Unleashing AI's Potential in Publishing: Insights from Atex Media Group's CTO Michele Alessandrini

Hello, Michela. Welcome to the show. How are you feeling?

Hi, Mike. I'm very good and I'm very glad to be here and thanks for having me.

Of course. Of course. I mean, we'll go into very interesting topics in the publishing industry and how AI is affecting all of that. Before we go into the topic, Michela, Give us a taste of you, like who you are, how you came into tech industry, and you know what you're currently doing.

Yeah, so I'm, uh, I'm Michele. I live, uh, I live in Italy, and, uh, currently I work for a company called ATEX Media Group, which is a company actually that have been around for 50 years. And, uh, and we produce, make, sell, and maintain software for the publishing and news industry around the world. Um, we are a tech company, and I've been in tech for all my professional life, which has been more than, you know, a 20-year journey. And, um, and I guess, you know, it's always been my, my passion, technology. And to me, it was a kind of like a natural path to kind of start in tech and move in tech. In the last probably 10 to 15 years, I've been moved more on the kind of like a cross between technology and business more and more, and, and on some of the technologies out there, including cloud computing and, and lately artificial intelligence. And I think where I am today is actually the where all this experience brought me. Because I mean, we, as a company, we can actually— we are producing technology that helps businesses to perform better. So I'm in a position as a CTO of the group to contribute to this technology, to this innovation, to this relationship that we have with, with our customers.

Okay, that's a, that's a, that's a deep, deep technological background, Michela. So I love that you working for an industry which I feel like is probably the mostly affected by AI advancement, and you guys utilizing this. So let's just go straight in the topic. So today we'll, we'll try to sort of see how you guys do this as a technology company, how you utilize AI for the publishing industry and give that benefit So the very first question I would have is, in general, in your view, how is AI transforming the publishing industry?

Yeah, I mean, I guess the publishing industry, as you rightly said, is probably one of the, you know, most evident industry in terms of exposure to AI. And the reason is pretty simple. They have content. So in the last year, we saw this subsection of artificial intelligence with generative AI exploding. And generative AI is all about content. And we work with an industry which have an incredible amount of content generated over the years. So they definitely want a probably the industry that can benefit the most from artificial intelligence today. And there are multiple ways probably in which AI is affecting them or will affect them more and more in the future. But I guess the bottom line is probably to help them to get in contact with the final users. So to be able to kind of like focus on what they really know how to do the best, which is writing, which is finding the news, which is dive deep into topics which are relevant and which matters for the people, and just make that relevance, you know, out there.

Okay, so let me ask, is it the digital content which they are reusing, do you see any requirements or any moves in the sense that people take an old content which is not yet digital, you know, maybe company has, you know, internal papers, maybe, you know, handwritten books or, you know, reports or whatnot, where this is getting lifted, put into digital and then getting given to generative AI to sort of, you know, reshuffle an existing content which is not digital yet?

Well, we saw that as a use case. Of course, I mean, many companies have been around for many years. And as you rightly said, I mean, many companies have archives which are not in digital form yet. So yes, it might be a use case. But actually, probably a broader use case is, generally speaking, the exploitation of the archives of the company, of the publishing companies. So many of our customers, but I would say the majority of the players out there, they sit over, you know, 1,000, 100,000, millions of articles, which are not really used as they should. So they do not represent an incredible value for the customers and for the publishers. Why? Because they probably, they've been written years ago, and they are not very easy to find. And, you know, probably the journalists that work today are not the ones that wrote the articles. So you lost track of what you produce over the years. And the AI is actually helping them through, you know, some specific use cases like smart tagging, to be able to recover the knowledge behind the content. And this is probably one of the most, I would say, relevant use cases for the publishing industry now. Archives and how to make that archives relevant again.

Okay, I can see, I can definitely see that. So, um, with the speed of advancement of AI technologies, and I mean, I, I'm, I'm just a user myself and I use ChatGPT, of course, um, and I can see how this product is evolving very, very quickly. And as well, I can see how the market takes these, you know, bigger products like ChatGPT puts, you know, API on top and makes a specific sort of problem solution services. And like it's advancing very, very quickly, you know, in the last couple of months or, you know, let's say a year or so, I've seen a bunch of different products coming out. How do you as a technology provider stay on top of these advancements to basically leverage them to make your technology better?

There is no magic formula here as just to decide strategically speaking to want to be there. And, and that means that from an organizational perspective, you need to have resources, you have to invest time, you need to hire the right people in the company, you have to gain, acquire, somehow capitalize the knowledge that you need in order to stay on top of a market as an— or as a technology segment which evolves every day, um, that is, uh, always moving, uh, and it's, uh, even if it's, uh, let's say, easy to access right now, uh, it requires many different skills in order really to, um, to be exploited completely. So I guess that, you know, the short answer is, uh, to gain the knowledge through, through the right people and through dedicating the right amount of time and, uh, and, and, and strategic focus.

