Inside a Frontier AI Lab
From First Check to Partnerships at Poolside

Introduction

Andrea [00:00]

Here, shipping something earlier at 90% is better than shipping it next week flawless.

Ceci [00:07]

We're seeing more and more people in this business operations title now. Do you think it's here to stay, or is it just a fad?

Andrea [00:13]

Whatever gets put in front of you, even if you have no clue, finding ways how to solve it and navigate the org to do that.

Bea [00:20]

When you put a constraint on something, people genuinely try to innovate.

Andrea [00:25]

Don't wait for the perfect moment. When your shoe cleaner tells you to buy a stock, it's too late. Same thing here. If everyone is talking about a company, it's probably too late.

Bea [00:41]

Welcome to Top of the Ops, the podcast where we have real conversations about what happens behind the scenes of a company. I'm Bea, ex VC at Lakestar and now Chief of Staff at Praktika.

Ceci [00:52]

And I'm Cecilia, ex VC at Talis Capital and now Financial Director at PortalOne. Most of our guests own operations at one company, but today's guest has spent the last year and a half operating at two.

Bea [01:05]

Andrea Badia is Director of Partnerships and Business Operations at Poolside, the AI lab building frontier models for software engineering. She did investment banking at JP Morgan until, may I say, I personally handpicked her to join Lakestar, where we worked together for two years. She was actually on the team that put money into Poolside and then she went and joined them. Since January 2025, she's been running capital markets, helping incubate a two gigawatt data centre company in West Texas, and, as of a couple of months ago, building their model partnership functions from scratch.

Ceci [01:42]

Welcome Andrea.

Andrea [01:42]

Thank you for having me.

Why She Invested in Poolside, Then Joined

Bea [01:43]

Of course. So at Lakestar you worked on both sides, Series B, and then you went and joined them. Take us to the moment you decided that. What did you see from the investor side that made you want to join them?

Andrea [01:56]

I think this deserves breaking it down into two parts. Why did I decide to invest, and then why did I decide to join them?

To take you back to the context, I was in VC in early 22, so right after the ChatGPT moment and right around the time that Anthropic was founded. It was the hot category. Bea, you will remember, but we had looked at every AI lab.

By then I received a tip from a friend who said there is another AI lab raising 500 million at a three billion valuation. I assumed this was just another one of them, but decided to look into it. It was quite a secretive company. There was nothing online, but something stood out. One of the founders, Jason Warner, was the ex-CTO of GitHub. I sent him a cold LinkedIn message and he replied in just a few hours. That is one of the first reasons why I decided to invest and then join. It's the people. It's very high density of talent, but also very kind-hearted and low ego. The second one is the way they work. I remember I was expecting a 10 page deck and I received a hundred page FAQ with every single one of my questions answered. That just shows how they work and their level of transparency. Obviously there were many other signs that an investor looks for, like they had the compute, they had the go to market, they had a distribution plan, they had a differentiated thesis focusing on code instead of going general purpose since day one. And then the icing on the cake was that they just killed the IC. We broke every internal rule, ownership threshold, revenue threshold, to write that check. It was pre PMF, pre revenue, pre everything, and we got full votes in the room. So that's why I decided to invest.

And then why did I decide to join? A few months after the investment, I had breakfast with the COO and we talked about Poolside, and I explained that I wanted to make the move to operator. At that moment I had an offer in hand. I had one of the hottest growth stage startups in Europe, repeat founders, in the AI application layer, 50 million ARR, 3x year on year, tier one VC, everything. Great offer on paper. And she kind of teased me and said, great role, right time, but why don't you join Poolside? My first reaction was, I don't have the background. I'm not ready for an AI lab. I'm not technical. However, I decided to do the process, and it was exactly the process that convinced me to join them.

