Latest / The Index Podcast / Bootstrapped to 320K Users, ZERO Funding: Agnese Jaunošāne (COO & Co-Founder, Ajelix) on Open-Source AI's Power Move
Transcript
- 0:12SPEAKER_02: Hey everybody, welcome to the Index.
- 0:13SPEAKER_02: I'm your host, Alex Kahaya, and today I am joined by Agnes Hornoshan, COO and co-founder of Agelix, an AI-powered productivity platform that turns data into finished, ready-to-use work, formulas, dashboards, reports, presentations, and more.
- 0:32SPEAKER_02: Built with her co-founder and husband, Artur, Agelix has grown to more than 320,000 users across 150 plus countries.
- 0:41SPEAKER_02: And probably the most interesting part, at least for me as an entrepreneur, totally bootstrapped.
- 0:45SPEAKER_02: They haven't taken a single dollar of outside funding.
- 0:47SPEAKER_02: Super impressive.
- 0:48SPEAKER_02: Agnes is going to walk us through how you guys did that and where AI-driven productivity tools go from here.
- 0:56SPEAKER_02: Thanks so much for being here.
- 0:57SPEAKER_00: Hi, thank you so much, Alex, for reminding me.
- 1:00SPEAKER_02: So really cool background.
- 1:02SPEAKER_02: Interesting that you guys were able to bootstrap this, especially in a day and age when, you know, so you guys started this around 2022.
- 1:09SPEAKER_02: And I mean, so many companies have raised so much money and you guys went to go went kind of a completely different route.
- 1:15SPEAKER_02: But before we get into that kind of Genesis story, like what's the reason you guys started the company?
- 1:19SPEAKER_02: What gets you out of bed in the morning to work on this?
- 1:22SPEAKER_00: I would say the first thing is the ability to build something on your own.
- 1:27SPEAKER_00: That was like the number one motivator for both of us.
- 1:31SPEAKER_00: Why we wanted to build something, like our own startup.
- 1:36SPEAKER_00: And for example, for me, I previously worked in, I had a like corporate background before Accenture.
- 1:43SPEAKER_00: I also worked in a previous, like one of our country's first unicorn.
- 1:48SPEAKER_00: And I had like this experience of what it means to build a startup and how amazing that is that you can serve your customers around the globe.
- 1:57SPEAKER_00: And so this is something that I actually fell in love with the idea of it, that you can be in one country, one specific location, you can serve like the entire world, basically.
- 2:08SPEAKER_00: And so I would say that was the first thing.
- 2:11SPEAKER_00: And yeah, and then we came together with uh with my co-founder Arthur, and he had like this experience, uh, more than 10 years of experience in automation, specifically in Excel and Google Sheets.
- 2:23SPEAKER_00: And he already knew that industry had this specific problem with spreadsheets because everyone hates spreadsheets, but everyone still uses them around the globe.
- 2:32SPEAKER_00: And then we came together and we decided, hey, I guess there is a fit, you know, you can do this, I can do that, and it all just added up very naturally.
- 2:42SPEAKER_02: What country are you based in?
- 2:43SPEAKER_02: I don't think I I mean I I think I know, but before I make an assumption, tell me where you guys started the company.
- 2:49SPEAKER_00: I am I am based in Latvia.
- 2:50SPEAKER_00: In Latvia.
- 2:51SPEAKER_02: Latvia, okay, cool.
- 2:52SPEAKER_02: Yeah, very cool.
- 2:54SPEAKER_02: And so what was the company?
- 2:55SPEAKER_02: You said you were uh part of one of the first companies to become a unicorn in Latvia.
- 3:01SPEAKER_02: What company was that?
- 3:02SPEAKER_00: It was printful.
- 3:03SPEAKER_02: Oh, cool.
- 3:04SPEAKER_02: Yeah, very cool.
- 3:05SPEAKER_00: It also operated, I think, mostly.
- 3:07SPEAKER_00: They had the biggest market is the US, and they did the print on demand.
- 3:11SPEAKER_00: And during the COVID wave, they just grew grew very fast, incredibly fast.
- 3:17SPEAKER_02: And um, as far as a gelics goes, how did you start out?
- 3:22SPEAKER_02: Like what was the ear what were the early days like?
- 3:25SPEAKER_00: The early days were like, well, we launched before ChatGPT was even like launched, and at that point, no one really knew what was AI, especially in our region.
- 3:37SPEAKER_00: It was like some word people knew there is that kind of technology.
- 3:41SPEAKER_00: And we actually started with like these super simple wrappers that we use like OpenAI API, and it was like seven productivity tools for these Excel and Google Sheets users.
- 3:52SPEAKER_00: It was super basic, it was our like first MVP, and then it kind of went viral because we launched before ChatGPT, and we were one of the first ones in the market that solved these Excel-specific problems with AI.
- 4:09SPEAKER_00: And when everyone started to uh search um Excel AI or Google Sheets, you know, they were starting to search for specific tools related that are AI, uh, they just found us because we were we were one of the first ones on Google Positions.
- 4:25SPEAKER_00: Then yeah, influencers started to talk about us, which was also free of charge because there were like just several companies that were doing this, and we were one of the first ones, and so it all kind of grew naturally.
