Antonio Mallia Podcast Transcript
Antonio Mallia joins host Brian Thomas on The Digital Executive Podcast.
Brian Thomas: Welcome to The Digital Executive. Today’s guest is Antonio Maglia. Antonio Maglia, founder and CEO of Seltz.ai, believes today’s web search infrastructure was built for humans, not AI agents. As organizations deploy increasingly autonomous systems, the quality, freshness, and speed of retrieval becomes a strategic differentiator.
Seltz explores why many AI teams are moving beyond traditional search APIs, and what it really means to build a retrieval stack from first principles, and whether search for agents could become as important over the next decade as search for people was over the last twenty years. Well, good afternoon, Antonio. Welcome to the show.
Antonio Mallia: Thank you, Brian. Thank you so much for having me.
Brian Thomas: Absolutely, my friend. I appreciate it, especially making the time today. You’re, I know generally your office is based in San Francisco, but you’re in Italy currently, so I know you had to traverse some time zones and calendars to get here, and I always appreciate that, so thank you.
And Antonio, let’s jump into your first question. Your path runs from the University of Pisa’s deep information retrieval tradition through a PhD at, at New York University to research roles at Amazon, Pinecone, and Bloomberg. And you’ve said the idea for Seltz crystallized while building Alexa’s question answering engine at Amazon when it hit you that the consumer of search was no longer a human, but a machine.
How did that realization pull you from research into founding a company, and why was that the right moment to build it?
Antonio Mallia: Yeah, that’s a very interesting question. Actually, I’m currently in Pisa because– And I wanna start from, from, where it originated. Pisa’s been a great place, a great university to pretty much have students kinda like becoming passionate about information retrieval and web search in general.
And that’s where, that’s where I started and where, where I k- come back to, you, find more people like me are passionate about information retrieval. And, you know, there are many young talents coming out of university, coming out of their, you know, research program from, from, from school that they want to join companies like Celonis, where we are actually going to apply these, these studies into something very, very concrete.
And through that, kinda like– I mean, starting from, from Pisa, from my original studies, going to, NYU, doing a PhD specifically on information retrieval and in particular on efficiency for web search. I kinda like realized that, this was a problem where, people were getting comfortable in a way.
It was kinda like what is called sometimes like a solved problem. So like nothing new was coming out, out of the, further optimizations until, transformers came out in two thousand and nineteen, more or less, and– or at least when they got more popular. And, and pretty much like search has changed dramatically.
Like many things have, have changed since then. At the time I was at Amazon and, we were building like a web search engine specifically for Alexa, which if you think about it, th-this is before LLMs, before the generative kinda like era. And Alexa was really reading sentences from, from a webpage to answer some of the user’s questions.
So it started to kinda like be clear that, the next user, the next tool that is gonna be leveraging search was not gonna be any more like a human. And so now at the time it was Alexa, like, this robotic device, right? But now things have moved towards being more LLMs and agentic workflows.
And so, after spending a, a couple of years at Pinecone, which was a very AI native company, I kinda like realized a-AI is really like the way web search is gonna be, fully exploited. And so decided to start Seltz with pretty much like past colleagues from Amazon past colleagues from, from research and academia.
And, now we are a team of fifteen people very, very passionate about this specific topic.
Brian Thomas: That’s awesome. Love the backstory. Obviously, the genesis of everything started in Pisa, which where you’re at right now, and your curiosity and interest in information retrieval I thought was interesting, and obviously that fuels your passion around this particular topic and, and industry and technology.
But you were determined to solve a bigger problem here in search, and I thought that was interesting. You’re making more search, obviously more accurate, more human relatable and really that improve that customer experience. So I appreciate that. Antonio, rather than wrapping Google being or brave like most AI search products, Seltz took the harder path of owning its own retrieval pipeline, crawling, extraction, indexing, retrieval, ranking, all that.
You’ve conceded that web scale search is one of the most capital intensive problems in software. Why was owning the full stack non-negotiable for you, and how do you justify that much harder path to investors and customers?
Antonio Mallia: Yeah, it’s been actually quite very nice experience to see a product, shaping out from nothing, like building everything from scratch.
And I think now it’s very clear to investors and customers why this makes a difference. And I can, I can tell you more about it soon. But like at the very beginning, obviously this wasn’t clear. And it was kinda like hard to justify why not do what everybody else was doing, which is, wrapping Google, Bing, or Brave.
Now regardless of like, if this is something that people should do or are allowed to do to wrap like an existing tool to sell a product, I think that poses like some limitation on, on the product itself. You’re pretty much like building as an output, like your interface is your ergonomic, right?
