SPEAKER_01: 0:10

Ladies and gentlemen, we have Bernard Acetuno. Hope I'm pronouncing that correctly. Co-founder of Stack AI. No pressure. Last one on the document. Thank you so much. No pressure.

SPEAKER_00: 0:24

There's never any pressure, Bernard. Thank you, George. I appreciate it. Nice, nice to see you in person, finally. Finally, right? By all means. Beautiful.

SPEAKER_01: 0:32

So here we are, Bernard on the sidelines of Unga and um at the NASDAQ today. Yeah. And uh we've talked a lot about the issue that you are apparently working on, which is AI. Very much, very much. And I'd like to ask you, how is Stack AI transforming? Well, let me go back. How is AI transforming enterprise operations? And tell us, because a lot of people are curious, how can AI tools be leveraged to solve complex global problems while at the same time, here's no pressure here, ensuring ethical and responsible implementation?

SPEAKER_00: 1:10

Absolutely, absolutely. So just in your risk, we have the same thing. It's a whoop. The whoop. It's a device that we're doing.

SPEAKER_01: 1:16

Rory McElroy.

SPEAKER_00: 1:17

Yeah, I know we are. Absolutely, yeah. So we live a healthy life. So the Whoop is consistently collecting data from our hands and from our vitals, giving us numbers and statistics on, you know, our different signals, you know, heart rate, sleeping, sleeping quality, whatnot. And the kind of data that we work with, this kind of device in general, is what people call structured data, historical, time series data, information that we always have collected over time. Historically, the computers and overall compute have managed to successfully interact and make analysis and insights from data that is well structured, well cleaned, well organized. However, that's only about 10% of the data that exists in the digital world. 90% of the data that exists in the world is what we call unstructured data. Scans of documents, Excel sheets, Word documents, you know, notes of documents, you know, audio recordings, videos, all sorts of, as we call it in our company, garbage that companies have. But unfortunately, most of the knowledge in the world exists there. Since the past decade, since 2010, we had a major revolution where AI researchers were able to create systems that could analyze and operate on this kind of data, on structured data. Eventually came to the point at the beginning of this decade where we could have systems like GPT analyzed that could generate text, could generate video, could generate images, which would be called generative AI. Generative AI overall can now unlock a series of new operations where now automation can reach that 90% of the data in the world. Now we can work with documents and we can find information in them. We can search for the things that we care about. We could collect payments, we can have conversations, we can generate poems, etc. We can take AI outside of the new with the world of numbers and tables and put it into the real world. Thanks to that.

SPEAKER_01: 3:07

Tell me just a practical question. How is that different from what we used to do with Google? Like we used to search, you know, I'm searching for a restaurant in New York.

SPEAKER_00: 3:14

Why is that different than great question? Because you know what happened with Google is that what Google would do was that they will take all the information in the internet, they will map it into a way that is structured, into what we call a graph, and then run a search over it. It wouldn't be AI in the sense that it's not like understanding what's under the information. It doesn't really extract things from what's on the website. It's just structuring the website into a way that is like searchable, and then it's trying to like traverse the web. And that's why many times you get things that, okay, they have the word you're searching for, but they don't really answer the question. Right. Because it doesn't really understand what you're looking for. Generative AI, on the contrary, does understand what you're looking for.

SPEAKER_01: 3:52

It does.

SPEAKER_00: 3:52

It does.

SPEAKER_01: 3:54

And what about agentic? We've heard about agentic. Yeah. So describe that in plain English.

SPEAKER_00: 3:58

Very good, very good. So, you know, it's a great question. Imagine you have someone that is a copywriter. Basically you tell you what something to write and it writes it. It can just write stuff. You tell it, okay, now write me a recipe and you know make me a via sandwich. It will go and write the recipe for a sandwich, but it will make you the sandwich, you can't do anything. If you give it the tool now to like go to a kitchen and cook, you know, take a pan, turn on the stove, you know, take bread, toast it, put cheese, you know, whatever, then the system is not just a copywriter, it can actually do stuff.

SPEAKER_01: 4:32

Create new new ideas.

SPEAKER_00: 4:34

Yeah, it can take actions over the world. More simply, if I go to the thinking. It is the definition of thinking is complicated. When you say thinking, what you're asking is, is there a logical reasoning behind every step of what it's doing? Over the past year, it has been a big breakthrough in the eyes. They were able to embed a logical reasoning change for every decision of a language model. That's where Jopen AI GPT 5 now comes to the market, Cloud Opus 4. They're able to like backtrace the logic they're following before they give an answer. That's the closest we can embed to thinking to a system like this. We humans do things that are much more complicated in our brain. But the point where we can understand why they make a decision and how they get to a conclusion. As they decide to, like, you know, when you tell them, send me an email, they decide to open Gmail, write something, put a subject, and then send that email, you can see the trace of actions they took. And that's why now agents or systems that can not only reason, but interact with its external systems now exist. What we do in Stack AI is that we have a software that helps enterprises build custom AI agents to automate internal back office operations. We're with banks, hospitals, large defense contractors, large government contractors and civil engineering as well, in order to automate key manual workflows that we have in collections, payments, business development, compliance, and the like.

SPEAKER_01: 5:48

So where do you see this going? You're sitting in the in a seat that's really unique because you know, as I like to tell people, and I and I so many of you in the room have educated me on AI, that it's not coming, that it's already here. How far are we from that sentient thinking, from the ability of AI to go beyond agentic? What's the next level after agentic?

