00:00:05.599
Humans are kind of lazy, and if the easy option is to let a computer make the decision, then we will.
00:00:40.479 --> 00:00:53.359
You may already know her from her book We Are Anonymous, but she has a new book out chronicling both Sam Altman and Demis Hassabis's Rise to AI fame through OpenAI and DeepMind, respectively.
00:00:53.600 --> 00:00:58.799
It's called Supremacy, AI, Chat GTP, and the race that will change the world.
00:00:58.880 --> 00:01:03.119
And it's just been shortlisted for Business Book of the Year by the FT.
00:01:03.280 --> 00:01:07.120
I've linked to all of her work and her books in the show notes, as usual.
00:01:07.359 --> 00:01:15.439
Now here we chat on the same day that Demis Hassabis's Nobel Prize for Chemistry was announced, with Palmi busily filing her story on that.
00:01:15.599 --> 00:01:17.439
So why not start there?
00:01:23.040 --> 00:01:28.560
Welcome Palmi Olsen to the future of you, and congratulations on your new book, Supremacy.
00:01:28.879 --> 00:01:29.359
Thank you.
00:01:29.599 --> 00:01:30.959
No, I'm excited to be here.
00:01:31.120 --> 00:01:32.079
Thank you for inviting me.
00:01:32.319 --> 00:01:33.760
Oh, it's fabulous to have you.
00:01:33.920 --> 00:01:39.280
Especially right now, it's all so topical, and um, we should give the book its uh proper title, I suppose.
00:01:39.519 --> 00:01:43.359
Supremacy AI, ChatGPT, and the race that will change the world.
00:01:43.519 --> 00:01:53.840
But we've caught you on a day where you are, I'm sure you're always busy, Palmy, but exceptionally busy because you're just writing up, or in the midst of writing up, something about one of the protagonists in your book.
00:01:54.079 --> 00:01:55.519
Why don't you tell us more?
00:01:55.920 --> 00:01:57.280
That's right, yes.
00:01:57.519 --> 00:02:11.360
Well, so the founder and CEO of Google Deep Mind, whose name is Demis Hasabis, is one of the two key protagonists in supremacy, AI, ChatGPT, and the race that will change the world.
00:02:11.599 --> 00:02:20.000
Um, and today he just won a Nobel Prize, the Nobel Prize for Chemistry, and he's sharing it with two other people.
00:02:20.159 --> 00:02:27.680
And this is for the work that he and the team at DeepMind have done on a process known as protein folding.
00:02:27.840 --> 00:02:49.680
And proteins are the building blocks of the human body, and the idea is that by using uh artificial intelligence to predict how these proteins will fold, they can scientists and medical researchers can use that process for drug discovery and make breakthroughs in biology.
00:02:49.919 --> 00:02:54.319
Um, and this is a project that's been going on for about eight years or so.
00:02:54.479 --> 00:03:06.800
And so it's really amazing that it's been recognized and had this incredible accolade, especially because, as I say in my book, Demis has craved a Nobel Prize for years.
00:03:07.360 --> 00:03:17.360
In fact, he even told his staff that he wanted DeepMind to win, you know, three to five Nobel Prizes over the next decade.
00:03:17.439 --> 00:03:20.000
And this was very much how he measured success.
00:03:20.159 --> 00:03:22.400
So that's a huge milestone for him.
00:03:22.560 --> 00:03:37.599
And I think it also adds a really big veneer of credibility to artificial intelligence and particularly to Google, because uh Demis is actually the first big tech executive to have ever won a Nobel.
00:03:37.759 --> 00:04:10.879
And I think that really is gonna, in a way, cast a very useful glow on uh on Google and perhaps even other tech companies who are kind of struggling at the minute to justify the benefits of artificial intelligence as being good for society, being good for humanity, and not just good for them, which is one of the big arguments I make in my book that actually right now the the wealth benefits, the influence benefits really are just being concentrated in the hands of these few very large companies.
00:04:11.280 --> 00:04:11.680
You're right.
00:04:11.840 --> 00:04:13.919
This takes us very nicely into the heart of your book.
00:04:14.000 --> 00:04:23.680
Because one of the things I wanted to ask you, maybe it's a jumping-off point for you to explain to the the audience who who haven't read it yet but should go and buy it and read it, um, because it's excellent.
00:04:24.000 --> 00:04:28.639
Is it a book about the hype versus the doomerism of AI?
00:04:28.800 --> 00:04:40.000
Or is it a biography about two potential geniuses or definitely, you know, amazing entrepreneurs and maybe engineers and scientists?
00:04:40.160 --> 00:04:45.279
Or is it both because it's very interesting, it's quite hard to characterize in that sense.
00:04:45.759 --> 00:04:46.240
Mm-hmm.
00:04:46.480 --> 00:04:51.759
Yeah, it's tempting to say both, but I would steer it more towards being more of a biography.
00:04:52.000 --> 00:05:08.000
Because I think it's a really important way to understand technology, which can seem so abstract and so invisible, it's just there, it's just woven into our lives and our infrastructure, is to remember that all technology is designed by humans.
00:05:08.160 --> 00:05:17.920
And there are people like you use your phone, every widget, every app, every piece of software you use has is a result of a design decision made by a human.
00:05:18.079 --> 00:05:31.839
And I think if we really want to get a sense of where AI is going, and especially the incentives that are fueling the building of AI today, it's really important just to understand the people behind this technology.
00:05:31.920 --> 00:05:47.600
Um, and there are so many people behind AI, but for simplicity's sake, and also because I think these two men are the two most important builders of AI, I did focus it on Sam Altman, the founder of OpenAI, and Demis Hisabas, the founder of DeepMind.
00:05:47.759 --> 00:05:59.199
And the other reason I focused on them was because they were really the two people who were the first to start companies aiming to build artificial general intelligence.
00:05:59.279 --> 00:06:11.120
And this is this really huge, ambitious goal to build AI that is as smart as a human, effectively, uh, that can do the same cognitive tasks as a human brain.
00:06:11.360 --> 00:06:15.759
We're not anywhere near that yet, and there's a big question about whether that's even possible.
00:06:16.079 --> 00:06:20.319
But it was it really started off as this humanitarian dream.
00:06:20.480 --> 00:06:27.360
Like Demis started this be in part in part because he wanted to make scientific discoveries.
00:06:27.519 --> 00:06:35.360
He believed that AI could accelerate scientific breakthroughs and that AI could be and computers could be an extension of the human mind.
00:06:35.600 --> 00:06:37.759
And he saw it as a way to cure cancer.
00:06:37.920 --> 00:06:40.399
If he built AGI, he could solve climate change.
00:06:40.480 --> 00:06:42.160
He was going to solve all these things.
00:06:42.399 --> 00:06:46.720
And Sam Altman, his he was more on the kind of economic benefit of AI.
00:06:46.879 --> 00:06:51.519
He kind of promoted prosperity, like if we have AGI, we'll have abundance.
00:06:52.079 --> 00:06:58.160
He used this word abundance all the time, um, where we'll kind of elevate the wealth of everyone on earth.
00:06:58.399 --> 00:07:03.759
Of course, that that that motivation changed over time, like those those uh objectives changed.
00:07:03.920 --> 00:07:06.560
But yes, that's why I that's that's how I would characterize it.
00:07:06.639 --> 00:07:08.399
It's really a kind of human story.
00:07:08.639 --> 00:07:16.480
And you're right, it is difficult to characterize the book itself, but you know, I I'm not easy to characterize as a person generally.
00:07:16.879 --> 00:07:17.120
Exactly.
00:07:17.279 --> 00:07:19.120
Well, it has different dimensions to it.
00:07:19.199 --> 00:07:25.519
You can come in it through several different ways, and it is a is a complex story, it's told very well, but I think that's nice.
00:07:25.600 --> 00:07:29.759
There's different levels or layers of interpretation or sort of storytelling in it.
00:07:29.920 --> 00:07:33.920
I have to ask you, I'm sure people have asked you before, which of them do you prefer?
00:07:34.240 --> 00:07:47.839
The first time I was asked that was like I was on stage at an event and I hadn't, I was not expecting the question, and I just said, well, I guess I prefer Demis because he he just jet yeah, he does just seem more genuine in his outlook.
00:07:47.920 --> 00:08:03.600
It does seem like he stuck a little bit more to the original principles of trying to and the original objectives of trying to you know use AI to solve health-related uh or scientific discoveries and uh to sort of support science.
00:08:03.839 --> 00:08:19.600
Whereas Sam, I feel like, you know, the idea of using AGI to to improve the wealth of everyone, that just that feels really like it's it's just so theoretical and abstract, and it really just feels much more like a like a marketing point for him.
00:08:19.920 --> 00:08:30.079
But other people kind of come away with different, like I had someone who read it say to me, Wow, I've read this book, and and it really feels like Demis is the bad guy and Sam is the good guy.
00:08:30.240 --> 00:08:32.879
So you can really come away from it with different interpretations.
00:08:32.960 --> 00:08:45.919
But I would say that actually the first half of the book is kind of biographical, and then and then it goes into kind of like the conflict and the tension and the warnings that were coming about this technology they were building.
00:08:46.000 --> 00:08:54.799
And then the final third is actually almost more like a polemic about the the risks of AI and just what the future could hold and why it all matters.
00:08:55.120 --> 00:08:56.639
Yeah, I mean you're right, you're right.
00:08:56.720 --> 00:08:58.639
Obviously it's your book, so of course you're right.
00:08:58.799 --> 00:09:13.360
But um, the first part of it is you are sort of exploring the character and the values of each of those individuals, I think, which then goes on to explain quite a lot of how the company uh companies look and feel and how they turn out.
00:09:13.440 --> 00:09:40.879
But one of the interesting bits in the middle, whilst you are discussing all of those ins and outs and you know, acquisitions and partnerships and all this sort of stuff, particularly with Google and DeepMind, is really interesting, I think, and that does give a lot of illumination into the character of Demis and how much does he want to stick to the original values and the original purpose and the social purpose of DeepMind versus, you know, particularly when you're talking about when they're looking to move into China.
00:09:41.120 --> 00:09:44.480
I think it's after AlphaGo, but maybe you can give a summary of that.
00:09:44.799 --> 00:09:49.600
I I found that a really interesting part of the story, actually, for lots of different reasons.
00:09:50.080 --> 00:10:06.879
Yeah, and I think it's this question about sticking to this, like sticking to the to the social purpose, I think is really interesting because I think DeepMind has done a really and Google have done a really good job of publicly framing themselves as sticking to these objectives.
00:10:06.960 --> 00:10:15.519
When actually, if you look at where they put the investment, it's so much more on just the Gemini AI model product, the product.
00:10:15.679 --> 00:10:31.279
You know, probably one of the best examples is the latest news about uh about AlphaFold, about the protein folding project, because actually the number of people who worked on that at DeepMind for all the publicity that it has received is tiny.
00:10:31.440 --> 00:10:36.320
There's something like 1,500 people who work at DeepMind.
00:10:36.639 --> 00:10:40.399
And actually the number of people who are working on AlphaFold is around 20.
00:10:40.559 --> 00:10:42.000
It's very, very small.
00:10:42.240 --> 00:10:48.159
So I think that the accolades that it has received actually is sort of overshadows.
00:10:48.320 --> 00:10:54.639
It doesn't reflect really where Google is putting all the investment and where DeepMind is putting a lot more of its time.
00:10:55.039 --> 00:11:10.000
And when I interviewed Demis for the book, I was asking him, you know, you started wanting to build AGI because you wanted to unlock the mysteries of reality and the mysteries of the universe and, you know, quantum entanglement and all this kind of stuff.
00:11:10.159 --> 00:11:14.559
And I said, Well, how many people at DeepMind have you actually got working on that now?
00:11:14.720 --> 00:11:16.399
And he said, Well, it's just me.
