Realities Remixed

The internet was built for humans. Search engines, websites and online business models were all designed around people reading pages, clicking links and consuming information. But what happens when AI agents become the primary users of the web and governance of the internet will need to evolve from human-speed interactions to machine-speed ecosystems

This week, Dave and Rob are joined by Antonio Mallia, Founder & CEO at Seltz to explore the rise of generative search, AI agents and the rapid transformation of the internet itself and how traditional search engines are being replaced by AI-powered systems.
 
TLDR
00:31 – Introduction
01:53 – Hang out: Rob is confused about how AI-generated complaints are overwhelming public services
08:55 – Conversation with Antonio Mallia, Founder and CEO of Seltz
52:57 – Advice for websites and biking with the kids
 
Guest
Antonio Mallia: https://www.linkedin.com/in/antoniomallia/
 
Hosts
Dave Chapman:  https://www.linkedin.com/in/chapmandr/
Rob Kernahan:  https://www.linkedin.com/in/rob-kernahan/
 
Production
Marcel van der Burg:  https://www.linkedin.com/in/marcel-vd-burg/
Dave Chapman:  https://www.linkedin.com/in/chapmandr/
 
Sound
Ben Corbett:  https://www.linkedin.com/in/ben-corbett-3b6a11135/
Louis Corbett:   https://www.linkedin.com/in/louis-corbett-087250264/
 
'Realities Remixed' is an original podcast from Capgemini

Creators and Guests

Host
Dave Chapman
Chief Cloud Evangelist with nearly 30 years of global experience in strategic development, transformation, program delivery, and operations, I bring a wealth of expertise to the world of cloud innovation. In addition to my professional expertise, I’m the creator and main host of the Cloud Realities podcast, where we explore the transformative power of cloud technology.
Host
Rob Kernahan
VP Chief Architect for Cloud and Cloud Realities podcast host, drives digital transformation by combining deep technical expertise with exceptional client engagement. Passionate about high-performance cultures, he leverages cloud and modern operating models to create low-friction, high-velocity environments that fuel business growth and empower people to thrive.
Producer
Marcel van der Burg
VP Global Marketing and producer of the Cloud Realities podcast, is a strategic marketing leader with 33+ years of experience. He drives global cloud marketing strategies, leveraging creativity, multi-channel expertise, and problem-solving to deliver impactful business growth in complex environments.

What is Realities Remixed?

From the award‑winning team behind Cloud Realities, Realities Remixed explores what happens when people, culture, industry, and technology collide, because real breakthroughs rarely come from technology alone.
In each episode, hosts Dave Chapman and Rob Kernahan dig beneath the surface of innovation to understand what’s truly shifting. They speak with leaders, practitioners, vendors and thinkers across industries to explore how technology reshapes human behaviour, organisational culture, and society, for better and sometimes for worse.

Web - https://www.capgemini.com/insights/research-library/cloud-realities-podcast/
Email - realitiesremixed@capgemini.com

Speaker 1:

What are you are you excited about going to a restaurant at the weekend or excited about

Speaker 2:

Oh, in that Okay. Yeah. Yeah.

Speaker 1:

It can be in any sense you would like.

Speaker 3:

The destruction of the human race because AI is rising and nobody's controlling it.

Speaker 1:

I don't know.

Speaker 3:

It could

Speaker 2:

be anything. Yeah. Yeah. Yeah.

Speaker 1:

I'm Dave Chapman.

Speaker 3:

I'm Rob Kernahan.

Speaker 1:

And this is Realities Remixed, an original podcast from Capgemini. And this week, a conversation show exploring the evolution of the Internet itself, generative search and how that is driving new models. Now joining us later in the show is Antonio Mallia. He's the Chief Executive and Founder at Seltz. Seltz are a web knowledge and retrieval infrastructure API built specifically for agents and large language models.

Speaker 1:

Now what that means is it is the sort of technology that's that's sitting underneath generative search. So when you're using Gemini Copilot, know, the the agent of your choice. It is using technology like Seltz to aggregate that information for you. So with Antonio, we're gonna dig into exactly how that happens. We might also get some insight into what happened with thenumbers.com.

Speaker 1:

So if you haven't heard our episode with Stephen Fellows on what happened at the numbers, maybe drop into that first and then this conversation with Antonio might give you a mirror image of of what's going on in the background there. And perhaps, you know, what you might do as a webmaster to avoid crash situations like the numbers found themselves in a year or so ago. Before we get to all of that though, what's confusing you this week, Rob?

Speaker 3:

Well, David, it's a complicated one.

Speaker 2:

It is

Speaker 1:

late on

Speaker 3:

a fry

Speaker 1:

I'm just gonna say it's late on a Friday. I had a bit of a late night last night and I'm about to do a long drive. So when you say

Speaker 3:

I'm gonna make you think for a bit though. I'm gonna I'm gonna pull that last bit of grey matter power out of your brain before you shut down for the weekend if that's alright. Do I have permission?

Speaker 1:

But Marcel might have to step in and and patch me.

Speaker 3:

Yeah. Here's a simple, it's it's it's complicated but it's simple.

Speaker 1:

I like that.

Speaker 3:

AI. Yeah yeah, a classic consulting AI is incredibly powerful. Free AI is now basically in the hands of every human on the planet if they have access to the internet, mobile phone, computer, etcetera. And there was a really interesting case in The UK where the waste collection department have become overwhelmed with really sophisticated complaints citing legal precedent associated with bin collection.

Speaker 1:

Brilliant.

Speaker 3:

So the interview with this department was the average complaint for your bin's not being collected was two paragraphs long and a bit of a rant and I'm gonna contact my MP. And now they're receiving 15 page letters citing legal statute about their bin and how they're gonna do all this stuff because AI In

Speaker 1:

hygiene and things like that.