I spent a couple of weeks working very hard on the case study and the interviews and I really, really enjoyed the process, and realised that you don't need to know everything on day one, but I could really learn what it took to be at an AI lab. And the decision moment was in Lisbon for Web Summit. I had breakfast with Eiso, one of the co-founders and today co-CEO. I had memorised every single line from his 20VC episode because I was so nervous, but I never needed it. It was a super laid back conversation until the last 30 seconds, where he said, okay, it's between two people, it's going to be the COO's call, but tell me, why do you think you're not the right person? There are many easy ways to get out of that question, but I decided to lay the cards on the table and I said, it's a technical gap. I will definitely have to step up and learn the intricacies of model building and an AI lab. I have no clue. But on the other hand, I pitched myself hard and I said, I'm a fast learner. I'm really hardworking. I can work under pressure. I can switch contexts fast and I want to get my hands dirty. So if you let me, you will not regret it. And I guess I convinced him, because a few days later I got the offer.

Bea [05:20]

Seems to be a repeat, because Andrea and I grabbed an Aperol Spritz in Canary Wharf while she was at JP Morgan. And she also had another process going on with another VC. And I said the same. I said, why don't you join Lakestar? So I guess that's like a repeat story going on in your life.

Why Leave VC and What She Looked For

Ceci [05:36]

It's such a beautiful story of being really self-aware of what your gaps were and not being afraid of showing your vulnerabilities. Why don't you take us back? Getting a role in VC isn't easy and it's quite a coveted role. But you said that becoming an operator had been in your mind for a long time. Why did you want to leave VC? Why operating? And what were you looking for specifically? And that clearly wasn't that 50 million ARR growth-stage startup.

Andrea [06:07]

Yeah, definitely not. My obsession from day one since I was very young was actually to become a founder and have my own business. And I just kind of drafted a plan on how to get there. And I thought, I'll go through investment banking, venture capital, I'll join a startup, and then one day I will build my own. And this was all thanks to my older sister, who is today a big boss at GS. She gave me the advice after uni, you're going to have time to start your own business whenever you want, just get the experience first at a recognised corporation and get the stamp on your CV. So that's what I did. And then when the right time came, I decided to start looking for startups to join.

And my checklist was quite simple. It was repeat founders, so people that have done it before. High talent density across every department, not only leadership. I wanted a global mindset, so not only Spain or Europe or US, but global since day one. A company solving hard problems. This is something I didn't know at the beginning and I realised a bit later, but I remember the founders at Poolside wrote that there are three problems humanity will climb in this century. It's space, energy, AGI. I joined Poolside to solve one of them. And on the way we decided to solve another one, but I don't think we're getting into space anytime soon. And the last one on my list was grit and determination. I wanted to be part of a team that works hard for a big goal and doesn't quit when things get hard. And this goes hand in hand with solving hard problems. I wanted to be part of a team with a winning mentality that doesn't mind being the underdog.

What BizOps Means at an AI Lab

Bea [07:42]

And so you joined in the BizOps team reporting directly to Margarida, the COO. What does BizOps mean at an AI lab? I imagine that BizOps at a SaaS company and BizOps at an AI frontier lab is completely different.

Andrea [07:56]

When I took this job, I didn't even know what BizOps meant, or the difference between BizOps, chief of staff, founder's associate. However, very clear to me from the beginning was I didn't really care about title. I mostly care about who I was working for, and in this case, I joined to support the COO and it was someone that I really admired and looked up to. At the same time, I wanted exposure to founders, which was something that was assured to me since day one.

And so on day one I didn't really know what I signed up for. The first three months was just shadowing, and when I mean shadowing I mean physically shadowing. I was in Madrid one week a month sitting next to her in her home, sitting next to every single conversation, and every other week in Paris together with the whole team. So two out of four weeks a month for sure together, and then there was always another week that we would meet for conferences or events. So the level of transparency and exposure I got was something that I'm extremely grateful for. There were no secrets or no conversation that was too hard or too senior for me to listen to. And that was extremely different from what I had experienced in banking and VC.

So I shadowed everything. Partnerships calls, leadership strategy, compute, investor calls, org challenges, legal reviews. And I realised how real context switching is. It was very different conversations in a single day. There was no job description and I was basically doing everything that was needed and more. And to be fair, many things I didn't know how to do, like my first token pricing exercise, or navigating the AWS cloud business that I had no clue about, or analysing different silicon options.

