Transcript
Speaker: Welcome to HSBC Global Viewpoint, the podcast series that brings together business leaders and industry experts to explore the latest global insights, trends, and opportunities.
Speaker: Make sure you're subscribed to stay up to date with new episodes. Thanks for listening, and now onto to today's show.
Speaker: Hello everyone. Welcome to the Perspective podcast. I'm Stuart Riley. I'm HSBC's Chief Information Officer. Today I'm delighted to be joined by Dmitry Shevilenko, who is the business officer the Chief Business Officer for Perplexity.
Speaker: Welcome. Dmitry leads all business growth for the company, including commercial and enterprise partnerships. and Today we're going to have a discussion on AI. There's no nothing like an AI discussion at the moment. um We're going to focus on sort of two things, the speed of AI, but also the commerciality of it. So how quickly we can realize commercial benefits in in enterprises today.
Speaker: Welcome. So excited for the conversation. Thanks for having me. Good. um look Let's start with a little bit. Not everyone will know Perplexity. I'm sure many of the audience will, but not everyone. So maybe tell us a little bit about the company, how it started and what you did.
Speaker: Perplexity was founded a lifetime ago in AI years in December of 2022. The original insight behind perplexity was that you would eventually have many different foundation models, many different LLMs.
Speaker: And the thing that you really needed, especially in 2022, is for them to not hallucinate. That was you the first big, hard problem in AI. And perplexity's insight was the thing that would stop that hallucination is if you ground the LLMs in real time knowledge and search.
Speaker: And so we built ah the first and best search infrastructure for AI agents. and that's what Perplexity was initially known for. In AI, your product evolves every you know two to three months. um We're now really focused on Perplexity Computer, which is an end-to-end agentic workflow product um that is effectively a digital worker. And so we've been really focused on bringing that to market the last few months, and it's going really well.
Speaker: I love this point about you know what an AI company does because like you've just described with yourself that I'm familiar with, you meet a company, they talk about their vision and what they're doing and three months later you've got a check that we're still on the same path or has the product evolved into something else. And you know I've been on that journey with Perplexity over the last few years and really seen seen how that's that's evolved.
Speaker: um I think some of the early work you did around grounding the model was really impressive. I mean, I know many people who really relied on your output for search and use particularly because they knew it was grounded in fact and that hallucination wasn't such such a problem. But hearing where you've moved on to into this perplexity compute obviously is like a fascinating model in terms of how we in enterprise think about getting commercial commercial value.
Speaker: um If I could start, you've always been multi-model. and I believe you still believe in that. Do you feel that they are becoming commoditized or, you know, how how do you see that multi-model landscape evolving?
Speaker: So I think the key concept is the models are becoming specialized. So, you know, the Claude Opus models are really good at planning. The GPT models are really good at writing. Gemini models are, you know, if you need to generate audio, they're really good at that.
Speaker: Now, everything I just told you will be different two weeks from now where I'll actually tell you, actually, it's the GPT model just gained an edge in VideoGen. You actually have a new open source model out of China that is the best at coding. Every benchmark that exists for LLMs, the leaderboards are changing you know on a weekly basis. You have ah the specialization as you know the diversity of tasks that models can handle is becoming more important.
Speaker: And so for us, that's actually even more purpose around being the orchestrator. Extend the analogy, you know you've got many different instruments. Those are the models. um And you need a conductor ah to make it make music together. right And you need to hit certain notes at certain times. And the the end user, ah they just want their work solved for them. right And they want to be able to focus on higher leverage problems. They're not in a position yeah to tell the drummer when to hit the beat.
Speaker: That's the role that we're playing as as the conductor of the orchestra of work. I'm glad you said that because actually that reinforces something we've done at HSBC. So we built effectively an an AI gateway that we've sat all of the models behind. So we are model agnostic, we work with pretty much every provider and and what we enable our internal teams to do is say you don't need to worry about the model. you know We've got this model garden and we will help you dynamically eval which model is most suited to your request or your agent that's running. And as you so as you said and we see, we're having to sort of very rapidly, dynamically reassess which is the latest model for the right type of work. And then sort of adjacent point to that is there's a lot of work that we're doing today. Like if I look at my own desktop setup, I have about a dozen agents on it.