Speaker 3:

In hygiene. And AI is now writing these massive documents, but of course, it's a legal requirement that they have to respond to complaints. So they have to go through this massive tome of information. And so there's this bit about the power is now with the citizen to be able to create this information, the system now has to be able to cope with the response because you can't ignore the response. You must respond, which means you have to go through the whole thing.

Speaker 3:

And the bit I'm confused about is where does this end? Because if the citizen keeps creating ever more information for companies to deal with that a human has to get involved in, I mean, that's just a huge cost. So Which means you're gonna drive up prices and all sorts of stuff. So I was like, I'm confused about where this is gonna go.

Speaker 1:

Good one. So it depends on where the human is in the loop, doesn't it? So if

Speaker 3:

I suppose If

Speaker 1:

the human in the loop The

Speaker 2:

part of it.

Speaker 1:

Is it has to be doing, the human has to be doing the processing and responding to all this stuff, then then you've got a problem on your hands. Isn't it more likely that agents will will do that, so you'll get agent to agent interaction around it?

Speaker 3:

Yeah. Yeah. So that's where I thought it would go. So you get agent to agent, you get this generation of token burn and you get response. But then that agent has to take onward action.

Speaker 3:

So if the complaint is valid, then they have to process it through the organization, dispatch a human, whatever, a truck. Right? So that becomes really complicated to implement.

Speaker 1:

Well, that's why we connect connect into an in like a like an intelligent operations backbone where the agents are basically planning all of that.

Speaker 3:

Yeah yeah but if you're a waste collection department that's a lot of cost just to process complaints when it might be a bit simpler and cheaper

Speaker 2:

to deploy.

Speaker 1:

No isn't you think about it the other way around which is you enable your business with an intelligent operations backbone and one of the agents that operate on that is the agent that responds to complaints.

Speaker 3:

But for a municipal county, that's a big thing to ask them to do when the world's largest organisations are struggling to do it. How do we expect the local system to do it? It's a lot of cost. And the other thing is, the onus is on them to make sure they respond to it correctly. So chat GPT writes me my 15 page letter, I don't care.

Speaker 3:

But if there's a right that you have to respond, do you rely on the agents to make sure they get everything right? So now you've put all this pressure downstream and potentially increasing costs for something as simple as say waste collection. And I know it's going on in HR, with firms and all of this sort of stuff and there's this there's this bow wave that's coming and yes, there are technology answers that could answer it but at what cost, to what end, to what price increase?

Speaker 1:

Well feels like, isn't it what you're describing a function of where we are in the adoption cycle? Yeah. Maturity. You've got, yeah, you've got one aspect of a system that is now empowered in the letter writing way that you describe and the rest of the system is yet to kind of modernize itself to be able to respond to that. And you're right in saying that at the moment, you know, there's only emergent knowledge about how these things are going to work, particularly when it is a, you know, kind of intra intercompany.

Speaker 1:

Intra company is difficult enough. Like when it's inter company and company to citizen interactions, and it's a citizen that's now not just using an agent to write a letter, but the agent might actually do the communication on their behalf directly to an agent that's sitting on the other side is the is the likely end game here, isn't it? And all that happens at machine speed. At the moment though, it's like how how long is it gonna take before that sort of tech becomes affordable, as you say, at like county municipal level?

Speaker 3:

Yeah. And and I think the the issue is it it may well drive costs. So when your tax demand for municipal taxes come in and it's gone up by 20%, it's because bob down number seven keeps using chat GPT to complain and it's like Yeah. Maybe we should put the way I thought about it was maybe we should put no AI in certain processes to prevent this type of thing from happening and detect it and refuse it and say you've got to, you know

Speaker 1:

Smash the looms. Smash the looms.

Speaker 3:

Stick the hammers in. But I mean, realistically, it it will drive cost in parts of the system where we probably don't need the cost. We need to think about it a little bit better than, you know, just thinking that we've empowered the citizen to do a load of good stuff. And I'm all for the power to the people and you know that, but there's sometimes where the people need to understand that if they take that power it might drive unintended consequences downstream.

Speaker 1:

Oh what comes with the great power Robert?

Speaker 3:

Great responsibility. Spider Man.

Speaker 1:

Exactly. Spy all lessons in life. You you go go to Spider Man, the great text, and you'll get some guidance.

Speaker 3:

I know. I mean Marvel had the answer for us. I mean we just need to pay more attention.

Speaker 1:

There we go. I think we fixed that one. Did we did we do that?

Speaker 3:

We didn't we haven't fixed it at all We spun it off into a weird place. Do you get my joke?

Speaker 1:

Anyway look well done. So on to the today's main subject and actually you choose the confusion well this week because it bridges perfectly into how on earth you start thinking about agent economy, what's going on in generative search, how agents are functioning and how the internet might modernize. So why don't we jump to our conversation with Antonio Mallia, and he is the Founder and Chief Executive over at Seltz. So whereabouts in the world are you today, Antonio?

Speaker 2:

Today, I'm in Pisa, which is a small city in Tuscany in Italy. Beautiful. I love Tuscany. And I moved here when I was 18 to actually study here besides famous for starting kind of the Internet revolution in Italy and I would say also in Europe because it was one of the first one connected to the Internet here in Italy. And so there is actually it's a vibrant city when it comes to information retrieval and search and in general, like, Internet.

Speaker 1:

I did not know that about Pisa. I did not know that.

Speaker 3:

And not only does it have a leaning tower and it's in one of the most beautiful parts of the world with excellent wine surrounding it, but it also was the first interconnected. There you go.

Speaker 2:

Look at that.

Speaker 3:

I've learned I didn't know that.

Speaker 1:

Now Robert's favorite wine is from that region, Robert, isn't it?

Speaker 3:

Yeah. Yeah. Brunello out of Oh. Tuscany and Umbria. I am a great fan of that.

Speaker 3:

Definitely peak for me. Yep. That's the one I go to.

Speaker 1:

Do you endorse that choice, Antonio?