Okay, that's a very, very interesting— so if I'm strategically positioning myself towards AI, and as I understand, your technology is exactly doing this, who do you go and hire for this? Like, like, how, how does the title sound? Is it like AI specialist? Is it, you know, machine learning people? Is it— I don't know, that would be my question is like, who do you actually hire to make sure that that strategic positioning is getting, you know, executed well?

Yeah, that really depends on what you want to do. So I would say, you know, there are job titles that sounds, sounds about right. Okay. So if we talk about machine learning, data scientists, they sound good. Okay, they sound aligned to what you want to do. But actually, there are more professions that might be more relevant for you to start within the AI and to be able to use AI for your business use cases. I'm talking about, of course, from a product manager that might be exposed to the meaning of the AI. We're talking about computational linguistics, which are the people who can understand or mix, you know, the study of language with computing. We are talking about data engineers who can help you to kind of like establish the right data flow to make, you know, and fine-tune the models, the AI models that are very easy to access to But they require actually access to the data and they require to be integrated with your own world. So I would say that probably product managers, linguistic and data engineers are probably the first one that you need. And of course, if you want to do more, if you want to develop your the next big thing, the next big model, then would, would you have, you know, the usual data scientists, the usual machine learning engineers, and so on and so forth.

So that would be your strategic, uh, team who would effectively allow you to build the plan.

Yeah.

Uh, and then you would just fill the gaps with, with, with an actual on-hands type of people. Okay, so I really like the linguistics things you touched, and you said that your, uh, you know, the business you work for is actually providing software for the whole globe. So how does this strategies differ, even if it differs, if you're providing the solution for an English-speaking company versus, you know, Italian or Spanish or just overall, you know, all different languages? Like, is the solution behind the hood is actually different, or do we have the modules which just are language independent effectively?

Um, I probably— it's a mix, it's a mix of both. So right now in the generative AI space, at least we have some huge models which are primarily trained on the English language because, of course, I don't remember by heart the statistics, but the vast majority of the content out there is in English. And therefore, our models which are very, let's say, perform very well in the English language and they try to do their best in all the other languages. So the more a language is exposed to content, the better that the model becomes in dealing with that specific language. From a user perspective of that model, then it becomes, let's say, I would say in the beginning, a multi-language type of product for you. So your linguistic approach to the model is just have little variation in terms of language. Okay, but it doesn't change completely the architecture of your application or the approach to the application. Then of course, moving forward, so the more you become mature in your journey, you can start a fine-tuning process within the model to be able to perform better in a specific language, a specific niche market, in a specific type of language, in a specific jargon, and so on and so forth. But I would say that the beginning, uh, the generative AI space is multi-language, um, by design, I would say.

Okay, so we got— so we got our team, they're thinking about what we're going to be building. Let's say, you know, we decided what we're going to build. What are the main challenges, you know, historically speaking, you guys had as a solution to integrate AI into your offerings? It's like, where is the biggest, you know, walls to climb, as they say?

Well, probably the biggest one is the fine-tuning of the model. So it is very easy to come out with, let's say, an API endpoint, just to be speaking technically, who is able to do a summarization of a text, right? And it's easy because, as you said, I mean, there are models out there which provide you very easy access to them. What is not trivial is to make sure that the quality that that endpoint produces is at the level that, in our case, like a publisher wants. So the outcome of the AI is the quality that the customers need. And this is, you know, we see this as a trend. So many customers see AI as magic, you know, you click a button, something happened, and I exactly get what I need and want just by asking. Sometimes it's true, but it's not the case of all the requests you can do. And therefore, I would say to find that balance, you know, find the right trade-off between performances, between output quality, about speed, and so on and so forth, is real, probably the biggest issue that the team the IT has to face when dealing with any product, you know, that includes AI.

Okay. Do you see that this understanding of quality differs when it comes to publisher to publisher, or is it just like a generic one understanding of quality? You know, giving your context, let's say I'm a publisher A and you give me the solution and I go like, you know what, I would like to see more of my brand language in this, you know, in this output, you know, some of these areas are not for me. So you go in and you fine-tune this, you know, for this publisher, but then publisher B comes out and they go like, you know what, that's actually not what we want. We want something else. And then, then you end up in some type of, you know, fine-tuning templates, which is per publisher. Oh, that's not true. And you know, you have just one generic solution, which actually fits everybody.

No, it's definitely around customization. And that is one of the design challenges of the application, of course. So you have like a universal model, but then it comes down to what customer or publisher A perceived needs wants in terms of quality. And I'll give you an example. I mean, there are publishers that, you know, they want all the titles of their articles with capital letters for each word, and there are publishers, they do not want that. But if you ask a model to produce, just to come up with a title for your piece of work, the model answers, "In a way," which is the way in which he believes is most likely to succeed. That way might be, you know, not well received by some of the customers or some of the publishers. So the fine-tuning of the model has to be done at the customer level in order to adjust the model to the needs of the customer. Of course, again, it's a trade-off. So sometimes you just have limits in the technology right now. So you need a little bit to adapt. To what technology provides today. But by design, uh, your application has to be capable to adapt, uh, and, and, and to kind of like to push down, to push through the customization that you need.