- 4:38SPEAKER_00: That's why we were able to bootstrap to so many uh users, and we got such huge uh features like on Forbes or like other like technical uh articles or or large media, and it was all like free of charge because we would just build a very specific problem, and we got a lot of reach by that, by doing that.
- 5:01SPEAKER_02: It sounds like you guys just kind of you know, it's part of just skills, experience in the industry, and making that move kind of right at the exact right time.
- 5:09SPEAKER_02: Yeah, which is just sometimes how it happens, you know.
- 5:12SPEAKER_02: Um, and very cool that you guys had this out in in the market before Chat GPT, because that GPT was like really the moment.
- 5:21SPEAKER_02: The ChatGPT was like the moment where everything exploded, but you guys were already in a pretty good position.
- 5:26SPEAKER_02: We we got connected on LinkedIn.
- 5:28SPEAKER_02: I saw a post that you wrote about open source AI, I think.
- 5:32SPEAKER_02: Talk about that post if you remember the right, the exact one, and tell me what you were talking about.
- 5:36SPEAKER_00: Yeah.
- 5:37SPEAKER_00: So the thing that we were talking about is the Anthropic CEO that that spoke around how open source is dangerous.
- 5:46SPEAKER_00: Open source, like uh open, yeah, frontier AI models are it's dangerous to open source them.
- 5:52SPEAKER_00: And so we disagree with that, of course, because well, for example, we are from the EU, from the European Union, and we don't have a lot of AI companies, specifically in Europe.
- 6:05SPEAKER_00: We don't have as much uh like this VC funding and everything.
- 6:09SPEAKER_00: We don't have this super, super developed infrastructure here in Europe.
- 6:12SPEAKER_00: We have like just a couple of like Mistral, if you know, the company from the France, and there are like other like small player players.
- 6:21SPEAKER_00: And right now, for example, we ourselves we use open source models that are available that China or your other European uh creators have developed and they have open sourced them.
- 6:32SPEAKER_00: And we use them to provide a safe infrastructure uh for, for example, our uh main customers.
- 6:41SPEAKER_00: Like what we work with mostly with uh enterprise and governments, which are super regul or regulated uh companies that are working in regular regulated industries, and for them it's super super important to understand where their data is getting sent to.
- 6:57SPEAKER_00: They can't just choose, like I, for example, a European company can choose just a provider in the US, and then they just send their data to Anthropic and they just can't track that.
- 7:08SPEAKER_00: They can just afford it.
- 7:10SPEAKER_00: And so open source is basically one of the ways, I would say one of the only ways how we can still provide this technology, which will inevitably become like a norm for our everyday work.
- 7:27SPEAKER_00: We are able to provide this open source technology to these institutions.
- 7:32SPEAKER_00: And if there's no open source technology, then we end up sending their data everywhere around the globe, and then everyone is training their models.
- 7:40SPEAKER_00: They're saying that they're not, but well, let's be honest, they will do that.
- 7:45SPEAKER_00: And then you just can't control anything.
- 7:48SPEAKER_00: There's just a big black box that just takes everything in, and no one can control that.
- 7:54SPEAKER_00: And I think that is very, very dangerous.
- 7:57SPEAKER_02: I couldn't agree more, and uh it that was that specific, just for people who maybe didn't see the clip, there were some clips going around of the CEO of Anthropic.
- 8:08SPEAKER_02: There was a a clip where the founder and CEO was in front of Congress in the United States, and he's talking about open source AI models being dangerous.
- 8:20SPEAKER_01: I think the path that things are going in terms of the scaling of open source models, I think it's going down a very dangerous path.
- 8:30SPEAKER_02: And this has kind of been a narrative of anthropics for a long time.
- 8:33SPEAKER_02: Like I I feel like they there it's partly true.
- 8:36SPEAKER_02: Like it's it is risky, these models, these models are racing towards superintelligence faster than we can keep up.
- 8:41SPEAKER_02: But also, I think partly marketing tactics on their part, like mythos is so dangerous that we had to pull it back.
- 8:46SPEAKER_02: You know, we've got the best model, use us.
- 8:48SPEAKER_02: Um The problem that I think both of us have with his take is that open like intelligence should be a human right.
- 8:58SPEAKER_02: And it should be something that you have a choice on where you run that software, where that your data goes.
- 9:02SPEAKER_02: And if you're using Anthropic or OpenAI or any really hosted provider, even if they're running open source models, your data is going to their servers.
- 9:11SPEAKER_02: With open source, you at least have the option to choose where that data goes and to retain the value you're creating by giving your data to that model and probably training that model on your data.
- 9:24SPEAKER_02: So it really is about ownership and control of your intelligence stack.
- 9:28SPEAKER_02: And the thing I keep hammering away at is if you're an enterprise and you're building an intelligence stack to run your business, it is the most important infrastructure investment you're gonna make probably in the next 20 years.
- 9:42SPEAKER_02: And if someone else owns that stack, you can't just port it wherever you want, change vendors, move things around, control what's happening, you're gonna get screwed at some point.
- 9:53SPEAKER_02: You're gonna find out really, really quickly one day your wake up and realize that uh your whole business runs on something that you don't have the ability to control because they're gonna do something that hurts your business.
- 10:04SPEAKER_02: And I mean, this I think there was a really big wake-up call when Mythos and Fable V got pulled off the market temporarily, right?
