AI Security, Slack Agents, and the Future of Work

Transcript generated automatically from the episode audio and lightly cleaned up (filler words removed, misheard names corrected where clear). Speaker labels were added afterward and may contain errors.

Rob May: Hi everyone, I'm Rob May. I'm the co-founder and CEO of Neurometric.

Ryan Eppley: And I'm Ryan Eppley, the co-founder CEO of Root Access.

Anna Kirk: Hi, I'm Anna Kirk, sales lead at Thread AI, joining you all remotely.

Rob May: This is the AI in NYC show. All right, so a couple of housekeeping things. First of all, Anna's remote today. This is our first show with a remote host. So, let us know what you think about it. Let us know how this works out and we'll see if the sounds good and everything else. So, this this is a fun experiment. The episode today is brought to you by Zo Computer. Zo is a personal cloud computer powered by AI. So, you tell it what you want, it'll build you a website, it'll make an automation, whatever. It just builds it and hosts it. There's nothing to do. It's easier to use than Claude Code or any of those things. There's no terminal, no DevOps. Super easy. It plugs into all the stuff you're already using. So, Zomper, there'll be a link on the show. Go check them out.

Before we got before we dig into the AI stuff today, I want to talk about I Anna said Anna said we should do a hot take on this because she and I got coffee the other day and had this whole conversation and I'm I'm really curious what the rest of you think, but I know this is very non-environmentally friendly to say, but I think coffee tastes best out of a styrofoam cup. And I know you can't even really get styrofoam cups anymore, but there I don't I don't know what it is about that. And like other like water doesn't taste great out of a styrofoam cup, but like I don't know. I What was your thought, Anna?

Anna Kirk: Okay. And this this cannot leave this room aka this like podcast audience. But I love styrofoam because there's something like tactile and like kind of like chewable about it that I really enjoy. It's like another like added benefit like piece of entertainment when you're sipping your coffee every single day. So yeah, not the environmentally friendly hot take, but I don't know, Ryan, what are your thoughts?

Ryan Eppley: I mean, not not usually the typical news that we cover here and NYC, but it reminds me of being a labor on like job sites growing up for extra cash. Like, you know, it'd be cold sometimes in Pennsylvania and they just fill up everyone's like styrofoam cup with like the worst coffee you've ever had. Yeah. And you know, that's the way you keep warm. So, I don't know. I think it's more of a nostalgia thing when you've had the styrofoam cup.

Rob May: So, that's what we were talking about, right? Which is like sometimes if you think the same drink tastes different out of different cups, right? Is it is it a nostalgia thing or is there actually some science to it or what? So I, you know, maybe somebody will weigh in. I'm more likely people are going to boycott us now that we talked about styrofoam, but we'll see.

Anna Kirk: There's there's a related hot take, Rob, that we also talked about. Maybe more of a lukewarm take. I like my coffee room temperature. Like I prefer my coffee like I don't I think it's more common than you think it is. Like if if you also share that opinion, feel free to write to us and let us know that I am not alone.

Ryan Eppley: Usually that just means you're drinking too much coffee and now it's replacing your water.

Rob May: Yeah, it's next question. [laughter] No, nobody nobody is going to write in and support that. So okay, write in and let us know what you think about the coffee the coffee and the star cup debate, but all right, I want to introduce our guest today. Boris Berenberg from Z0.ai.

Boris and I met a couple years ago through a group called PEF PostExit Founders. The New York chapter is the largest chap. We were just talking before and I think it was a whiskey tasting event or something like that. But yeah, Boris, welcome to the show.

Boris Berenberg: Thank you. Appreciate you having me.

Rob May: So I want to get started. So you so our sponsor today is Zo Zo. You're Zero, which is different. You should check out both. [laughter] walk us through a little bit what Zo does, what ZZ does, sorry. And like what prompted you to pursue the idea? Okay. You're obviously multi-time founder. I'm always interested in like why what drew you to this idea once you had the experience with startups.

Boris Berenberg: Yeah. So, I after I sold my last company, I was hanging out in private equity land. Busy getting married, renovating my house, you know, all the other things you do when post sale. And I was like, "Oh, I'm going to go do something good for the world." Started building an edtech company. Realized why people don't do that. And while doing that, I started doing in-person AI tutoring for executives. And I would go in and I would be sitting with like CEO, COO, CFO, whatever. Teach them something, be like, "Here's your homework." Come back a week later. Did anyone do their homework? No. Do you remember anything? No. Like, you're behaving like children, right? And I was like, "Okay, there's going to be a huge like AI uptake problem, especially amongst nontechnical people, right?"