Speaker 2:

Oh, yeah. Absolutely. I mean, Brunello is definitely one of the most popular and most protected also type of wine because obviously, you know, there's always the people always try to kinda copy the the flavor or the process and some. Like, I definitely definitely endorse Brunello.

Speaker 3:

There you go, Dave. You see? You see? External validation. Happen has to happen to me once or twice in my life.

Speaker 1:

Robert generally keeps his light under a bushel and he's quite a humble guy. But when it comes to wine, he's he's the hilarious wine snob. And it's thoroughly it's thoroughly entertaining.

Speaker 3:

There is that has been the odd occasion I've turned my nose up at some offerings, but yeah.

Speaker 2:

Well, I would say that wine is not the only good thing that you can appreciate here in in in Tuscany. It's Now? It it's definitely like Italy in general. It's kinda like less under the radar when it comes to when it to technology. But I would say that things are changing, fortunately.

Speaker 2:

And yeah, that's definitely one of the reasons why I keep promoting Italy also when it comes technology and and research around some of these topics that I am passionate about.

Speaker 1:

Let's spend a moment on that. And, you know, how do you characterize the sort of growth of the tech industry in Italy over, say, the last ten years or so?

Speaker 2:

Yeah. I I think it's kinda like the smaller example within Europe, and Europe itself is a smaller example within the world, right, where obviously there is The US kinda like leading it, although we have seen kind of like also other countries kind of like coming in and having a role, obviously, like China is definitely one of them. But at the same time, there is kind of like Europe has been a little bit more silent. And within Europe, there are also other cases like, again, as Italy, where they've been even more silent. But I would say that, you know, things are changing.

Speaker 2:

And one of the reasons, in my opinion, why things will be changing is because of like the universities, hubs that we have. And, I think there are some there is excellence essentially in in Europe as well as in other places. I'm trying to say it's like there's not such a big difference even if it's less if even if Europe has been less vocal on AI these days. I think, you know, the type of research that happens, the type of, like, talents that universities produce is definitely, like, aligned with what they've seen in in The US.

Speaker 3:

I think one of the things that be interesting in your view on is the one of the criticisms of Europe is the way investment flows in to allow the great ideas that are created here through that sort of university network to thrive. It's often America or China comes in and nicks the the skill. Do you see that changing at all or do you think it's an area we still need to improve to, you know, get better?

Speaker 2:

Yeah. I mean, it's definitely changing. We are seeing Europe scaling up when it comes also to investments. And I think if I look back five, ten years, I couldn't name like an Italian VC, but like these days, on the other hand, it seems that even kind of like venture capital is becoming a thing in Italy and obviously more and more in Europe. In particular, you know, France and The UK are pretty vibrant when it comes to companies from The US opening offices, opening hubs in Europe, like Paris.

Speaker 2:

I've seen so many companies in the tech space, in the AI space in Paris now, London as well, clearly. Yeah, I think we are hearing about like Anthropic and other foundation labs opening their offices. I recently heard of Milan, actually.

Speaker 1:

Oh. Right. And PISI, you mentioned you mentioned PISI at the beginning, is where you are. But you mentioned it as like a a tech scene in Pisa?

Speaker 2:

Yeah. And I would say more than a tech scene kinda like the way you would expect in SF. The type of like tech senior, it's more about it's more related to research and academia. Right. That's also how I started in a way, my kind of like my path, my journey into what I do today.

Speaker 2:

I was a student and I wasn't really interested in anything in particular when it comes to information, when it comes to computer science. It's like, you know, everything that relates to computer is interesting to me, was interesting to me. But then I kinda like got engaged I engaged with, you know, this specific field which is called the information retrieval and it's kinda like the discipline of finding information through query intents, through specific requests. And I kind of got super passionate about this specific topic. I think, you know, since then I've always been doing this and I've always been very, very kind of like intrigued by how many different concepts, how many different type of like exercises you have to do to optimise, to make it work, like a search engine, which is in the end kind of like the de facto implementation of an information retrieval mechanism and system.

Speaker 1:

Well, so let's use that as a as a perfect bridge into Seltz itself. So just tell us a little bit about what Seltz is for those who don't know. And what was the moment of inspiration that was powerful enough to get you on founding the company?

Speaker 2:

Yeah. So Seltz is the web search engine for the future user of the Internet. There is there is always this idea that, you know, we are gonna be searching information online by, you know, putting queries, putting putting questions and getting answers from search engines. But we think that in the future that's not going to be the case anymore. So if you think about like you have to do like a very difficult task or something that you don't do on a daily basis, so you don't know how to do it, you know, we used to pretty much go on Google and search for, okay, what do I need to do to submit my to do my taxes, for example.

Speaker 2:

It's always like one of these tasks that happens once a year. You never remember what you did last year, it usually gets complicated. So go and search online. But the way that we are doing it today, for example, is slightly different, where we go on one of these general purpose LLMs, let's say, HRGPT, Pod, etcetera, and we try to ask this new technology. We say, hey, how do we do this this operation?

Speaker 2:

How do we do this task? I think and that's and that these models have some generic knowledge because they have been trained on on tutorials, they've been trained on some of these on documentation around how to do these operations. But like they are not they're not up to date because they've been trained, they have a cut off, and, you know, all the training goes back to a year or two, and and at the same time, it's not given that they have memorized all the important data. So what they do, they they learn how to search the web. And they search the web through the same tools, the same search engines that humans use today.

Speaker 2:

Right. And that's that's very inefficient from machine to actually use a UI that wasn't built for for the machine itself. It's kinda like, you know, you have to write a letter or actually, Charles GPT needs to write something but then wants to use the keyboard. It's like, why do you want to

Speaker 3:

use the keyboard?

Speaker 2:

Just go digital And to so the same thing is happening now with with web search. These agents have the ability to go and search directly through an API, through a more direct protocol, and don't need the interface that used to be popular. And so what we are building is we're building like a web search engine that can connect directly with an agent such that the next time you're gonna ask the agent, it's gonna be up to date regarding that type of information that you were looking.