- 10:11SPEAKER_02: They started to realize, oh, you know, we need to make sure we build systems that can port models, right?
- 10:16SPEAKER_02: So I can go from Fable to GPT to Kimmy to DeepSeek to name another open source model.
- 10:24SPEAKER_02: So yeah, this is why we connected, because I think we share uh the same ethos around open source.
- 10:29SPEAKER_00: Yeah, I 100% agree on this because there's another thing.
- 10:34SPEAKER_00: I know this because uh maybe you will also know this, but uh in Europe, there are a lot of recent debates, you know, the the geopolitical situation and the US and EU relationships right now aren't at the at its best.
- 10:48SPEAKER_00: And so it's like a wake-up call for a lot of governments that are starting to realize wait a minute, where do we run our entire IT systems?
- 10:57SPEAKER_00: And it's all your uh US tech, like the Microsoft and everything.
- 11:02SPEAKER_00: And then there's another thing that's okay, but but all our data is like stored in EU, so it must be safe.
- 11:10SPEAKER_00: But then there's a thing that you there is a US Cloud Act, which if you know there was like uh this uh this Microsoft CEO in France, uh just uh he was brought in for court and he asked uh he was he had to testify if US government asks for data, for example, for from some kind of European country company data, they have to provide this data.
- 11:35SPEAKER_00: And so it doesn't even matter anymore if the data is stored in Europe, but who owns the data, who owns the infrastructure.
- 11:44SPEAKER_00: And so Europe right now is just going away.
- 11:47SPEAKER_00: A lot of governments are already agreed that they will either build on open source or they will uh use European alternatives for and businesses that work under European legislation and in jurisdiction of the Europe.
- 12:04SPEAKER_00: So this is another concerning thing that just supplements the story about uh the Anthropic CEO telling that open source is dangerous.
- 12:13SPEAKER_00: And then there's another thing uh regarding the Midas and Fable case.
- 12:18SPEAKER_00: If you, I'm sure you already know this uh in news that Chinese are also, government are also trying to figure out how they could uh limit the open source models that are getting like uh launched in the market, because America already pulled out their best frontier models from the market, and the China right now is also trying to limit the open source use, and it's just like kind of war.
- 12:47SPEAKER_00: You know, everyone's just pulling the blankets uh to their side, and it's um fascinating.
- 12:52SPEAKER_00: Also, for us as a European company, it's a very huge advantage because for the last four years, um our clients have compared us to ChGPT.
- 13:03SPEAKER_00: Like we're nowhere close to that, but I I mean thank you for for uh showing this uh this kind of yeah.
- 13:11SPEAKER_00: But uh for us right now it's a huge opportunity because we are a European company and we can provide this uh uh these factors that, for example, regulated or governments require.
- 13:21SPEAKER_02: Well, and I think you guys are a case study in positioning from a tactical level, right?
- 13:28SPEAKER_02: Because your software is architected in a way that I'm I'm making an assumption here, but just from hearing you speak, right, in some of our chat on LinkedIn, your architecture supports multiple models, right?
- 13:42SPEAKER_02: So you guys can so you guys can change, you can use the frontier models for things that that they're especially good at.
- 13:48SPEAKER_02: And you can pivot to any open source model.
- 13:51SPEAKER_02: The next best thing, thinking machine, just came out yesterday with an open source, open weights model that seems amazing.
- 13:57SPEAKER_02: And I haven't dug too deep into it, but like you guys could use that.
- 14:00SPEAKER_02: And that's that's why as soon as these open models came out, and you you could switch off of all of open AI stuff and start routing requests to these other models.
- 14:09SPEAKER_02: And what we found, and this is this was interesting, there was a there was a big headline about open router, and they have, I forget what it's called, but they have this way of like routing models to different for different tasks.
- 14:21SPEAKER_02: They help they and I haven't looked at the data myself, but they claimed to get as good or comparable results to Fable 5 by combining a bunch of different open source models.
- 14:30SPEAKER_02: So that's the other thing, is like the orchestration of models has turned out in the harness that the agent runs in, has actually turned out to be as important, if not more important, than the model itself.
- 14:41SPEAKER_02: Because you can fine-tune how you interact with the model to get the bet to even get better results, be more cost effective, right?
- 14:48SPEAKER_02: And so I think you guys are a case study.
- 14:51SPEAKER_02: Right now, your customers use your service for specific data manipulations, but if they're gonna build their own intelligence stack in-house that maybe does other things than what your service provides, they need to look at you as like a case study of how to do this properly.
- 15:06SPEAKER_02: Because you've essentially built those systems for a very specific use case the way that they should.
- 15:10SPEAKER_02: That's what I'm hearing anyway.
- 15:11SPEAKER_00: Yeah, that's that's true.
- 15:13SPEAKER_00: And also just to add to this that uh we actually just invested a pretty huge investment for us as a booster company in our own hardware.
- 15:22SPEAKER_00: So it means that we will be able to provide a specific piece, you know, a specific computing power to a specific enterprise if they don't want to share like the space.
- 15:31SPEAKER_00: So we will be able to host like our own models, open source models on our own hardware.
- 15:36SPEAKER_00: And that is a huge selling point because for the last years we've been highly focused on B2C and self-serve, you know, consumer, consumer product.