I do feel— I mean, in my experience, you know, walking into customization area is always, you know, it creates a lot more for you. You have more teams you have more different clients to manage, you know, customer success needs to grow as well. If you think about fine-tuning, you're probably paying extra for the, you know, computing power to fine-tune for exactly that client. How does that, you know, come back into the cost of the product? Is it, is it that you need to sign a longer contract? Like, I'm just interested in the pricing module of how all of this affects the pricing module of your platform.

Fortunately, very little. So of course there is a cost of the design of the solution. So there are costs of the people that you need in order to design and produce and maintain the solution overall. But in terms of, I would say, an end user, it will be just a matter to pick the right product in your case. So to pick the product that offers you the customization. Because as you rightly said, customization comes with some cost around it. But in terms of technology, that cost can be absorbed quite easily by the technology behind it. So fine-tuning a model, for instance, is the cost of fine-tuning a model is more on the data engineering part rather than fine-tuning per se. So just to give you an example, I mean, it's not really the computing power to fine-tune a model and retrain specifically a model, at least if we talk about generative AI, but it's more on how do I integrate the data source with the model and be sure that the fine-tuning actually give you the results that you expected, that you hoped.

Okay, so you do this once effectively, right? You just fine-tune it to the point where the output is matching against your requirements, your quality, and it's done. The work is done. After that point, there's, there's really no more labor cost which goes into that.

Yes and no. For some part, probably yes, or at least you will not fine-tuning, you know, a model every day or every week. But what we saw is that models change as well. So it's not probably any magic formula, any silver bullet, but we saw that every 2 to 3 months you have to review Okay, what you have, what you did, what the technology state of art is in the market, if there are things you can do differently, and if you need to change prompts, for instance, or instruction sets, because models change and because language evolves, because feedback— yeah, you need to integrate feedback from the customer. So I would say that once you start, you know, labor comes in every 2, 3 months, if we're talking about fine-tuning a model.

That's very interesting. So it's effectively, do you see this changing? I'm just looking from the perspective of how AI is advancing. And, you know, we're getting a cheaper computing cost now, you know, NVIDIA, AMD, you know, the market which is driving the, you know, the, the hardware itself for AI, and then, you know, the software layer— everything is moving so, so quickly. So if you want to adapt to this, you're probably gonna have that change management around, you know, looking at new tech, putting in new tech, you know, trying new things. It's just like, it's just gonna be very, very constant, isn't it? It's not a technology you create and forget and then you concentrate on selling it.

Oh, absolutely not. At least not right now. So what I'm expecting is that it will stabilize in, you know, over the years, probably. But right now, you need— it's one of the reasons to have a team, to have the right skills within the company, or to rely on a company that has the right set of skills, because it's not a one-off, but it's a perpetual let's say, journey into the AI, into this technology, which brings you quality, which brings you the result that you need and want. Otherwise, to me, is playing, okay, is experimenting, is make it proof of concept, which is all good, okay, which has been the last year that we saw basically in our industry. But at some point, if you want to, you know, if you want really to kind of like to integrate and to benefit from the AI, you need to play following the rules. And the rules today, C, working every day and B, and run rather than, you know, walk and see.

It's a race right now, isn't it? Yeah, it is.

It is.

If you want to participate, you can't stop. Otherwise, you're just going to get outrun very quickly. Okay, so if I take us back to the quality, I'm interested to understand, can you expect the output fully, or is there a human-driven, you know, some type of sanity checks which still take place into the publishing cycle?

We consider to have those sanity checks. So within our software offering we provide the journalists ways to report back the quality of the output of what we provide. Because ultimately, the quality of many of the use cases of AI today is human-driven. So it is in our industry, the journalists in the end, that's to say, okay, this is a title I wanted. This is the quality I expected, or this quality is good enough for what I need to achieve. And in that case, so yes, within the process, there are ways for, for human to send their feedback and help through, again, that engineers and workflow and the work of people to become better.

Okay, so there's a effectively a feedback loop which takes place. I, I use the, I use the solution, I get the output, I, I, I provide my feedback, goes back in, and it's just constant improvement, you know, continuous effectively. Absolutely. Innovation which takes place. So do you measure the quality of output which is based at the editorial level, as in compared to based on, on user engagement level? So for example, If I'm putting in a query, I'm getting some output, I'm just going to put this out there and see how people react to it? Or do I actually do editorial level sanity checks and then I put that out?

It's the second one. So at the editorial level, you accept the quality that will go out to your users. Then of course you can measure, you know, at different levels, uh, the quality that you do, but let's say the gatekeeping is when, well, at least in the use case, you know, related to articles, in when the article is written before it gets published. You can measure the effectiveness of the AI also by looking and relating the, let's say, the business metrics. So the very end of your of your business goals towards the user, you know, the user, not of the AI. This is something you can do. This is something that is done. But I would say that, you know, the first gatekeeping is before an article gets published.

Okay.

Because in the end, there's a publisher, it's a journalist that is the ultimate gatekeeper, except the quality. Accept the article.