00:11:16.639 --> 00:11:23.759
He just, you know, when he comes home from work, he reads research papers late into the night until two, three in the morning.
00:11:23.919 --> 00:11:36.799
So again, these I I think like these professed goals, which sound really good and which Demis and Sam still talk about in public, are actually not really the focus of the companies.
00:11:36.960 --> 00:11:52.000
The, you know, even if you look at DeepMind's website, a few years ago, the homepage said a lot more about its efforts in healthcare, its efforts in you know, AI, machine learning to solve problems around climate change and new energy sources.
00:11:52.240 --> 00:11:53.200
That's all gone now.
00:11:53.519 --> 00:11:55.279
Wide used to cover that all the time.
00:11:55.679 --> 00:11:56.639
Okay, yeah, yeah.
00:11:56.960 --> 00:12:00.000
Yeah, it was it was there was always a story coming out about DeepMind.
00:12:00.159 --> 00:12:01.120
I didn't read so much anymore.
00:12:01.200 --> 00:12:02.799
It's interesting, it's true actually.
00:12:03.039 --> 00:12:03.279
Yeah.
00:12:03.440 --> 00:12:06.320
And you and it's reflected just even uh you just look at the website.
00:12:06.399 --> 00:12:11.039
I mean, there's so much more there about it's a product-based homepage now.
00:12:11.200 --> 00:12:15.600
Uh it's talking about Gemini, it's talking about YouTube, it's talking about a new video generation tool.
00:12:16.000 --> 00:12:28.720
Deep Mind was always framing itself as this kind of a research, AI research lab for scientists who could make discoveries that would lead to making the world a better place and get paid incredibly well because they had Google as a benefactor.
00:12:28.879 --> 00:12:36.000
And now I think the reality has just come a little bit more to bear now because we're in this AI arms race between the tech companies.
00:12:36.080 --> 00:12:40.960
So it can't really sort of pretend to be this do-gooding AI research lab anymore.
00:12:41.039 --> 00:12:42.080
It really is just a company.
00:12:42.720 --> 00:12:49.039
And do you see this arms race as very much sort of Microsoft versus Google?
00:12:49.360 --> 00:12:49.679
Yes.
00:12:50.080 --> 00:12:50.399
Exactly.
00:12:50.480 --> 00:13:01.519
I wanted to ask you, you know, why you chose them, you focused on them, and what happens to, or what you consider about, you know, Meta, Apple, exactly.
00:13:01.679 --> 00:13:02.799
Where are they in the fray?
00:13:02.879 --> 00:13:08.159
Because obviously the way you have posited it is this battle between these two big giants.
00:13:08.480 --> 00:13:21.840
So I was in Silicon Valley a few weeks ago and I met with a venture capital investor named Reed Hoffman, who was on the board of OpenAI, and you know, he he's played a big role in all the things that have happened with OpenAI and with DeepMind too.
00:13:22.080 --> 00:13:28.799
He and I are very different, like in our perspectives of you know, the benefits of AI to society and of and big tech's role in that.
00:13:28.879 --> 00:13:36.799
And his argument is like, when it comes to AI, I would be concerned if there was like, is this was only a one or two horse race, but this is a five-horse race.
00:13:37.039 --> 00:13:38.799
And so that's why I'm not worried.
00:13:38.960 --> 00:13:40.480
But I don't agree with that.
00:13:40.639 --> 00:13:47.519
I think it actually is more of a two-horse race, because to your question about Apple, I don't see Apple really being in this race.
00:13:47.679 --> 00:13:53.519
I don't really see Facebook being in this race because its model is open source.
00:13:53.679 --> 00:13:56.639
I mean, again, I should be careful on using the word open source.
00:13:56.960 --> 00:13:57.600
So it says.
00:13:57.919 --> 00:13:58.879
Yes, exactly.
00:13:59.120 --> 00:14:05.600
Strictly speaking, and you know, the open source um community does has kind of taken issue with that definition.
00:14:05.919 --> 00:14:10.799
But it's not benefiting monetarily in the same way that Google and Microsoft are.
00:14:11.039 --> 00:14:16.639
You know, even Amazon, maybe we can't completely discount Amazon, but Amazon is sort of up there.
00:14:16.799 --> 00:14:20.399
I think it tracks really well with who the biggest cloud providers are.
00:14:20.559 --> 00:14:25.039
There's three big cloud providers: Amazon, Microsoft, and Google.
00:14:25.200 --> 00:14:38.399
But in the case of AI, you know, really at the top, I think, is Google and Microsoft just dominating the market for consumer-based AI products and also AI products for businesses.
00:14:38.559 --> 00:14:43.519
And DeepMind is a proxy of Google, and OpenAI is a proxy for Microsoft.
00:14:43.679 --> 00:14:46.159
Microsoft owns 49% of OpenAI.
00:14:46.480 --> 00:14:55.519
It's very hard to see how even a even a brilliant, sort of smaller organization can cut through because even if they do brilliant work, they end up getting acquired, don't they?
00:14:55.759 --> 00:14:56.639
Absolutely, yeah.
00:14:56.879 --> 00:15:01.279
And then they're not doing the challenging innovation that you would expect them to do in the market.
00:15:01.440 --> 00:15:03.759
They just become another, just another acquisition.
00:15:04.000 --> 00:15:14.240
It's one of the questions I had for you, actually, in like 20 years' time, do you still think it will be Microsoft and Google, or do you think there will have been an upstart that has disrupted the market?
00:15:14.399 --> 00:15:19.360
Or do you think actually these are just really an oligopoly or potentially even a monopoly in the end?
00:15:19.440 --> 00:15:20.080
I don't know.
00:15:20.639 --> 00:15:24.639
Yeah, I mean, I think we're just this is so unprecedented.
00:15:24.799 --> 00:15:27.279
I was just looking at some market data the other day.
00:15:27.360 --> 00:15:44.399
And if you look at the SP 500 index, like back in 1960, and then you look at the 1970s, then you look at the 1980s, the companies who were in the top five stocks in the SP 500, it was a real diverse array of industries represented there.
00:15:44.559 --> 00:15:52.480
You had ExxonMobil, you had Exxon, you had General Electric, you had healthcare companies, so it was you know, oil, energy.
00:15:52.720 --> 00:15:58.879
But then something really odd has happened in the last decade, which is that the top five stocks are all tech companies.
00:15:59.039 --> 00:16:06.879
The tech industry, there's never been a time where the entire like top five stocks in the SP 500 are all from the same industry.
00:16:07.120 --> 00:16:11.919
And not only that, but the index is weighted increasingly towards those stocks.
00:16:12.159 --> 00:16:18.240
So I think in 2020 it was something like around 12% of the value of the S P 500 was tech companies.
00:16:18.320 --> 00:16:19.440
Now it's 25%.
00:16:20.080 --> 00:16:24.000
So this is when you ask me like what will it be like in 20 years?
00:16:24.159 --> 00:16:41.679
Like I just think like statistically, like the most likely scenario is it just stays the same that these tech companies will continue because that's this is we're in such an unprecedented situation where the power is concentrated so much among those companies, it'll be quite hard to loosen that grip.
00:16:42.480 --> 00:16:44.559
Well, yeah, two things could loosen that grip.
00:16:44.720 --> 00:16:56.480
One, uh major war, which ends up reprogramming the economy either intentionally or accidentally, or two, one thing I wanted to ask you about the sustainability angle of all of this.
00:16:56.720 --> 00:17:05.279
Because for a long time I've been looking at sort of the resources that are required for sort of training some of these models, and in particular, sort of water.
00:17:05.519 --> 00:17:07.920
Obviously, lots of futurists look at these sorts of resources.
00:17:08.079 --> 00:17:15.039
I mean, it takes as much water to, I don't know, train GPT, uh chat GPT to as it does to call a nuclear reactor, etc.
00:17:15.279 --> 00:17:28.960
So these sorts of you know basic resources that we need for human survival, are they going to get triaged towards actually the digital public infrastructure, for example, that's all, you know, running on and dependent on AI?
00:17:29.119 --> 00:17:37.440
Did you look at, because I've read I've read a lot of the book, but I haven't finished it yet, this angle about sustainability and whether that is an obstacle in the future.
00:17:37.759 --> 00:17:38.960
Do you know it's funny?
00:17:39.119 --> 00:17:48.400
I see it as absolutely a valid point of concern, but I'm actually not sure if how big a problem it's going to be in the next five to ten years.
00:17:48.559 --> 00:17:55.680
And the reason for that is we keep hearing from people in the tech industry that they are working so hard on efficient.
00:17:55.920 --> 00:18:06.880
Like it's in their interest, not just for the good of climate, but it's genuinely in their interests to make these models as efficient as possible and not so data hungry because that's expensive for them.
00:18:07.119 --> 00:18:13.359
So if you go to an AI conference, you know, just in the past year, that's what every all the researchers are talking about.
00:18:13.440 --> 00:18:19.359
That's what they're working on is how do we make these models more efficient, less data hungry, less power hungry?
00:18:19.519 --> 00:18:32.480
You know, I was talking to someone at Google Cloud, one of the executives there, and they, you know, of course they would say it, but this guy was very, very confident, like, yeah, they're absolutely the these things are gonna become more efficient over time.
00:18:32.559 --> 00:18:34.319
The chips are gonna become more efficient.
00:18:34.559 --> 00:18:35.759
Will that actually happen?
00:18:36.000 --> 00:18:38.720
Will the human ingenuity be there to make it happen?
00:18:38.880 --> 00:18:39.519
I don't know.
00:18:40.240 --> 00:18:43.359
Well, they're gonna have to make it happen, I guess, um, in a way.
00:18:43.839 --> 00:18:52.720
Netting out today, uh, where do you think open AI is versus DeepMind or or Google really or Microsoft and Google?
00:18:52.960 --> 00:18:59.440
Obviously, there have been some recent changes with OpenAI saying, you know, they are literally going moving from non-profit to profit.
00:18:59.599 --> 00:19:00.720
Where are we now?
00:19:01.039 --> 00:19:02.240
Yeah, it's a great question.
00:19:02.319 --> 00:19:09.119
And you would think maybe OpenAI because they were the first, you know, at the starting gun with Chat GPT.
00:19:09.200 --> 00:19:17.599
I mean, they were just way ahead of everyone else with this large language model that you could chat to, and it's like literally like talking to a human.
00:19:17.759 --> 00:19:20.240
It could generate this incredible human-like prose.
00:19:20.480 --> 00:19:25.440
But since then, I think Google has done a really good job of catching up.
00:19:25.519 --> 00:19:26.640
It was caught on the back foot.
00:19:26.880 --> 00:19:28.079
Given that they invented it.
00:19:28.559 --> 00:19:29.039
That's right.
00:19:29.279 --> 00:19:32.319
Yeah, which I think was probably a little bit humiliating for them.
00:19:32.480 --> 00:19:42.880
Yeah, because they they'd invented it and then they hadn't released it because they were worried about what this, you know, this AI model could say, what weird things it could say when the public start using it.
00:19:43.039 --> 00:19:51.599
But since then, we now see the the big advantage that Google has, which open AI does not have, is distribution.
00:19:51.839 --> 00:20:03.839
So even though OpenAI is effectively like a product arm of Microsoft now, so you can say, well, Microsoft, you know, it's everywhere, everyone uses Word, but not in the same way people use Google.
00:20:04.000 --> 00:20:06.160
It just doesn't have the same kind of reach.
00:20:06.400 --> 00:20:15.359
Like you, if you have an iPhone, the default search engine on your iPhone is Google, because Google pays Apple billions of dollars to allow it to be on there.
00:20:15.680 --> 00:20:21.920
And it's Android phones are used by are on the are on billions of smartphones around the world.
00:20:22.079 --> 00:20:32.880
And and and the Google's new Gemini model, which is effectively its response to Chat GPT, that's being baked into every uh Android phone now.