- 15:45SPEAKER_00: And our main like also the the customers came also from the US.
- 15:50SPEAKER_00: And right now we are like you noticed, we are shifting very hard and we are speaking with the enterprises and with governments.
- 15:57SPEAKER_00: And the things that you just outlined are the things that hurt for them the most.
- 16:02SPEAKER_00: And that is exactly what we're what we're solving right now.
- 16:05SPEAKER_02: Very cool.
- 16:06SPEAKER_02: Massive, important bet investing in your own hardware.
- 16:10SPEAKER_02: You guys clearly know the market and what's happening.
- 16:12SPEAKER_02: Talk about that.
- 16:13SPEAKER_02: Why buy hardware right now?
- 16:15SPEAKER_02: Why do that?
- 16:15SPEAKER_02: Because you just now like again, you guys aren't sitting on$100 million.
- 16:18SPEAKER_02: Like this is expensive.
- 16:19SPEAKER_02: Hardware is very expensive.
- 16:21SPEAKER_00: Yeah.
- 16:21SPEAKER_00: Well, the first thing is that we want to build an autonomous business.
- 16:27SPEAKER_00: And at some point when we started this, the API costs were like super, super small.
- 16:33SPEAKER_00: And we had like this huge margin.
- 16:36SPEAKER_00: But as the time moves on, we also want to provide a qualitative service to our customers, and the API costs just keep raising and it's just keep on eating on our margin.
- 16:47SPEAKER_00: So at some point, we just we just came together with my founder, we just we were just trying to figure out how do we increase our margin, how do we actually build a sustainable business where everything, all the entire infrastructure, we own the entire infrastructure.
- 17:03SPEAKER_00: And there were also at one point the thing that when we used the the external like API, like third-party providers, a lot of people were using these APIs, and so our customers weren't able to use our service because the data centers just couldn't uh you know process everything.
- 17:22SPEAKER_00: And that also impacted our business because we were able to provide the usage for our customers.
- 17:28SPEAKER_00: And and yeah, and then we just rounded the numbers, did a lot of spreadsheet work, calculated the costs, the ROI, when this could potentially pay off.
- 17:40SPEAKER_00: And then we made the decision.
- 17:42SPEAKER_00: Another factor here is that as we are a European uh company, if you know, we have a lot of these government uh funds and grants from the European Commission or local grants, and we some part of this infrastructure will be uh paid, not paid, but supported by the government.
- 18:02SPEAKER_00: So we are reducing the risk of our money because some part percentage will be funded by the by the government.
- 18:09SPEAKER_00: Oh nice.
- 18:10SPEAKER_00: So that's that's like VC, but only we don't give give away our equity, but we do have some kind of support.
- 18:17SPEAKER_00: Yeah.
- 18:17SPEAKER_02: That's great.
- 18:18SPEAKER_02: Yeah.
- 18:18SPEAKER_02: And I mean, so basically what you're talking about is like at some point your usage gets so high of these models that as a company it made sense to self-host the open source models.
- 18:27SPEAKER_02: And so you needed you needed hardware to do that.
- 18:30SPEAKER_02: And I do think that's a an analysis that every, especially like I would say medium to large enterprises, so you know, people doing like tens of millions to hundreds of millions to billions of dollars a year in revenue, they're gonna want to start looking at those investments in the near future.
- 18:45SPEAKER_02: And I say in the near future because I actually think that hardware and access to hardware is where like the dominoes start to fall over for like the collapse of this bubble we're all kind of in.
- 19:02SPEAKER_02: Right.
- 19:02SPEAKER_02: Everybody talked about like an AI bubble, and I and I think we're probably are in one.
- 19:06SPEAKER_02: The AI bubble, I don't think, is on the application end, like where you guys sit.
- 19:11SPEAKER_02: It's actually on the model and infrastructure side of things, where like the physical hardware, like where the bits meet the atoms, the metal in the data centers, because you've got trillions of dollars in data center commitments just from open AI alone, right?
- 19:22SPEAKER_02: And my theory is that at some point these open models become so useful and so effective and so cheap that it hits the revenue of some of these closed sourced companies pretty heavily, and then they can't meet their commitments for their infrastructure investments.
- 19:38SPEAKER_02: At the same time, between now and then, access to that hardware is very difficult to get.
- 19:44SPEAKER_02: We just had another data point from Anthropic where they took or said they were going to take at least temporary, fail temporarily Fable 5 out of their subscription offering and into only API usage.
- 19:56SPEAKER_02: So you have to pay per call.
- 19:58SPEAKER_02: And the reason they sit dated for doing that is hardware.
- 20:01SPEAKER_02: You can't access it.
- 20:02SPEAKER_02: Right.
- 20:03SPEAKER_02: And this is the other trend that I was talking about when I was like, this is a very smart move for you guys.
- 20:07SPEAKER_02: It's not just cost, it's long-term thinking about access to the compute and the energy.
- 20:12SPEAKER_02: And this is another thing people in Europe really need to start thinking about.
- 20:15SPEAKER_02: The government's there.
- 20:16SPEAKER_02: Like, there are there have been um concerted efforts that people are talking about of like sort of misinformation campaigns, disruption campaigns to like stop people from building data centers.
- 20:26SPEAKER_02: And on one hand, I get it.