Let's say I accepted the quality, so I generated an article, I accepted it, I put it out there. Is there any statistics where, where AI-generated content with least amount of changes performing better or worse compared to an article which, let's say, is purely human written or has very little of AI involvement in it?

Currently, I feel we are still at the beginning of that journey, so I would not rely on any statistic in February 2024, especially because as today, probably the majority of the AI has been used for articles not on the major publishers out there, but rather in different contexts. It's hard at this point to actually to say, to have, you know, a relevant number, a relevant number to show. But this is what every publisher wants to know. Okay, so, and performance in that case, what is it? What is the definition of success? Is, you know, the amount of new subscribers, maybe if you are, you know, a journal that sells subscriptions, maybe can be, I don't know, the amount of times that your article becomes first in Google. I don't know if we're talking about technologies that helps your position on the search engines. So your definition of success change also, you know, the perceived and the real quality or the real, let's say, contribution that AI can give to an article or a piece of content in I like that.

I like that. It's gonna be different and it's gonna be contextual to every, every publisher, isn't it? It is based on what they actually expect from the tool. Um, do people actually notice if the article is AI-generated or not?

Oh, again, it's very hard to say because it really depends on the article. Some articles probably are not, um, uh, will be very hard for an AI to contribute in a meaningful way. So if you're talking about a very recent event or a very local event, so an event that requires data that are not out there anywhere, will be very difficult for AI to generate or to contribute with relevant information. So in that case, probably yes, as a user, you would see something very generic. Okay, or would you miss something? Other types of articles, probably you as a user would not see, you would not perceive a difference, or would be very, very, very hard. And that is where, you know, some of the ethical, you know, issues or questions are lingering, you know, right now.

Well, let's ask that question then. Let's ask that question. What are ethical considerations? You know, if you're producing content and you're effectively using AI, you're putting this out there. It will not be about real-time news. It will, you know, it will be probably, you know, something up, I don't know, a year, a year before, and it's probably going to be generic topics. But that's still good content to put out. And if you're aiming at just generating more content, you know, to compete better you know, in, in Google, let's say, are putting this out there, at which point do I raise the question, you know, it's like, is, is this okay? You know, is it— because I'm not writing this, right? So I'm just effectively reshuffling what's already likely online in different words.

We are still at the beginning of that discussion, but I would say that, you know, globally, the general consensus that is building up, at least amongst the major publisher is to try to be as transparent as possible with the users. So one general approach that seems to kind of like, again, building up is for the publisher to say explicitly how the AI got into that piece of content through a specific and explicit maybe wording at the end of the article through a little table saying exactly how the AI contributes. But as a user, you should be able to trust the publisher or at least understand that what you are reading, what you are seeing, what you're watching, what you are listening has been produced by a human or by a computer or by an algorithm. And this is the first step. So transparency is probably the key. That generally speaking is seen as probably the first step to mitigate this issue, which is part of the solution, if you wish, right? So you cannot really get away from that issue, that ethical issue. But what you can do is to mitigate it by explicitly saying explicitly, kind of like open up your AI strategy with your users.

I like that. I mean, worst case you can do is lose trust with your readers.

Yeah.

And they go like, oh, I don't know, like, you know, is this generated? Someone wrote, you know, I don't trust your brand anymore and I'll go look for news somewhere else. That's probably the worst outcome you can get. So being transparent around how you use the tool And as well, probably not using the tools is not a choice because you just gonna fall behind in just sheer amount of content you can put out. Because automation which comes with the AI, it's very high level, right? You know, the numbers are quite strong. So if you're not using it, you're just probably not going to compete well in the current market. So you want to use it. It's just going to tell the users how exactly that, you know, influenced the article. Okay, that's a sort of like a natural flow of order which deliver that solution, right? You know, publishers just being more open. So one of the biggest fears is either misinformation or just, you know, wrong facts, or, you know, AI hallucinating around some stuff and people get this they read this as a, you know, as a piece and they take that as knowledge. Is there any like consideration of how you do this? Like in a sense of not being labeled as, okay, this is fake news.

Yeah, this is probably the biggest, the famous elephant in the room. So hallucination is part of what the model provides today and there is no clear answer technically speaking. So you have to deal with it. Of course, hallucination and wrong information is a potential risk for the adoption of the technology, is a risk that all the publishers are aware of. So fact-checking, which is probably the technical answer or the technical holy grail that everyone is trying to achieve, is, uh, is definitely a big topic. It's a big topic for us, it's a big topic for industry. So right now, the approach at least we are trying to follow is, um, is to be able to verify facts in, outside the model itself. So the approach we are taking is using AI to understand and isolate the facts within a given text, because of course not, not, not everything you write is a fact. And once we have isolated the facts, we check those facts against a different amount, a different amount of sources. Uh, could be like Wikipedia, could be a search in Google, could be your own archive as a customer. So you start to trust yourself. You can build your own database of facts, or you can build your own database of sources. So right now is a bit, okay, the model says that, let's understand if that is real or not by checking around.