Speaker: I can run in multi-agent mode, you know, they work autonomously. That's fantastic. But I imagine if I think forward, the models are going to change so dramatically that my agents are going to become in their own way pretty useless pretty quickly or at least unnecessary. So how do we make sure that we're investing today in the things that will prove to be worthwhile in the longer term and not worry about the things that are going to get deprecated very quickly?
Speaker: I think most of the time it's not worth it for large organizations to try to build their own internal AI applications um because they're they're not going to be able to be at the at the bleeding edge. it's not you know It's not their core competence. ah What actually is really and where I would invest in engineering is um data hygiene.
Speaker: A big part of yeah what we're we're living through now with the Snowflake and Databricks integrations is everyone now you know is a data scientist, right? And can operate in real time with all the best contexts when they're making critical business decisions.
Speaker: i the biggest constraint to the usefulness there is that the way the organization has historically cataloged ah data is a mess. And you no AI can reason through it because no human can reason through it. It's just kind of yeah a bunch of, you know for the same category of event, like say like a user sign up, you'll have 20 different definitions across 20 different products and organizations. And there's no one way to rationalize through that.
Speaker: that's kind of a evergreen investment of if you have excellent data hygiene, um as the models get more capable, you're only going to get more leverage out of that.
Speaker: Versus over optimizing for a specific use case today, ah the you know the model improvement might just make whatever agentic enhancement that you guys supported less relevant. Yeah. actually it's very aligned to what we are doing. In fact, our our data strategy has become almost a more pivotal part of our overall technology strategy than the AI tooling itself, because yeah we realized that out the plot the part we play is bringing that unique set of data that only we have to the things that providers like yourself can bring in terms of the the intelligence, the models, the the mechanism to route that, that sort of thing. And that goes back to the security point, right? If you, you in that data hygiene process,
Speaker: you're giving yourself confidence like, yeah, we're, you know, our user level permissions are tight so that when this does get exposed to a external AI product, you know, and we have the right commercial terms in place and data protections that we feel comfortable about how this works end to end, and that's where you get real leverage. yeah I think security is an interesting point because certainly in my role, we are worried about the security of a bank. of Fundamentally, yeah our job as a bank is to make sure our customers trust us and that's by securing their money, their data, their wealth, their assets. um So cybersecurity is like critically important. And obviously, the controls that we have in place are probably good for a human and good for AI.
Speaker: um The ai when you're running in this sort of agentic type space, can just do many things much faster. So if there is any control gap or any vulnerability, you know you're much more susceptible to it.
Speaker: How much do you guys think about that? you know what What's your approach to that? i mean I think some like core principles there is, We look at it from how do we empower your your employees to you make better, smarter decisions that are more customer-centric. 99% of that is actually just about read access to data and not write access, right? So I think the agentic part of you know the agent will then go and make changes in your system. If you're a bank, like
Speaker: do that last, right? Because that's that's really where your nightmare scenarios, you know, so, but in terms of empowering your your teams to have access to, you real-time information, so again, they're they're making smart decisions, they're not losing a week waiting for some report that they can just generate dynamically on the fly. ah that's That's the part where I think you need to be more aggressive and leaned in. Let's take another scenario, like customer support, right, where you do want you know certain types of right access. You want to be able to take certain actions. um
Speaker: i think there it's almost you know you should approach it more in a classical machine learning model where you you just you actually want to remove the probabilistic nature of AI and really just have the models be very smart, rule-based deterministic systems. And I think that's another way of gaining confidence. Yeah, yeah, makes sense.
Speaker: Do you have any views one, two, three years out where you might be, what might change, yeah how your strategy evolves? I mean, we want to be serving um people that are excited to use AI its full capacity, right? And i that that means a dual purpose, right? So it's both the 5% of like the power users, but then bringing the 95% along, right? And so our strengths, yeah I'd say like the three things that have been constant during you know three very wild years in AI that feel like 30
Speaker: is accuracy will always be foundational, right? For people to trust AI, they need to believe it's accurate. And so all the investments we've made in search and browser capabilities, that's what, you know, that's what, without that and the connectors, everything else kind of becomes irrelevant to the conversation.
