Speaker 5: [00:00:00] You’re listening to a new episode of The Brave Technologist, and this one features Fred Gifford, who is an AI strategy lead based in London. He spent the past five and a half years at Sony AI applying frontier AI research to solve challenges in gastronomy, AI ethics, and biomedical. He’s also an author under the pen name Fred Lunzer.

His debut novel, Sike, focuses on a young couple navigating life with an AI therapist. In this episode, we discuss Sony AI’s research on building biomedical knowledge graphs, the importance of open research in AI, how writing about fictional AI features compare to the real thing, along with AI ethics and accountability.

And now for this week’s episode of The Brave Technologist.

​

Speaker 5: Fred, uh, welcome to The Brave Technologist. How are you? Very well. Thank you for having me. Yeah. Thanks for coming. I mean, uh, we’re here at London at the AI summit. I know you were, uh, moderating a panel earlier, , about AI and film. , What was kind of the biggest takeaway, uh, that you had from the panelists?

Speaker 6: So, so I think [00:01:00] takeaways when you’re talking about AI and art, the- or AI and the arts, uh, it’s always interesting thinking about what, what the limitations are today, and maybe in this instance, the question was kind of, uh, what- which parts of AI filmmaking are still deeply human, and I think that’s a great question.

Uh, um, and I- it, it… The consensus is around, you know, all the things you’d expect it to be, things like storytelling, , narrative, those, those, those parts that AI can’t really get close to, and I think there’s something there to do with discovery. There was a great Rich Sutton talk for SARE where he’s talking about generative AI and how much or how little it can do in terms of discovery and really, like, finding new things.

I won’t paraphrase the talk, but it’s a great one, and I think that sort of applies to, to elements of storytelling and narrative as well. Like, AI can do certain things, but actually, and, and I’m biased because I, ‘cause I’m a writer, but, um, AI can do some things but can’t quite get that far into, into that- Yeah

human space. Yeah. I hope that makes [00:02:00] sense.

Speaker 5: Humans seem to have a way of, like, smelling it out a little bit too- That’s right … where it’s like, okay, it looks really awesome, but people can still kinda tell, you know? Yes. It’s still got that shine to it or whatever- Yes … you know, to where, like, humans are just good at sensing that stuff.

Speaker 6: Completely.

Speaker 5: Until we’re not, and then by then, you know, whatever. Uh, but, you know, Sony’s been kind of building… Sony AI’s been building on top of, like, biomedical knowledge graphs. Can you walk us through what that product actually does and, and- Yeah … what the problem’s solving?

Speaker 6: Yeah, for sure. so obviously in pharmaceutical, uh, and in drug development, data’s still extremely scarce, failure still at 90%, cost of getting a drug or asset to clinical trial rising and at around $300 million, and that might still fail.

Uh, we tried to come at that and, and, and support that process from the knowledge space, so really thinking about you have the natural world, science explains some part of that, scientific knowledge explains some part of science, and scientific knowledge has this great data source, which is academic [00:03:00] papers.

And, uh, they’re kind of considered quite fusty. Um, everyone knows about academic papers. we tried to take a new approach to that, and so what we do is we build knowledge graphs out of academic papers, knowledge graphs built of, of triples, of, um, uh, an entity, say a gene, and its relationship to another entity, like a disease or a drug.

Um, we look at how the, the knowledge evolves over time, so we timestamp these knowledge graphs, look at how they evolve, uh, and then use that to predict new relationships in the future that might come out, and that’s a really interesting- Uh, system for loads of reasons. We also have explainability functions built in there.

But in particular, I think it gets interesting not only because you can come up with new hypotheses, but because you can then bring into drug development a new data source- Mm-hmm … which is a kind of k- knowledge graph model-created dataset. Uh, and, and with that comes a score. And so if you think about drug development today, there are loads of ways you score candidates going through the pipeline.

But as far as we’re aware, no one [00:04:00] is taking the weight of scientific literature, of everything everyone’s discovered, and building that into a score and applying that to the candidates. And so that’s really, that’s really what we try to do.