Okay. And that's a manual process?

No, no, no, it's an automation. What becomes manual is kind of to go through the outcome of this automation and see, okay, actually what the automation found proves or doesn't prove that the fact is, is real or not. In the end, there is also always a human there to say, okay, uh, even if you found the 3 references in Wikipedia that says that this date is wrong, I still believe that it's true, they are wrong. And I, as a human being, accept the final outcome. I think that, you know, that is the real core of the discussion. You need for some of these big issues a gatekeeper, which is still human. The more we are progressing, the more models will advance, the maturity of the companies will advance, maturity of the readers will advance, maturity of the human species towards AI will advance. Some of these components, some of this behavior will change. But right now, I think we cannot consider AI without humans. Uh, that is, uh, at least not in the publishing industry.

It's a wild, wild west otherwise. And, and if you lose trust with the technology, you know, the adoption sort of falters because people just don't trust the technology. So we need the gatekeeping to get— to make sure basically what's getting outputted and that automation benefit is actually delivers the full value without losing the trust. Let me just postulate on something, Michele, I have. So would it be possible to create a very, you know, a very non-trusting fact-checking module which tries to find things which are wrong in an article? So effectively, it has this bias leaning to prove that the facts are wrong in an article. And just sort of, you know, once you get output from one module, give it to another whose pure purpose is, is, is, you know, to prove that the facts are wrong.

Yeah, on paper, yes. But on paper, yeah, on, on paper, yes. But the variety of, of, of, of topics, matters that you can write an article is, uh, is, is, is, is huge. Is huge. So, uh, and I'm thinking about For instance, everything that is really local, because if you're talking about an article that says, you know, talk about Newton and the gravity formula or math or statistics or history, that becomes, I wouldn't say easy, but understandable how you could do that. But if you are describing the local game of a local town and then you you said that specific match ended with a specific result, understanding that is real or not, that becomes an issue. You know, it's not always possible, or it's not a given that an automation can actually say, yes, that was a fact, or yes, it was wrong. You know, right or wrong is not always, you know, they carry different level of, uh, of, uh, of difficult, um, compared to, to the type of fact that you want to check.

Okay, I can absolutely see that. Yeah, so it's just a trade-off which we're paying off right now by utilizing the technology. On one hand, we're getting this, you know, good boost of automation, but on the other hand, there's still, there's still things which is just not, you know, perfect, and, and there's still a human being involvement necessary. How do you think that changed?

Let me say, just, just, uh, just to finish, let me say that even now human, human beings, uh, do things wrong, you know. They might be on purpose if they want just to steer the conversation in one direction, or by mistake because maybe they were reported like a wrong piece of information. So even now we are exposed, you know, to wrong facts. Okay, what we need to avoid is to kind of make things worse by using AI. Okay. But the level of wrong facts we are exposed to today will probably not change that much. You know, with the AI, it's— it can improve on something. Of course, if you write a wrong date, this is something very easy for an AI to spot. But, you know, overall, the reassurance that all the facts are in an article or in a piece of text or a piece of content are real is still a long journey to go.

How long is it? And let's ask this question: how will AI evolve in the next 5 years? And how would will this affect publishing industry? How will it improve publishing industry?

I would say that, you know, um, in the next 5 years we will see a real adoption from the publishing industry. So this first year has been, uh, I stop and look, okay, because I really want to understand if it's, if it's real, it's just a fad, if, uh, if will be relevant for me, yes or no. Then of course there are players more open to innovation, more open to risk, and then they moved before others. But the first year has been, let's see, right now they are all moving into, okay, now I understand that there are benefits for me, so I want to do more. Okay, in the next 5 years, what I see is an industry completely embracing it. Which means from their perspective, being able to focus on the only one thing they want to focus today, which is writing the best quality article they can have. Why? Because that content represents the key for the businesses to get the trust of the customer, to get more subscriptions, to bet, you know, to create business out of of a piece of content. So in the publishing, in the publishing industry, at least what I see is a complete embrace of the technology and the reduction of the overheads that today, let's say, the journalists as a category have to spend outside writing, beyond writing, you know, beyond their job, beyond their core of the job. Which is writing article research news, understanding the truth behind some news, research the facts.

And, uh, okay, do we see AI in publishing industry coming a lot closer to real-time news in 5 years, or that's just way too soon?

Yeah, I can see that. Uh, the fact that today some models have to be trained every 1-2 years because costs are too much largely depends by algorithms and how much computing resources they need in order to be trained originally. So if I look at other technologies that we had in the past and I'm looking and I'm thinking right now at the decoding of DNA. When we started decoding DNA, to generate a sequence, to decode a sequence was costing probably close to a million. Very, very, very, very expensive. 20 years later, it was around, I think that today is $100. Yeah. Okay. And the same will go with the AI. So I can see the real-time training an option. Probably would be near real time, not really real time, uh, but you can, uh, in 5 years at least, but will be definitely an option that will be all exposed to.