So the most important quality I would say in this type of role is whatever gets put in front of you, even if you have no clue, finding ways how to solve it and navigate the org to do that. Over time, obviously, it has shaped into a different role. Then I realised what I enjoy doing and what I don't enjoy doing, but at the same time, what I'm good at and what I'm not. And it doesn't mean that now I'm only doing what I enjoy or what I'm good at. Don't get me wrong, I still get to do a lot of legal stuff that is not my favourite part of the job. But I had an opportunity to shape my role a bit more and double down on the things that I'm good at, like investor relations and capital markets.

Shadowing, and Learning Compute From Zero

Ceci [10:06]

So shadowing is such an important part of starting any ops or Chief of Staff role. Knowing what you know now and having done it the way you've done it, what would you tell anyone starting this role now about shadowing? How do you make the most out of that one or three months that you can actually shadow your principals?

Andrea [10:26]

I would say there is no homework that you can do to be ready for this. I remember when interns are joining investment banks or venture capital and they ask you the summer before, what can I do to prep? And you tell them, just enjoy. There's really nothing that can prepare you for this. If you are in these shoes, you are plugged into the ecosystem, you're listening to podcasts, you're getting up to speed on the AI ecosystem, but really there's nothing you can do to get up to speed. So I would say don't go with any pre-settled ideas, try to keep an open mind and take every call and every task at the same level of importance. It's very easy at the beginning to say this call is not so important because we're talking about people organisation, and in this other one we were talking about a big compute deal or a big partnerships deal. Go into every call with the same mentality of, this is as important, and trying to learn and take the learnings into the next one.

Bea [11:18]

You also mentioned jumping into something that you know nothing about, like negotiating compute, and just very, very hard problems that you haven't been exposed to, neither in banking nor VC. Do you have a mental framework or support around you when you say, hey, now I have to deal with AWS, and I know nothing about that. Where do you start? Do you have any process for that?

Andrea [11:40]

I would say that it took me a few months of just shadowing and smaller tasks before I was thrown into the lions of having to do full tasks end to end on my own with a high level of ownership or responsibility. So I was able to get up to speed in all of the big buckets before getting there. I don't think I have a standard mental framework for tackling this type of task, but it is true that I learned this thing in banking of never ask the same question twice, and I'm trying to apply that. During the first months of the job you can ask all the questions you want, so I force myself to ask questions. I remember in the first months in Paris I would sit down with everyone to ask every question I had during the calls in the day, and I would write them down so that I wouldn't ask them again in a month from now. So this way I was building my library of topics that I learned, and then trying to figure things out, but also putting my hand up when it's something that I clearly don't know and I shouldn't spend too much time trying to figure it out if it's not efficient.

What She Had to Unlearn

Bea [12:42]

And was there anything from banking or VC that you had to unlearn when you work at Poolside?

Andrea [12:48]

Yes, the zero error standard from banking, so the fact that you don't ship things until they're perfect. Even before sending a deck, even if it's 3am, you will read every footnote, check every decimal. Here, shipping something earlier at 90% is better than shipping it next week flawless, because decisions will be made without you otherwise.

And from VC, I think it's about the optionality, or keeping all doors open. You don't want to make mistakes and you don't want to commit to a solution, so you keep many doors open and you have multiple bets. Here it's about making one bet with full conviction, going blindly at it, and then being able to adjust quickly. So if you make mistakes, you determine it and change and pivot quickly. Running fast and not being scared of making mistakes once you've done it.

Two Jobs and a Data Centre

Ceci [13:40]

I love that. I love that a lot. Let's go back to your beginning at Poolside. Back then you held two jobs for about a year. Can you walk us through how that worked and what a typical week looked like for you, high level?

Andrea [13:56]

It was around March 2025 when Eiso came up with the insane idea at the time of, should we build our own data centre. At that moment, we were only six people on this special project. And so we had to divide and conquer a little bit. The first thing we did is we had to look at every land option that we had on the table and run it side by side with every silicon option. So it was like a puzzle, trying to match and see if it made sense and which one was the most viable combination, if any, for us to own a data centre. At that moment, it was just leadership and then my colleague and I at a similar level, who were tasked with trying to analyse the viability of this project.

And so the first few months were just building an Excel to analyse if this was something that made sense or not. However, neither of us two were ever in infrastructure or the data centre space. So the first person we called was actually this investor from an infrastructure fund who had been in multiple data centre Excel models, to walk us through how that worked. So for the first two months since we started the idea of what today is Project Horizon, we were focusing 90% of our time on building this model and trying to assess the viability. So I kind of parked a little bit the work I was doing on the AI business.