Speaker 5: Super interesting. has Sony been in that field for a while?

Like, uh, ‘cause I feel like-

Speaker 6: Yeah … uh,

Speaker 5: it’s a, it’s kind of a different angle than what one might expect. Like, I have Sony stuff in my house or whatever, you know?

Speaker 6: Yeah, totally. Totally. Uh, no, and, and biomedical is not an area that Sony’s done a bunch in. Uh, Sony AI was set up in 2020,

It’s 250 people now. Focus 80% of its work on, or 70%, on projects that are related to Sony, like movies, music, and gaming, and then 30%, which was more blue sky areas like gastronomy. Nice. Can you give Michelin-starred chefs data tools and, and robots? Um, we just released a huge project, which is a table tennis-playing robot.

Speaker 5: Wow.

Speaker 6: If you haven’t watched the videos, this is the most exciting- … project I’ve ever seen. I didn’t work on it. I sat beside it. Um, but it’s incredible. And biomedical’s another one of those outside of [00:05:00] Sony’s kind of, uh, main areas.

Speaker 5: It’s awesome, though. I mean, like, it’s a huge area where it can have a huge impact on humanity, right?

Like- For sure … and, and it’s cool.

The Brave Technologist is brought to you by the Brave Search API. Access billions of indexed web results from a simple API call with the Brave Search API. Join the leading names in AI and tech using the Brave Search API to power agenic search, keep LLMs current with real-time data, train foundational models, and bring the best of the web directly to the leading edge.

Luke: Get started today at brave.com/api.

Speaker 5: So Sony AI publishes a lot of open research too, right? What’s a strategic argument or, or a bet that, that Sony’s taking, , uh, on that with that kind of position?

Speaker 6: Yeah. It’s a great question. I mean, I can only offer my, my own opinion on it.

Um- But I think, you know, I think there are a few elements where Sony has always kind of built for the community. Mm-hmm. It’s, it’s always thought about building technology for creators. Obviously, it’s trying to, trying to do that in loads of different ways. I think y- there’s also elements of kind of [00:06:00] standardization that Sony has always done.

It’s always looked at how you, how they can build standards for the entire industry. And so with some of our projects, like our AI ethics work, it’s like a, a simple no-brainer that that stuff would all be public. And so we- Yeah … so, so we’ve done loads of stuff there. I think in other ways, being a bit more cynical, there, there are probably ways in which it’s just not too dangerous.

Because if you build a kind of model for gaming, PlayStation is a kind of closed circuit. Right. So if you, if you release some of that stuff, well, it’s kind of only PlayStation that can play with it, if that makes sense. Sure. Sure. Sure. So there’s not maybe too much risk of, of competition. Um, that’s kind of cynical.

I think anyway, there is a strong, , push. , Sony’s from Japan. In Japan, my understanding is that open sourcing has been, like, relatively peddling, basically. Mm-hmm. Sony has got a bit of a push to, to open source stuff, and definitely at Sony AI, we’re a research group. Everyone’s researchers.

A lot of them come from academia. It, it’s kind of in our DNA a bit.

Speaker 5: I just totally commend you guys for doing that, too. I mean-we, we… Everything we do at Brave’s all open sourced, and it’s, like, one of those things [00:07:00] where it takes a few months to rewire your brain for it a little bit, but once you do, it’s like, “Oh, my gosh, I wish everything was like this.”

Because, you know, you, you basically have, like, the world building with you, or also s- finding things, and y- It’s a r- such a, a better way to build. Like- Yeah … and, and with what you guys are doing, uh, around these graphs, I think it’s, like, super cool to have that open, you never know whatever other people can run with, with these things- For sure

or even the, the ideas- Yeah … right, that can kinda come from that. Mm. Super cool. , You’re also an author, right?, Was it Psyche? Is that what your novel- That’s right. Yeah, yeah. Yeah, okay. I, uh, so, um, and it kind of centers on a couple navigating life with an AI therapist, which I thought was super interesting.