Okay, that would be interesting. I mean, I can— I, I see some non-publishing industry AI advancements, you know, around, uh, like video and, and audio, being able to create your avatars and then just sort of, you know, put your voice in, give some text, and it speaks like, you know, it's you. That, you know, you can still see it, it's not you. But all these AI-driven avatars, you know, reading the news, that feels like something which is coming for me. At least I see that on, you know, on YouTube and on social media. And then if you can take the content as real as possible to real time, you suddenly, you know, raise this question. It's like, like, do you— do even a person needs to do that? So that's an interesting future we're looking at. Um, in the spirit of the show, Michela, one question from your perspective as a CTO: what is one most important factor for a company's growth, tech company's growth?

I want to say keep innovating, but keep innovating sounds like too much of a buzzword. So I want like to rephrase a little bit that in, in, in in establishing, say, a culture of learning. But learning means stay curious, uh, in the market at all levels. So from the CEO down to, to, to, to, to, you know, to, to all the other employees of the company. And, uh, I mean, what we've seen in the last 10 to 15 years is a technical world, a technology world that changes at the pace which is very hard for a single person to keep up. Okay. We saw like a huge amount of different languages, technologies, frameworks. Some of them lasted for a year, maybe less, you know, and being able to process all that is very hard. So the risk for a company is just to freeze completely, say, okay, I pick, you know, my, my toolchain, I pick some of the some of the technologies that I need today. And I become very good in these. And I just focus on my core, which is fine. Okay, the 90% of the company should be, should be around that. But you cannot be blind on the fact that while you are focusing on the core, there's a whole world moving, you know, a completely different pace. So it's very important starting, you know, from, from a you know, from the strategy from the very top to define a strategy of innovation, to define a strategy that allows people to look outside the core. And to me, this is the key to understand, to have visibility, to understand and to make the right decision, which possibly are this piece of innovation, this piece of technology, this is not for me, I don't care. I will skip it. Okay. But what you have to avoid in order to be relevant, in order to grow as a tech company, is not knowing. This is, this is unfortunate. This is a situation of many, many, many companies, especially small to medium companies, which probably do not have, you know, they have to be focused on the core. They struggle to find the time, to find the resources, to find the right strategy. To give, uh, you know, that perspective to people.

I love that.

To me, in the end, is, uh, is, uh, the ultimate recipe to, to grow and to, to, to stay competitive in the medium to long term.

In general, growth is propelled by doing things in different ways, or doing things in a better ways rather than different. You, you constantly need to find how can you do better. Which is effectively, you know, the growth curve. So do you have a strategy or a way of getting this done within the company? Do you think it should be embedded in every single employee? As a result, it should be a company's culture, or do you think it's more of a dedicated team responsibility which could be, you know, I don't know, led by a head of innovation or whatever, and they have a certain people which, you know, just 8 hours a day they look at the market, you know, in, in the, in the industry they compete in, and they try to source ideas and they try to innovate. So is it like per-person cultural or dedicated team?

Probably is a trade-off, uh, and it really depends on the topic. There are topics that will require someone to, uh, have his or her head within the topic every day, okay, because they are too complex to manage per persona, okay, per, per, per. But I strongly believe that everyone should be empowered to feel that innovation can help what they are doing during the job. So generally speaking, the per-person model is what I chase, what I would like to see in the company. And that, that means in the end, uh, for everyone to feel owners of what they do, from one side. So you are the owner of your job and you are the owner of your results. And if I can manage to do that, and if I can manage maybe from a central perspective to give you the tool to understand what's around you. And if I hire you correctly, if I, you know, help you to grow, the result will come, you know, that person will find the right tools for the job. Then of course, that innovation can be helped by a central team like AI, for instance, you know, the central team look around every day, they scout technologies, they understand technology, but then they're trying to reflect and push down that technology to all the different people, to all the different teams, say, hey, this is something that we found. Now it's your responsibility, kind of like, to understand if it's a good fit for you. Okay. And that is a virtuous cycle to me, um, that maybe for some topics start from the center, started from the top, but has to be fueled by every person in the organization. Otherwise we'll be dying. Okay, if I, if I have the genius, if I have the, you know, the head of innovation that everything that, uh, you know, comes out is perfect, and, you know, it's very innovative, but actually it's not making any, anything that is relevant for, for my core as a company. What I'm doing is just to pleasing myself.

Okay, okay.

And this is a company, something you need to avoid as much as possible.

Okay. I love that. Great answer, Michele. So I switched the gears, you know, from our AI discussion to this.

Yeah, no, no, no.

So coming back to AI, right? So we did discuss and we mentioned many benefits for publishing industry to embrace AI as a, you know, as a tool. Can we zoom in those benefits? Like if we've, you know, just look at the order of most impact, like what is the most impactful benefit a publishing company can get by utilizing AI tooling?