And you know, I know we're going to talk about some politically adjacent topics later, but to me, this is like almost a political company that I'm building because our like main mission is avoid a massive layoff of American knowledge workers. Like that's like the northstar like as the AI transition happens, how do we avoid massive layoffs of American knowledge workers?

And the method that we're going like what we're doing to achieve that is we're saying hey we're going to put AI into visible places so that today like if you have a power user who's wonderful at X and nobody sees them doing it because they're using Claude or ChatGPT or whatever even Zo right not to criticize but like it's a it's a private interface and there's no cross trainining happening you know and so I say like if you have a kid is your kid ever going to learn to do a cartwheel? If they don't see somebody else doing a cartwheel and the answer is like yeah of course there's a small tiny percentage of kids who will like figure it out from nothing but the majority have to see it and so we make that AI visible in Slack and then you know I guess again later we're going to be talking about what cloud released today but we're very focused on the risk security compliance side of it and really making sure that this works for regulated industries so you know agents can process PHI they can process financial data etc.

Ryan Eppley: Did you start with a specific industry because I saw on your website very compliance focused. I even saw the CISO called out really really in the top. Was it finance first? Was it healthcare? Was there a specific lane to start?

Boris Berenberg: Not really. We knew we needed, you know, and again, like we can kind of talk about this in part during the cloud portion, but, we knew we needed to have something that is going to be significantly differentiated from the core model providers as well as from what we'll see in adjacent organizations. So, you know, Google, I guess core model provider, also hyperscaler, also workspace platform, right? And so like obviously they're going to build in the space. Slack themselves are investing heavily in Slackbot from what I'm hearing. There's a whole new like Slack future that is shortly to be released. You know, Microsoft, nobody really loves Copilot, but hey, it's probably the most used AI product in the world right now, right? So, it's there's a little bit of everything and we needed to kind of find our niche and so we focused on kind of that compliant cluster. We're excluding Govvete, so we're not going to compete with you guys at all. We just don't know how to kind of effectively do procurement there.

Rob May: We should, Anna, before we I don't want to ignore you since you're on. Do you have a question for Boris before we keep going?

Anna Kirk: Yeah, I mean, I would I'd be curious to hear a little bit more about kind of why you are excluding GovTech and kind of like why you took that approach from the beginning.

Boris Berenberg: Yeah, I think like so my previous company, we tried to sell to government and we just very quickly realized that this wasn't my competence. You know, we tend to succeed in sales cycles where we are, you know, heavily connected to decision makers that can top down make a decision and once they've made it, everything else kind of flies out of the way. Right? If we're having to go and do, you know, 18-month RFP process for a home improvement company, then it's a lot tougher. You know, and it's just not my personal skill set. And as a tiny, you know, startup, we have to kind of pick and choose our battles. And this is this is the one of the compromises we've made to narrow the ICP and make GTM easier, faster, and more aligned with kind of like like each of each of the each of the verticals we go after is close enough together that it makes sense. And GovTech is like almost a step too far. Even like we're considering legal, but even that could be a step too far, I think, because their problems around data privacy are very different than the compliance regimes of health tech and fintech.

Anna Kirk: I mean, I'd be I'd be curious to hear a little bit more about kind of how you think about you know, like the horizontality versus kind of the like vertical specific motion. Obviously, you could continue going down and becoming more and more specific to a given channel, but how are you thinking about that trade-off? Like, yes, there are definitely specific security requirements for the legal vertical or the fintech vertical, but kind of where where do you see yourselves landing?

Boris Berenberg: We have landed from like it's not so much landed, it's where we built from day one, which is just be the most possible secure option out there. Because it's always easy to disable security features whereas it's tougher to add them down the line and you know even even having done that I can tell you like I like to joke like I'm you know people buy from me because I tell them what's wrong with our product right like I will be honest about like where we fail I can tell you one of our security vulnerabilities right which is if you can somehow prompt inject the model to respond to the user We use Slack as our primary user interface, which means that if you get the model to respond with an arbitrary URL that you control, Slack will helpfully try and unfurl that URL and will then exfiltrate data. And any AI that lives in Slack has the same vulnerability and you simply cannot do anything about it. Right. Right. I mean, kind of, right? Like for example, you could have a deterministic parser which takes any URL and then like markdown quotes it to reduce it, right? This type of thing. But you know, loosely if you want working URLs, you're you cannot solve this.