Once a decision was made, my colleague and I started splitting the time, because we didn't know so much the likelihood of this project going ahead at 100%. The founders decided that it was not worth having everyone working on just this project alone. So my colleague was taking care of the day to day operations of building the data centre, so negotiating the land lease, securing the long lead equipment items, securing all of the suppliers, negotiating the silicon options and so on. And I was mostly focused on the fundraising and investor relations for both companies as a whole, because back then it was just one idea.

And so my involvement in the infrastructure business has come and gone in waves. Further on we decided to do a full spin-off of the business and fundraise for both companies separately. So then I was tasked to build all of the fundraising materials for the company as a standalone. So then suddenly I had to step on and spend more time with all the people that we had hired for that business alone, and understand the business and be able to pitch it to investors. A few months later, I went back fully to the AI business and that's how the partnerships motion came around. It really went and came back in waves, depending on the needs of the business for each one of the two businesses. But I would say today I'm mostly in the AI business and I'm picking up the infrastructure business, waiting for it to close the financing of the first phase and for me to close the loop.

Who Buys the AI Business, Who Buys the Infra

Bea [16:49]

That's peak context switch. Not only are you context switching for one team and one function, but you're also context switching for two different companies. So kudos for that. You ended up pitching both companies. Could you explain to us who buys the AI business and who buys the infrastructure one? Do they want to hear different things?

Andrea [17:10]

Completely different risk and return profiles. So the AI business is venture and growth capital, as you might be familiar with. They basically underwrite the upside and they tolerate ambiguity, mostly because they have seen it before. We are mostly looking for investors who have invested in other AI labs, so our competitors, and are familiar with the fact that it's very hard to make revenues until you get to the frontier. So they need to understand that you have to underwrite a company's ability to get to the frontier before it does, which most likely means pre-revenue. And they know the ingredients it takes to get there. It's the right team, resilience, compute, making the right bets. That's a perfect fit for Poolside and the type of investor that is interested in that business.

For the infra side, it's infra private equity. It's pure project financing. They care about having a lease signed and the quality of your tenant, so they can then derive the cost of capital. They want to be completely downside protected by the real assets, so the land and power that you own. So it's very tangible. And they will stress test every assumption, so the spreadsheet has to work to the decimal. And this, in all honesty, at the beginning it was quite confusing and not very rewarding, because I felt like I was not good at either. But eventually it taught me real discipline and I realised that you cannot pitch an infra investor with a vision slide and you cannot pitch a growth investor with a DSCR table. So I learned how to speak both languages fluently.

Staying Organised Across Two Companies

Ceci [18:36]

In the answer before, you were talking about moving from one company to the other and all of the work that you were doing. I wonder whether you have a system or a mental framework or something that enables you to be organised within the chaos of not just one startup, but two startups in two completely different business models. How do you deal with that?

Andrea [18:59]

To be fair, it wasn't until January 2026 that we decided to fully spin off the business into its standalone company called PIC, Poolside Infrastructure Company. So before that, in my mind, it was all one business and it was part of the page of, okay, we want to be fully vertically integrated from dirt to intelligence. This is just another part of the stack.

From January 2026, when I was told that we are doing a full spin-off of the business, that's when I started thinking about, okay, now these are two different businesses. We did a full system separation. So we created different Slacks, different CRM, different everything. We hired a fully independent team for that business and so on. And then it became naturally a bit easier. I had two emails suddenly for a while. I have my PIC email, I have my Poolside email. So I had to adapt a little bit. And similarly, I have two calendars. So for a while it took me a little bit to adapt to this, but the reality is in April, when I started off the partnerships role, I started winding down on the infrastructure business and today it is very minimal. I'm only three hours a week. But it definitely got easier because my level of involvement has gone down. It was a bit chaotic at the beginning, I will not lie.

Bea [20:13]

Yeah, I also have two email addresses. I have one email address for myself and one email address, I don't know if I should say this out loud, but as a B2B sales person, my name is Sarah Johnson. And so this is literally two different hats. This is also a major context switch for me. But the things you do for your job, you know?