00:20:33.119 --> 00:20:36.000
So I think that is like a really big advantage.
00:20:36.079 --> 00:20:56.799
And not only that, but when you use an Android phone and you use Gemini on the system, you know, because Gmail and GDocs and the camera and the calendar, all these things are linked, and you can sort of use Gemini as a much more kind of comprehensive assistant that can talk between all those different apps.
00:20:56.960 --> 00:20:58.319
OpenAI can't do that.
00:20:58.400 --> 00:21:01.440
It doesn't have access to all those different apps that people use.
00:21:01.839 --> 00:21:09.920
It's relying on other businesses to build it into their businesses to sort of do the same thing, but not as effectively.
00:21:10.079 --> 00:21:14.880
So it's kind of, I mean, it's hard to say like who who's really like ahead right now.
00:21:15.359 --> 00:21:22.319
It's so close, but if I really had to put money on it, I would say Google, because they're the one that have the the the most distribution.
00:21:23.039 --> 00:21:23.599
It's very interesting.
00:21:23.680 --> 00:21:24.160
It's true.
00:21:24.240 --> 00:21:25.519
They're like the Coca-Cola, aren't they?
00:21:25.599 --> 00:21:27.599
It's like the Pepsian Coca-Cola, I always think.
00:21:28.079 --> 00:21:32.640
So um when we chatted uh originally, we talked a little bit about agency.
00:21:32.880 --> 00:21:42.640
So I did want to ask you about that because it's obviously something that comes up when you're just talking to anybody about the future and what they think the implications of some of these AI tools are.
00:21:42.799 --> 00:21:44.799
I mean, obviously they're incredibly convenient.
00:21:45.039 --> 00:21:50.079
You can be really creative, maybe you can be productive with them, we're not really sure yet.
00:21:50.240 --> 00:21:59.759
But um, people are worried about how much agency it might rob from them, given that it's working on the basis of their own personal data as inputs, and actually it's making decisions.
00:22:00.160 --> 00:22:02.960
Decisions or managing outcomes for them.
00:22:03.359 --> 00:22:04.400
What's your point of view?
00:22:04.640 --> 00:22:08.559
Having researched this so in depth as you have, what's your point of view on this?
00:22:09.279 --> 00:22:14.079
I think there's definitely the risk of an erosion of human agency.
00:22:14.640 --> 00:22:18.559
The problem is this is a very, very difficult thing to measure.
00:22:18.880 --> 00:22:22.960
It's kind of an abstract, squishy subject.
00:22:23.200 --> 00:22:27.119
Like, how do you measure that my agency as a human has declined?
00:22:27.359 --> 00:22:31.200
It's not even really something that people talk about that much, but they recognize it.
00:22:31.279 --> 00:22:38.799
Like when you just talk about how addicted we are to our smartphones, for example, does that represent a loss of human agency?
00:22:39.039 --> 00:22:53.759
I think when it comes to how we make decisions, I think there are going to be more situations where companies and organizations and people just outsource their decision making to generative AI, to AI.
00:22:53.920 --> 00:23:03.519
And what's really interesting is when you hear companies talk about this and AI makers talk about this, they say, but we don't want AI to take over that to make the decisions.
00:23:03.599 --> 00:23:06.240
We want them to augment human decisions.
00:23:06.480 --> 00:23:10.000
But actually, if you think about it, humans are kind of lazy.
00:23:10.079 --> 00:23:13.920
And if the easy option is to let a computer make the decision, then we will.
00:23:14.079 --> 00:23:15.599
And I'll give you an example of this.
00:23:15.839 --> 00:23:24.079
So London's Metropolitan Police have been trialing facial recognition, which is a form of AI, for a number of years.
00:23:24.240 --> 00:23:39.599
And there was a study done of officers who were using facial recognition cameras from a police van in the middle of a busy street, and it was just the camera was looking for people as they walked down the street to see if any of them matched a watch list.
00:23:39.839 --> 00:23:45.920
Now, the problem with facial recognition is that sometimes it's wrong, particularly when it's looking at people of color.
00:23:46.240 --> 00:23:59.680
But what was happening inside the police van is that when the face flashed up on the screen as a match, the officer assumed that the facial recognition system was correct.
00:23:59.920 --> 00:24:05.200
Rather than sort of double check, they would go out and just kind of detain the person.
00:24:05.680 --> 00:24:15.839
And the scientists who were studying this, who were in the van observing the whole phenomenon, actually gave a name to this and they called it deference to the algorithm.
00:24:16.000 --> 00:24:33.200
And this was a big concern because we assume that computers are often or more often right than humans are, because they are backed by these networks of neural networks and uh and their machines, and they're just like a calculator, it's always correct, right?
00:24:33.359 --> 00:24:35.519
Um so you just assume it's correct.
00:24:35.680 --> 00:24:49.119
And I think that's uh that to me is just an example of the possibility that we are actually just going to delegate a lot of decision making to computers because we trust them and because it's easier for us.
00:24:49.599 --> 00:24:58.720
So whether that happens in the criminal justice system, in businesses, in marketing, in customer service, I could see this happening in all sorts of different parts of life.
00:24:59.039 --> 00:24:59.200
Yeah.
00:24:59.359 --> 00:25:08.799
So, and you do cover in the book elements of the kind of trust and safety teams and what has happened to those over time, um, like Temerick Gibr and people like that.
00:25:08.880 --> 00:25:13.200
And what's your point of view on where we are with trust and safety?
00:25:13.359 --> 00:25:17.119
Even, you know, even broader than maybe the two companies you're talking about.
00:25:17.359 --> 00:25:18.720
Has trust and safety helped?
00:25:18.880 --> 00:25:20.799
Has it just been a massive distraction?
00:25:21.039 --> 00:25:22.559
Do we need to start again with it?
00:25:22.720 --> 00:25:23.599
What do you think?
00:25:24.000 --> 00:25:28.960
Oh, it's definitely helped for the little that resources that these teams have.
00:25:29.200 --> 00:25:50.960
But when you talk to people who work on these teams and people who have knowledge of these teams, the overwhelming consensus is they're tiny and they are under-resourced, and they're largely drowned out in all the product meetings by the people whose greater priority is to get these systems bigger, um, get them to launch faster.
00:25:51.200 --> 00:25:54.400
And of course, you know, I'm not going to dismiss all these efforts.
00:25:54.480 --> 00:26:04.559
Like Google, I think, really does put a lot of time and effort into trying to build its algorithms responsibly, but at the end of the day, it's not really being regulated for this stuff.
00:26:04.799 --> 00:26:08.799
And so it's very much its own decision how much investment to put into these.
00:26:08.880 --> 00:26:11.680
And by and large, these teams are small.
00:26:12.000 --> 00:26:19.519
You know, the other thing is if I went to Google now and said, how big is your trust and safety team, they don't have to tell me, and they probably wouldn't tell me.
00:26:19.680 --> 00:26:21.440
I've asked these kinds of questions before.
00:26:21.599 --> 00:26:26.000
I can't tell you the number of times I've asked Facebook, how many content moderators do you have?
00:26:26.079 --> 00:26:27.200
And they never tell you.
00:26:27.359 --> 00:26:29.359
They never tell you because they don't have to.
00:26:29.519 --> 00:26:42.720
And so the hope is that with new regulations coming down the pipe, from especially from the European Union, uh, there's going to be more transparency from these companies over the kinds of risk assessments that they're doing for their algorithms.
00:26:43.359 --> 00:26:57.279
This is interesting because at the moment we've also got Drag's report coming out saying that, you know, well, actually, we're regulating, I mean, talking generally, not about this specific case, but generally we're regulating so much that we're stifling innovation.
00:26:57.440 --> 00:27:05.680
So we need to, you know, maybe think about that and potentially not reverse it, but, you know, is there too much regulation that's stifling innovation?
00:27:05.759 --> 00:27:09.599
So I wonder actually what's going to happen with trust and safety in general.
00:27:10.079 --> 00:27:14.319
Yeah, well, I think the approach by the European Union is quite good.
00:27:14.559 --> 00:27:17.359
They have the AI Act, which is coming into force.
00:27:17.440 --> 00:27:23.599
It's sort of coming into force now, really, but I don't think we're gonna see the effect of it for another year or so.
00:27:23.839 --> 00:27:42.079
But part of the way that act is gonna work is that companies like Google, like Apple, um, OpenAI, DeepMind, they're going to have to be doing these risk assessments on their algorithms, and they're gonna have to share the results of that with auditors, with people from outside the company.
00:27:42.240 --> 00:27:50.799
This is unprecedented because for years, the only people that these companies were accountable to was themselves.
00:27:50.960 --> 00:27:53.279
They literally self-regulate.
00:27:53.519 --> 00:27:54.240
It's so funny.
00:27:54.319 --> 00:28:00.079
Like companies like Uber and Facebook, they have this term transparency report.
00:28:00.160 --> 00:28:02.319
They all put out these quarterly transparency.
00:28:02.799 --> 00:28:12.000
It's so funny that they call them transparency reports when it's literally produced by them and you just trust that the numbers are true.
00:28:12.160 --> 00:28:20.480
They're not, there's no like auditor or like a big consultancy firm like anyone from the big five actually checking their homework.
00:28:20.640 --> 00:28:23.680
It's just the the the companies doing it themselves.
00:28:23.920 --> 00:28:29.359
So it sounds, it seems like this is on course to finally change at some point in the next year.
00:28:29.680 --> 00:28:32.319
Now, two questions very specific that I wanted to ask you.
00:28:32.400 --> 00:28:36.720
One from me and one from somebody who I asked, what would you love to ask Palmy?
00:28:37.119 --> 00:28:43.599
My question is really about Eric Schmidt and what he talked about recently, about agent-based systems.
00:28:43.839 --> 00:28:59.119
Actually, you know, you almost be able to replace the programmer because you'll be able to give it some kind of prompt or instruction to, you know, I think in the case he was giving, it was copy all of TikTok and do the, and if it, you know, in half an hour it doesn't give you the results, start again and redesign the system.
00:28:59.359 --> 00:29:03.599
Where do you think we're going with these AI agent-based systems?
00:29:03.759 --> 00:29:11.599
Are they going to be in the hands of users so that we can personally affect things so they can go out and work on our behalf?
00:29:11.839 --> 00:29:25.920
Or actually are they more centralized and they're almost the property of these big companies, the tech giants that you're talking about, and actually the consumer's just on the end of whatever it's been programmed or prompted to do.
00:29:26.319 --> 00:29:32.000
Yeah, that's a really good question about whether this is something that consumers are going to have that much control of.
00:29:32.319 --> 00:29:38.079
I think there's probably going to be products that come out, like ChatGPT, for instance.
00:29:38.240 --> 00:29:44.880
I think this will like ChatGPT will probably be given I hate this word, but agentic properties.
00:29:45.200 --> 00:30:02.960
And so the latest version of ChatGPT, which is called O1, has these reasoning capabilities, which allows it to do tasks or carry out sort of reasoning tasks where it actually takes a few seconds to quote unquote think about something before it does something.
00:30:03.039 --> 00:30:12.880
And these are kind of more mathematical in nature than the sort of language-based text prediction that it was that it was doing before.
00:30:13.119 --> 00:30:20.880
And this has been touted as a precursor to Chat GPT eventually being able to carry out tasks like an agent.
00:30:21.039 --> 00:30:22.799
So that's probably coming.
00:30:23.039 --> 00:30:43.839
One of the first examples of agent-based AI has actually come from Salesforce, which is this big software as a service company that sells uh enterprise software to all sorts of companies like banks, um, British Airways, you know, it's like oftentimes they use their customer relationship management software.
00:30:44.079 --> 00:30:59.759
But what their agents can now do is like in customer service, if I'm going onto, let's just say, for example, British Airways's website and I want to book something, like the agent isn't just going to give me information, it could actually book something for me, or it could lodge a complaint.