Like it’s really pretty much like shaped around something that existed before and that was built actually for humans. Again, Google, Bing, or Brave, they, they, they, they literally built a web search engine in the traditional way. Meaning like ranking is extremely important. So the position of every document in the top ten results that, people retrieve is actually significant.
Meaning that, a, a person would go top to bottom will, will, read the first snippet, go down to the second one, and then eventually, like at some time, at so- at some point stop and, and decide that, maybe the query wasn’t right and they’re gonna like re- re- rephrase it then.
So there is a concept of kinda like, patience in a way patience in, in, looking at these results in ranked order. For– with LLMs, this is not a thing anymore. Ranking doesn’t really matter because an LLM is gonna read all the context regardless if it’s at the beginning or at the end, in kinda like the same way, in a very eager way of, of, extracting information, understanding the context.
And so the way we built and the way we are still improving and building our engine is more, focused on for that type of customer. Meaning like, you know, ranking is obviously important in, in a global way, meaning like how do you get to extract, let’s say the top ten documents that are, more relevant for generating some context So, ranking is important to pretty much like consider those the top documents that would be able to answer to provide context and, allow the LLM to answer the question.
But ranking within the results set is not, is not a thing anymore ’cause everything is gonna be consumed
More, moreover, like one other thing that we, we looked at when we were sort of like deciding how the shape of our API should, should look like is like the content itself. Snippets which is what traditional search engines return to user they’re, they are extremely valuable for humans in the sense that they can be very short, as long as they are informative about what the user, the the person is gonna find inside the webpage after clicking that link.
So, truncated sentences are fine ’cause it– they give you like an idea of what is gonna be inside, even, even without giving the full information, even without answering to your query or question. But this is not okay for an LLM because, it’s a truncated sentence. So something that ends up with dot, dot, dot, right, is gonna invite the LLM to hallucinate in a way, to kinda come up with the rest of the sentence without actually knowing that information.
So, you know, a solution to that is clearly to provide either the full content or very dense snippets or paragraph where, pretty much like you don’t have sentences that are truncated. You make sure that, they– the sentence itself contain the information that the LLM is looking for.
And so it’s really like a different way of thinking about, you know, ranking, the way you present the results, the way an LLM surface these, these results. Now another interesting aspect about this is that LLMs, AI models in general, robots, if you wish they query in a very different way.
Like the input to the, to our search API, for example, is, is actually quite different from an input that a user, a human could insert. In fact, humans are usually bounded by the number of tokens that they insert in a way. Like they– People like to be to write short queries to be a little bit ambiguous in the way they formulate the query.
And maybe they will refine it afterwards, right? But like the first attempt is like a very, very simple generic query that is hard to kinda like, fully understand from a retrieval perspective. LLMs are completely different. They like to write very long queries. Sometimes these long queries, they even look like a passage, a sentence.
They also try– They are, they are also very, very specific. They pretty much know what they need, and they will go and, and, and nail the specific topic. So, if you think about like, a we-a web search engine, building a web search engine, there are things that you have to prioritize over others.
To give an example is like you probably don’t really need spell correction if the input is coming from an LLM because it’s rarely gonna me-mess up with the you know, with the, with the query. But it’s way more important to produce the right content for the LLM, the right context, such that it can extract the, the right information to, take eventually like an action, produce an answer, et cetera.
Finally, the other interesting thing to consider when when, when we are building like Salts, while building Salts, we thought a lot about this, is like, what about the content? Most of these LLMs are being trained already with the content that is present on the web. In fact, the pre-training phase of an LLM is actually done by looking at open source crawled version of the internet.
So now if they had a chance to, look at these documents in, in details what is the point or of repurposing it to them at runtime when a query comes in? Most of the time, they in fact memorize some of these infor– the information contained in the more kinda like know-knowledge dense parts of the web.
I think and, you know, we are seeing this with our customers these LLMs care about a fraction of the web more than everything. So, building a very scaled up version of a search engine like as it could be Google today with, trillions of pages indexed won’t make a lot of sense for an LLM in the sense that there is probably no need to have every single page on int-on the internet to be indexed, right?
And instead, we try to focus on the things that bring more value to our customers. So to give an example, LLMs are not great at memorizing tables, numbers he-headcounts any type of, like, value that is A, subject to change or change, or B it’s it, it comes within some context. So it’s really, really difficult to to memorize, and it’s in a structure that is not easy for LLMs the way they are architected, right, to, to be memorized.
So we, we tend to, we tend to index and, and, and leverage a lot like these type of data that contains, either tables or, or, financial reports or, or things that are harder to be embedded in the model weights or documents that usually web crawlers in the traditional sense don’t reach don’t, don’t crawl.