SPEAKER_00: 6:13

So what agent is more of a framework of AI, right? And that what people want to go beyond agent is not really towards actually adding more capabilities. AI today is as complete as it gets to achieve arbitrary tasks. The real next level of success, uh the next big leap, the way we get to what people call artificial general intelligence. Meaning AGI, J artificial intelligence, AGI, which means a system that is able to, given a set of tools, perform a task at the same reliability that a human would. Given, let's say I take I take an AI like Chai UBT, and I give it the ability to write emails, I give it the ability to collect payments, I give it the idea to write content, I give it the ability to like, you know, respond to emails, I could tell her, become my content marketer, research the web every week, give me a nice post that is creative, it's different we see, and then send an email with it every Monday. If it could do that at the reliability that we have for a human, then we will have what we call AGI. And this is the example I would do for marketing, but this will apply for anything. I could have an AI that is a financial analyst, I tell it, write me a model that forecasts the performance of the Google stock, giving it to assumptions, and it could research the information, find what it needs to do. So, the same level of an analyst, then we will have what we call AGI. That's one level at the level of performance of a human. The next unlock, and this is what people are actually concerned about, AGI, we're, people, some people say we're quite close. I think that we are missing many steps together, but we're, it's tangible how to get there. The part that is really concerning for many, it's SSI, ASI, artificial superintelligence, meaning intelligence that transcends the ability of a human per of a human. Meaning if I ask a system, go and you know, figure out everything you can in order to break it to this computer, it could, at a scale beyond what 100 humans could do, you know, iterate with many algorithms, tries, you know, different different systems, create email scans on, et cetera, in order to do so, to a level that is beyond the ability of an individual human. And that's where most concerns are. Nevertheless, it was also where most opportunities. You could tell AI, okay, now run a bunch of experiments over these molecules in order to like cure this disease. And it could do it at the scale beyond what you know a major university and state life would do so, all just by itself. And that's quite promising for many fields, in medicine and in sciences especially.

SPEAKER_01: 8:20

Isn't that already somewhat happening?

SPEAKER_00: 8:22

It's happening to a small degree, but the problem is that AI language models, which are probably what most people are use as AI today, they have essentially two limitations. First of all, the degree of accuracy at which they can reason, meaning like at what if we all play with ChatGPT, we all have seen that at some point. After trying enough questions, we can always get him. We can always like get it to like you know make a mistake or like skip something, you know, and correct it. So a degree of accuracy it has. The second thing that they can work with is what we call context window, meaning how much text they can and how much they can memorize of a conversation. Language models, by being systems that live inside of a GPU, they can only handle so much information. And once they cross that information, it just stops working. Maybe you have seen that you know if you give ChatGPT a document that is way too large, like multiple thousands of pages, it just doesn't work. Tells you this is too much text, I can work with this. So in order for a language model to get to the proper superintelligence level, it must be, first of all, very accurate, and second of all, be able to work with like a very vast amount of data, way beyond what we can do with existing GPUs and existing compute. That's why projects like StarGa is so important, because it allows to take this technology to the next level beyond the physical barriers we have today.

SPEAKER_01: 9:28

And so, what what do you see uh the world looking like? Um, I won't go to 2030 because that's too far, just in a year or two. How different are things gonna be?

SPEAKER_00: 9:38

So, of for sure the underlying technology will continue improving at the same exponential rate. I have no doubt about it. You know, it's uh the people working on it haven't even advertised on it most of the things they have already developed. So the technology will keep improving and we'll become more accessible and of course more secure and more affordable. But more than the technology becoming more available, the biggest the biggest leap I will therefore see is organizations and individuals understanding where to apply the technology. I think that over the past three years there's been this kind of like a promise of AI that has been very exciting, and those that have managed to leverage correctly have seen tremendous gains. Very few organizations have managed to do that. But if you manage to achieve tremendous gains with very little payroll, or the contrary, like you know, take an exponential organization to where they are today. I think as the years come through and organizations find the specific places within an industry where with an organization where AI can be very impactful, this will be very transformational. The problem with Gen AI is that it's a very transformational software. In order to like properly use AI, a company must be properly using cloud, first of all, which is already a hard conversation for many banks. You know, oh, you're using you're not using cloud, okay. You must be using softwares that have open APIs, you must be using softwares that are that have you know systems that you can interpret code from, MCC service, why not? And that's challenging for many organizations. So there will be blockers to get in there that transcend the abilities of technology. They just rely on the infrastructure we have today. But towards getting to a point where those first few wins that really help the first batch of AI users and creators, they'll you know, we will see a time, a time in the world where most organizations and people will be able to benefit from those.

SPEAKER_01: 11:18

Incredible. What's your final call to action based on what you're seeing in AI?

SPEAKER_00: 11:22

Yeah, if you're working for an organization that is always asking, oh, what do I do with AI, why is AI good for? Go to Stack.ai and click on book at demo and you'll you we'll we'll help you figure it out. We'll help companies. Yeah, for sure. We'll help companies like Nubank, BAE systems, LifeMD, YC retirement all over the globe to use this. Well, are you are you bullish on the future? You think it's gonna be positive? Absolutely. I think I think we are under hyping the future. Wow. The best is yet to come.