Why did you do a book specifically about that, uh, AI therapy, and, and what were you trying to kinda… Were you, were you trying to work something out with that, with your own- Yeah … head? Yeah, yeah. Absolutely. So, so, so as you say, it’s about a young, a young rap ghostwriter. He’s, he’s a kind of rap obsessive, and, uh, he’s fallen in love with a young venture capitalist, and he has an AI therapist that sits on his smart [00:08:00] glasses, which in this world are pretty ubiquitous.

Speaker 6: And the AI therapist watches everything he does, how much he laughs, cries, uh, how much sunshine he sees, birds singing, things like that, and gives him kind of , mental health advice, and guides him towards contentment and calm. And I think the thing I was trying to do there is… I mean, the, the book is really not a dystopia.

It’s, it’s about AI. I kinda wrote it speculatively before ChatGPT broke, um, and before AI therapy became much more mainstream. It’s really not a dystopia, and I was trying very hard to, to push away from the kind of surface level hype, big hopes, big fears, um, that we have for AI, and think about what happens when AI is ubiquitous- Mm-hmm

is as ubiquitous as, say, our smartphone or electricity or something like that. And that’s kind of the place I was trying to get to. And, um, the question there was, you know, what, what does that look like? What, what, what do we need to be worried about then, or what can we be optimistic about then? Um, I was really against, you know, as we just alluded to, I was really [00:09:00] against a kind of polarized view on this.

I don’t wanna be, polarized towards or against AI. And so that’s where I was trying to get. I was trying to focus on this kind of middle ground, what happens there. Um, in terms of AI and therapy, I think they’re an interesting pairing because I think we load them with loads of expectations.

Yeah. So we hope AI will answer all our questions. We also fear it will answer all our questions and take over the world. Um, we also hope psychology will answer all our questions, and, and we overload it with too much. Yeah. And it’s, it’s a bit, it’s a bit too much in the age of kind of, especially, um, social media therapy.

I think it can go overboard, and so th- that pairing was interesting to me as well. When

Speaker 5: did you write the book?

Speaker 6: So I wrote it kind of between 2021 and 2023 was the main bulk of the writing.

Speaker 5: It’s interesting too, like when you think about like how, you know, uh, Meta has these like Ray-Bans.

I think they’ve sold millions of these things now. Yeah. Like I just went and got glasses a few months ago, and I was asking, you know, the woman that was [00:10:00] helping me like how well they… She’s like, “They keep moving faster than we can keep them in stock.” Yeah. And, and how has that kind of juxtaposition, uh, a- against what you were thinking about when you were wr- i- is there any connection there?

Like are, are you, do you feel like you see it, it play out in a different way?

Speaker 6: Yeah.

Speaker 5: Uh, w- I mean, ‘cause it’s still, I guess it’s still a relatively small sample size of people using those, but it’s in the millions. Yeah. , How does that stack against what your, your perception was at the time?

Speaker 6: Yeah. So I mean, I, obviously it’s fiction, so I, I cheated and I made the smart glasses incredibly good.

Speaker 5: Mm-hmm.

Speaker 6: No offense to a Ray-Ban pro- product. But, but you know, I th- you know, they’re still developing, and they’re still- Sure … reaching dominant design, and like there, there are loads of things that they’re still doing.

And so I think in the smart glasses as well, I was making them as ubiquitous in the book as our smartphones, you know- Mm-hmm … which by this point kind of just do it all, right? Our smartphones, like we, there’s not really so much that we’re still expecting them to get better at.

Speaker 5: Right.

Speaker 6: Um, and so I sort of put smart glasses in the same boat.

On the AI [00:11:00] therapy side, though, I think looking at the book now, what, what surprises me about how much I got wrong was that in the book, the AI app is specifically focused on mental health. Mm. It’s an AI therapist app, and it’s run by a company that only focuses on mental health and AI therapy. So the mission is to improve users’ mental health, take it or leave it.

That is obviously not what the mission of OpenAI is, let’s say, and when we use ChatGPT for AI therapy, that’s a very different mission that they have. And so- Oh,

Speaker 5: yeah …

Speaker 6: this fascinates me, that I, I think I got it wrong- And , maybe it’s just really obvious or, or maybe it’s a useful literary - sort of technique anyway to do it like this.