- 20:28SPEAKER_02: And I don't know if those are true, if it's actually happening from China.
- 20:30SPEAKER_02: Like, I haven't, you know, this is like conspiracy theory stuff, but um, it's not impossible that it is happening.
- 20:36SPEAKER_02: Regardless of what you think, not having enough compute and and power to support intelligence inside of your like sovereign nation is a bad thing.
- 20:46SPEAKER_02: Like it's a disadvantage because I think I think open source superintelligence is coming, it's coming faster than you think, and the tide is shifting.
- 20:54SPEAKER_02: It's gonna be here regardless of what kind of restrictions you put.
- 20:58SPEAKER_02: You know, you to you talked about like pulling the blankets back.
- 21:00SPEAKER_02: You just can't stop it.
- 21:02SPEAKER_02: The the world will build open source software, people will adopt it, it'll get better and better.
- 21:07SPEAKER_02: And at some point, you're gonna need hardware to run it on.
- 21:10SPEAKER_02: The only thing that could stop that from happening, but it will still be the kick the dominoes, right, to cause this kind of crash moment is a different kind of AI that doesn't require so much hardware, like a different way to train and a different way to do inference.
- 21:23SPEAKER_02: And I do think that that is something that we will see.
- 21:26SPEAKER_02: Like I just logically, just that has to happen.
- 21:30SPEAKER_02: And and maybe from a computer scientist would tell me it's impossible, but I don't think it is.
- 21:34SPEAKER_02: Like there's a lot of research happening around that exactly.
- 21:38SPEAKER_02: Anyway, that's the other strategic part that I think is really smart.
- 21:41SPEAKER_02: If you own it, even even like some of the older models of GPUs are still really good with these open models, and they're still super expensive because you can't get them.
- 21:49SPEAKER_02: The supply chains are completely wrecked.
- 21:51SPEAKER_02: There's no capacity.
- 21:53SPEAKER_00: That's that's another bonus because we are located in this region.
- 21:57SPEAKER_00: Uh we are one of the first ones.
- 22:00SPEAKER_00: That actually order these specific GPUs.
- 22:03SPEAKER_00: So there's not a lot of competition in our region, but there are like quotas, how many GPUs, for example, you can get for each region for many media, et cetera.
- 22:12SPEAKER_00: And yeah, right now we are one of the first ones who have this kind of just like compute.
- 22:19SPEAKER_00: So that's a huge advantage because of our like regional uh where we are located.
- 22:25SPEAKER_00: But but yeah, I agree.
- 22:27SPEAKER_00: Uh the future, I think the models and everything that that will definitely uh keep on uh developing.
- 22:33SPEAKER_00: And regarding that, you mentioned about the the campaigns and the power that is needed.
- 22:39SPEAKER_00: Actually, we we just realized that in our region there are not a lot of data centers.
- 22:44SPEAKER_00: And actually, like 70% of the entire European like cloud infrastructure is owned by US, uh US like tech giants like Google Amazon.
- 22:55SPEAKER_00: Like we don't have like our own, there are there's like a super, super piece of pie of our own market, and that's also a huge problem.
- 23:04SPEAKER_00: And Europe has a lot to do here, and we're just falling behind in like every area.
- 23:10SPEAKER_02: Well, doesn't Europe have pretty good sustainable energy?
- 23:13SPEAKER_02: Or I guess they have a lot of gas there, like different countries like Germany relies on gas from Russia, which is a really bad thing.
- 23:19SPEAKER_02: But then I think they woke up to that when the pipe like the that got closed, right, during the beginning of the Ukraine war.
- 23:24SPEAKER_02: So they're making investments.
- 23:25SPEAKER_02: And I mean, I don't think it's impossible for them to to change, you know, in in in sort of catch up, you know.
- 23:32SPEAKER_00: But as you mentioned, there's a lot of propaganda regarding the wind generators, that they are bad for the environment and everything, and that is really slowing down because we're also trying to get rid of the the the Russia, like in terms of power and everything, and that just really doesn't help.
- 23:50SPEAKER_02: Yeah.
- 23:51SPEAKER_02: So here's here's a question I have for you.
- 23:53SPEAKER_02: You're on stage at a conference, I don't know, pick one, and you have to, you're gonna give like one take, and that take is probably gonna get you laughed off stage.
- 24:03SPEAKER_02: Like, what is that your prediction for AI that you think like other people would be like, you're crazy.
- 24:10SPEAKER_02: That's not, you know, that's that's nuts.
- 24:12SPEAKER_00: That's a tough question.
- 24:14SPEAKER_02: I know I just sprung it on you, but I want to see, I want to see what comes out.
- 24:18SPEAKER_00: I wanted to say that that everyone will have their own like server in in the room, you know, because the like AI will become like, I don't know, like spreadsheets or Excel, like calculate, like calculator.
- 24:32SPEAKER_00: And but it's already happening, so it's not something that people would laugh at laugh about me, right?
- 24:38SPEAKER_02: Well, I think I think you're right, it is already happening, but I think I don't think as many people realize it as you would say.
- 24:45SPEAKER_02: Like I was talking to my parents about this.
- 24:47SPEAKER_02: I was like, hey, you guys need to get a Mac Mini and or something, and you need to run some open source models on it.