And so but beyond that, yeah, I think we, you know, when I see kind of other people like getting caught out on security issues, I'm very fortunate because we're often like, "Oh, cool. Yeah, we thought of that from day one." and there's there's a standards body that's working together right now called AARM. Was originally founded. So, it's autonomous agent brie forget I'm going to butcher it, but ARM.dev. It was founded by a guy named Herman. He's a product guy at Vanta and it's now been contributed to the cloud security alliance. And it's the question of like okay hey either you're you're inference plat or I guess like agent runtime platform and or your product or you know one of a number of categories how do you run an agent in a way where it can be externally verifiable that it is safe enough let's say and so the ARM spec v1 tries to do that the v2 spec we're part of the working group right now we're like actually iterating on how to make better and it's a lot of things like where where you know where does it have deterministic escalations where does it have probabilistic escalations where does it look at things like semantic drift in conversation intent you know all of these types of things and then again like how can that actually be externally verifiable

Rob May: so I think it's super cool that people are learning AI by watching other people use AI have you seen any data yet or do you have any concern that there are certain things that people won't want to do in the open that might suppress certain use cases and how do you get around those or have you thought about that?

Boris Berenberg: Yeah, I think the most interesting one of that is a very boring one which is I don't want people to watch me like flail around until I get something to work. Yeah. And I'm like no, but like if you don't show other people that it takes iteration, they won't iterate.

Ryan Eppley: Right. Right. Like I remember when I was like 16 and I took like a Java class and I was like programming is dumb. Like it's so hard. Everything's always broken all the time. And I didn't understand that it's always broken for everybody all the time. It's not even like I'm not dumb. It's like this is just like you're really a debugger if you're Yeah. Yeah. Yeah. Like like that's the job. And yes, a new issue. Great. Great. I've I've made it to the next issue and that's like just your life. Yeah. Back then we didn't have loop/goal.

Boris Berenberg: But the Yeah. I think that you know this is kind of even the very boring things are are where we get kind of push back and then you get kind of the more interesting stuff which is less individual preference and more around UX of like okay cool if we want to have a conversation about people's salaries right like I'm on the HR team you're the CEO we have a HR channel we put the bot in there we can ask the bot about people's salaries fine okay like should HR people be able look up each other's salaries.

Rob May: Yeah, that's right. Like very simp that policy varies at different companies.

Boris Berenberg: Yeah. Yeah. It's a very simple and so you very very quickly get into these questions of like or hey I want to explain something to you. I've created a Google doc in my Google Drive. You should have access to it but you don't today. Should I be able to ask the bot to talk about that in a channel where you are a member? Yeah. Right. And it's like that's really interesting. Like this is actually kind of the complexity of not only like do you do that thing but like let's say you can answer that question easily. You're like no we deny by default fine whatever. How do you present that to the user? Do you explicitly say like oh yes we found information but we refuse to disclose it.

Rob May: But you're you're you're you're hitting on a broader societal issue I think with AI which is when you get these intelligent models they can know lots of things they can do lots of things. How do you control access given that there's some natural access controls in the world by virtue of the limits of things that people know the tools they have access to or whatever and so suddenly you give this thing all the knowledge all the tools you know how does it know what

Boris Berenberg: yeah I this is going to be fascinating we take the perspective that I think that broadly that there's there's kind of a religious split right now which is should agents only ever act with a human identity and human authorization Or should agents have their own delegated authoriz identity and its own authorization. And so we are closest to that last category.

We say that when you put Z and we pronounce it zero and my wife loves the Zo computer sweatshirt that we have at home. So shout out to them. And I have the clicky swag that they have. So, so great great company. The you know we kind of take this perspective of you take you hire zero into your company and then you put it in each like channel or whatever, however you want to segment and that should be the last hire for that team ever, right? It's like hiring an intern that is going to join the HR team, that is going to join the ops team, that's going to join the finance team, and it will simply be useless on day one. And I think to the question around like hey you know horizontal versus narrow use cases etc. This is honestly where we struggle right is we have to go to each company and we have to say great what use cases do you want and you know as we gain customers in a given segment and it's not even like it's pretty narrow right like you're like oh health tech great okay but like tellah health is completely different from you know you know claims processing right they have different use cases different everything so as we gain depth in each of these verticals we're able to bring more use cases out of the

But broadly we have to work with customers with their existing partners to do the implementations and to do those implementations that are correct for them and for the organization for the industry right we I mean I can tell you in our seed deck I was kind of making the point of you know there was the Gartner I think it was a Gartner report or Mackenzie report last year about like 90% of AI implementations fail everyone's like oh no and if you read it like they're talking about like it's because systems aren't self-learning. I was like, I don't think that's why, man. Like, that sounds like somebody paid you to sell like I think it's mostly because people don't take consulting approach to it, right?