00:31:00.000 --> 00:31:02.640
So it can it can actually take action.
00:31:02.720 --> 00:31:04.559
And some people are quite concerned about this.
00:31:04.640 --> 00:31:08.960
Like, well, these large language models have been making mistakes, they've been hallucinating.
00:31:09.200 --> 00:31:17.839
What's to say they won't, you know, accidentally, you know, make an incorrect booking, book the wrong flight, or take some kind of dangerous action.
00:31:18.160 --> 00:31:21.039
Then you can can go and hire an AI lawyer to sort it out.
00:31:21.359 --> 00:31:22.319
Yes, that's right.
00:31:22.480 --> 00:31:23.119
Exactly.
00:31:23.359 --> 00:31:25.920
We'll really feel like we're in the matrix at that point.
00:31:26.160 --> 00:31:30.240
I think my big question is what happens when these things make mistakes?
00:31:30.559 --> 00:31:31.359
How will we know?
00:31:31.599 --> 00:31:32.079
Yeah.
00:31:32.319 --> 00:31:34.799
You want to book a book an appointment with the doctor?
00:31:35.039 --> 00:31:36.640
This is what Salesforce were telling me.
00:31:36.720 --> 00:31:40.079
Like this is it'll be able to actually book the appointment for you.
00:31:40.319 --> 00:31:45.279
So just what happens when these things uh actually mess up?
00:31:45.839 --> 00:31:46.720
It's a new system.
00:31:46.799 --> 00:31:48.319
We don't have any rules for it yet.
00:31:48.480 --> 00:31:50.079
And it's gonna be very interesting.
00:31:50.319 --> 00:31:59.200
The question that Chris wanted to ask you was and he works with a lot of clients, he's obviously trying to um bring them with him into the world of AI.
00:31:59.279 --> 00:32:08.319
Uh, but he's saying, I find the most impactful AIs are based on small proprietary models versus general purpose LLMs trains on the public web.
00:32:08.559 --> 00:32:14.480
So in the battle for supremacy, the title of your book, um, how will that square in the future?
00:32:14.720 --> 00:32:18.000
Small versus large, you know, private versus public.
00:32:18.640 --> 00:32:19.519
Do you know it's funny?
00:32:19.599 --> 00:32:25.200
I've heard people say the same thing that these smaller, more specialized models will be more effective.
00:32:25.440 --> 00:32:34.240
I have yet to see that in practice, though, because I have seen examples of small specialized models just not be as good as the generalized models.
00:32:34.559 --> 00:32:46.000
You know, like the likes of just GPT-4 from OpenAI is better than even a model that a financial company is just a real-world example.
00:32:46.240 --> 00:32:50.000
A financial company made their own model based on their own financial data.
00:32:50.160 --> 00:32:59.440
It was a large language model, they spent like a year on it, and it just even based on their own proprietary data, it still wasn't as good as OpenAI's general purpose model.
00:32:59.759 --> 00:33:01.119
Why did they think that was?
00:33:01.599 --> 00:33:03.519
It just wasn't as accurate.
00:33:03.759 --> 00:33:07.839
It it just didn't, wasn't as effective, it wasn't as capable or lucid.
00:33:08.400 --> 00:33:10.720
So I don't know how clear it is.
00:33:10.799 --> 00:33:18.400
I I like maybe some of these smaller models will still be useful, particularly for companies who don't have the resources to pay for the big models.
00:33:18.640 --> 00:33:25.519
It's like like open source, it's an alternative to to having to pay for access to the API of of OpenAI or Gemini.
00:33:26.079 --> 00:33:31.359
But I think I don't know, I I think the the question is still an open question about whether people will use these.
00:33:31.519 --> 00:33:37.200
And you know, one thing I would love right now, which I haven't been able to find, is market share data.
00:33:37.359 --> 00:33:42.240
Like in just in just the market for large language models, who's actually dominating?
00:33:42.400 --> 00:33:53.200
I I just anecdotally from what I've read, it seems to be Google's Gemini and OpenAI's GPT, but who who's behind them?
00:33:53.359 --> 00:33:54.319
We just don't know yet.
00:33:54.400 --> 00:33:58.319
So that's hopefully something we'll find out in the next you know months, few months to a year.
00:33:58.720 --> 00:34:02.000
I don't really know anyone who only uses one.
00:34:02.240 --> 00:34:10.079
So I will use ChatGPT and then I'll put that output into Claude, for example, and get Claude to improve on it and then put it back to whatever.
00:34:10.239 --> 00:34:13.039
And then on my mobile, I use perplexity a lot, actually.
00:34:13.199 --> 00:34:22.559
I don't know why, but when I'm on the go and I've got smaller questions, sometimes I use perplex and I like to use like all of them because they all do, they all have different strengths and weaknesses, I think.
00:34:22.719 --> 00:34:31.119
So it's quite good to get a a bit of an adversarial relationship going between like they have come up through different ways and they're so they've got different strengths and weaknesses.
00:34:31.360 --> 00:34:36.400
Yeah, and perplexity was started by academics, and which is why they're really big on sources.
00:34:36.639 --> 00:34:37.280
Yeah, I love that.
00:34:37.440 --> 00:34:38.960
Yeah, and it is really good.
00:34:39.199 --> 00:34:45.599
But I in a you know, another wrinkle to this whole dynamic is none of these really have a strong moat, right?
00:34:45.920 --> 00:34:56.159
And just like you, I was talking to a startup, an AI startup, who were just like, yeah, we use cloud, and then sometimes we use we use OpenAI, you know, for their own product.
00:34:56.480 --> 00:35:00.239
Their product is like built on top of those things, and it's really easy to switch.
00:35:00.400 --> 00:35:02.159
And that's very unusual in tech.
00:35:02.239 --> 00:35:10.800
Like with if you're using a cloud provider, for example, like Amazon's AWS, it's really hard to switch to, say, Microsoft Azure.
00:35:10.960 --> 00:35:13.840
But it's super easy to switch between foundation models.
00:35:14.079 --> 00:35:20.719
And apparently the CEO of Anthropic was asked about this at a dinner, at an AI dinner, like, what's your moat?
00:35:20.880 --> 00:35:23.840
And he was like, Well, this is sorry, this is hearsay, by the way.
00:35:23.920 --> 00:35:26.079
This is what I heard from one person who was at the dinner.
00:35:26.239 --> 00:35:33.360
And he said, Well, our moat is that nobody else can really afford to spend, you know, X billion dollars on compute like we can.
00:35:33.599 --> 00:35:38.079
So I mean, yeah, I guess that's a moat, but that's not that great a moat.
00:35:38.239 --> 00:35:39.039
It's this is just money.
00:35:39.199 --> 00:35:39.920
The moat's money.
00:35:40.159 --> 00:35:47.280
I think just try and use as much of it as possible all across the board while it's still relatively affordable.
00:35:47.840 --> 00:35:50.480
Great point, because this is the history of tech, right?
00:35:50.559 --> 00:35:58.719
Like you start off cheap and then they just jack up the prices once they've got once they've really established economies of scale and have users who are utterly dependent on them.
00:35:59.039 --> 00:35:59.440
Yeah.
00:36:00.000 --> 00:36:01.519
Whilst we're both using each other.
00:36:01.760 --> 00:36:02.000
Yeah.
00:36:02.159 --> 00:36:07.119
Um so as we close then, and it's been brilliant to spend some time with you and listen to this.
00:36:07.280 --> 00:36:10.159
What should I have asked you about that we haven't covered?
00:36:10.639 --> 00:36:18.639
Like I often hear, I worry when I'm talking about my book with people that I just leave people feeling kind of depressed about the future.
00:36:18.960 --> 00:36:22.639
Maybe I don't I feel like we haven't gone into that too much with this discussion.
00:36:22.800 --> 00:36:26.639
But I would say, like, I am hopeful about all the regulatory efforts that are happening.
00:36:26.719 --> 00:36:29.920
And it's not just regulators, it's also efforts in the courts.
00:36:30.000 --> 00:36:41.199
You know, there's some really interesting court cases out there, like the one in San Francisco that's fighting uh websites that produce deep fake porn and are trying to shut those websites down.
00:36:41.440 --> 00:36:43.920
That's something that I'm hoping to write about in the next couple of weeks.
00:36:44.159 --> 00:36:45.920
I think they're doing amazing work.
00:36:46.239 --> 00:36:54.639
And, you know, there are state laws in the U.S., there are unions, like think about the Writers' Guild and the Writer's Strike in Hollywood.
00:36:54.719 --> 00:37:00.159
Um, and they manage to change the contract language about using AI for script writing and things like that.
00:37:00.320 --> 00:37:10.480
So I worry a little bit about civil society and sort of campaign groups with names like AI Now and the Institute for Strategic Dialogue and Open Markets Institute.
00:37:10.559 --> 00:37:21.360
And these organizations are very poorly funded uh compared to tech companies and are just trying to fight back against some of the monopolistic practices of tech companies and the AI harms.
00:37:21.599 --> 00:37:28.000
A lot of these civil society groups, I feel like their voices are being drowned out a little bit by all the hype in tech.
00:37:28.480 --> 00:37:36.800
So I what I would love to see is just hearing, just see more of those voices in the public and just so we're just a little bit more aware.
00:37:36.960 --> 00:37:47.920
You know, going back to the Nobel Prize for Demis and Jeffrey Hinton, you know, two people linked to Google, two or three people actually, this is great casting a really positive glow on Google.
00:37:48.079 --> 00:38:07.280
I worry a little bit, it kind of muddies the debate, like it kind of eclipses some of the risks that we're trying to talk about with AI and just spotlighting the great stuff that AI does when actually this is such a tiny part of Google DeepMind's real investment and effort.
00:38:07.519 --> 00:38:13.679
So yeah, I think we just we do have to kind of talk about some of these, uh, some of these risks and and how they're being mitigated.
00:38:14.159 --> 00:38:17.440
I think the thing is there's just so many risks at the minute in the news.
00:38:17.599 --> 00:38:19.760
Like there's a lot of threats, there's a lot of risks.
00:38:19.840 --> 00:38:30.960
It's very difficult for people, the general public, to get their heads around the risks of AI, whether it's AGI or whether it's bias or what whatever it is, because there's so many threats at the minute.
00:38:31.119 --> 00:38:33.280
And it's kind of um kind of interesting.
00:38:33.599 --> 00:38:34.800
That's brilliant, Palmy.
00:38:35.039 --> 00:38:38.639
Thank you so much for spending a bit of time with us and um congratulations on the book.
00:38:38.800 --> 00:38:40.639
We shall put all the links in the show notes.
00:38:40.800 --> 00:38:45.519
Is there anything else you'd like to add about the book apart from the fact that everybody should go and buy it?
00:38:45.920 --> 00:38:53.440
Uh well, it has just been shortlisted for the Financial Times Business Book of the Year Award, which is which is very, very exciting.
00:38:53.519 --> 00:38:56.559
Um, so we'll find out in December with the result of that.
00:38:56.719 --> 00:39:02.239
It's been great just being able to talk about the book and tell the story of the people behind the the technology.
00:39:07.599 --> 00:39:11.039
Thank you for listening to The Future of You, hosted by me, Tracy Follows.
00:39:11.119 --> 00:39:15.039
Be sure to check out the show notes for more info about the topics we covered today.
00:39:15.199 --> 00:39:19.760
If you enjoyed this episode, please like and subscribe wherever you get your podcasts.
00:39:19.920 --> 00:39:23.599
And if you know someone who would love this episode, please share it with them.
00:39:23.760 --> 00:39:29.599
For more on the future of identity in a digital world, visit futuremade.group slash the future of you.
00:39:30.000 --> 00:39:34.719
To explore the future of everything else, head over to future made.group.
00:39:34.800 --> 00:39:38.480
The Future of View podcast is produced by Big Tent Media.