Ceci [20:33]

It's an identity switch.

Bea [20:36]

Full identity, but it's because I conferred with Claude and we concluded that Sarah Johnson was the least threatening name for a B2B salesperson. So I went with that.

Back to Andrea. A couple of months ago you moved into model partnerships alongside the capital markets role. Very interesting move, because you spent almost two years behind the scenes and now suddenly you're completely on the distribution side, very exposed to the other side of the game. What changed and what pulled you in that direction?

Moving Into Model Partnerships

Andrea [21:13]

So on April 28th, we released the first models. We had been in the shadow for a bit over three years and exactly that date we decided to come out in the world and release our first models. And this changed the job to be done at Poolside. Before our first model launch, the big constraint was capital and compute. Post launch, it was distribution and mind share.

Our models were suddenly live on OpenRouter. We were getting all types of traffic from many platforms and nobody was properly owning that surface or trying to distinguish signal from noise. We have some superstar engineers on our team who took care of having us present in those platforms on day one, but that was really not their job and that was beyond the capabilities of the job that was needed. And leadership quickly realised that they needed someone to own those relationships going forward, not only on launch date, but really between model launches and forward. And so I realised that gap internally and I basically volunteered myself to take care of it.

I pitched it to my boss, the COO, and I said, hey, I've noticed this gap. I think I would be a great fit for this role. Why don't you let me give it a try? And I was actually in a trial period for a month, taking care of it without it being actually my role, and people were confused because they were thinking of me as investor lady, data centre lady, and then now suddenly also partnerships. But I'm very happy I did. I think I ended up doing an okay job because I was actually officially given the job internally, and it's been great so far.

Building the Function From a Blank Spreadsheet

Ceci [22:49]

That really shows how varied you can be. And it looks like you built this function from the ground up. Can you walk us through that? How did you approach it? What was day one of, I do model partnerships now? Talk us through the first few things that you did and tackled in your mental model.

Andrea [23:10]

So the first thing I did was I created a spreadsheet and I tried to map the ecosystem. I started talking with engineers at Poolside and I asked, hey, which apps or agents or gateways do you use on your day to day? Because I was actually so far removed from the job itself. I knew the obvious ones, but I was missing so many. So I tried to create a comprehensive list of the different platforms that matter for us and divide them into categories.

So there were the inference providers, like Baseten is our sole provider today. There were the gateways like OpenRouter, but also Vercel. There are also the agent harnesses like Cline, Kilo Code, OpenCode. Then there are also, somewhere close to this, the benchmarks and evals providers, the third party validation that people care about, like Artificial Analysis, LM Arena, evals. Then there are also fine-tuning platforms like Prime Intellect. And so, similarly to what you would do in VC, I tried to create a market map of the ecosystem and then a comprehensive list of every player in it. This was step number one.

Step number two was prioritising them. So I tried to look at ClickHouse, which is our internal database. I remembered I did an SQL course at uni, but now thanks to Claude, I don't need to actually write the queries myself. And I started querying the data to understand, okay, where are tokens flowing today and which surfaces matter more than others, and which other platforms are we not present in that we should be present. So I prioritised that list into tier one, two and three, similarly to what we would do in VC.

And then step number three was actually instrumenting this motion. I'm a big fan of Attio. I created our first investor CRM in my first month at Poolside. And so I took that and I replicated it for partnerships. So now we have a well-functioning CRM with every single partner. I convinced everyone in the org to start using it. I created a weekly partnerships meeting. We have a partnerships Slack channel where I share a summary of every note and I give official updates in every all hands. So now it's a bit more standardised.

What a Frontier Lab Actually Feels Like

Bea [25:12]

And actually VC is helpful, right? Because you go through the same exercise. You look at thesis formation, look at all the segments, try to break it down and make informed decisions on each of them. So I would say the VC training was pretty spot on in this case. I have a very general question: how is it to work at a frontier AI lab? I suppose the pace is intense, everybody's super smart, you're very hardworking. But for everybody that doesn't know how it is to operate at a frontier lab, how is it?