I think to implement AI effectively, you basically you need to be a business analyst who can go in and quickly understand a given domain and a very narrow domain, right? It's not like accounting. It's like accounting in teleaalth at a$10 to50 million annual run rate, right? Like like it's that narrow. And then you need to be technical enough to implement some kind of systemic solution. You need to know enough about AI like you don't need to be an expert honestly but if you need to understand context management sub agents you know context rot blah blah blah and then finally you need to be able to deliver change management within an organization and I don't think most people can do that.

Rob May: Yeah definitely not. When you're talking about people flailing around. So I have a really interesting story that I tell in my companies now which is so I didn't learn to ski till I was 40 which is a hard time to learn to ski but I did it because I so I so I have a daughter and a son and they got this age where I mean my daughter was like 12 and my son was a couple years younger and so they they you know when they were little they both like to do the same things and now they're teenagers they like to do the same things but they were drifting apart and skiing was the only thing they both like to do. So I was like okay dad's got to learn to ski so I can do this with them.

So, here I am like 40 years old and I'm like out on the bunny hills like I can't even turn left and like little kids it's their first day and they're whizzing around me and like I always told the story to people because I was like it's really embarrassing like I'm I'm trying to turn I'm crashing into the walls but like almost everybody that I told the story to was like you know I bet it was great for your kids to see you struggle with something like that at your age and then get better and everything else and like set this thing for lifelong learning. I was like, well, I was telling it to be funny because it was embarrassing, but like they, you know, it's I think you do have to set the standard with people, and this is why I tell the stories at my companies now, which is like it's okay for people to see that. As long as you're not constantly flailing on everything, right? It's like I tell people all the time, it's okay to make a mistake. That's how you learn. That's how you get better. If you're making the same mistakes all the time, you're fired. But

Anna Kirk: Yep. Yeah. And how can you make a contained mistake? Like I think there's almost across like the enterprise community I and I'd be curious what you see as well, Boris, but I feel like I see two extremes. Number one is every single AI decision, every single agentic decision has to be fundamentally centralized and it has to go through a trillion different gates of review processes and like strategy decks before the first single thing is vetted. And then I see the other end of the spectrum, which is like, okay, we're we're willing to make these like isolated tiny failures and accept that maybe that means that our kind of AI landscape is a bit fragmented at first before we like get into the motions of actually like repeating that strategy. Is that also what you're seeing?

Boris Berenberg: I might Well, my advice to people that are buying AI, just give the startups money and go with it, man. Like just come on, [laughter] you know? I think it's a great question. I would say that we have a strategic decision in how we build our product and how we go to market to say hey we sell top down we want to sell to an org. They don't have to all implement it at once but we want the buy in at the org level and part of our strategy is to say okay you've got you know zero running with HR you know now legal wants to go out and buy Harvey like is Harvey going to offer you something like materially different other than us just connecting to your Google drive right and you being able to ask questions there right and will it offer you let's say I mean I don't know how much Harvey costs but my guess is a lot right like will it will the cost of procurement of the net next tool, right? In terms of vendor selection, procurement across security, legal risk, etc., right? And then implementation, like how much could you spend on iterating on building something that's actually right for you? And so our our our pitch is basically buy us once and then reject other vendors.

Yeah. But you know oftentimes companies you know we're definitely having trouble dislocating vendors right part of our pitch is like hey you went out and you bought a bunch of claude 30% of people use it just cut your spend right and people like well you know like claude's big you guys are small you know so to your point just give give the startup [laughter] some money exactly for a reason

Rob May: Let's talk a little bit about doing this in New York. So, how long have you been in New York? Did you do your first company here? And if so, what made you decide to either stay or come here as opposed to, you know, moving to Silicon Valley or Austin or some other place?