Speaker 3:

And it's a I remember it as being a massive watershed moment in the AI model interaction where you'd ask it a question, it couldn't tell you today's date and things like this. It would say, I've been trained to have and then suddenly it got connected to all the the wider ecosystem and the utility of it leapt massively very, very quickly. So it a major change in the way people started to interact with it and then I think that was one of the moments where people started to adopt it day to day because it became relevant and up to date and useful as opposed to something that could sort of analyse structure. So it's a key, I suppose I'd say it's like an event that occurred that really I think caused a hockey stick, especially for the consumer to be able to, you know, embed it into day to day.

Speaker 2:

Yeah. And I would say this still happen today because, you know, most of the time these machines are still connected to tools that do not expose all the information that they need. So if you think about it, web search engines, they expose snippets because they are basically trying to trying to tell you like, hey, this information might be contained in this website, go and open it. And so that is not sufficient for an RRM to actually, you know, generate an answer. And so although it's not saying anymore, sorry, I can't give you that answer because I've been trained a year ago, what's happening today is that it looks at one of these snippets that says, oh, the age of this person is dot dot dot.

Speaker 2:

It's gonna hallucinate and it's gonna actually, place, data that's not written in the snippet because it was supposed to open the website to find out which date in that case, the person was born. And so even the way these models consume the data is different from humans. Like, we are patient and we go from top to bottom and decide what links to open and whatnot. An LLM can actually go and read these 10 documents in a blink of an eye. And so, again, the interface is different.

Speaker 2:

And I think I want to also mention that at some point, you know, we're not going to be asking chatbots how to do things. They are gonna be doing things. And so making them more accurate is really key to exploit AGI.

Speaker 1:

Right. And just for the benefit of listeners who haven't perhaps interacted with it with this stuff in the way that you're describing like, we all know what traditional search looks like. So the human experience of traditional search whether it be Google or whatever it might be. What's the actual interface and experience in what you're talking about? Is it simply that you're using a particular AI at the front, whether it's Gemini or Copilot and and you just that's the search you're talking about, which is like you're just kind of iterating through prompting to get to to get to better answers or is it something different?

Speaker 2:

No. That's that's it. You iterate through prompting to get better answer, and the model under the hood knows how to actually go and call an external tool to either search the web or fetch the web, to fetch a specific page, so get the full content out of the page.

Speaker 1:

Over the summer period, unfortunately, went through a period in my family of varying different health scares and health situations. I was using generative search extensively actually to ask questions of that and and there was some sort of hyper specific stuff that I was trying to find out because my dad, you know, had a, has like a lung issue And he worked in the Northeast Of England at a certain time in history which was it was a very industrial period. And and and what I kind of loved about it is you could get incredibly discreet about the parameters that you were searching for and get back some very very specific ranked kind of causes and potential issues which then had treatment paths attached to them. So it might sound obvious in the in this day and age that you can use AI like that, but I was even I think about it every day, you know, when it when I was actually applying it to quite a sensitive and very specific situation, I thought it was absolutely amazing.

Speaker 2:

Yeah. No. It is. And I I I think that's very interesting. I would be curious to know what was kinda like the iteration process, like, long it took to get to the final information.

Speaker 2:

Because, obviously one of the important thing is actually to make it faster such that you don't have to wait kind of like the agentic period, which sometimes it's even like thirty minutes to get an answer because it has to go over and over and over certain search operations and it's gonna go deep.

Speaker 1:

It was so I was using Gemini and it was returning it was returning ranked information in, oh, like less than thirty seconds. Yeah. It was it was it was mildly slower than just a normal traditional web search. But but but way more contextful and rich in information.

Speaker 3:

It's the it's the when you Google something, there's an art form in how you use traditional Google search with site and this word and that word and the LLMs take all that pain away from you. But there's a thing when you do a Google search, who goes past page two? I mean, your hand if you've ever been to like page seven of a Google search. You know, when you get down there, you're in the dark depths of things that, you know, the the the the and it's just like that whole LLM takes all that pain away from you, so it's so much more convenient and you get Well

Speaker 1:

it also aggregated. It it it aggregated and contextualized the information in in such a way that it saved me, know. So if I I was trying to correlate the specifics of a medical condition with the age of a person, with their geographic location and the industrial history of that geographic location during the working life of that person. And it was doing all of that aggregation for me and then presenting that back in an extremely integrated and very easy to understand way. I mean, the toil that that moved, I mean, can you imagine the amount of searches you would have had to do then the work you would have had

Speaker 2:

to To do do reading it.

Speaker 1:

To get to that position. Yeah?

Speaker 2:

Yeah. If you bring essentially everything if you bring everything into the context, nice part is that information coming from two, three, four different documents can actually end up in the same answer and the same result set.

Speaker 3:

AI is not perfect, right? Still You get hallucinations and things need to be checked, it's an awful lot better than it used to be. I would go through an AI query, say, a year and a half ago and I'd find loads of errors. And now you go through and you find one or two a bit nuanced, but there's nowhere near the same issue. They've done a lot of heavy lifting in the engineering and the processing to to get the accurate or to maintain accuracy as well.

Speaker 3:

So it's pretty reliable off the bat. I mean, obviously, serious stuff you need to absolutely check. But, yeah, for the vast majority of things it it works extremely well, doesn't it?

Speaker 1:

So so do we think, Antonio, that that the only reason that, you know, classic web search is still even bothered with is because people just aren't aware of the fact that you can use generative search? Is that, like is is the end of it, know, kind of absolutely written at this point, do you think?

Speaker 2:

Yeah. That's actually a great question. I've been thinking a little bit about it. I mean, I definitely talk to people that tell me like I don't search in the traditional way anymore. It's like zero per day.

Speaker 2:

That's kind of like, that's surprising in a way because there are sometimes like very simple things for which, you know, why would you go and use a powerful tool like a large language model to answer some of the basic stuff.

Speaker 1:

Yeah.