I would say more focus. Uh, probably the most certain benefit is, uh, for them to really focus more time than what you do today, what they do today, on their core job, which is writing articles, uh, and, and everything that is behind it. Because, uh, currently behind the article there is a lot of things that are not strictly related with the article. I'm talking about, you know, okay, coming up with a title might be one, but everything that maybe is SEO-related, so search engine optimization-related, yes, it is related with the article, yes, it's functional for your business, but it's not strictly related to the quality of the content. It's more about being effective when you distribute the content out there, right? And, you know, publishing on social media, publishing on the newsletter, translating an article into another language because maybe you have a multi-language website for your media. Those are all activities that are not strictly related with the, you know, with the quality of the articles. And this is an immediate benefit, immediate gain that a publisher can can get from embracing AI. Those are the level step zero kind of benefit when a publisher moves into AI, having more time, having more focus in what they know how to do the best, which is writing articles.

Okay, I love that. I mean, that's a huge time saving and cost saving. So outside of focus of of your core capabilities, which is writing an article, all of these around you can effectively just delegate. So you're getting time back, again putting more articles out, and you're probably getting, you know, more consistency around these outputs, around the SEOs and, you know, and translations. And as a result, you're, you know, saving on labor costs because no one— like, someone doesn't need to do that, right? You can again reutilize that That's a huge, huge impact.

But even more, probably it's about being able to do the job. Compared to like 30 years ago, the amount of information that a journalist or a publisher needs to process in order to write an article is immense. Okay, so there are sources coming from all hundreds of social networks you can imagine. There are images, audio, video, press releases. There are, you know, a thousand different sources. So being able to manage is beyond what a human being can do. And the result is that you either kind of like spend your time looking around or you're spending time kind of trying to write an article with a limited amount of information. Because it's too much. Okay, so the AI can actually allows you to be able in the first place to process the world around you, which is, is as today volume of data and variety of that in the world of big data, if you wish. It's, uh, it's probably beyond what a single human being can do, um, to be able to process the world and, and write a piece of content that reflects that word. Or at least maybe you could do, but with a very limited amount of, you know, output volume. So if you write one article per month, that is fine, but that will be unsustainable, of course. So AI can help you to focus again, but also to process the big data world which is around you. Which is around the journalist.

Okay, that, that's, yeah, that, that's another one which I didn't really thought about. I mean, I myself use tools all the time when it comes to just understanding market trends, which are, you know, a huge abstraction on top of social media and what people are talking about. And I never thought about from like publishing perspective that that's, you know, that's a huge influence in that one as well. So thanks for, thanks for this one. That was new one. Right, so I read books, right? And often I like books written in certain way. I like— I didn't even know that that's a thing until I read a huge amount of books, right? And then I sort of found the ways of people writing styles, and I found that that information how they give to me in that style actually sticks with my head much better. And it applies to not like novels. I'm, I'm talking about like, you know, job-related books, technology books, uh, any type of books. So if I have interest as a writer now to have my style of writing in AI-generated content, like how well is this achievable?

It is achievable, not probably 100% of what you had in mind. So again, it's not magic. You know, AI today doesn't reflect you and your brain, your emotion completely. Okay. But with the right amount of quality in terms of content and within the fine-tuning process, which includes different approaches to it, you can achieve results which are definitely encouraging and at the quality level that can be enough. To produce something that is valuable for the readers. So having your own model, if you wish, and we're going back to the customization we were talking before, right? So having the model of your publisher, so the publisher, the publisher A has the model A that is able to produce a piece of content that follows the tone of voice the type of angle, the type of words, the specific language and wording and punctuation that the publisher uses and wants to use, that is achievable. It's not magic because sometimes it will be not consistent 100%. Okay, right now there are no predictive algorithms when it comes to generative AI. Okay. Okay, there is always a level of hallucination, there's always a level of, of changing the output, and that's where human and then quality control eventually comes in.

Okay, but you find—

Unic is an answer. Yes.

So there's a, there's a step you can take towards that direction, but it wouldn't probably be likely what you expect. But there's definitely a set of brand language rules which you can bake into the output.

Yeah.

Okay. So, um, so this is like, for me, like when I try to think about technology behind it, it's like, for me, that feels like there's a lot of moving parts, you know, which goes into all of that tech. You know, we talked about customization, which is pretty much like the last layer, but then, you know, beneath you have those different modules, then you have all the different data coming in, maybe from this brand, maybe from that, maybe from generic internet, social media. Like, that's a, that's a really big system in my head. Can you, can you touch on, like, scalability of, of the solution you guys have? It's like, you know, I'm happy if you go technical by then, Michela. It's just like, you know, are you in the cloud? You know, which cloud provider? How you manage your data lakes? You know, your training modules? Like, all of that, you know, the whole picture, if you can.