But AI bots are kind of becoming a one-stop shop for everything. Mm-hmm. I don’t think they’re gonna be as specialized as I had kind of envisioned, envisioned them being, and I think that causes big problems. I think there’s then a misalignment between… and it’s probably the root of why, uh, we’re, we’re butting up against real issues of what we do with chatbots and what we take from [00:12:00] them.

Speaker 5: Yeah, yeah. There seems like a, a, a real tension too between therapeutic AI needing deep personal data to be useful and, and the privacy risks that creates even to the point of, like, Sam Altman saying, “Hey, like, people are using ChatGPT as a therapist or a lawyer or whatever.”

We still have to hand over those things if they’re asked of us, right, by authorities or whatever. Um, uh, has your view on, on that evolved at all? Like, were you… Was that part of your thinking when you were writing the book? are we… . How do you feel about this now, too?

Speaker 6: Do you mean in terms of the relationship between, um-

Speaker 5: Well, I think, like-

like

Speaker 6: data

Speaker 5: privacy? The, the, the prompts are kind of a, the form factor. It, it, it’s something where people kind of unconsciously will give it more information- Yeah … than if they’re doing a form field or something where they’ve got to… The input, just the way that you’re doing it and, and, and how personable it, it is, right?

, It seems like, uh, people are putting their… letting their guard down a little bit more a- and putting really personal stuff- what’s your take on, on the state [00:13:00] of that now?

Speaker 6: Yeah. So I think this… Absolutely. And I think, you know, in the book, , no data is ever taken out of the app. Mm. Okay. So that, that was like a… That was a way to slightly solve that privacy issue- Yeah, yeah, yeah … because I believed it had to be solved. And that point about AI chatbots being a one-stop shop for everything, I think that, that is a huge problem.

Like, you don’t necessarily want to share the data that you share with your insurance provider with your… that same data with your doctor, that same- Right … data with your boss. And so that, that really does seem to, to, to cause issues. Um, I mean, I just wouldn’t say it’s that new, right?

Speaker 5: Right.

Speaker 6: Data privacy is…

Or, or, , the surveillance capitalism, uh, world and, and challenge has been going on since before AI. Yeah. And I, I wonder if we’re, like, too immune to the fact that our data is going everywhere and-

Speaker 5: Yeah, I think there’s a lot of truth to that, and that’s a fair point. Yeah. I mean, , it’s basically proliferated, you know, a, a while ago.

But, um- Mm … but yeah, there’s just something about, too, like you’re getting into the, the thought process a little [00:14:00] bit differently- Right … it seems like. , AI ethics are also kind of, uh, listed as one of the core focus areas at Sony too, right? , What does an AI ethics function actually look like inside a company like Sony or a large company?

Speaker 6: So, um, so of course, across all of Sony, there is an AI function and an AI governance function, and I can’t speak too much to it. I don’t do that much with that function. And Sony AI itself is a research group, though, does huge amounts in AI ethics. Mm. And there’s some overlap between that team and the governance stuff.

A- and, and, um, that happens. I don’t know too much about the governance, but on the, on the AI ethics research side, uh, this is a team of, you know, complete superstars in the kind of ethics policy world doing the most amazing work. They’re really, like, a phenomenal team because they have… It’s completely unfair.

They kind of have every skill available. They’re, they’re kind of economists as well as, as, as well as policymakers, as well as AI experts. Um, they’re doing amazing work. Their big focus and the big thing that they released, which got front cover of Nature, was an ethical [00:15:00] image data set.

Speaker 5: Okay.

Speaker 6: So image data sets, obviously very grubby ethics in how the- they generally have been made- Oh, yeah

in the past. And they tried to build an , an ethical one w- building it with ethics and for ethics. They succeeded in this and with ethics, so all the photos, um, the images, human images given to them are, you know, the people are properly remunerated. Their consent has been given. They can do consent withdrawal.