00:00:00.000 --> 00:00:05.599
Humans are kind of lazy, and if the easy option is to let a computer make the decision, then we will.
00:00:40.479 --> 00:00:53.359
You may already know her from her book We Are Anonymous, but she has a new book out chronicling both Sam Altman and Demis Hassabis's Rise to AI fame through OpenAI and DeepMind, respectively.
00:00:53.600 --> 00:00:58.799
It's called Supremacy, AI, Chat GTP, and the race that will change the world.
00:00:58.880 --> 00:01:03.119
And it's just been shortlisted for Business Book of the Year by the FT.
00:01:03.280 --> 00:01:07.120
I've linked to all of her work and her books in the show notes, as usual.
00:01:07.359 --> 00:01:15.439
Now here we chat on the same day that Demis Hassabis's Nobel Prize for Chemistry was announced, with Palmi busily filing her story on that.
00:01:15.599 --> 00:01:17.439
So why not start there?
00:01:23.040 --> 00:01:28.560
Welcome Palmi Olsen to the future of you, and congratulations on your new book, Supremacy.
00:01:28.879 --> 00:01:29.359
Thank you.
00:01:29.599 --> 00:01:30.959
No, I'm excited to be here.
00:01:31.120 --> 00:01:32.079
Thank you for inviting me.
00:01:32.319 --> 00:01:33.760
Oh, it's fabulous to have you.
00:01:33.920 --> 00:01:39.280
Especially right now, it's all so topical, and um, we should give the book its uh proper title, I suppose.
00:01:39.519 --> 00:01:43.359
Supremacy AI, ChatGPT, and the race that will change the world.
00:01:43.519 --> 00:01:53.840
But we've caught you on a day where you are, I'm sure you're always busy, Palmy, but exceptionally busy because you're just writing up, or in the midst of writing up, something about one of the protagonists in your book.
00:01:54.079 --> 00:01:55.519
Why don't you tell us more?
00:01:55.920 --> 00:01:57.280
That's right, yes.
00:01:57.519 --> 00:02:11.360
Well, so the founder and CEO of Google Deep Mind, whose name is Demis Hasabis, is one of the two key protagonists in supremacy, AI, ChatGPT, and the race that will change the world.
00:02:11.599 --> 00:02:20.000
Um, and today he just won a Nobel Prize, the Nobel Prize for Chemistry, and he's sharing it with two other people.
00:02:20.159 --> 00:02:27.680
And this is for the work that he and the team at DeepMind have done on a process known as protein folding.
00:02:27.840 --> 00:02:49.680
And proteins are the building blocks of the human body, and the idea is that by using uh artificial intelligence to predict how these proteins will fold, they can scientists and medical researchers can use that process for drug discovery and make breakthroughs in biology.
00:02:49.919 --> 00:02:54.319
Um, and this is a project that's been going on for about eight years or so.
00:02:54.479 --> 00:03:06.800
And so it's really amazing that it's been recognized and had this incredible accolade, especially because, as I say in my book, Demis has craved a Nobel Prize for years.
00:03:07.360 --> 00:03:17.360
In fact, he even told his staff that he wanted DeepMind to win, you know, three to five Nobel Prizes over the next decade.
00:03:17.439 --> 00:03:20.000
And this was very much how he measured success.
00:03:20.159 --> 00:03:22.400
So that's a huge milestone for him.
00:03:22.560 --> 00:03:37.599
And I think it also adds a really big veneer of credibility to artificial intelligence and particularly to Google, because uh Demis is actually the first big tech executive to have ever won a Nobel.
00:03:37.759 --> 00:04:10.879
And I think that really is gonna, in a way, cast a very useful glow on uh on Google and perhaps even other tech companies who are kind of struggling at the minute to justify the benefits of artificial intelligence as being good for society, being good for humanity, and not just good for them, which is one of the big arguments I make in my book that actually right now the the wealth benefits, the influence benefits really are just being concentrated in the hands of these few very large companies.
00:04:11.280 --> 00:04:11.680
You're right.
00:04:11.840 --> 00:04:13.919
This takes us very nicely into the heart of your book.
00:04:14.000 --> 00:04:23.680
Because one of the things I wanted to ask you, maybe it's a jumping-off point for you to explain to the the audience who who haven't read it yet but should go and buy it and read it, um, because it's excellent.
00:04:24.000 --> 00:04:28.639
Is it a book about the hype versus the doomerism of AI?
00:04:28.800 --> 00:04:40.000
Or is it a biography about two potential geniuses or definitely, you know, amazing entrepreneurs and maybe engineers and scientists?
00:04:40.160 --> 00:04:45.279
Or is it both because it's very interesting, it's quite hard to characterize in that sense.
00:04:45.759 --> 00:04:46.240
Mm-hmm.
00:04:46.480 --> 00:04:51.759
Yeah, it's tempting to say both, but I would steer it more towards being more of a biography.
00:04:52.000 --> 00:05:08.000
Because I think it's a really important way to understand technology, which can seem so abstract and so invisible, it's just there, it's just woven into our lives and our infrastructure, is to remember that all technology is designed by humans.
00:05:08.160 --> 00:05:17.920
And there are people like you use your phone, every widget, every app, every piece of software you use has is a result of a design decision made by a human.
00:05:18.079 --> 00:05:31.839
And I think if we really want to get a sense of where AI is going, and especially the incentives that are fueling the building of AI today, it's really important just to understand the people behind this technology.
00:05:31.920 --> 00:05:47.600
Um, and there are so many people behind AI, but for simplicity's sake, and also because I think these two men are the two most important builders of AI, I did focus it on Sam Altman, the founder of OpenAI, and Demis Hisabas, the founder of DeepMind.
00:05:47.759 --> 00:05:59.199
And the other reason I focused on them was because they were really the two people who were the first to start companies aiming to build artificial general intelligence.
00:05:59.279 --> 00:06:11.120
And this is this really huge, ambitious goal to build AI that is as smart as a human, effectively, uh, that can do the same cognitive tasks as a human brain.
00:06:11.360 --> 00:06:15.759
We're not anywhere near that yet, and there's a big question about whether that's even possible.
00:06:16.079 --> 00:06:20.319
But it was it really started off as this humanitarian dream.
00:06:20.480 --> 00:06:27.360
Like Demis started this be in part in part because he wanted to make scientific discoveries.
00:06:27.519 --> 00:06:35.360
He believed that AI could accelerate scientific breakthroughs and that AI could be and computers could be an extension of the human mind.
00:06:35.600 --> 00:06:37.759
And he saw it as a way to cure cancer.
00:06:37.920 --> 00:06:40.399
If he built AGI, he could solve climate change.
00:06:40.480 --> 00:06:42.160
He was going to solve all these things.
00:06:42.399 --> 00:06:46.720
And Sam Altman, his he was more on the kind of economic benefit of AI.
00:06:46.879 --> 00:06:51.519
He kind of promoted prosperity, like if we have AGI, we'll have abundance.
00:06:52.079 --> 00:06:58.160
He used this word abundance all the time, um, where we'll kind of elevate the wealth of everyone on earth.
00:06:58.399 --> 00:07:03.759
Of course, that that that motivation changed over time, like those those uh objectives changed.
00:07:03.920 --> 00:07:06.560
But yes, that's why I that's that's how I would characterize it.
00:07:06.639 --> 00:07:08.399
It's really a kind of human story.
00:07:08.639 --> 00:07:16.480
And you're right, it is difficult to characterize the book itself, but you know, I I'm not easy to characterize as a person generally.
00:07:16.879 --> 00:07:17.120
Exactly.
00:07:17.279 --> 00:07:19.120
Well, it has different dimensions to it.
00:07:19.199 --> 00:07:25.519
You can come in it through several different ways, and it is a is a complex story, it's told very well, but I think that's nice.
00:07:25.600 --> 00:07:29.759
There's different levels or layers of interpretation or sort of storytelling in it.
00:07:29.920 --> 00:07:33.920
I have to ask you, I'm sure people have asked you before, which of them do you prefer?
00:07:34.240 --> 00:07:47.839
The first time I was asked that was like I was on stage at an event and I hadn't, I was not expecting the question, and I just said, well, I guess I prefer Demis because he he just jet yeah, he does just seem more genuine in his outlook.
00:07:47.920 --> 00:08:03.600
It does seem like he stuck a little bit more to the original principles of trying to and the original objectives of trying to you know use AI to solve health-related uh or scientific discoveries and uh to sort of support science.
00:08:03.839 --> 00:08:19.600
Whereas Sam, I feel like, you know, the idea of using AGI to to improve the wealth of everyone, that just that feels really like it's it's just so theoretical and abstract, and it really just feels much more like a like a marketing point for him.
00:08:19.920 --> 00:08:30.079
But other people kind of come away with different, like I had someone who read it say to me, Wow, I've read this book, and and it really feels like Demis is the bad guy and Sam is the good guy.
00:08:30.240 --> 00:08:32.879
So you can really come away from it with different interpretations.
00:08:32.960 --> 00:08:45.919
But I would say that actually the first half of the book is kind of biographical, and then and then it goes into kind of like the conflict and the tension and the warnings that were coming about this technology they were building.
00:08:46.000 --> 00:08:54.799
And then the final third is actually almost more like a polemic about the the risks of AI and just what the future could hold and why it all matters.
00:08:55.120 --> 00:08:56.639
Yeah, I mean you're right, you're right.
00:08:56.720 --> 00:08:58.639
Obviously it's your book, so of course you're right.
00:08:58.799 --> 00:09:13.360
But um, the first part of it is you are sort of exploring the character and the values of each of those individuals, I think, which then goes on to explain quite a lot of how the company uh companies look and feel and how they turn out.
00:09:13.440 --> 00:09:40.879
But one of the interesting bits in the middle, whilst you are discussing all of those ins and outs and you know, acquisitions and partnerships and all this sort of stuff, particularly with Google and DeepMind, is really interesting, I think, and that does give a lot of illumination into the character of Demis and how much does he want to stick to the original values and the original purpose and the social purpose of DeepMind versus, you know, particularly when you're talking about when they're looking to move into China.
00:09:41.120 --> 00:09:44.480
I think it's after AlphaGo, but maybe you can give a summary of that.
00:09:44.799 --> 00:09:49.600
I I found that a really interesting part of the story, actually, for lots of different reasons.
00:09:50.080 --> 00:10:06.879
Yeah, and I think it's this question about sticking to this, like sticking to the to the social purpose, I think is really interesting because I think DeepMind has done a really and Google have done a really good job of publicly framing themselves as sticking to these objectives.
00:10:06.960 --> 00:10:15.519
When actually, if you look at where they put the investment, it's so much more on just the Gemini AI model product, the product.
00:10:15.679 --> 00:10:31.279
You know, probably one of the best examples is the latest news about uh about AlphaFold, about the protein folding project, because actually the number of people who worked on that at DeepMind for all the publicity that it has received is tiny.
00:10:31.440 --> 00:10:36.320
There's something like 1,500 people who work at DeepMind.
00:10:36.639 --> 00:10:40.399
And actually the number of people who are working on AlphaFold is around 20.
00:10:40.559 --> 00:10:42.000
It's very, very small.
00:10:42.240 --> 00:10:48.159
So I think that the accolades that it has received actually is sort of overshadows.
00:10:48.320 --> 00:10:54.639
It doesn't reflect really where Google is putting all the investment and where DeepMind is putting a lot more of its time.
00:10:55.039 --> 00:11:10.000
And when I interviewed Demis for the book, I was asking him, you know, you started wanting to build AGI because you wanted to unlock the mysteries of reality and the mysteries of the universe and, you know, quantum entanglement and all this kind of stuff.
00:11:10.159 --> 00:11:14.559
And I said, Well, how many people at DeepMind have you actually got working on that now?
00:11:14.720 --> 00:11:16.399
And he said, Well, it's just me.