Andrea [25:52]

So the funny thing is I don't know how it is anywhere else, because this is the first startup I have ever joined. So many times I stop and I ask the veterans at my company, hey, is this normal or is this because we're at an AI lab? So there are many things that are not normal and are not the standard in other startups. The flywheel is very different. You need to prove model capabilities, so you need to get to the frontier before you can make revenue, and you need the revenue to get investment from VCs that gets you the compute that then gets you the model capabilities. So it is a crazy flywheel and things are happening sometimes in a weird order, and you need to just be comfortable with it and the high level of uncertainty.

There is a big dependency on the research team. At the end of the day, we're all dependent on the work that they're doing. Sales cannot sell if we don't have a good enough model and so on and so forth. So as we say at Poolside, and our founders have said this many times, it all collapses into the model and the model is the most important thing. Until we get to the frontier, everything is going to be a bit confusing and it's going to feel in a bit of an unnatural order, but everything will make sense once we get there and then we will become more of a normal company.

Bea [27:14]

And I have one curiosity. Do you think the ability to be comfortable with uncertainty was something that you had intrinsically, is it part of your personality, or you had to learn it through the years?

Andrea [27:26]

I definitely had to learn. My boss always says, I need to get the banker out of you, in the best way, because I want to have everything under control. And I didn't deal well with high levels of uncertainty at the beginning, definitely. And I also had to be reminded constantly, hey, you're not a founder. Yes, you can be involved, and I'm super dedicated and I bleed the colours of my company and I want us to win. But also, I should be able to sleep at night, because at the end of the day it is not my company. So there is a fine line and I had to learn how to deal with high levels of uncertainty. At the end of the day, this is a rollercoaster and you have to be fine with it.

Talent Density, Culture and the Paris Office

Ceci [28:06]

It's the healthy level of detachment. You talked earlier about talent density as one of the most important criteria for you when you were looking about where to move next. And you also mentioned the team and the research team that you have at Poolside. Is the talent density still there? What do you do to maintain it as you're scaling as a company? How would you define that talent density? How would you define the culture at Poolside?

Andrea [28:34]

In order to compare apples to apples, I think you need to compare companies in the same industry. So when I think about Poolside, I'm comparing it with other AI labs and where they are based and their talent. You have seen the crazy talent wars that are going on in the West Coast. They're hiring researchers for more money than they're hiring footballers these days. So it is very competitive. There is no loyalty. And the truth is we cannot compete at those price points, although we are American and we do pay good salaries.

So the founders took an approach since day one, which is we're not going to hire in the West Coast. So they built a global team since day one. They hired people in the East Coast, in Canada, a lot of people from Asia, scattered around Europe. And because we have one founder from Canada and the other one from Europe, they were able to understand the different countries in Europe and around the world and find the right talent in each one of them. And the reality is, by assembling these types of people, we managed to get one of the lowest churn rates, if you may, of the industry. So we have actually very few people that have left Poolside, and most of the ones that have left have actually not gone to competitive labs.

So they want to be part of a place where they can make a difference. They could have a very shiny salary and stock option plan at other labs, but maybe they would not have the agency to drive the experiments that they wanted as they do at Poolside in other places. And so we made that choice very deliberately and I think it has paid off. And at the same time, when the team is hiring, Eiso always says we're looking for kind-hearted, low-ego, very ambitious and hardworking people, and that is something that I think describes the culture at Poolside very much. If something we fail, it is that we are a bit conflict-averse, so it's something that we need to go after a little bit more, but for the rest, yeah, extremely nice people and talented at the same time.

Ceci [30:38]

I love that, and maybe a quick follow-up question. You're obviously a global organisation. Are you fully remote or do you have hubs? In a world that is now reverting to let's see everyone face to face all the time, how do you maintain a high level of collaboration, being global and potentially remote?

Andrea [30:56]

Yes, this is something that the founders put a lot of thought into since day one. They wanted our team to be fully remote. However, it is completely mandatory to get together three days every month. And this is in Paris. And it is not because we're a French company. The news got this completely wrong. We are American since day one. We just happened to have an office in Paris that we fill three to five days every month. And that means that we fly people from Brazil, Canada, every part of Europe, every month for a few days. And we get together and work all in a room. And the rest of the month, we rent the office out to other startups in France.