Speaker 2:

And others don't really use LLMs in general, like they just keep searching in the traditional way. I think that's kinda like more an habit at this point.

Speaker 1:

An habit. Yeah.

Speaker 2:

Yeah. But I guess what's probably also missing is is, you know, the ability for taking actions. Like, if you can search something on the on one of these chatbots and then you have to go back into your browser to actually execute based on what you've learned, I think that creates kind of like probably a preference in, okay, let me just search on the traditional engine so I go to the destination where I can then take the action. And so I think, you know, is definitely like still kinda like something some something like a gap that needs to be filled before we can pretty much move all the workflow into a Right.

Speaker 1:

And if we just maybe step behind the experience for a bit and just talk a little bit about what's what's going on behind to to generate and then return such very specific contextual information. And I wanna bring in something that happened to the Numbers website recently. Are you aware of the the conversation going around the Numbers? And Numbers is a film industry, like a movie industry website that sort of basically tracks the performance of movies like box office and all that kind of stuff. And it's been there.

Speaker 1:

It's one of those, you know, websites that's been there since IMDB type of thing and it's been around for decades. And what it found at the beginning of last year and we we had an author on recently called Stephen Follows who had done some writing on this and we had him on talking about it from the perspective of the numbers. And this is an interesting opportunity to be able to get into what was going on there. So what happened on the numbers? Old site, legacy technology.

Speaker 1:

When agents were starting to crawl over the numbers, the numbers couldn't sustain it and they've had to take it down and think about kind of re engineering that website. So maybe you could give us some insight, like what's going on with the crawling? Like how often does that happen? What is it trying to do? And then we'll and then we'll maybe explore from there.

Speaker 2:

Yeah. Actually, that's one of the the current issue of like chatbots. They tend to, let's say, use the big tool every single time. And when I say the big tool, mean like live patching the web page. And so that's the number that the numbers that you were mentioning.

Speaker 2:

Like now every time you ask something to a chatbot, it will open up the website and there's no caching happening. There's always like a live, you know, fetch of the web page. So essentially it looks like a visitor and now these old websites have 1000x more visitors and they cannot sustain the load. That's clearly an issue. And and the way I see this shouldn't be happening.

Speaker 2:

Most of the time, these websites don't change constantly. They they they they are stale and and there's no need to go and create this additional pressure on top of the webmasters. And so in my opinion, like, a search engine, what a search engine does, it goes into this website once a week, once a day, once a month, depending on the frequency of the updates, keeps a copy that is then optimised and made searchable by other tools. So you don't really need to go and disturb these web sites, right, don't need to create this pressure, this load every time you ask an information because the information was has been extracted already and I would keep only a pointer to the website. So if you really need to go expand or navigate or take an action on the website, the link is there.

Speaker 2:

But just for the information, that's something that can be extracted in advance. And that's what search engines have been doing. So one of the reasons why we're building sites in this way.

Speaker 1:

The increase in traffic, what's it gonna do to the infrastructure of the internet do you think? Because we're now going to kind of machine speed reading, not all of them are going to be as well behaved as what you're describing. You know, so there's going to be a massive increase of machine to machine noise on the internet. What's going to break first do you think?

Speaker 2:

I mean the very first thing that is going to break is as you said, sort of like the availability of the website you know, being online and so on. So the very first thing is like, obviously, as the owner of the website, like, I probably would like to block these scrapers, these live scrapers because they are kind of like, you know, they are creating an issue to my real users. Right. Right. And that's what has been in conversation, like, Cloudflare blocking a bot and a bunch of like others companies that pretty much like are gating the the web and it's becoming always more closer and closer.

Speaker 3:

And what about the economic model? So the thing that we the internet got paid for was our attention, the advertising revenue was essentially what what paid for the internet. Now if the bots are going, they don't do anything with adverts, they skip them. Are are we gonna have, I mean Cloudfair just released a sort of using standard coin to charge a bot to scrape a website type thing. Where do you think that's gonna go?

Speaker 3:

Because this is a this is quite a big change for how the Internet has been paid for, being structured, being set up. So we've got this load increase in this machine to machine and then you've not getting the ad revenue in for the the sort of business model. What what happens there?

Speaker 2:

Yeah. And I'm having a lot of conversation around this with publishers, content producers, and and so on. Because the main issue that they see is, well, first of all, some of the popular publishers, they care a lot about like the fact that these LLM, these chatbots need to kinda somehow understand that this is coming from an authoritative source, that just the just the claim coming from another website is not as important or as it shouldn't be it shouldn't be weighted in the same way. So they wanna maintain kinda like a posture, a status. And and the business model that they have seen, they've used and leveraged so far has been usually like the licensing business model.

Speaker 2:

Pretty much like, hey, I'm I have this content and I have this data and you think that's valuable. I can license it to you and then you can you can do whatever you want with this data. You can use it however you you want. And if your business grows, that's good for you. But that's the licensing part is something that has been arranged in advance.

Speaker 2:

Now they think that this needs to change, and we are seeing this with, for example, a couple of couple of companies. Like, I remember Reddit sort of, like, thinking about, like, stopping the licensing agreement with with with Google. And so there are definitely like companies in particular also in the news industry where they want to actually own an index, own their search capability. So they don't need to license and give up their data, but own their own way, their endpoint on the web. So essentially now chatbots can connect to their search engine and can use their search engine without having the raw material.

Speaker 2:

Yeah. And that's one of the reasons why we invested in building the technology, because now we can offer that as a as a service to to companies that want to expose their their their data without giving it away.

Speaker 3:

And I suppose that's the then they can potentially monetise access to that data if they need to but it's me being my data, my asset. Where the internet was open, like the numbers, people would hit the website to nick the data. I suppose now it allows you to change your commercial model as an entity to try and sort of counteract what's going on, which which makes good sense.

Speaker 2:

Yeah. And you also don't lose access to the most important part, which in my opinion, it's the other type of data. So pretty much like, okay, if users are coming and reading this article, why is that? What are they searching? So I can stay ahead of like what is valuable for my for my users, for my customers.