Boris Berenberg: Yeah. So, I end I moved here before I started my previous company. I moved here because I was bluffing trying to get rejected from a job. This was when Uber was like going through a lot of like Uber bad public media and I didn't necessarily want that on my resume and I was interviewing with a company. I won't name their name but they were much slower in their hiring process and Uber gave me a very good offer very very quickly. It was triple what I had made like eight months prior. Wow. And I was trying to use that to negotiate with the other company. And I was trying to be unreasonable to Uber. And I was like, "Well, if you won't, you know," I was like, "If you won't do what I want, then I want to work 100% remote from New York." And they were like, "Well, no, but like if you show up to work or not, that's between you and your manager." And it was kind of like a wink wink, like go to New York. And I was like, "Oh I guess I'm moving to New York." Yeah. And you know, I quit after like six or eight months from there. Just like wonderful team. Everything was great. But it was like a career dead end for me.

And so I built my last company entirely in New York. I loved it. It's wonderful. I was on another podcast recently and I was like I need a word for like nationalist but for a city like like like what what is it? What is it when you're like you know I like to troll people sometimes when I'm traveling abroad. I'll be like, "Oh yeah, I live in the best city in the world." And they're like, "Oh, what does London have to say about that?" I'm like, well, it's not the best city in the world, so I don't care what it has to say. Yeah, I love it here. I'm very I'm I'm very grateful to be here. I have no intention of moving. I grew up in California. So I know exactly what I'm not getting there.

Rob May: Yeah. These two both spent a bunch of time in California, so in everyone's here. Yeah. Yeah. [laughter]

Boris Berenberg: But, you know, it's funny because like my dad grew up in Moscow, right? Like I'm from Russia and like I was born there, moved here when I was a kid. And every time he comes visits he's like how can you live in this like hell and I'm like you know it's like grass is greener type of thing. Like you grew up in Moscow and then you moved you know to Los Altos, right? Like I grew up in Certino and now I'm in New York.

Ryan Eppley: You know who else is crazy about where they they grew up? Surprising people from Maine. Maine. The state. People love Maine. They love Maine. And you're like, it's winter nine months of the year in Maine. It's a ter I mean, it's beautiful when it is nice out.

Boris Berenberg: That is true. That is true. But yeah.

Ryan Eppley: And so what's your what's what's your take on the New York tech scene overall? You know, having spent time in California and seeing that, do you know, what do you feel like it has? What do you feel like it lacks?

Boris Berenberg: I think that it has significantly more breadth of ideas especially as it relates to sectors, right? For better or worse, New York has advertising. It has television and film. We have pharma across the water in Jersey, not that far away in Boston. We're like the financial capital of the United States. You know we have significant presence in fashion and you know like like there's industry after industry after industry here and so you have a lot of you know noncore tech thinking which I think is quite lacking in Silicon Valley. Conversely, I think the engineering here is much worse. Like, you know, we're we're hiring. We're only interviewing people that already live here. Like 90% I think of our candidates already live in New York. And it is significantly worse than what is in available in Silicon Valley in California.

Now the question is if the talent were here, could I afford them? Right? Like because a lot of those people are there because guess what? They're getting paid, you know, 750k a year. Yeah. And we can't offer that. So, unless somebody wants to spend more money with startups. [laughter] but the but you know, my last company was built with a 100% remote team mostly of you know, whoever I could get my hands on in the lowest cost of living locations. So, you know, like you're not going to I don't mean this as a criticism of anyone, but like these are not the A++ of the world. And so I think that what you see in New York is better business and worse technology if that makes sense. And I mean obviously we have here, we have Datadog, you know, we have a number of excellent technology companies.

Rob May: Yeah. I mean there's great engineers everywhere. The question is how many are there, right? Yeah. And I what do you got because you hire different probably kind of engineer than they hire. What have you seen in terms of the talent?

Ryan Eppley: It's tough. And maybe it's just like because we're such an international hub as well. Most of our applicants are not US citizens or green card holders. Yeah. Whereas a lot of my founder friends in Silicon Valley, they they have obviously international applicants, but like there's a handful of applicants that are just living in like literally their zip code, right? Like there's so many technologists in the Bay Area. So I think there's a lot of talent. It's just there's not a ton of homegrown talent in New York. That's what I've experienced when I see 4,000 applicants for one job.

Boris Berenberg: And for you, like I don't know enough about kind of your business. I know very high level from what I saw on the website, but are you like ITAR regulated or kind of what constraints do you operate under?