- 24:54SPEAKER_02: And the I do think your answer is accurate for like a mainstream audience to my question in the sense that they may not laugh at you, but they might be like, you're crazy, I'll never do that.
- 25:05SPEAKER_02: But what happens is one day you can't get access to intelligence.
- 25:09SPEAKER_02: Like everyone, uh like a bunch of people have super intelligence, or a few companies or governments have superintelligence and you don't.
- 25:15SPEAKER_02: And then other people who saw it coming, I mean, I know we sound I sound like a crazy person, right?
- 25:19SPEAKER_02: Like I sound like a like a prepper, but it's true.
- 25:22SPEAKER_02: I'm gonna have a computer at my house that has a ton of GPUs in it, and it's gonna run like a trillion or whatever, build a huge, a huge open source model on it, and I'm gonna still have superintelligence, and it's gonna give me a massive edge against everybody else.
- 25:37SPEAKER_02: And I like I know there are a bunch of entrepreneurs, it's kind of starting small, right?
- 25:43SPEAKER_02: People like us, people in like Silicon Valley, these entrepreneurs that are kind of like more libertarian types, they're doing this.
- 25:50SPEAKER_02: They are looking at what's happening and they're like, I am setting up a small data center.
- 25:55SPEAKER_02: They're going more than just a computer.
- 25:56SPEAKER_02: They're like, I have a small data center with its own solar power, and they're running their own agents with their own models.
- 26:04SPEAKER_00: That's that's also one thing that we also discussed that at what point will we have our own data center?
- 26:10SPEAKER_00: Because when you buy hardware, you realize, wait a minute, I have to pay like thousands per month for data center, you know, everything.
- 26:18SPEAKER_00: And we're like, okay, so we will buy and and build our own data center at some point because it just makes sense.
- 26:25SPEAKER_00: And then there will be solar batteries that will power everything, and it just makes really sense.
- 26:32SPEAKER_02: Sounds like you guys could solve Europe's problem with not having enough data centers.
- 26:36SPEAKER_00: Yeah, also, also true.
- 26:38SPEAKER_02: Very cool.
- 26:39SPEAKER_02: I want to talk about governance and like guardrails.
- 26:44SPEAKER_02: This is something we chatted a little bit about on LinkedIn as well.
- 26:47SPEAKER_02: It's something I've been working on at my company, Corbett's, but you know, one of the things we've seen, this big headline, the most recent one, GPT 5.6, was just like deleting headline, deleting random files for people out of their projects without without being asked to, like just randomly.
- 27:04SPEAKER_02: What are some of the systems you've built inside your product to prevent it from doing that, to give you the visibility of when something's happening?
- 27:13SPEAKER_02: Like, how are you managing this kind of thing?
- 27:15SPEAKER_02: Because this is another big, especially with enterprise, right?
- 27:17SPEAKER_02: This is another big thing that they're looking for.
- 27:19SPEAKER_00: Yeah.
- 27:20SPEAKER_00: Well, the first thing is guardrails, but also in terms of what employees are submitting to the chat.
- 27:27SPEAKER_00: Is there like any sensitive information or thing or things like that?
- 27:31SPEAKER_00: So we we are right now, all of this that I will right now explain is all in development as we are speaking with these companies, but the things that we're seeing is also the problem with the credentials and with management and with audit trails.
- 27:48SPEAKER_00: And every company that we speak to, and we explain that there will be audit trails, detailed usage, uh like uh analysis where you can understand who used what at what point, uh, how much each project, for example, costed, or how how much each employee uh spent.
- 28:09SPEAKER_00: And the same applies to audit trails where you can see who accessed what at what point and and all of that.
- 28:17SPEAKER_00: Regarding the things that you mentioned, I actually need to check with my co-founder, who's also the CDO, because I'm not sure about the fact how we will control the AI from not deleting all of the files.
- 28:30SPEAKER_00: So that's one thing that I should check.
- 28:33SPEAKER_02: My team has a solution for you.
- 28:35SPEAKER_02: We built an open source agentic operating system, this platform called Corbettz.
- 28:39SPEAKER_02: And um you can think of Corbutz as like Red Hat.
- 28:42SPEAKER_02: So Red Hat's like an open source software company.
- 28:44SPEAKER_02: They all they did was offer stable versions of Linux for people.
- 28:50SPEAKER_02: And you could just pay for pay them to manage it for you, or you could just take Linux and run it yourself, just like we talked about with the open source models.
- 28:57SPEAKER_02: Interchange is the operating system we built, so that's like Linux itself.
- 29:01SPEAKER_02: And the way we prevent actions from agents is at the runtime.
- 29:06SPEAKER_02: So it's like at the system level.
- 29:07SPEAKER_02: So we can block any tool call.
- 29:09SPEAKER_02: Whereas like what happened with GPT 5.6, probably, and I haven't done the I haven't like looked at the forensics, but this has just sort of been like a repeated thing that's happened with all these models, is they have system prompts that say things like, don't ever delete production code, you know, or like customer data.
- 29:27SPEAKER_02: But the thing with the system prompt that is not good is it's it the analogy I use is like, I have three kids, okay?
- 29:34SPEAKER_02: Eight, five, and three, and they're the best.
- 29:36SPEAKER_02: They're awesome.
- 29:37SPEAKER_02: They're high energy, they're a lot, you know.