Speaker 2:

Right? So as a publisher, I wanna know what people want to read and how they read it and how they use it.

Speaker 3:

That's like the toothbrush with the camera sending data back about what's going on. You start to understand what people are interested in, then you can focus your business model against it, which is like you using the data in reverse to get an advantage. I quite like that. That's sort of like turning it into your your advantage for better business.

Speaker 2:

And if you just license your data to a third party company, right, to another company, you completely lose, you detach from your own product and and so you don't know how it is used and how it is and how valuable it is, right?

Speaker 1:

Just as an aside, been thinking about getting one of those toothbrushes

Speaker 3:

I have. I absolutely have. I knew it. So for the listeners who maybe didn't catch up, there's a new toothbrush on the market, it's got a little camera in it and when you're brushing it jets jets water between your teeth to auto floss. And me and Dave have had a debate about is this a useless thing or is it something and I went, it's just a very early iteration of something that's going to be commonplace.

Speaker 3:

But then we got into the data mining point which is where's the data of the video of the inside of my mouth going and who's mining it because somebody will be the dental. I'll start to Absolutely be being able leverage. Or flossing or something. They'll start start to punt me stuff trying to sell me services. But yes.

Speaker 1:

And and and you know, they'll be making you know, like customized dentures. So you know, when you need to get, you know, you you when all your teeth fall out and stuff like that, the internet will already know everything about your mouth. So you'll be able to go on to denture manufacturing sites get it sent to you. Haven't done that. Without having to go to

Speaker 3:

the dentist. It'll also send it to me. An agent will go, Rob's teeth have fallen out, send him a set of dentures.

Speaker 1:

You're right.

Speaker 3:

I I wouldn't even have to think about it Dave, a package will arrive in the post, I'll let him go, oh I need one of these.

Speaker 1:

It's like lifetime dental care.

Speaker 3:

I mean that's it.

Speaker 1:

Because you saw your data.

Speaker 3:

I know, I was thinking about it and not just a hat rack my friend, was I think three iterations of that and it's gonna be everybody's gonna have one.

Speaker 1:

No. So have you

Speaker 3:

you already No I haven't because it's a bit clunky but I'm like just improve, it's like the foldable phones, The first ones were crap and now they're in quite, I mean I've had a fold for ages and it's brilliant, so it's that you have to get through three iterations I reckon before it becomes Sounds right. Utilitarian in a way it's good. But yeah, I suspect a lot of We wanna see Sorry, go on, Antonio.

Speaker 2:

We wanna see product market fit first and then we're gonna decide if it's worth to evolve with

Speaker 3:

the product

Speaker 1:

enough. Whether it's worth €500. Well,

Speaker 3:

that's the other thing, right? Early adopters often get bitten or burnt, so I'm waiting for the price to come down and the design to be iterated. I reckon it's gonna be that's it's not a fad that I'm convinced it's gonna stay, it's too convenient.

Speaker 1:

I absolutely love how much cool air you've drunk on that throw Things are things are only been out for like a week.

Speaker 3:

I'm marketed to Dave. I'm easily marketed to. I'm a marketer's dream. I believe it all. Love You saw

Speaker 1:

it. That's the future. Take my money. I've seen the future in it.

Speaker 3:

Shut up and take my money.

Speaker 1:

And it's a camera on a toothbrush.

Speaker 3:

Anyway, we digress.

Speaker 2:

And speaking of early adopters Yeah. I I've always thought in a way like why people just, you know, buy product early. And, I think only now I'm kinda like realizing why that's that's the case. You were basically saying like, let me actually wait for the two, three iterations. But you know what?

Speaker 2:

Only the first early adopter, the one that buys the version one that is gonna send the feedback to the company that produces it is gonna shape the product the way they want it.

Speaker 3:

Ah, very good point. Yes. So I have to buy it. You get advantage, don't you? So I can make the product like I want the product.

Speaker 1:

You do realize, Antonio, you are speaking to one of the world's foremost Google reviewers. Of pubs. So Rob enjoys giving a bit of feedback to the internet. Now. So Rob, you wanna get on this train now,

Speaker 3:

don't you? Maybe that's a good point because I and you know what gives you a little bit of glee is when the location you've reviewed responds back. You get the little email and you can have a little chat with them. It's good. I was in the retro music.

Speaker 1:

Give give us your favorite example.

Speaker 3:

Oh, well, retro music. I went there, yeah, up north, and I ended up talking to the founder of it for about an hour about computers from the eighties and the nineties. He had loads of them. It was brilliant. I loved it.

Speaker 3:

Just I was in I was in heaven, David. I was heaven. I was speaking to him for ages and ages and ages, a real passion of his and he maintains this thing, it's a labor of love and it was like, I walked in, it was my childhood in front of me. And then so I reviewed it and I gave it five stars and he responded with a chatter and he remembered me and all that sort of stuff. It's just a nice social touch online that says this is an environment.

Speaker 3:

If you're if you're like me, that is one of the best places to go. It's not very big but oh my word, is it just filled with fun and memories.

Speaker 1:

Really? I mean, we really tapped into something there,

Speaker 3:

didn't No. But the I mean, the other good example I was talking earlier is just like the Google Glasses. They came in early iteration, complete disaster, social rejection, clunk implementation, but now we're into meta releasing them with Ray Ban and things like that and they've got into an and we tried them out at re:Invent and they do act they are actually really really good experience. They're probably third, fourth iteration and for me that's the key is they've refined it down to a point where it's just whether or not society will accept the fact that everybody's recording everyone all the time. But the new Apple Watch apparently listens to everything you do all day and will summarize it at the end of the day and I'm like, so now I have to stop talking around people with watches now, it's like

Speaker 1:

crazy. Anyway,

Speaker 3:

back on track.