Ryan Eppley: Yeah. So, we have a ton of compliance constraints. So, like hiring is much harder for us. But on the flip side, a lot of people want to work in this if they're like into that as well. But you're asking about like the talent and like I would definitely say it's funny because we get a ton of applicants from the Bay Area. They'll take the call and then like I think they're just like thinking about moving to New York. But I haven't had success like having someone want to move to New York. Even though there's a ton of applicants from the Bay Area, I don't know. If you're in New York and you want to be in the startup community, I think there's great talent there. There's just not enough of it yet. That's the challenge we found.

Rob May: I think the interesting thing is like there's like I call it the prisoner exchange, right, of like all the New Yorkers move to California and then all the California people come here and like we're kind of just apartment swapping. But getting somebody at the right time of their lives is just really tough, right? And there needs to be some like just like when you're doing sales, right? There needs to be some kind of like time pressure or something some kind of something changed or whatever, right? And you know, maybe we need to be organizing, you know, some kind of hiring events here and flying out a bunch of engineers and, you know, giving giving them the opportunity to experience, you know, a couple couple weekends in New York. Yeah. Well, we have some good news. Let's hop into some good news.

Let's let's let's hop into some of the news of the day. So I actually want to start with so maybe one of the most controversial pieces. So Steve Yegge, who's a very well-known, you know, programmer, wrote this piece called the flat curve society. And his argument is that since the government is now regulating some of these high-end models like cloud mythos and everything else that that is that the government is going to regulate super intelligence and the rest of us are going to basically be stuck with the level of AI that we have today for most of the rest of our lives unless you've got a you know special clearance or something like that. And I a lot of mixed feelings about number one whether this should be happening and then also whether this will happen. But I'll I'll I'll go last. What do you guys What you have thoughts on this Anna?

Anna Kirk: Yeah, my immediate reaction after reading this is like I don't think if you look at like technological breakthroughs that have been accessible to a wide array of builders like I'm still an open source optimist. I don't think that like the technical capability or the modding model capability will be gated and I think it will be very difficult to do that. What I found more interesting about that article was kind of his like separation between like the demand horizon, i.e. Like you can't tell the models apart if your problems aren't hard enough, and like the discernment horizon of deployment isn't actually bound by capability. It's bound by verifiability, like your ability to actually like discern what performance looks like. I think I agree more with the latter of like maybe there's something to this like verifiability point of like that actually being the ceiling.

Ryan Eppley: Yeah. I mean, you have to see this all the time with how your business model works, right? Like you don't need the bazooka for what a small knife can do every time.

Rob May: Yeah. Yeah. I don't know. It's true. I mean, I actually think probably frontier intelligence is probably overused by a couple, you know, an order of magnitude or something. I mean I would say probably at scale well I mean here's the other thing even for these small models the intelligence density so the number of things you can do for a 10 billion parameter model 50 billion like whatever it's getting better every year and so even the stuff we have is going to get smarter for you know various you know reinforcement learning techniques they have and post all these post- training things. So, it's interesting. I just I kind of agree with Anna like I think it's going to be hard when super intelligence gets here to regulate it in a way like you can do it with the big companies, but there's so many tools out there to train these models and there's so many people like it's just it's going to be hard to be I mean you can get, you know, sort of jailbreak models out there that do things that the frontier models won't aren't supposed to do. You know, create illegal propaganda and whatever.

Boris Berenberg: So, I'll go a step further, right? Like if you consider that the argument for locking it down is a national defense one, right? Then isn't releasing one a national offense one? Right? So if we say, "Hey, we have super intelligence. We don't want anyone else to have it. Like we will gate global access to it." Fine. Right. Great. Like doesn't that encourage a global competitor to now release an open source one to enable all of our enemies essentially, right? Like it's like an enemy of my enemy is my friend, but in this case it's like the opposite, right? So there's a huge incentive to make access to this available whether that's open source or if it's proprietary, right? And like China has been very big in releasing like let's say open weight models. I don't want to call them open source, right? But like open weight models like what's to stop them from developing, you know, something let's say equivalent to mythos and saying hey API gated but we don't care who uses it.

Rob May: Yeah. Right. So there's just a lot like and people could get around like you could ban it in the United States, ban the API and somebody could open a Hong Kong subsidiary and get around it.

Boris Berenberg: I mean there's all well there's the whole thing of like like right now a huge percentage of token purchase of open AAI and claude is being used in Southeast Asia through discount proxies. I don't know if you guys have heard about these things, but basically Chinese firms are like using shell companies to sign up for accounts and then they sell tokens for pennies on the dollar, but in exchange they they scrape absolutely everything, right? And you know in the let's say the countries where the their money doesn't go as far for token economics like they don't care, right? And I mean like let me ask you this, right? If you, let's say, are the vibecoded unicorn one man story that all the VCs talk about, like, do you care if the Chinese read all of your code? Like, does it actually matter? No.