00:11:16.639 --> 00:11:23.759
He just, you know, when he comes home from work, he reads research papers late into the night until two, three in the morning.
00:11:23.919 --> 00:11:36.799
So again, these I I think like these professed goals, which sound really good and which Demis and Sam still talk about in public, are actually not really the focus of the companies.
00:11:36.960 --> 00:11:52.000
The, you know, even if you look at DeepMind's website, a few years ago, the homepage said a lot more about its efforts in healthcare, its efforts in you know, AI, machine learning to solve problems around climate change and new energy sources.
00:11:52.240 --> 00:11:53.200
That's all gone now.
00:11:53.519 --> 00:11:55.279
Wide used to cover that all the time.
00:11:55.679 --> 00:11:56.639
Okay, yeah, yeah.
00:11:56.960 --> 00:12:00.000
Yeah, it was it was there was always a story coming out about DeepMind.
00:12:00.159 --> 00:12:01.120
I didn't read so much anymore.
00:12:01.200 --> 00:12:02.799
It's interesting, it's true actually.
00:12:03.039 --> 00:12:03.279
Yeah.
00:12:03.440 --> 00:12:06.320
And you and it's reflected just even uh you just look at the website.
00:12:06.399 --> 00:12:11.039
I mean, there's so much more there about it's a product-based homepage now.
00:12:11.200 --> 00:12:15.600
Uh it's talking about Gemini, it's talking about YouTube, it's talking about a new video generation tool.
00:12:16.000 --> 00:12:28.720
Deep Mind was always framing itself as this kind of a research, AI research lab for scientists who could make discoveries that would lead to making the world a better place and get paid incredibly well because they had Google as a benefactor.
00:12:28.879 --> 00:12:36.000
And now I think the reality has just come a little bit more to bear now because we're in this AI arms race between the tech companies.
00:12:36.080 --> 00:12:40.960
So it can't really sort of pretend to be this do-gooding AI research lab anymore.
00:12:41.039 --> 00:12:42.080
It really is just a company.
00:12:42.720 --> 00:12:49.039
And do you see this arms race as very much sort of Microsoft versus Google?
00:12:49.360 --> 00:12:49.679
Yes.
00:12:50.080 --> 00:12:50.399
Exactly.
00:12:50.480 --> 00:13:01.519
I wanted to ask you, you know, why you chose them, you focused on them, and what happens to, or what you consider about, you know, Meta, Apple, exactly.
00:13:01.679 --> 00:13:02.799
Where are they in the fray?
00:13:02.879 --> 00:13:08.159
Because obviously the way you have posited it is this battle between these two big giants.
00:13:08.480 --> 00:13:21.840
So I was in Silicon Valley a few weeks ago and I met with a venture capital investor named Reed Hoffman, who was on the board of OpenAI, and you know, he he's played a big role in all the things that have happened with OpenAI and with DeepMind too.
00:13:22.080 --> 00:13:28.799
He and I are very different, like in our perspectives of you know, the benefits of AI to society and of and big tech's role in that.
00:13:28.879 --> 00:13:36.799
And his argument is like, when it comes to AI, I would be concerned if there was like, is this was only a one or two horse race, but this is a five-horse race.
00:13:37.039 --> 00:13:38.799
And so that's why I'm not worried.
00:13:38.960 --> 00:13:40.480
But I don't agree with that.
00:13:40.639 --> 00:13:47.519
I think it actually is more of a two-horse race, because to your question about Apple, I don't see Apple really being in this race.
00:13:47.679 --> 00:13:53.519
I don't really see Facebook being in this race because its model is open source.
00:13:53.679 --> 00:13:56.639
I mean, again, I should be careful on using the word open source.
00:13:56.960 --> 00:13:57.600
So it says.
00:13:57.919 --> 00:13:58.879
Yes, exactly.
00:13:59.120 --> 00:14:05.600
Strictly speaking, and you know, the open source um community does has kind of taken issue with that definition.
00:14:05.919 --> 00:14:10.799
But it's not benefiting monetarily in the same way that Google and Microsoft are.
00:14:11.039 --> 00:14:16.639
You know, even Amazon, maybe we can't completely discount Amazon, but Amazon is sort of up there.
00:14:16.799 --> 00:14:20.399
I think it tracks really well with who the biggest cloud providers are.
00:14:20.559 --> 00:14:25.039
There's three big cloud providers: Amazon, Microsoft, and Google.
00:14:25.200 --> 00:14:38.399
But in the case of AI, you know, really at the top, I think, is Google and Microsoft just dominating the market for consumer-based AI products and also AI products for businesses.
00:14:38.559 --> 00:14:43.519
And DeepMind is a proxy of Google, and OpenAI is a proxy for Microsoft.
00:14:43.679 --> 00:14:46.159
Microsoft owns 49% of OpenAI.
00:14:46.480 --> 00:14:55.519
It's very hard to see how even a even a brilliant, sort of smaller organization can cut through because even if they do brilliant work, they end up getting acquired, don't they?
00:14:55.759 --> 00:14:56.639
Absolutely, yeah.
00:14:56.879 --> 00:15:01.279
And then they're not doing the challenging innovation that you would expect them to do in the market.
00:15:01.440 --> 00:15:03.759
They just become another, just another acquisition.
00:15:04.000 --> 00:15:14.240
It's one of the questions I had for you, actually, in like 20 years' time, do you still think it will be Microsoft and Google, or do you think there will have been an upstart that has disrupted the market?
00:15:14.399 --> 00:15:19.360
Or do you think actually these are just really an oligopoly or potentially even a monopoly in the end?
00:15:19.440 --> 00:15:20.080
I don't know.
00:15:20.639 --> 00:15:24.639
Yeah, I mean, I think we're just this is so unprecedented.
00:15:24.799 --> 00:15:27.279
I was just looking at some market data the other day.
00:15:27.360 --> 00:15:44.399
And if you look at the SP 500 index, like back in 1960, and then you look at the 1970s, then you look at the 1980s, the companies who were in the top five stocks in the SP 500, it was a real diverse array of industries represented there.
00:15:44.559 --> 00:15:52.480
You had ExxonMobil, you had Exxon, you had General Electric, you had healthcare companies, so it was you know, oil, energy.
00:15:52.720 --> 00:15:58.879
But then something really odd has happened in the last decade, which is that the top five stocks are all tech companies.
00:15:59.039 --> 00:16:06.879
The tech industry, there's never been a time where the entire like top five stocks in the SP 500 are all from the same industry.
00:16:07.120 --> 00:16:11.919
And not only that, but the index is weighted increasingly towards those stocks.
00:16:12.159 --> 00:16:18.240
So I think in 2020 it was something like around 12% of the value of the S P 500 was tech companies.
00:16:18.320 --> 00:16:19.440
Now it's 25%.
00:16:20.080 --> 00:16:24.000
So this is when you ask me like what will it be like in 20 years?
00:16:24.159 --> 00:16:41.679
Like I just think like statistically, like the most likely scenario is it just stays the same that these tech companies will continue because that's this is we're in such an unprecedented situation where the power is concentrated so much among those companies, it'll be quite hard to loosen that grip.
00:16:42.480 --> 00:16:44.559
Well, yeah, two things could loosen that grip.
00:16:44.720 --> 00:16:56.480
One, uh major war, which ends up reprogramming the economy either intentionally or accidentally, or two, one thing I wanted to ask you about the sustainability angle of all of this.
00:16:56.720 --> 00:17:05.279
Because for a long time I've been looking at sort of the resources that are required for sort of training some of these models, and in particular, sort of water.
00:17:05.519 --> 00:17:07.920
Obviously, lots of futurists look at these sorts of resources.
00:17:08.079 --> 00:17:15.039
I mean, it takes as much water to, I don't know, train GPT, uh chat GPT to as it does to call a nuclear reactor, etc.
00:17:15.279 --> 00:17:28.960
So these sorts of you know basic resources that we need for human survival, are they going to get triaged towards actually the digital public infrastructure, for example, that's all, you know, running on and dependent on AI?
00:17:29.119 --> 00:17:37.440
Did you look at, because I've read I've read a lot of the book, but I haven't finished it yet, this angle about sustainability and whether that is an obstacle in the future.
00:17:37.759 --> 00:17:38.960
Do you know it's funny?
00:17:39.119 --> 00:17:48.400
I see it as absolutely a valid point of concern, but I'm actually not sure if how big a problem it's going to be in the next five to ten years.
00:17:48.559 --> 00:17:55.680
And the reason for that is we keep hearing from people in the tech industry that they are working so hard on efficient.
00:17:55.920 --> 00:18:06.880
Like it's in their interest, not just for the good of climate, but it's genuinely in their interests to make these models as efficient as possible and not so data hungry because that's expensive for them.
00:18:07.119 --> 00:18:13.359
So if you go to an AI conference, you know, just in the past year, that's what every all the researchers are talking about.
00:18:13.440 --> 00:18:19.359
That's what they're working on is how do we make these models more efficient, less data hungry, less power hungry?
00:18:19.519 --> 00:18:32.480
You know, I was talking to someone at Google Cloud, one of the executives there, and they, you know, of course they would say it, but this guy was very, very confident, like, yeah, they're absolutely the these things are gonna become more efficient over time.
00:18:32.559 --> 00:18:34.319
The chips are gonna become more efficient.
00:18:34.559 --> 00:18:35.759
Will that actually happen?
00:18:36.000 --> 00:18:38.720
Will the human ingenuity be there to make it happen?
00:18:38.880 --> 00:18:39.519
I don't know.
00:18:40.240 --> 00:18:43.359
Well, they're gonna have to make it happen, I guess, um, in a way.
00:18:43.839 --> 00:18:52.720
Netting out today, uh, where do you think open AI is versus DeepMind or or Google really or Microsoft and Google?
00:18:52.960 --> 00:18:59.440
Obviously, there have been some recent changes with OpenAI saying, you know, they are literally going moving from non-profit to profit.
00:18:59.599 --> 00:19:00.720
Where are we now?
00:19:01.039 --> 00:19:02.240
Yeah, it's a great question.
00:19:02.319 --> 00:19:09.119
And you would think maybe OpenAI because they were the first, you know, at the starting gun with Chat GPT.
00:19:09.200 --> 00:19:17.599
I mean, they were just way ahead of everyone else with this large language model that you could chat to, and it's like literally like talking to a human.
00:19:17.759 --> 00:19:20.240
It could generate this incredible human-like prose.
00:19:20.480 --> 00:19:25.440
But since then, I think Google has done a really good job of catching up.
00:19:25.519 --> 00:19:26.640
It was caught on the back foot.
00:19:26.880 --> 00:19:28.079
Given that they invented it.
00:19:28.559 --> 00:19:29.039
That's right.
00:19:29.279 --> 00:19:32.319
Yeah, which I think was probably a little bit humiliating for them.
00:19:32.480 --> 00:19:42.880
Yeah, because they they'd invented it and then they hadn't released it because they were worried about what this, you know, this AI model could say, what weird things it could say when the public start using it.
00:19:43.039 --> 00:19:51.599
But since then, we now see the the big advantage that Google has, which open AI does not have, is distribution.
00:19:51.839 --> 00:20:03.839
So even though OpenAI is effectively like a product arm of Microsoft now, so you can say, well, Microsoft, you know, it's everywhere, everyone uses Word, but not in the same way people use Google.
00:20:04.000 --> 00:20:06.160
It just doesn't have the same kind of reach.
00:20:06.400 --> 00:20:15.359
Like you, if you have an iPhone, the default search engine on your iPhone is Google, because Google pays Apple billions of dollars to allow it to be on there.
00:20:15.680 --> 00:20:21.920
And it's Android phones are used by are on the are on billions of smartphones around the world.
00:20:22.079 --> 00:20:32.880
And and and the Google's new Gemini model, which is effectively its response to Chat GPT, that's being baked into every uh Android phone now.