Speaker 1:

Yeah. So just coming back to search, I wanted to talk about trust and accuracy in the sense of, as Rob touched on briefly earlier, some you know kind of early uses of Gen AI were were returning hallucinations and misinterpreting information. First of all in terms of that, Antonio, how are you guys fixing for that? Like what what what's the actual process of of of kind of of trying to drive to the most accurate representation? Particularly if you're bringing from, you know, multiple sources in the background.

Speaker 1:

And then and then just so you've got it in your head, then I wanna come on to like, you know, what's the potential attack vectors here? So does this open up, you know, like security concerns in any way? But let's let's start with accuracy if you can.

Speaker 2:

Yeah. Yeah. Oh, absolutely. So the way just to just to be also transparent here, I I personally still think that we are not there yet in terms of accuracy. It's true that the models have been evolving and they're But getting when it comes to more complex tasks, we are still trying to figure it out, how to improve these models.

Speaker 2:

And one way to, you know, when I mean, we are investing a lot on this is like these models definitely don't know how to memorise things and they don't know how to, you know, do this continual learning. And so all the news or, you know, numbers in general, tables are extremely hard for them. They always need these these external context to be provided together with the prompt. And so the way we are trying to improve that is is by having a different interface, which means they can query our index in in in a massive in a massive way, pretty much like sending tens, if not hundreds of queries all in parallel, so that such that all the different angles are covered for a specific topic. But at the same time, these models get back content that is shaped in a form that they like, which right now it's typically in markdown, for example.

Speaker 2:

So this type of format includes, you know, the data that these LLMs can consume and use in a way that they couldn't otherwise. So imagine getting back a table, right? How can you get back a table, from one of the standard, the traditional search engines built for humans? We don't like tables or snippets. We don't like or short descriptions of the website.

Speaker 2:

Right? But LLMs would love that. And so we are really changing the the paradigm by pretty much exposing it to multimodalities and and yeah, things that normally they wouldn't get access to.

Speaker 1:

And with the just this the sheer amount of sort of traffic and activity that it's gonna drive, what if any is the increased cyber risk that could come along with this that organizations should maybe be cognizant of?

Speaker 2:

Yeah. I'm not too sure about, like, what kind of, like, cybersecurity problem that could cause, you know, given that the data is available online, it's publicly available anyway. So, you know, these LLMs could potentially access this data. I think the risk is in general to have a system that is, you know, way more powerful than a human, but that then, I mean, that's the type of things that we are aware of. Oh, yeah.

Speaker 1:

Yeah, absolutely.

Speaker 3:

I do think there is a point around sort of like in security, there's always a principle known as needle in a haystack, where if you have a small bit of data that might be sensitive but it's surrounded by enough, it's hard to find, right? It's not perfect, but it's kind of a thing. So if you go into Google and a human tries to find it, it can be hard but I think people are gonna have to get a lot better at not leaving data on the internet because the the this type of system will be much better at at finding it.

Speaker 1:

Routing it out.

Speaker 3:

Yeah, whereas it might have been hidden in the back corner of Facebook somewhere or something like that. Now something can root it out much quicker and I do wonder if there's a personal point here around the sort of the users having to become even more savvy. Mean, it's difficult enough already about phishing and scams and all that, but you've got this sort of your data is is probably easier to find if somebody really put their mind to it with an agent.

Speaker 2:

Yeah. I mean, that's that's definitely true. But I also see that at the same time, you know, when we leave these traces essentially online or this sensitive data, etcetera, like, right now we are doing it because it's us, you know, writing the code when we leave, for example, like a password, the written, aren't coded in the code or, you know, the sensitive information because we write a ticket or something, etcetera. Like, as soon as we end off all these tasks to AI, these agents are going to be smart enough not to write an hard coded password or not to write like a phone number or kind of sensitive data in one of the tickets. So I think it's it's the way I look at agents is there is gonna be, like, the the bad agent, but there is also gonna be the good agent that is gonna be working with you.

Speaker 2:

So it's gonna be agent versus agent. So anyway, like the one that tries to find

Speaker 3:

Agent trying to find your data, an agent trying to delete it on your behalf. Excellent.

Speaker 1:

The thing that that boggles my mind a little bit about about that, and I wonder if you've got a position on this, which is yes you'll have that's almost certain to happen isn't it? And it will happen at machine speed not human speed. So is it you know kind of what do you what do you do you have any kind of mental model of what a machine speed internet looks like from that perspective? Because you know humans are good at thinking about things in you know kind of the way that we perceive time and activity. But you know kind of say e commerce between agents is gonna happen in a fraction of a second.

Speaker 1:

Not like it did maybe between humans where it might have taken hours, days, weeks for something to get to get done. You know, what's your what's your wait, you know, what's your mental model of that?

Speaker 2:

It's fascinating how we designed the Internet in with trade offs built for humans. Mhmm. Right?

Speaker 1:

Yeah. Yeah. Exactly.

Speaker 2:

Oh, a web page is gonna load in half a second because half a second is nothing for a person, and that's acceptable. But it's a lot of time for an agent, right? Like in half a second maybe able to read 300 pages and we're only opening one because there is a network latency that we need to respect because there's a physical limit, right? So the entire infrastructure is built for humans and needs to evolve if we want to make these agents more efficient. And I think, you know, going back to the position of sort of like, hey, we need to actually keep a copy that is closer and closer and closer to the model, a copy of the web obviously, right?

Speaker 2:

I think that's one of the reason why, you know, why we're building like search engine that is completely independent from the cloud.

Speaker 3:

And in your sort of cached version, how do you deal with data that doesn't age quickly? So certain scientific data will always be true like atomic numbers on carbon or whatever versus something that might age very quickly like, you know, news stories or things like that. Do you have a way of understanding recency and tagging stuff to say, yeah, this needs to be refreshed much faster than that, or is it just cycling everything all the time? What's how does that work?