Right? Like, so no. So, and I think even if you think about like, okay, great. So, somebody could hypothetically create a billion dollar company while giving away all their IP and it still would be fine to them. So that you know then the question become again and you know this is why I just think about like I'm I'm I'm generally against this type of policy and I don't think it'll work across again like what everyone's kind of brought up right whether it's verifiability whether it's you know somebody producing SLMs whether it is increasing competence of existing medium small models whether it's just you know third party providers in other countries like it's seems fool's eron to me and is reminiscent of bad export policy around compute to China now giving China a domestic chip industry that is going to be highly short

Rob May: I always I always point out to people the story where like because the United States had a landline infrastructure Africa jumped us in mobile because when mobile came around they adopted it much faster because they didn't have an alternative and we were like I still got my landlines you know whatever and it was much slower and so the mobile app innovation and everything there was much much higher for a while and yeah I was with you when we put all those export controls on you high-end GPUs. I didn't think it makes sense. I think it was counterproductive. Yeah. I think you want other people dependent on your industry as long as you can so you always have that card to pull. When you actually pull it, it changes changes the game.

Anna Kirk: Yep. Yeah. I am really curious around so one of his predictions where I don't even know if I necessarily agree but he's predicting kind of a return of buyover build for SaaS where bas like SaaS where basically everything will become much more specialized because there's like a there's that like gap of verifiability and there's a gap of AI enablement and education.

Rob May: It's not even that, you know, you know why I quit? Because like I did I went through the whole thing where from like whatever the Claude Code was released around Thanksgiving last year up until about April 1st this year, I coded up tons of apps. I did all this stuff and I was like, man, I got to maintain this stuff. And the other problem is, let's say I want to build my own CRM. I can I can do that. But yeah, I don't talk to a bunch of co like I know what I need in a CRM and if I just need something lightweight, I would code it up. Like I hope the product managers at CRM companies are thinking about cool things I haven't thought about that they're just going to drop on me and I'm going to be like yes awesome. So I am I think a year ago I would have been on the SaaS apocalypse train and I'm not anymore. I they might be overvalued, right? So I don't know if the market's going to come back to that level because maybe they're they're not as defensible as they used to be. I think you might see fragmentation in that market. But fragmentation because it's easier to go create a CRM competitor if you're a CRM expert now is different than you know I'm going to build my own CRM. So that that's my current take.

Boris Berenberg: I think the one thing that's interesting is pricing and packaging, right? Like everyone always talking about like oh like can you build it? Should you build it? This and that, right? But I think you know is it just going to be the case where you know I'm going to make up names here. Don't take these like somebody like IBM goes to Atlassian and it's like hey we're spending 5 million bucks on Jira today, right? Like we can vibe code our own Jira and maintain it for a million. Yeah. We don't want to do that. Sell it to us for a million. [clears throat] we're happy to pay you a million and like we can both be happy because we don't want to be maintaining Jira and you guys are the experts at this and you guys have figured it out like and so I think there's there's a quite plausible future where we see you know price compression as a function of renegot of like the ability to have a negotiation over buy verse build

Rob May: yeah and also because I do think you'll see more competition from fragmentation where where you know you would go before and you'd be like And if I can't get to at least, you know, maybe 20 million in revenue, it does make make sense to try to start this CRM company even as a lifestyle business and it's so long to build. But now, if you're like, I'm an expert, I could build this relatively quickly. Yeah, you could do that. And you could be a, you know, twoerson $5 million a year CRM shop and it's a nice lifestyle business.

Boris Berenberg: Well, look, I'll even tell you this, right? Like I literally come from the Jurro world, right? Like I'm I'm the Atlassian guy. And the we internally as part of our compliance have to do like tracking of you know hey what are decisions policies blah blah blah and I was trying to figure out what's the best way to you know do all of that ticket tracking and you know policies I was like okay cool so like to this day jur and confluence are actually kind of the best option and I was like this sucks for like a twoperson team where everything is being done by agents this sucks and so I started basically building my own jur competive because why not, right? Everyone does this and I don't think it actually has a viable future as a business as a jur competitor, right? The only possible future that this thing has is actually being a free tool with an entirely different business model. And I don't know what that's going to be, right? Like it's just like currently it's an internal tool that we just kind of made public for shits and giggles. But if anyone wants to try it out, it's called audit otter. [laughter] Okay. Yeah. Had a domain laying around.