- 29:39SPEAKER_02: And you and I will be having a conversation, and then all three of my kids are like over here yelling, dad, dad, dad, dad, dad.
- 29:45SPEAKER_02: And I just forget everything you're saying to me.
- 29:47SPEAKER_02: Like I won't be able to keep track.
- 29:49SPEAKER_02: That is what happens to agents, right?
- 29:50SPEAKER_02: Their context gets completely blown out, and they forget that one line in their system prompt that said, do no harm.
- 29:56SPEAKER_02: And then they do a lot of damage, and it happens like at the speed of light.
- 29:59SPEAKER_02: If you block it at the runtime, then they can't do it.
- 30:02SPEAKER_02: They try to make the call, and because you have audits in, you can see them making that call and you know they're messing up, but no damage was done.
- 30:09SPEAKER_02: And why it's important that you know they're messing up is because then you can train it to get better to avoid wasting those calls.
- 30:15SPEAKER_02: You know, because every time it makes a tool call, it's using inference and it's costing you money.
- 30:19SPEAKER_02: And so you can kind of like make it more effective that way.
- 30:21SPEAKER_02: But that's that's our our solution to that, is how we architected it.
- 30:25SPEAKER_02: We're getting to the top of the show, and I always ask this at the end of the show, what have I not asked you that I should have asked, or that you wanted to talk about?
- 30:33SPEAKER_00: I knew I should have prepared an answer to this because last week I was uh on another podcast, and they asked me the same exact question.
- 30:41SPEAKER_00: And at that point I was you know blindsided.
- 30:44SPEAKER_02: It's a hard one, it's my favorite one.
- 30:46SPEAKER_00: Oh, okay.
- 30:47SPEAKER_00: I had actually I had prepared notes for this podcast because there were like things that I wanted really to speak about or explain, but now I need to go through all of it.
- 30:58SPEAKER_02: Just look at them.
- 30:59SPEAKER_02: You got time.
- 31:00SPEAKER_00: Yeah.
- 31:00SPEAKER_02: We're not in a rush.
- 31:02SPEAKER_00: Okay.
- 31:02SPEAKER_00: One of the topics that also came to my mind was also about shadow AI, and I actually did some research on uh that you were you were also speaking about these topics on your own like uh podcast or YouTube other channel.
- 31:17SPEAKER_00: I I noticed that you spoke about this, and uh yeah, maybe we could talk about this because for the what is shadow AI?
- 31:26SPEAKER_02: Sounds scary.
- 31:27SPEAKER_02: I'm terrified.
- 31:29SPEAKER_00: Really, you are so I should explain to you.
- 31:32SPEAKER_00: Yes, because that is like one of the things that we're also seeing like in our like users uh that are like from our database, and also when we speak to the companies, they just don't invest in AI because of all the data concerns, and there's no like solid alternative.
- 31:49SPEAKER_00: But at the same point, their employees are just chatting with you know, ChatGPT, Cloud, sending all the government data into the chats, right?
- 31:57SPEAKER_00: And doing everything that they want to do.
- 31:59SPEAKER_00: And I think there were there was also a research, I think, by Mackenzie that said that around 79% of all enterprise employees use Shadow AI.
- 32:10SPEAKER_00: And I just think it's it's it's such an insane number.
- 32:15SPEAKER_00: And there's definitely, I'm sure there are already solutions to that, but I think there's a huge market potential.
- 32:22SPEAKER_00: And yeah, I would love to actually hear your thoughts on that on this and how do you, if you solve it.
- 32:28SPEAKER_02: Yeah, I mean, I think so.
- 32:30SPEAKER_02: Shadow, shadow IT is a term from decades old, right?
- 32:34SPEAKER_02: And shadow AI is just an extension of that.
- 32:36SPEAKER_02: Um, but it's basically all the things that are running that your team is using to run the business that you don't even know about.
- 32:43SPEAKER_02: And your example of this is something I know happens.
- 32:46SPEAKER_02: Everybody is guilty of this since the day ChatGPT came out, people started using it for work because the productivity gains are just too big.
- 32:54SPEAKER_02: People want the productivity gains.
- 32:55SPEAKER_02: They'd rather spend time with their family and having coffee with their friends at work than just like doing tedious stuff.
- 33:01SPEAKER_02: Like all the drudgery kind of work that we do in spreadsheets where you guys started to manipulate data, all that stuff goes away.
- 33:08SPEAKER_02: But if you haven't invested in the infrastructure to support AI at scale, and I say infrastructure because we're not talking about personal agents.
- 33:17SPEAKER_02: We're talking about infrastructure that runs in a cloud, whether it's yours, on-prem, AWS, GCP, whatever, if you haven't invested in that infrastructure, your team's just doing it anyway.
- 33:27SPEAKER_02: They're not just not doing it.
- 33:28SPEAKER_02: They're not using, you know, they're gonna do it anyway, and you really have no way of preventing it.
- 33:34SPEAKER_02: I mean, you can put policies in place and tell them no, but they're still they're still gonna happen.
- 33:38SPEAKER_02: Still gonna happen.
- 33:38SPEAKER_02: So the best thing you can do is invest in the right infrastructure, is put those that infrastructure in place with the guardrails, permissions, credentialing, all the systems that unify all the best practices.