Yeah, um We are in the cloud, so technically speaking, I think you are right. Scalability and being able especially to adapt fastly to changes and new integration, new complexity, or the complexity which differs from the complexity you had like a month ago, uh, is key. And it's also key to be able to test and try, fail and retest and try and fail again until you find the right, the right way to do things. So cloud is definitely the answer to these, um, to these aspects. And we use the cloud, um, and we decided to use a serverless approach as much as possible. So our platform born in a serverless way and in a managed component way. So we write the code, but when it comes to data flow integration, when it comes to flow of the request, of the response, everything related to the auditing part, to the governance part, to the reporting part is all managed through serverless components. And the reason is pretty simple, because serverless is scaled by design, first of all, and doesn't require to have our focus on the infrastructure management per se. So at least not at, you know, not in the quantity needed in, in, in the traditional infrastructure, even in the cloud. So being today 100% serverless allows us to think and not to worry too much, besides the architectural design decision that we took, about the scalability of the model. So the bottlenecks today are the models themselves, which have limits, which do not answer real time, or anyway, they have their limits that today are— it's not possible just to pass. Okay, we, from a publishing industry perspective, we saw that as a real non-problem as today. So the publishing industry doesn't require to be, you know, 100,000 requests per millisecond because otherwise, you know, you don't get any value. So let's say that the limits of the model are not really a problem today and the scalability of our system is taken care by using the cloud and we use AWS for this specifically and to go completely serverless.

Okay, so you can effectively— yeah, serverless means that it's quite dynamic load type of system, right? Like, if you need anything, you just— it's just going to blow up as much as you need, and if you don't, it will just shutting down.

That's the idea. That's the idea. And also, that will, uh, has an impact on cost, because if you don't use a model, your, um, costs are zero. Or very close to zero. And if you use the model, you will, you know, you will generate costs. And that is a very kind of, I would say, nice model to have within AI in particular, and within AI when you deploy to a customer that might be having, you know, for a few months testing phase, and then you roll out a few journalists and then to the whole newsroom, and then you add more use cases as along. That will be very hard to— not impossible, but hard to manage with traditional infrastructure. And, and traditionally, again, would, would solve— you would solve that problem by over-provisioning, typically, which, you know, means more cost. So the fact that if you don't consume, you have zero cost, and then you, you increase that as you use the service It represents a good strategic choice also from a cost perspective, which means being able to offer better pricing to our users.

Did you consider microservices as opposed to serverless?

Well, the two things are not in contrast. So we use microservices in terms of design pattern overall. Our microservices are implemented through serverless components.

Okay, that makes sense.

This is, uh, so the two of them for us in, in, in, in the AI product that we sell, it's called AI Atex platform, combine.

Okay, that's a quite up-to-date tech you guys have, which is, you know, good to hear. Otherwise you wouldn't be playing with AI, I guess.

Yes.

So Maybe just before we close, one last touch on the personalization and privacy. I feel like that's— it's definitely out there, and I would just like to understand how you guys do this. So giving you context, you know, I use mobile phone, I use Google News, uh, I touch off on couple of articles, you know, I click through, maybe I read them fully, and, and instantly it just picks up on that. And it starts sending me articles around what I'm actually engaging with, which, you know, supposedly drives my engagement, right? So when you bake this into your machine learning or AI solutions, there's a question of privacy because that's my data, right? You're taking my data and you're putting in your machine learning, which is effectively you are third party, uh, to the publisher to whom I actually give my data. It's like Are there any real challenges around that situation?

They are, but the technical option available to you allows you to solve those challenges if you want. So, for instance, we do not deal with— so, we do not own the data. Okay, in that sense, we are just processing the data. So we are data processor by all means. So the data stays within the customer environment. So the customer owns the data, it's the customer, the publisher, that is responsible, if you like, to have all the right agreements in place in terms of data. But again, the real bit, at least for the publishing industry today, is that the type of use cases do not have an impact up to now really on personal data. So it's more on the journey that the news goes through. So we're talking about, you know, discovering new news, like news gathering, and there is a production of the news, news production, then you have the news distribution that goes out there in your website and, you know, in the social media channels everywhere. So the news is out there, and then you have the news monetization, which is the last bit of the journey of the news. And, and through all the journey, then I, I would say probably not the 100%, but the vast majority of use cases do not get into personal data. But again, as a technology provider, we are not there to exploit specific data, but to create a you know, less friction for the publisher to get value to the customer.

Okay.

So, and then this is done by, uh, at its core, keep the data as private as possible and, and, and, and restricted to the customer domain as much as possible. And today it is possible, luckily.

It's easier from GDPR perspective as well if you're just a data processor you're instantly exposed to much less regulation requirements around that framework.

And I think it is right, uh, start by being a data processor, uh, because in the end, uh, you are providing— you are an enabler for the value of the publisher, and that value is between the publisher and the customer. You are actually helping that value to happen. With technology, you know, by being an expert in, in a certain type of technology. So your role is a processor, an enabler. You're not owner of that relationship. That relationship stays within the two, within the two main leading actors, which are the publishers and, and, and the users.

Okay, I love that. And this is where we end. Thanks, Michele, to coming to the show. I mean What you just shared is, is, is just great. You understand how the most impactful or most impacted industry, which is publishing, they— how you guys do this, you know, the backends and all the ethics and all the shebangs in between. Thanks for coming to the show, Michela.

Thank you very much, Mike. I really enjoyed the conversation today, and I really hope that, you know, to have provided some, some insights for, for whoever will listen to this.

You absolutely did, Michela. Thank you again.

Thank you.