Oh. Um, it’s, it’s diverse in all the important ways, and then you can use that to evaluate ethical, uh, or, or kind of, say, computer vision models, let’s say. , You can use this to evaluate them for bias, et cetera. Um, that, that’s a huge piece of work and really not easy, but it’s that type of stuff they really focus on and various other projects around that as well.

But yeah, for a company like Sony AI, that, that is a kind of meaningful project, I’d say.

Speaker 5: Yeah, and it’s not easy, right? Like, and it seems like if you can crack it on a research thing, right, like there’s potential with that framework to probably extend it out to other, other things, too.

Speaker 6: That’s exactly the aim, yeah.

Speaker 5: [00:16:00] Yeah, yeah, yeah. Um, yeah, it’s, it’s really interesting. Accountability al- also kind of comes up related to that, but also more generally speaking,, from your own point of view, like obviously not on the behalf of Sony or whatever, but like how do you see, us holding tech leaders and companies accou- or what would be ideal f- for accountability for these things?

‘Cause when things go wrong, they can go really wrong, right? Yeah. Like, and it doesn’t seem like there’s like… Even if there are laws sometimes selectively enforced, or maybe they weren’t written by people that actually understand how things work, right? You know, I… What, what’s your take on, on accountability with this stuff?

Speaker 6: Yeah. I mean- If I knew that

Speaker 5: It’s not an easy one,

Speaker 6: right? Like, uh- Yeah, I think, I think it’s such a difficult question, but it’s obviously such an important one, and-

Speaker 5: Right …

Speaker 6: um, I think also, like, being able to predict the issues, that’s really difficult. Mm-hmm. I remember when scraping was totally normal. Right.

When that was, like, considered, like, a completely legit thing to do, was to scrape the internet, and maybe people even had the philosophy that [00:17:00] scraping the internet was okay because that’s what the internet was there for. It was to kind of give information to people, and then that obviously for l- all sorts of good reasons, we go through these ethical epiphanies and we realize that’s really wrong.

And, um, there, there’s, there’s loads of stuff there and that, you know… Then also the sad thing about scraping now being really not okay is that, you know, all the small research groups, they’re the ones who need to be able to scrape because they can’t afford the data. So then- Right … so, like, it’s really, really tough.

And meanwhile, yeah, so holding the, the tech leaders accountable, I think more needs to be done. I’m, like, astonished at how little regulation seems to be in place. I’m astonished at how little, like, um, reaction seems to happen when these kind of ethical transgressions or legal transgressions are made over and over and over again.

I don’t, I really don’t have a good answer on that. No, no, that’s fine. , But I’m really kinda curious too, like, when we talk about ethics and, and, you know, you say you’ve got, like, a great team Do you find that it’s easy to find consensus around ethics with, with that group? ,

I think it’s difficult, [00:18:00] but, um, I think there is consensus, and I think the, I think that the consensus is found sort of relatively easily. There’s something funny about AI in general, is that we’ve gone kind of in the space of a couple of years, we’ve done sort of what happened to the, the factory from the Industrial Revolution to today.

And if you think about the factory floor, there is still, like, bad workplace stuff happen- ha- happening. Right. There’s still all these issues in factories, and that’s hundreds and hundreds, like… Sorry, not hundred. Hundreds of years later. Um- With AI, we’re trying to do all of that overnight. And so on the one side, I think there’s loads and loads of low-hanging fruit.

So around things like, , racism and sexism in models, I think there’s consensus there because it’s, it’s really obvious. Like, some of the stuff is just so big and so important, it needs to be done. I’m sure there’s loads and loads more stuff where it’s still- Yeah … there are gray areas. Mm-hmm. And definitely, you know, when it comes to the legal side of stuff, there are loads of gray areas as well.

Oh,

Speaker 5: yeah, yeah, yeah. It’s very, very dicey in, in, in jurisdictional and yeah, like [00:19:00] different, different senses of y- of what’s, what’s what, for sure. Um, what’s one, uh, bold prediction you have for the future tech that you’d be willing to bet on today?

Speaker 6: Quantum, I would say. Is that a really lame answer? That might be a very boring answer.

No, it’s not

Speaker 5: a lame answer. It wasn’t agentic. I, I think, yeah.