Speaker: ah The second piece is you always want access to the best intelligence, right? And that's where the multi-model orchestration comes into effect. Yeah. And then the third is any enterprise AI transformation strategy requires your enterprise to actually use AI. And and the the thing that is most subjective is what is great user experience and user interface and how do you use that to educate people and drive adoption.
Speaker: And yeah that's always been a strength of ours and we just need to keep being at the edge there. So those are the things that won't change. thing that that you know I like to say is I'm most confident in the fact that six months from now, I'm going to have a top three priority that today I don't know what it is.
Speaker: And the reason I frame it that way is we all need to remain ah very mentally flexible that the emergent capabilities that come from new models, even the model makers themselves don't know what they are. Using a cooking analogy, like every time they bake a model,
Speaker: they have no idea whether it's going to taste like chocolate or vanilla or some new flavor that you don't even know that it's a flavor until you've tasted it, right? And that is, you know, I think for us to be successful, we need to hold ourselves to that standard of being the best and first at identifying what those novel, useful applications are and bringing those to market, you know, grounded in accuracy, using all the best intelligence, and then, you know, really being empathetic to the user. Yeah, yeah, makes sense.
Speaker: I think many of our viewers will hear this. They'll be thinking to themselves, the ai or the the current phase of AI has probably been going on for a few years. um And when they look at the economic impact they've had in their company, maybe they're looking at it thinking it's been quite limited.
Speaker: And yet you and I sit here thinking, this is going to be absolutely revolutionary and game-changing to pretty much everything we do. You know, when I look at that evolution, the way I think about that is,
Speaker: we're going from the development of the underlying models and tools to the embedding of those models and tools into actual commercial application.
Speaker: and And of course the advancement of the tools got better. So when I think about something like like coding, you know i I was pretty blown away a year ago with the tools that I had to assist me to code. But they look like absolute dinosaurs compared to the code.
Speaker: coding tool I have today in that last year. Do you think that is going to be the next one year in a broader sense? Like ah is is the next year the time where everyone starts to really see the the the productivity of their domain?
Speaker: I mean, there's still people today that would argue oh, the internet was kind of overhyped, right? So I think there is a bit of a ah boiling frog kind of syndrome ah that we're living through with AI where like, you know there's a lot of amnesia because, you know, a lot of people don't even have the,
Speaker: memory of like, oh yeah, this is actually a lot more powerful than it was, you know, 18 months ago. Societies and cultures evolve slower than technological capabilities, right? And so if you work for a larger organization, ah the reason AI hasn't been, know, as transformational as you think it, as the hype merits is it's less due to technology, it's more the, know, again, the the process constraints around adopting it and you need everyone to adopt it for it to have the maximal impact.
Speaker: Where I see the bleeding edge already transforming is with startups, right? So the amount of people you need, the amount of capital you need to start a new business ah has been radically shrunk.
Speaker: And yeah we actually launched something last week called the Billion Dollar Build, which is a contest to see if you can build a startup just largely using perplexity computer, And we have ah a sister corporate venture fund that will, know, you can win up to a million dollar investment if you're kind of on that trajectory of of building a one, two person billion dollar business.
Speaker: It would have been an absurd concept, you know, five years ago. ah But we see people realizing this now. Right. And so we we think the future is going to be smaller, nimbler teams.
Speaker: ah you know making up smaller, nimbler organizations. And that's where the transformation will will really kick into effect. So that leads me brilliantly to my final question.
Speaker: Net job growth because we become more productive, more innovative, new products, new services, or net jobs decrease because we fundamentally do the things we do today, but they get automated via AI.
Speaker: Massive boom in entrepreneurship. So, you know, many, many more companies, but they'll all be much smaller. ah And very large organizations that ah don't have the C-level bought in on using ai will will you know have no choice to stay competitive. i mean, everyone's going to end at the same place. It's just a question of, you know do they up-level their workforce or do they have to downsize you the up-level capabilities or downsize the the size?
Speaker: Brilliant. Dimitri, thank you for joining me today. That was fantastic. Thank you. I really enjoyed that. Cheers. Thank you for joining us at HSBC Global Viewpoint. We hope you enjoyed the discussion.
Speaker: Make sure you're subscribed to stay up to date with new episodes.