And those are like we, you know, I was mentioning financial re-records like typically like SEC filings, for example, are not leveraged in-depth or at least not enough. And so LLMs, they clearly need to access them at runtime because they, they are subject to change, they are fresh, and so, they wanna stay up to date.
But also by searching, just by searching on, on a standard search engine and looking at the snippets, it will be really, really hard to extract the information that, you know, a potentially very long and complex financial report record could contain. And then finally, obviously, as I was mentioning, freshness.
Freshness is the number one feature that search engine or built for AI should focus on because this is really sort of like something that there is a cutoff, something that the LLM hasn’t seen in training ever, and so they need to stay up to date. It’s really the only way to do this continual learning as of now, at least in, in…
And, and so pretty much like, keep the LLA, the LLM and the model up to date. So for all these reasons, we decided to build it from scratch. We also decided that we are never gonna have a fallback in one of the traditional– using one of the traditional search engines, and those are, like, our core principle.
Brian Thomas: Thank you. Appreciate that. You unpacked quite a bit there. But what some of my takeaway notes here, you’re proud and excited about building this robust platform from the ground up, taking a separate path that most others typically wouldn’t take. The traditional search is all about ranking and the things that list items in a numerical rank sequence, but you’re, you’re doing this differently, obviously, using search AI.
Everything will be available, making this much more robust and a better user experience. And again, just love what you’re doing, building this again, just from scratch, doing this the right way and, and ensuring that searches are relevant and meaningful when users are doing that. So I appreciate that.
And Antonio, wrapping up here, you framed the big question as whether search for agents could become as important over the next decade as search for people was over the last twenty years. If you’re right, what does the world look like when most web retrieval is machine to machine, and what has to be built technically and commercially for the future to fully arrive?
Antonio Mallia: Yeah. I mean, I personally see a world where yeah, most of this is gonna be machine to machine and as we are seeing the number of– so like the query load, the number of queries issued by machines are, growing up very rapidly. In particular, because if you think about it, like the population using web search is in the end limited and we have certain we can, we can submit a certain amount of queries per year.
And, and pretty much we reached a plateau. In fact, in the past five years, there hasn’t been like a consistent growth as it used to be in the early 2000, right? Where everybody was approaching the web and, and, always more people, even more people were using search engines in a more extensive way until there was kinda like a spike.
But that pretty much reached a plateau in, in, in the past years. What we are seeing on, on, on the other hand is, is, AI systems pretty much submitting always more queries. And, and, and we predict that this is gonna surpass the number of queries submitted by humans in an exponential way.
Meaning that, you know, as of today, maybe we are, or we are using, you know, one agent per person and, we are slowly gonna be offloading all our, our knowledge work to these AI agents. But I think, in, in the early future, like in the, in the close future the– we are gonna be spawning like many agents per person, right?
With all all of them being specialized and experts in different, in different topics. So all these agents will, will absolutely need external knowledge. They will need fresh information coming. And so that is when web search will, will need to be built Specifically for these systems with all these different expertise and with all these different requirements.
I think the also number one feature that we will need to have in the future to fully arrive is really the ability to scale up web search in in an unseen ever, in, in an unseen way, something that we have never seen before. I think, the sort of like query pressure that, even large and co- large commercial search engines have seen until now, it’s it’s only gonna be a fraction of what is gonna be web search in, in a few years when there’s pretty much gonna be no limits in the number of users.
‘Cause again, while we are limited as, as, number of people using a number of humans using the the, the, the web I think that’s not gonna be the case for, for agents. Meaning like it will pretty much scale horizontally as we do with systems. And so efficiency is pretty much gonna be the number one priority for most of the companies working in this space.
Brian Thomas: Thank you. Appreciate that. That’s awesome. So the web retriever world obviously is gonna evolve. I liked your insights here. It’s gonna be a, a lot of machine to machine in the future, as you mentioned. These LLMs, these AI systems they’ll be doing the queries. These queries will certainly surpass human searches exponentially.
We know that. And today, you know, we may have one agent working for us, but in the future, humans will be leveraging many agents making today’s search only a tiny fraction of what search will be like in the future. I thought that was interesting. But efficiency will be key, obviously. That’s the key now, but it’s gonna really be the key in the future for companies.
So, so thank you. And Antonio, it was such a pleasure having you on today, and I look forward to speaking with you real soon.
Antonio Mallia: Brian, really appreciated the times. And thank you so much again for having me. This was a, a very interesting conversation. Hope to have you– to, to, to, to be again here soon.
Brian Thomas: Bye for now.
Antonio Mallia Podcast Transcript. Listen to the audio on the guest’s Podcast Page.