So that was for a very long time the only hub, if you may, that we had. That is why people thought that we were French. But it was actually logistically the best place to be on the map, because many people in Europe can get to it by train versus flying, including people in the UK. And also some people might have issues with visas. Getting to the UK or the US versus France is very easy in terms of visas if you have a very global team. And so that's what we did. And then only after a year that I had joined, we had a big enough hub in London and enough people pushing to have an office that we actually opened a very nice office in London, but that is really the only place where we have a full-time office with people going in and out every day. In the future, we might open other offices in other locations, but I don't think we will ever force people to go in.

The Stack, the Excels and the Windows Laptop

Bea [32:25]

Super nice. I actually love the efficiency of renting out the office when you don't use it. I'll take that. I would like to talk about AI a little bit. Obviously, you work for an AI lab. I would imagine you have AI processes and you had to achieve a certain level of technical depth now that you're in the model partnerships role. How did you get this level of technical depth? And what do you have open on your laptop, let's say, on any single day?

Andrea [32:53]

As I mentioned, big synergies from VC. I learned that you never go into a call with a partner not having tried their product. And so that really only happened as I picked up the partnerships motion, that I started living in the terminal and trying some of our agents with our models, with other models, to understand the difference. I had to really understand what people mean when they say that the Hermes agent has a different attitude or personality than Kilo Code or OpenCode. So I had to try it for myself. And at the same time, some of these players even have their own desktop app. So I had to go in and check, okay, do they have a model picker? Do they have a recommended list where I can get Poolside listed there?

And so I started playing with some of these surfaces and that's how naturally I became a bit more proficient on the technical side. So right now in my bar, I have Hermes desktop app, OpenCode desktop app, and then I have the terminal, where I usually run a few agents in parallel to compare the same question, how it is answered by different agents. And then obviously I have Wispr Flow, I'm a big fan. I try not to type too much and I dictate a lot. And then your usual Granola, Claude and ChatGPT. And obviously I play with our own models in the terminal directly. So that's where I have Laguna up and running.

Ceci [34:16]

Can you give us an example, during your day to day or during your week now in your new role, of a process or a workflow that you've automated end to end with AI?

Andrea [34:25]

So I find myself looking at data very often. Something I have open every day is ClickHouse, where I have my dashboard with at least 20 graphs showing me daily different metrics of token consumption of our models by source, by paid versus free endpoint. This is driving a lot of the conversations I have with partners. And so I realised I was writing a lot of queries, fixing a lot of queries. And so I basically created a skill on Claude that breaks down every single customer ID within our data lake and classifies them into the four categories I care about, be it gateways, direct, and others. And basically every time I say, hey, can you write a query about this or that, it ensures consistency to make sure there are no errors.

Bea [35:22]

Controversial question. Does capital markets, so the model that you're building, does Claude or AI have any input in the model? Or do you still do it manually? And I know that for an ex-banker, it's a very difficult question, because we are all very controlling of our own models and Excels.

Andrea [35:43]

So just to tell you how old school I am, I am one of only two people at Poolside of 180 that doesn't own a Mac. I'm still running on a Windows and I'm heavily punished for it. So that means yes, I build my Excels by hand and I do not use AI for that. However, whenever I have questions, like I'm getting an error here, I will copy and paste it into Claude separately, but I don't want it messing with my actual Excel. However, on the other hand, for PPT, I do run it on the side and I ask it to help me improve the slides. But on Excel, I'm quite old school, I would say.

Bea [36:17]

I was bullied into getting a MacBook, pretty much.

Andrea [36:21]

I actually have a Windows Surface just because it looks less of a Mac than a Lenovo. And so when I'm in the company onsites in Paris, people will think I have a Mac, so they will treat me like one of them. But really, when they realise that, when I ask an engineer to fix something on the terminal and they're like, wait, this shortcut is not working, because apparently the terminal has shortcuts in the same way that Excel does and it's different in Windows versus Mac. And so they get all freaked out when they realise I've been using a Windows all along. So yeah, it is a big topic at Poolside. I'm still surviving, but maybe now with the new role, I will have to move to a Mac, who knows.

Model Factory

Ceci [36:57]

I'm so excited for you.

Bea [36:58]

One last question before wrapping everything up. Can you tell us a little bit more about model factory, what it is at Poolside and how are you involved?