00:20:33.119 --> 00:20:36.000
So I think that is like a really big advantage.
00:20:36.079 --> 00:20:56.799
And not only that, but when you use an Android phone and you use Gemini on the system, you know, because Gmail and GDocs and the camera and the calendar, all these things are linked, and you can sort of use Gemini as a much more kind of comprehensive assistant that can talk between all those different apps.
00:20:56.960 --> 00:20:58.319
OpenAI can't do that.
00:20:58.400 --> 00:21:01.440
It doesn't have access to all those different apps that people use.
00:21:01.839 --> 00:21:09.920
It's relying on other businesses to build it into their businesses to sort of do the same thing, but not as effectively.
00:21:10.079 --> 00:21:14.880
So it's kind of, I mean, it's hard to say like who who's really like ahead right now.
00:21:15.359 --> 00:21:22.319
It's so close, but if I really had to put money on it, I would say Google, because they're the one that have the the the most distribution.
00:21:23.039 --> 00:21:23.599
It's very interesting.
00:21:23.680 --> 00:21:24.160
It's true.
00:21:24.240 --> 00:21:25.519
They're like the Coca-Cola, aren't they?
00:21:25.599 --> 00:21:27.599
It's like the Pepsian Coca-Cola, I always think.
00:21:28.079 --> 00:21:32.640
So um when we chatted uh originally, we talked a little bit about agency.
00:21:32.880 --> 00:21:42.640
So I did want to ask you about that because it's obviously something that comes up when you're just talking to anybody about the future and what they think the implications of some of these AI tools are.
00:21:42.799 --> 00:21:44.799
I mean, obviously they're incredibly convenient.
00:21:45.039 --> 00:21:50.079
You can be really creative, maybe you can be productive with them, we're not really sure yet.
00:21:50.240 --> 00:21:59.759
But um, people are worried about how much agency it might rob from them, given that it's working on the basis of their own personal data as inputs, and actually it's making decisions.
00:22:00.160 --> 00:22:02.960
Decisions or managing outcomes for them.
00:22:03.359 --> 00:22:04.400
What's your point of view?
00:22:04.640 --> 00:22:08.559
Having researched this so in depth as you have, what's your point of view on this?
00:22:09.279 --> 00:22:14.079
I think there's definitely the risk of an erosion of human agency.
00:22:14.640 --> 00:22:18.559
The problem is this is a very, very difficult thing to measure.
00:22:18.880 --> 00:22:22.960
It's kind of an abstract, squishy subject.
00:22:23.200 --> 00:22:27.119
Like, how do you measure that my agency as a human has declined?
00:22:27.359 --> 00:22:31.200
It's not even really something that people talk about that much, but they recognize it.
00:22:31.279 --> 00:22:38.799
Like when you just talk about how addicted we are to our smartphones, for example, does that represent a loss of human agency?
00:22:39.039 --> 00:22:53.759
I think when it comes to how we make decisions, I think there are going to be more situations where companies and organizations and people just outsource their decision making to generative AI, to AI.
00:22:53.920 --> 00:23:03.519
And what's really interesting is when you hear companies talk about this and AI makers talk about this, they say, but we don't want AI to take over that to make the decisions.
00:23:03.599 --> 00:23:06.240
We want them to augment human decisions.
00:23:06.480 --> 00:23:10.000
But actually, if you think about it, humans are kind of lazy.
00:23:10.079 --> 00:23:13.920
And if the easy option is to let a computer make the decision, then we will.
00:23:14.079 --> 00:23:15.599
And I'll give you an example of this.
00:23:15.839 --> 00:23:24.079
So London's Metropolitan Police have been trialing facial recognition, which is a form of AI, for a number of years.
00:23:24.240 --> 00:23:39.599
And there was a study done of officers who were using facial recognition cameras from a police van in the middle of a busy street, and it was just the camera was looking for people as they walked down the street to see if any of them matched a watch list.
00:23:39.839 --> 00:23:45.920
Now, the problem with facial recognition is that sometimes it's wrong, particularly when it's looking at people of color.
00:23:46.240 --> 00:23:59.680
But what was happening inside the police van is that when the face flashed up on the screen as a match, the officer assumed that the facial recognition system was correct.
00:23:59.920 --> 00:24:05.200
Rather than sort of double check, they would go out and just kind of detain the person.
00:24:05.680 --> 00:24:15.839
And the scientists who were studying this, who were in the van observing the whole phenomenon, actually gave a name to this and they called it deference to the algorithm.
00:24:16.000 --> 00:24:33.200
And this was a big concern because we assume that computers are often or more often right than humans are, because they are backed by these networks of neural networks and uh and their machines, and they're just like a calculator, it's always correct, right?
00:24:33.359 --> 00:24:35.519
Um so you just assume it's correct.
00:24:35.680 --> 00:24:49.119
And I think that's uh that to me is just an example of the possibility that we are actually just going to delegate a lot of decision making to computers because we trust them and because it's easier for us.
00:24:49.599 --> 00:24:58.720
So whether that happens in the criminal justice system, in businesses, in marketing, in customer service, I could see this happening in all sorts of different parts of life.
00:24:59.039 --> 00:24:59.200
Yeah.
00:24:59.359 --> 00:25:08.799
So, and you do cover in the book elements of the kind of trust and safety teams and what has happened to those over time, um, like Temerick Gibr and people like that.
00:25:08.880 --> 00:25:13.200
And what's your point of view on where we are with trust and safety?
00:25:13.359 --> 00:25:17.119
Even, you know, even broader than maybe the two companies you're talking about.
00:25:17.359 --> 00:25:18.720
Has trust and safety helped?
00:25:18.880 --> 00:25:20.799
Has it just been a massive distraction?
00:25:21.039 --> 00:25:22.559
Do we need to start again with it?
00:25:22.720 --> 00:25:23.599
What do you think?
00:25:24.000 --> 00:25:28.960
Oh, it's definitely helped for the little that resources that these teams have.
00:25:29.200 --> 00:25:50.960
But when you talk to people who work on these teams and people who have knowledge of these teams, the overwhelming consensus is they're tiny and they are under-resourced, and they're largely drowned out in all the product meetings by the people whose greater priority is to get these systems bigger, um, get them to launch faster.
00:25:51.200 --> 00:25:54.400
And of course, you know, I'm not going to dismiss all these efforts.
00:25:54.480 --> 00:26:04.559
Like Google, I think, really does put a lot of time and effort into trying to build its algorithms responsibly, but at the end of the day, it's not really being regulated for this stuff.
00:26:04.799 --> 00:26:08.799
And so it's very much its own decision how much investment to put into these.
00:26:08.880 --> 00:26:11.680
And by and large, these teams are small.
00:26:12.000 --> 00:26:19.519
You know, the other thing is if I went to Google now and said, how big is your trust and safety team, they don't have to tell me, and they probably wouldn't tell me.
00:26:19.680 --> 00:26:21.440
I've asked these kinds of questions before.
00:26:21.599 --> 00:26:26.000
I can't tell you the number of times I've asked Facebook, how many content moderators do you have?
00:26:26.079 --> 00:26:27.200
And they never tell you.
00:26:27.359 --> 00:26:29.359
They never tell you because they don't have to.
00:26:29.519 --> 00:26:42.720
And so the hope is that with new regulations coming down the pipe, from especially from the European Union, uh, there's going to be more transparency from these companies over the kinds of risk assessments that they're doing for their algorithms.
00:26:43.359 --> 00:26:57.279
This is interesting because at the moment we've also got Drag's report coming out saying that, you know, well, actually, we're regulating, I mean, talking generally, not about this specific case, but generally we're regulating so much that we're stifling innovation.
00:26:57.440 --> 00:27:05.680
So we need to, you know, maybe think about that and potentially not reverse it, but, you know, is there too much regulation that's stifling innovation?
00:27:05.759 --> 00:27:09.599
So I wonder actually what's going to happen with trust and safety in general.
00:27:10.079 --> 00:27:14.319
Yeah, well, I think the approach by the European Union is quite good.
00:27:14.559 --> 00:27:17.359
They have the AI Act, which is coming into force.
00:27:17.440 --> 00:27:23.599
It's sort of coming into force now, really, but I don't think we're gonna see the effect of it for another year or so.
00:27:23.839 --> 00:27:42.079
But part of the way that act is gonna work is that companies like Google, like Apple, um, OpenAI, DeepMind, they're going to have to be doing these risk assessments on their algorithms, and they're gonna have to share the results of that with auditors, with people from outside the company.
00:27:42.240 --> 00:27:50.799
This is unprecedented because for years, the only people that these companies were accountable to was themselves.
00:27:50.960 --> 00:27:53.279
They literally self-regulate.
00:27:53.519 --> 00:27:54.240
It's so funny.
00:27:54.319 --> 00:28:00.079
Like companies like Uber and Facebook, they have this term transparency report.
00:28:00.160 --> 00:28:02.319
They all put out these quarterly transparency.
00:28:02.799 --> 00:28:12.000
It's so funny that they call them transparency reports when it's literally produced by them and you just trust that the numbers are true.
00:28:12.160 --> 00:28:20.480
They're not, there's no like auditor or like a big consultancy firm like anyone from the big five actually checking their homework.
00:28:20.640 --> 00:28:23.680
It's just the the the companies doing it themselves.
00:28:23.920 --> 00:28:29.359
So it sounds, it seems like this is on course to finally change at some point in the next year.
00:28:29.680 --> 00:28:32.319
Now, two questions very specific that I wanted to ask you.
00:28:32.400 --> 00:28:36.720
One from me and one from somebody who I asked, what would you love to ask Palmy?
00:28:37.119 --> 00:28:43.599
My question is really about Eric Schmidt and what he talked about recently, about agent-based systems.
00:28:43.839 --> 00:28:59.119
Actually, you know, you almost be able to replace the programmer because you'll be able to give it some kind of prompt or instruction to, you know, I think in the case he was giving, it was copy all of TikTok and do the, and if it, you know, in half an hour it doesn't give you the results, start again and redesign the system.
00:28:59.359 --> 00:29:03.599
Where do you think we're going with these AI agent-based systems?
00:29:03.759 --> 00:29:11.599
Are they going to be in the hands of users so that we can personally affect things so they can go out and work on our behalf?
00:29:11.839 --> 00:29:25.920
Or actually are they more centralized and they're almost the property of these big companies, the tech giants that you're talking about, and actually the consumer's just on the end of whatever it's been programmed or prompted to do.
00:29:26.319 --> 00:29:32.000
Yeah, that's a really good question about whether this is something that consumers are going to have that much control of.
00:29:32.319 --> 00:29:38.079
I think there's probably going to be products that come out, like ChatGPT, for instance.
00:29:38.240 --> 00:29:44.880
I think this will like ChatGPT will probably be given I hate this word, but agentic properties.
00:29:45.200 --> 00:30:02.960
And so the latest version of ChatGPT, which is called O1, has these reasoning capabilities, which allows it to do tasks or carry out sort of reasoning tasks where it actually takes a few seconds to quote unquote think about something before it does something.
00:30:03.039 --> 00:30:12.880
And these are kind of more mathematical in nature than the sort of language-based text prediction that it was that it was doing before.
00:30:13.119 --> 00:30:20.880
And this has been touted as a precursor to Chat GPT eventually being able to carry out tasks like an agent.
00:30:21.039 --> 00:30:22.799
So that's probably coming.
00:30:23.039 --> 00:30:43.839
One of the first examples of agent-based AI has actually come from Salesforce, which is this big software as a service company that sells uh enterprise software to all sorts of companies like banks, um, British Airways, you know, it's like oftentimes they use their customer relationship management software.
00:30:44.079 --> 00:30:59.759
But what their agents can now do is like in customer service, if I'm going onto, let's just say, for example, British Airways's website and I want to book something, like the agent isn't just going to give me information, it could actually book something for me, or it could lodge a complaint.