Speaker 6: I think quantum, I mean, like quantum, it’s refreshing because the science checks out.

Speaker 5: Yeah.

Speaker 6: There’s so much talent in the industry. There’s, um, a lot of investment. These kind of quantum companies are flush with cash. Uh, there’s enterprise interest from kind of pharmaceutical banking, et cetera, et cetera, and there’s enterprise commitment.

Um, it just, it just feels exciting, I would say.

Speaker 5: It’s awesome. And i- if the current trajectory of AI development holds, uh, what does the relationship between humans and AI systems look like in 10 years?

Speaker 6: So I think the, the answer is somehow, um, it’s somehow colored by the idea that, [00:20:00] uh, uh, and, and you know, this is re- this is obvious to all of us, and maybe it’s like the root of lots of discussion.

But AI is doing lots in terms of product, so it’s taking over a process, many, many processes, and coming out with a product, and spitting out a product, and that product is sometimes great, sometimes not, et cetera. But if, wh- you know, when that happens, and maybe to use the Industrial Revolution as, as an example, if, if a chair is factory-made, um, and all the chairs we start using are factory-made, it doesn’t mean the chairs are bad.

But we also begin in some way to, to really put a lot of emphasis on process. So then we talk about things like handmade and- Mm-hmm … and we care about process the more that stuff gets automated. And I wonder if there’s something that’s gonna happen with AI in the same place that we’ll begin to think about, uh, like human thought or like human-made as a little bit kind of cheesy and clumsy.

But there is this, there is this idea that we’ll begin to focus loads on, um, on the kind of process that goes into a product again. And, um, maybe to bring it back to literature, I’ve been… I t- I just did a review on a [00:21:00] Brandon Taylor book, um, uh, his new novel called American Black Figure. It’s all about an artist in, in New York, uh, and also on a book by the kind of avant-garde writer Can Xue, a Chinese, Chinese author.

And both of them, I was kind of intrigued that in both of them, you know, a key thrust of their, of the book is about how important the physical process of stuff is, physical making art, physical reading, you know, that, that type of thing. And I can really see that our, our relationship with AI will, will… I don’t know if that affects our relationship directly with AI, that, but that that will be an offshoot, that we’ll, we’ll begin to think much more about what is like physically done versus what’s done by AI.

Yeah. , That is not a clear thought in my head

Speaker 5: at all. Like craftsmanship- Yeah … or ar- artisanal, right, like, uh, element. No, I, I, I battle with this when I know that AI could work a spreadsheet a certain way, and I just- Yeah … know that, oh, yeah, but also, uh, I don’t know if I trust it- Yeah, yeah

fully on some of these things. Like, uh, surprisingly, uh, charts and things like that still are, got room to improve. Yeah. Um, are you working on another [00:22:00] book?

Speaker 6: Uh, I am, yeah.

Speaker 5: Oh, yeah? Yeah. Is it… Can you tell us anything about

Speaker 6: it? Yeah. It’s not about AI, but it’s about, uh, it’s a kind of retelling of a, of an opera called Rigoletto, and , it’s all set in expat Japan.

Oh, wow. And it’s a bit of a like, yeah-

Speaker 5: Wow . It’s awesome …

Speaker 6: kinda gritty, grimy expat Japan basically.

Speaker 5: Very cool.

Speaker 6: Yeah.

Speaker 5: Well, we’ll have to have you come on and, uh, and talk about it

Speaker 6: when it comes out. I would, I would love that.

Speaker 5: Yeah, yeah. Thank you so much, uh, for, for making the time today, Fred. It has been really interesting conversation, and, uh, yeah, I really appreciate you dropping by.

Speaker 6: Thank you so much for having me.

Speaker 5: Awesome. Cheers. Thanks.

Luke: Thanks for listening to the Brave Technologist Podcast. To never miss an episode, make sure you hit follow in your podcast app. If you haven’t already made the switch to the Brave browser, you can download it for free today at brave.com and start using Brave Search, which enables you to search the web privately.

Brave also shields you from the ads, trackers, and other creepy stuff following you across the web.