Andrea [37:08]

Many people are not aware of this, but most of the compute spent for big AI labs goes towards experimentation. So there is this article I really like from Epoch AI that says that OpenAI spent 7 billion in compute in 2024. 2 billion of that went to inference, so people consuming tokens from their models. The other five went to R&D. Out of those five, only 400 million went to the large training run for ChatGPT 4.5. The remaining, up to 5 billion, was experimentation.

What does this mean? The large training run is fixed by physics. There is only so much compute you need for a specific size of a model, and it's going to be the same for every lab. The big difference is experimentation. And this is because model building is essentially a science of trial and error. And the more compute you have, the more experiments you can run, the faster you can iterate and the quicker you can get to the right recipe, as we call it.

So given that we are a much smaller team than some of our competitors and have much less compute than some of our competitors, we had to find smart ways on how to be very efficient with experiments. And the way experiments are done at other labs is artisanal. So they run ad hoc experiments one at a time. They're bottlenecked by people. So the more people, more experiments. And so what we have done is very early on, we industrialised the way that we do model building and we call it the model factory. So this is literally an automated assembly line for model building. So every experiment is tracked and reused and can be used for the next model run.

Another big secret is that model building is 90% an engineering problem and not a research problem. And at Poolside, apart from having great researchers, we have great engineers. And so they have been sitting side by side with the researchers, looking at how they build models and helping them do it faster. So the result today is that we are running 10,000 to 20,000 experiments a month. We trained our model from scratch in eight to nine weeks with only 4,000 H200s, with a team of less than 100 people. And our previous model was trained in only five weeks. So the asset right now for us is that speed and efficiency on how we train models.

Bea [39:41]

For me, it's incredible when you put a constraint on something, people genuinely try to innovate off the back of that. And it's similar across many different things, how constraint can create innovation. Many founders that really managed to do something innovative, they say that they would have been unemployable, so that was the only thing that they could do. Or companies, when they have less capital they make it work better, or if you have less compute you make the compute work better. So I think that's very powerful as a statement and I really like it.

Rapid Fire

Ceci [40:12]

Is it time for rapid fire? Let's do it. So besides Excel shortcuts, what's the most useful thing you took from banking that you still use every week?

Andrea [40:22]

It's not a skill but more of a mentality. It's definitely hard working and the ability to work under pressure.

Bea [40:29]

Agree. What is one thing from VC that VCs maybe get wrong and you found out on the operating side and now you know?

Andrea [40:37]

I think VCs think that joining a startup, it's all about strategy. And it's very hard, if you come from a VC and you've never been in a startup before, that you will be deciding any type of strategy. I think strategy comes from the founding team and the leadership and the rest is all about execution. You grind and you all follow the same direction that the founders have set. That's the job.

Ceci [41:01]

We're seeing more and more people in this business operations title now than we've seen before. Do you think it's here to stay or is it just a fad?

Andrea [41:09]

I think it is here to stay, as people are trying to move a bit farther away from more generalist chief of staff, founder's associate types of role and more like, okay, I want a title with a bit more specificity.

Bea [41:25]

Do you have any advice for people in VCs that want to make the jump, how can they prepare? There's no preparation recipe, but is there anything that they should think of before making the jump?

Andrea [41:40]

I think definitely diligencing the company like it was an investment of your own money, because it is one, it's your time, which is the most expensive investment. The team is everything, so more than the idea, more than the market, just diligence the team, focus on what they've done before. And also don't wait for the perfect moment. When your shoe cleaner tells you to buy a stock, it's too late, same thing here. If everyone is talking about a company, it's probably too late. So do it early, otherwise you probably miss the upside.

Closing

Bea [42:07]

Fair enough. Well, Andrea, thank you so much. We really, really enjoyed the chat. And it's very powerful that you were part of the team that wrote the first check into Poolside and then you ended up joining them. This is the first time that we have a frontier AI lab on the podcast. And I think it was super interesting to understand the intricacies that are very, very different from the application layer. So thank you so much for joining us.

Andrea [42:34]

Thank you for having me.

Ceci [42:35]

Thank you. And if you enjoyed this episode, don't forget to like and subscribe and send it to your friends. This was Top of the Ops. See you.

Bea [42:42]

See you next time.