00:31:00.000 --> 00:31:02.640
So it can it can actually take action.
00:31:02.720 --> 00:31:04.559
And some people are quite concerned about this.
00:31:04.640 --> 00:31:08.960
Like, well, these large language models have been making mistakes, they've been hallucinating.
00:31:09.200 --> 00:31:17.839
What's to say they won't, you know, accidentally, you know, make an incorrect booking, book the wrong flight, or take some kind of dangerous action.
00:31:18.160 --> 00:31:21.039
Then you can can go and hire an AI lawyer to sort it out.
00:31:21.359 --> 00:31:22.319
Yes, that's right.
00:31:22.480 --> 00:31:23.119
Exactly.
00:31:23.359 --> 00:31:25.920
We'll really feel like we're in the matrix at that point.
00:31:26.160 --> 00:31:30.240
I think my big question is what happens when these things make mistakes?
00:31:30.559 --> 00:31:31.359
How will we know?
00:31:31.599 --> 00:31:32.079
Yeah.
00:31:32.319 --> 00:31:34.799
You want to book a book an appointment with the doctor?
00:31:35.039 --> 00:31:36.640
This is what Salesforce were telling me.
00:31:36.720 --> 00:31:40.079
Like this is it'll be able to actually book the appointment for you.
00:31:40.319 --> 00:31:45.279
So just what happens when these things uh actually mess up?
00:31:45.839 --> 00:31:46.720
It's a new system.
00:31:46.799 --> 00:31:48.319
We don't have any rules for it yet.
00:31:48.480 --> 00:31:50.079
And it's gonna be very interesting.
00:31:50.319 --> 00:31:59.200
The question that Chris wanted to ask you was and he works with a lot of clients, he's obviously trying to um bring them with him into the world of AI.
00:31:59.279 --> 00:32:08.319
Uh, but he's saying, I find the most impactful AIs are based on small proprietary models versus general purpose LLMs trains on the public web.
00:32:08.559 --> 00:32:14.480
So in the battle for supremacy, the title of your book, um, how will that square in the future?
00:32:14.720 --> 00:32:18.000
Small versus large, you know, private versus public.
00:32:18.640 --> 00:32:19.519
Do you know it's funny?
00:32:19.599 --> 00:32:25.200
I've heard people say the same thing that these smaller, more specialized models will be more effective.
00:32:25.440 --> 00:32:34.240
I have yet to see that in practice, though, because I have seen examples of small specialized models just not be as good as the generalized models.
00:32:34.559 --> 00:32:46.000
You know, like the likes of just GPT-4 from OpenAI is better than even a model that a financial company is just a real-world example.
00:32:46.240 --> 00:32:50.000
A financial company made their own model based on their own financial data.
00:32:50.160 --> 00:32:59.440
It was a large language model, they spent like a year on it, and it just even based on their own proprietary data, it still wasn't as good as OpenAI's general purpose model.
00:32:59.759 --> 00:33:01.119
Why did they think that was?
00:33:01.599 --> 00:33:03.519
It just wasn't as accurate.
00:33:03.759 --> 00:33:07.839
It it just didn't, wasn't as effective, it wasn't as capable or lucid.
00:33:08.400 --> 00:33:10.720
So I don't know how clear it is.
00:33:10.799 --> 00:33:18.400
I I like maybe some of these smaller models will still be useful, particularly for companies who don't have the resources to pay for the big models.
00:33:18.640 --> 00:33:25.519
It's like like open source, it's an alternative to to having to pay for access to the API of of OpenAI or Gemini.
00:33:26.079 --> 00:33:31.359
But I think I don't know, I I think the the question is still an open question about whether people will use these.
00:33:31.519 --> 00:33:37.200
And you know, one thing I would love right now, which I haven't been able to find, is market share data.
00:33:37.359 --> 00:33:42.240
Like in just in just the market for large language models, who's actually dominating?
00:33:42.400 --> 00:33:53.200
I I just anecdotally from what I've read, it seems to be Google's Gemini and OpenAI's GPT, but who who's behind them?
00:33:53.359 --> 00:33:54.319
We just don't know yet.
00:33:54.400 --> 00:33:58.319
So that's hopefully something we'll find out in the next you know months, few months to a year.
00:33:58.720 --> 00:34:02.000
I don't really know anyone who only uses one.
00:34:02.240 --> 00:34:10.079
So I will use ChatGPT and then I'll put that output into Claude, for example, and get Claude to improve on it and then put it back to whatever.
00:34:10.239 --> 00:34:13.039
And then on my mobile, I use perplexity a lot, actually.
00:34:13.199 --> 00:34:22.559
I don't know why, but when I'm on the go and I've got smaller questions, sometimes I use perplex and I like to use like all of them because they all do, they all have different strengths and weaknesses, I think.
00:34:22.719 --> 00:34:31.119
So it's quite good to get a a bit of an adversarial relationship going between like they have come up through different ways and they're so they've got different strengths and weaknesses.
00:34:31.360 --> 00:34:36.400
Yeah, and perplexity was started by academics, and which is why they're really big on sources.
00:34:36.639 --> 00:34:37.280
Yeah, I love that.
00:34:37.440 --> 00:34:38.960
Yeah, and it is really good.
00:34:39.199 --> 00:34:45.599
But I in a you know, another wrinkle to this whole dynamic is none of these really have a strong moat, right?
00:34:45.920 --> 00:34:56.159
And just like you, I was talking to a startup, an AI startup, who were just like, yeah, we use cloud, and then sometimes we use we use OpenAI, you know, for their own product.
00:34:56.480 --> 00:35:00.239
Their product is like built on top of those things, and it's really easy to switch.
00:35:00.400 --> 00:35:02.159
And that's very unusual in tech.
00:35:02.239 --> 00:35:10.800
Like with if you're using a cloud provider, for example, like Amazon's AWS, it's really hard to switch to, say, Microsoft Azure.
00:35:10.960 --> 00:35:13.840
But it's super easy to switch between foundation models.
00:35:14.079 --> 00:35:20.719
And apparently the CEO of Anthropic was asked about this at a dinner, at an AI dinner, like, what's your moat?
00:35:20.880 --> 00:35:23.840
And he was like, Well, this is sorry, this is hearsay, by the way.
00:35:23.920 --> 00:35:26.079
This is what I heard from one person who was at the dinner.
00:35:26.239 --> 00:35:33.360
And he said, Well, our moat is that nobody else can really afford to spend, you know, X billion dollars on compute like we can.
00:35:33.599 --> 00:35:38.079
So I mean, yeah, I guess that's a moat, but that's not that great a moat.
00:35:38.239 --> 00:35:39.039
It's this is just money.
00:35:39.199 --> 00:35:39.920
The moat's money.
00:35:40.159 --> 00:35:47.280
I think just try and use as much of it as possible all across the board while it's still relatively affordable.
00:35:47.840 --> 00:35:50.480
Great point, because this is the history of tech, right?
00:35:50.559 --> 00:35:58.719
Like you start off cheap and then they just jack up the prices once they've got once they've really established economies of scale and have users who are utterly dependent on them.
00:35:59.039 --> 00:35:59.440
Yeah.
00:36:00.000 --> 00:36:01.519
Whilst we're both using each other.
00:36:01.760 --> 00:36:02.000
Yeah.
00:36:02.159 --> 00:36:07.119
Um so as we close then, and it's been brilliant to spend some time with you and listen to this.
00:36:07.280 --> 00:36:10.159
What should I have asked you about that we haven't covered?
00:36:10.639 --> 00:36:18.639
Like I often hear, I worry when I'm talking about my book with people that I just leave people feeling kind of depressed about the future.
00:36:18.960 --> 00:36:22.639
Maybe I don't I feel like we haven't gone into that too much with this discussion.
00:36:22.800 --> 00:36:26.639
But I would say, like, I am hopeful about all the regulatory efforts that are happening.
00:36:26.719 --> 00:36:29.920
And it's not just regulators, it's also efforts in the courts.
00:36:30.000 --> 00:36:41.199
You know, there's some really interesting court cases out there, like the one in San Francisco that's fighting uh websites that produce deep fake porn and are trying to shut those websites down.
00:36:41.440 --> 00:36:43.920
That's something that I'm hoping to write about in the next couple of weeks.
00:36:44.159 --> 00:36:45.920
I think they're doing amazing work.
00:36:46.239 --> 00:36:54.639
And, you know, there are state laws in the U.S., there are unions, like think about the Writers' Guild and the Writer's Strike in Hollywood.
00:36:54.719 --> 00:37:00.159
Um, and they manage to change the contract language about using AI for script writing and things like that.
00:37:00.320 --> 00:37:10.480
So I worry a little bit about civil society and sort of campaign groups with names like AI Now and the Institute for Strategic Dialogue and Open Markets Institute.
00:37:10.559 --> 00:37:21.360
And these organizations are very poorly funded uh compared to tech companies and are just trying to fight back against some of the monopolistic practices of tech companies and the AI harms.
00:37:21.599 --> 00:37:28.000
A lot of these civil society groups, I feel like their voices are being drowned out a little bit by all the hype in tech.
00:37:28.480 --> 00:37:36.800
So I what I would love to see is just hearing, just see more of those voices in the public and just so we're just a little bit more aware.
00:37:36.960 --> 00:37:47.920
You know, going back to the Nobel Prize for Demis and Jeffrey Hinton, you know, two people linked to Google, two or three people actually, this is great casting a really positive glow on Google.
00:37:48.079 --> 00:38:07.280
I worry a little bit, it kind of muddies the debate, like it kind of eclipses some of the risks that we're trying to talk about with AI and just spotlighting the great stuff that AI does when actually this is such a tiny part of Google DeepMind's real investment and effort.
00:38:07.519 --> 00:38:13.679
So yeah, I think we just we do have to kind of talk about some of these, uh, some of these risks and and how they're being mitigated.
00:38:14.159 --> 00:38:17.440
I think the thing is there's just so many risks at the minute in the news.
00:38:17.599 --> 00:38:19.760
Like there's a lot of threats, there's a lot of risks.
00:38:19.840 --> 00:38:30.960
It's very difficult for people, the general public, to get their heads around the risks of AI, whether it's AGI or whether it's bias or what whatever it is, because there's so many threats at the minute.
00:38:31.119 --> 00:38:33.280
And it's kind of um kind of interesting.
00:38:33.599 --> 00:38:34.800
That's brilliant, Palmy.
00:38:35.039 --> 00:38:38.639
Thank you so much for spending a bit of time with us and um congratulations on the book.
00:38:38.800 --> 00:38:40.639
We shall put all the links in the show notes.
00:38:40.800 --> 00:38:45.519
Is there anything else you'd like to add about the book apart from the fact that everybody should go and buy it?
00:38:45.920 --> 00:38:53.440
Uh well, it has just been shortlisted for the Financial Times Business Book of the Year Award, which is which is very, very exciting.
00:38:53.519 --> 00:38:56.559
Um, so we'll find out in December with the result of that.
00:38:56.719 --> 00:39:02.239
It's been great just being able to talk about the book and tell the story of the people behind the the technology.
00:39:07.599 --> 00:39:11.039
Thank you for listening to The Future of You, hosted by me, Tracy Follows.
00:39:11.119 --> 00:39:15.039
Be sure to check out the show notes for more info about the topics we covered today.
00:39:15.199 --> 00:39:19.760
If you enjoyed this episode, please like and subscribe wherever you get your podcasts.
00:39:19.920 --> 00:39:23.599
And if you know someone who would love this episode, please share it with them.
00:39:23.760 --> 00:39:29.599
For more on the future of identity in a digital world, visit futuremade.group slash the future of you.
00:39:30.000 --> 00:39:34.719
To explore the future of everything else, head over to future made.group.
00:39:34.800 --> 00:39:38.480
The Future of View podcast is produced by Big Tent Media.
