Prof G Markets
00:48:17 · Ed Elson, Luca Ferrari
[00:48:17] S1: It sounds quite similar to the strategy that Ray Dalio used at Bridgewater for many years where there was a lot of he he wanted to track everything. He wanted to track every meeting, every interaction, sort of put everyone on a social graph, which I think in a lot of ways probably worked. But, also, he got a lot of criticism for it, and a lot of people did not like it because it feels like... It felt like to a lot of the workers in that culture that they were being surveilled and that every interaction was a test. And it sounds like in a lot of ways with the level of tracking that you find... That that that you are employing that you might run into potentially similar issues? I guess, how do you think about that issue and and do you think about it? Do you think of it as a potential issue? [00:49:05] S2: I didn't work at Bridgewater. I... I've read a few things about it, so I'm slightly familiar with it. What I just described only applies before you get hired. [00:49:15] S1: Okay.
00:49:15 · Ed Elson, Luca Ferrari
[00:49:15] S1: Okay. [00:49:15] S2: On the job, we actually don't do any, there's a just a review at the end of the year, which is based on input from your colleagues. We don't do any tracking of working hours, of anything really. So we... And precisely for what you mentioned, B, are staunch believers in the fact that most people react really well to being trusted. Yeah. And so we try to hire people who believe we believe are amazing or can be amazing given enough time, and then we give them uncomfortably high levels of trust and freedom, to the point that sometimes we've been criticized quite the opposite, sometimes we've had these situations where we acquired a business and then we established as a general manager, someone who's like 27 years old, and sometimes acquired team members tell us, are you crazy like this person, how can you trust she can do the job? And historically she has by the way, but I think if we were to monitor or surveil people, that would really destroy that trust that's such a motivating factor for people to come to work and do their best. [00:50:20] S1: I guess it makes sense. It's like you're extremely discerning before they're hired and then extremely trusting once they're hired, which I think makes a lot of sense, is probably the right strategy.
00:50:20 · Ed Elson, Luca Ferrari
[00:50:20] S1: I guess it makes sense. It's like you're extremely discerning before they're hired and then extremely trusting once they're hired, which I think makes a lot of sense, is probably the right strategy. [00:50:32] S2: Over time, you assess people's outcomes over a range of scenarios. You don't need to to surveil altogether if someone is consistently delivering, delivering quickly, high quality work. I don't I don't care to know really how you got it done, know, like if you use AI or didn't use AI, did you work late or you didn't like, ultimately we wanna work with people who are effective. And so we can just measure how we contribute to projects over time. And I don't think surveilling people would add in very much if anything to be honest. But pre hire you don't have, like pre hire you need to, you don't have any information, you need to look at in every niche and cranny to be able to figure out whether extending an offer makes sense. [00:51:09] S1: One of the big themes with your business is sort of using AI and using modern computer science and infusing it into these businesses. What does that look like? What what does it look like to take a company like Eventbrite and then give it whatever the AI bending spoons magic actually is? What are you attaching onto these businesses?
00:51:32 · Luca Ferrari
[00:51:32] S2: So I think we try to use AI broadly speaking for two things. One is there are sometimes features you can build for customers that are only possible, say today, thanks to AI, that were maybe not possible in the past. You mentioned Eventbrite, think. We haven't done it yet, it's very early, but I think one of the ways AI could could help make Eventbrite better is by greatly enhancing the recommender systems. When you go to Eventbrite, there's roughly 80,000,000 people who go there monthly to find interesting events near them. What you show them as as a recommendation makes a huge difference on whether they are satisfied and whether organizers sell tickets. So it's really like a win win. And that's a difficult problem, especially because you don't have, you know, the data that meta does. I mean, you have very limited data clearly because this... You know, the people only go there for the events. So and AI, we believe could be quite helpful in making that better, and even marginal improvements in that recommender system can be quite valuable for both parties of the marketplace. And the other big ways, and that's completely bespoke to the product obviously. And then there is something much more, say general purpose that applies to almost any business that's using AI and technology more generally, cutting edge technology to be operationally much more effective and efficient. For instance, again, told you we have dozens of proprietary technologies. Most of these are are based in AI, especially these days where if you don't use AI, you're probably giving a lot of value on the table. But some are really AI centric. For instance, we've got something called Alternative Spooner, old ALT Spooner. Each one of us, and Spooners by the way are basically the members of this core team, you know, highly selective recruiting process, and people will move around their businesses fluidly. But we are working now to extend this technology, by the way, to all team members, it takes a little bit more work, but you have this personal assistant that lives in Slack, we use Slack for communication, and it automatically has access as exactly the same level of access you do across all of our toolkit. And you just talk to this agent and you ask it to do pretty much anything.
00:53:39 · Luca Ferrari
For example, you could say, I actually saw this live as I was in a channel on Slack where we post feedback on Evernote, ideas to improve the product, was there to, I was just writing about a bug I found. And I saw our general manager for Evernote, she wrote that she had experienced the bug and asked her alternatives for her to, first of all, investigate it by going to to this tool we built to collect and and categorize customer support tickets and and check whether that bug was prevalent among customers or was this something she she was a corner case, and then go into the code base, identify the root cause, code a fix, and then message the technical lead for that project, Marco, so that it could review the pull request, and if fine, basically approve it so that that fix would go and benefit all customers. And she did that in, I don't know, like five, less than five minutes, and the agent went off and did it. The bug was fixed the day after, because Marco apparently didn't have time to review it the same day, but probably took, I don't know, thirty minutes of a person's time to do something that would have taken optimistically ten hours, even just two years ago. But actually we would probably never have solved that bug because it would never probably have reached the level of priority that we would go after it with the cost. So this is just one example. And interestingly, that technology, you don't even get to choose the model that's used, that's been used. So there is a gateway that the harness is built in house of open source components, and then the gateway is also built in house. So then when a task comes in, we have an algorithm that will determine the best model to, or combination of models to implement the task, to trying to optimize for quality and cost. So we have of course access to the APIs for the frontier models, but also we have self hosted over probably 15 or something different open weight models and combinations thereof. And and for 99% of the tasks, we actually go up and wait, spent almost nothing. We... We're currently, like I said, close to $4,000,000,000 in run rate net revenue.
00:55:37 · Luca Ferrari
We are, as far as I can tell, one of the most aggressive users of AI of any company writing 94% of our code with AI, almost all of our data analysis, almost all of our design, and on a run rate annualized basis, we're spending $15.01, $5,000,000 in all AI related spend, because it's almost sold down with super efficient open weight stuff, and the frontier models are used only for supervision, which is super cheap, it's like if you have a senior engineer just overseeing 10 junior engineers, just taking a look, but not doing the work themselves, that's way cheaper, And then for some super critical tasks, but those are quite rare. So when you bring this to an acquired team, it can make a massive difference in the efficiency with which things get done. And it's not just you cut costs, but also you just launch more features, you can run more experiments in monetization. It's all a virtuous cycle of compounding over time, which is quite powerful, especially if you can deploy it across new acquisitions, because then you can invest more in building these tools because you amortize across a growing base of revenue.