- 33:50SPEAKER_02: So the people that are actually using it are getting good at it too.
- 33:53SPEAKER_02: And so if you invest in them and give them the tools and let them use them, then everyone else gets better at it too.
- 33:58SPEAKER_02: Especially if you use open source, because open source means you can you can improve your system over time.
- 34:04SPEAKER_02: You have control over that.
- 34:06SPEAKER_02: And so it's compounding growth on your intelligence stack over time.
- 34:11SPEAKER_02: And and then you'll be able to see everything.
- 34:13SPEAKER_02: Now, yes, we do have some tools that allow you to sort of audit what's happening in your company, but that's only as good as the things that we're able to connect those tools to, right?
- 34:25SPEAKER_02: Because like if you're if you're using OpenAI or Claude in your company, you still probably don't have the visibility on all the tool calls.
- 34:32SPEAKER_02: You don't have audit logs of like the actions that they're taking.
- 34:35SPEAKER_02: So I'll g I'll give you an example, I'll give you an example.
- 34:38SPEAKER_02: Claude code deleted the customer production database for this little SaaS company that did like car rental SaaS.
- 34:46SPEAKER_02: This was on Twitter about three months ago.
- 34:48SPEAKER_02: And they didn't have a backup of their database for like three months.
- 34:52SPEAKER_02: It was like 90 days old.
- 34:54SPEAKER_02: And they were screwed, right?
- 34:56SPEAKER_02: Like it was a really bad day for that founder.
- 34:58SPEAKER_02: And so he asked the AI, like, dude, what did you do?
- 35:02SPEAKER_02: And it was like, oh, you're right, I made this mistake.
- 35:05SPEAKER_02: I thought that I was thought this was the solution.
- 35:06SPEAKER_02: I hit the delete button.
- 35:08SPEAKER_02: Well, the thing that's happening there is he like, you're sending a prompt to your AI asking it what it did.
- 35:12SPEAKER_02: It's inferring the answer.
- 35:14SPEAKER_02: It's not actually showing you a log of what it did.
- 35:17SPEAKER_02: It could be lying, it could be just straight up hallucinating.
- 35:20SPEAKER_02: They did a forensic audit and it turns out it did actually do those things, but that's because they like actually were able to go look at the database and see that that API call happened.
- 35:29SPEAKER_02: You still don't have all the reasoning, all the inference calls, all the tool calls that it attempted to get to that point where it made the mistake.
- 35:36SPEAKER_02: The other half of that coin was like the database provider really had a bad architecture that allowed the delete button to be pressed by an agent via API.
- 35:45SPEAKER_02: Like that should never be allowed.
- 35:46SPEAKER_02: But the agent also shouldn't have had access to that tool at the runtime.
- 35:50SPEAKER_02: So, anyway, my long, long-winded point here is you got to build the infrastructure.
- 35:55SPEAKER_02: You gotta make the investment.
- 35:56SPEAKER_02: Otherwise, Shadow AI will take over everything and your data is just gone.
- 36:01SPEAKER_02: It's just out there in the wild for anybody, and it's like publicly, I mean it's essentially publicly available, right?
- 36:06SPEAKER_02: Like their servers could get hacked, like wherever it's gone, whatever.
- 36:09SPEAKER_02: You might as well consider it completely compromised at that point, in my opinion.
- 36:13SPEAKER_02: So that's what I that's all I got on that.
- 36:17SPEAKER_00: Yeah, I 100% agree on this.
- 36:19SPEAKER_00: That's that's exactly why I think you and and I and uh our company are also uh solving this because it's a super, super huge problem, and and there needs to be a solution for that.
- 36:33SPEAKER_02: So what's what's next for AGLIX?
- 36:34SPEAKER_02: Like where are you guys going and how can people like get in touch with you and learn more?
- 36:39SPEAKER_00: Okay, so I think I already spoke about the part that where we are going because like infrastructure, we still need to start working on it because we just today received a message that the entire server servers are complex uh complicated and we need to insert them in the data centers, and we're also right now trying to figure out the model training, and so that's also another thing that we will definitely approach.
- 37:06SPEAKER_00: But yeah, we need scientists and we need universities, and we're already speaking because we can't do this like bootstrapped.
- 37:12SPEAKER_00: We need like uh a lot of a lot of people around us.
- 37:16SPEAKER_00: That's one thing, and uh and yeah, we will still go uh try to understand our niche and not niche but our market for this specific uh solution that we're we're uh uh right now developing.
- 37:29SPEAKER_00: And where can we contact me?
- 37:32SPEAKER_00: Definitely on LinkedIn because that's the place where I am like every day I'm posting.
- 37:36SPEAKER_02: You go back to me right away.
- 37:38SPEAKER_00: Yeah, yes, because I'm I'm there, you know.
- 37:40SPEAKER_00: Yeah, I'm living on LinkedIn, so uh I do post a lot every day almost, and and I answer to almost every DM that doesn't start with something AI generated.
- 37:52SPEAKER_00: So definitely give me give me a message and I will answer to it.
- 37:56SPEAKER_02: Awesome.
- 37:56SPEAKER_02: Well, Agnes, thanks so much for being here.
- 37:58SPEAKER_02: I appreciate you.
- 37:59SPEAKER_00: Thank you.
- 37:59SPEAKER_00: It was a lovely conversation.