Speaker 2:

Yeah. Cycling everything all the time, unfortunately, doesn't work because the web is so big, it's like trillions of documents and you can't really sort of like access billions of documents at the same speed, right? So you need to make a decision and the decision is essentially, is this type of information that is subject to change quickly or not? And, you know, if you're looking at published research papers, right, they're published. They're gonna stay there unchanged almost forever.

Speaker 2:

And so there's no need to go and refresh this data once you have acquired it once. But in news, in news is subject to changes lately. There are live news sometimes, and so you need to, you know, constantly recrawl them. So we we build models essentially to understand when web page is classified as something that you need to recrawl faster than something else. That's really the that's really the only way.

Speaker 2:

And there are signals that basically tell you, like, if this is, like, something that, you know, subject to change in fifteen minutes or in in, I don't know, like, two months.

Speaker 1:

And maybe just to bring today's conversation to a a little bit of a conclusion, let's maybe gaze ahead a little bit like the work you and your organization doing Antonio is like right on the cutting edge of how this stuff is working. What are you thinking about and you know, what can you share with us when you think about maybe eighteen months or three years out from now?

Speaker 2:

Yeah. I think the way we think we the way we think about, like, web search in general is subject to change in the next years exactly because of the the the motivation that that that we discussed before. So, like, the entire the entire system, entire infrastructure was built. Think about trade offs that humans are fine with but not agents. And so it's really like a kitchen sink of ideas right now where people try so many different things and only a few of them obviously will become the new standard.

Speaker 2:

And we are kind of like playing this game as well where we bring in new ideas that are the nice thing that I like is that they are influenced by where the academic community is going also. And so we try to stay in touch a lot with the information retrieval community. We are trying to bring in essentially new ideas on how these agents will need to surface the web in the future. And it's fascinating because it's constantly changing. It's like every single day there is a new piece.

Speaker 2:

And I think that's also part of part of the reason is because of like AI. AI can accelerate a lot the way you discover new things. So it's almost self improving yet. Not yet, but like almost self improving.

Speaker 1:

Now, as you know, we end every episode of this podcast by asking our guests what they're excited about doing next and that could be a great bottle of Brunello that's waiting on your your kitchen table or it could be something in your professional life or maybe a bit of both. So Antonio, what are you excited about doing next?

Speaker 2:

It's it's a Friday, so I'm excited about like spending my weekend with my kids. Oh, yeah. I have a six year old and a three year old that are learning to to bike. So, yeah, it's gonna be a lot of fun. What I'm really excited next is we are we are gonna be working on a couple of papers actually.

Speaker 2:

So I'll probably like we're as a team, Seltz, we are gonna be spending some time doing research. So back to the origins.

Speaker 3:

And

Speaker 2:

I think that's something that always keeps me keeps me very motivated and excited.

Speaker 1:

Well, enjoy that. Three and six are such beautiful ages. You know, you've got an amazing ten years ahead there with with your little ones. Mine are a bit older now. So they'll be other the other side of that ten years.

Speaker 3:

But

Speaker 1:

yeah, enjoy that. It's gonna be lovely.

Speaker 2:

Thank you.

Speaker 1:

I wanted to just maybe pull a couple of threads together because as we mentioned in the intro, we've done a couple of shows exploring this new type of search and what's going on on the internet and how the internet is changing and things like that. Yep. As we said we had we had Stephen Follows on talking about what happened with the numbers. So from the from the perspective of a data provider and a webmaster. I know Antonio has taken us through like a really a really good deep dive into actually what's going on, what's the state of the art and what you should do.

Speaker 1:

So Antonio, in the in the scenario of the numbers just to pull threads together, what would you advise websites to do to sort of ensure that, you know, when this inevitable amount of crawling is going to happen that they're controlling that appropriately?

Speaker 2:

Yeah. So I think, you know, we are in a point in time when the industry has to evolve the way they they treat their data, the way they the value that they give to this data. It's kinda like different these days. And Mhmm. I think shut shutting the doors and or or, you know, closing the Internet as it's happening is probably not the solution.

Speaker 2:

We have seen that in music and and in the past. And and the evolution was Spotify as a platform, right, where artists can can share more openly their, you know, their songs and and podcasts. And but at the same time, users still pay for for listening, it's just in a different way. And I think, you know, that is probably, like, the evolution of data in general that is on the Internet. I think, you know, the content providers, publishers will still own their data and they will make it available in different forms.

Speaker 2:

So it's really the distribution that's going to be changing. You're going to definitely see something happening, something revolutionary. And this is like going back into the early two thousand when Google came out as a search engine and it was kind of like a very different type of approach, an innovation that we didn't see before. And I think now this is happening again. It's like, you know, we are living that world and we are evolving into something new, so But

Speaker 3:

the strategy you talk about, Antonio, is exactly what the numbers needs to do, take control of their data back, produce that engine, understand what data people are actually interested in as opposed to be there, I'm not saying they're a victim but you know, they don't have control over their own destiny, they need to change the way they think about it, change their architecture and bring in the types of strategies you're discussing. But actually there's a big advantage for them as well, So they know what data has high value and what data might not be needed, so then they can refine their processes themselves. So it's, as you say, a natural evolution of the system.

Speaker 2:

Exactly, it's essentially owning your own data and maintaining the control of your own data without giving it away as a license, but like still making it available for your customers.

Speaker 1:

Well, look, Antonio, thank you so much for sharing your insights today and helping us do a thing. It's been a real pleasure talking to you.

Speaker 2:

Thank you so much for having me. It was a great conversation, really appreciate it.

Speaker 3:

Cool. Anything on this week's show resonate or look like it's heading for your business, get in touch at realitiesremixedcapgemini dot com. We're on LinkedIn too and we'd love to hear from you. Connect and DM us. The awkward questions are the one we enjoy the most.

Speaker 3:

We love watching the world change while we're talking about it. We'll track it and tell you what we spot as we go. Please rate and subscribe small thing but a big help. Huge thanks to our guest Antonio, our sound and editing wizards Ben and Louis, to our Machiavellian producer Marcel and to every one of you listening. See you in another reality next week.