But I think that we will see people building these tools internally and then maybe commercial like there's going to be a lot more shoot I'm bling on the word when like random stuff happens but it's pleasant serendipitous.

Rob May: Serendipity.

Boris Berenberg: Thank you. I think there's just going to be a lot more serendipity around these things.

Anna Kirk: Yeah. There is interesting like I'm I'm curious to your point around like pricing pressure too. Like how much will that factor impact what kind of pricing negotiations look like at the enterprise versus like the kind of opposite calculus of I'm IBM and I don't want to pay as much for Atlassian but the opportunity cost of me building this in house is higher than the actual material cost of me like building this in house now that like the kind of like ceiling on productivity is much higher. So, like I'm curious kind of like where that'll net out.

Rob May: Yeah. So, related to some of this flat curve stuff that we were talking about is also this this article from Reuters that Alibaba has sued the defense department because the defense department has branded it a Chinese military company and Alibaba says, you know, we have no Chinese affiliated government directors or any of that kind of stuff. This is a really complicated issue, I think, right? Because like I am a free trader. I'm a globalist. I like all this kind of stuff. I recognize the security issues. You know, I always go back to I asked Matty earlier to put up the that famous Google WTF slide when they found out that the NSA was like, you know, breaking the Google circuit in the middle and figuring out, you know, being able to intercept traffic and stuff like that. And so I know these things happen and I know that they're probably, you know, sometimes we find out years later that there were subpoenas and back doors and all this kind of stuff, but yeah, you know, what do you guys think? Is this good for AI? Is this bad for AI? Like, should we embrace Alibaba here more or not? I mean, what's the You're closest to this industry. What's the What's your read on how people feel?

Ryan Eppley: I don't know. People really worried. I think I think you have to imagine that there's a lot of concern because we're already doing export controls on our own stuff and now we're starting to see other stuff be worthwhile considering as export controllable in the future. So, I think I think there's a lot that we don't know. I when I read that article, my first thought was, what are they not including in this article that I would need to understand before I can make an opinion? There's so much going on behind the scenes. But, I think it is a big first step into, you know, our public companies on the docket. And I think because how the CCP works like a lot of a lot of overlap does happen between the largest companies and the government whether they say it explicitly or not in a way that is pretty accurate. I just don't know what information we don't know.

Boris Berenberg: Yeah. Like the article towards the end mentions that this only applies to procurement within the department of war. Right. Which is like okay cool. Like that is very different from saying we are like like like blanket like like Airbnb uses what is it? Qwen, right? Yeah. So, but like let's be transparent like Airbnb uses Qwen on some US inference provider most likely, right? So like is it now like like like could the Department of War use Qwen on a US inference provider? Right. Like is that prohibited?

Ryan Eppley: Let's not forget like the whole TikTok thing happened not too long ago. I guess I think it's still mankind, right?

Rob May: Well, and like there was like, oh well, where is the data centers going to be? Oh, Oracle's going to come in and try to, you know, mitigate some of this.

Ryan Eppley: And that wasn't just Department of War, though, right? Or Department of Defense at the time. That was also just an infiltration with the citizenry and people were concerned, you know, some type of company based in China having access to influence over the citizenry, not just the department.

Boris Berenberg: Yeah. I think before we started talking you'd brought up kind of this question of can you have a back door in an LLM, right? And you know we were kind of starting to go down the conversation of like well it's nondeterministic like how often does it actually pop up blah blah blah and it's like well like if you think about influence campaigns right then like if you have a 100,000 people out of out of you know like you don't actually need like it just needs to be high impact. It doesn't have to be high frequency.

Rob May: Yeah. Right. That's true. So, I don't know. I think this is one of those areas where if we can ever figure out, you know, cost-effective explanability in LLMs, like it helps solve like areas of concern like this. Cool. And then last article to cover today. So, Anthropic launched TAG, the agentic AI co-orker in Slack, related a little bit to what you guys do, and we were talking before the show. This is one of the hardest things about raising money, right, which is everybody is like, is Anthropic or openi just going to do what you do? They're smart people with good technology and lots of money. They they can do many many things, but you can't be good at everything. It's fundamentally different. You actually brought this up earlier that you're privacy architected from the beginning that requires you to make a series of trade-offs in your design decisions that you can't be like the easiest to use, the easiest setup, the most free flowing, integrate with everything, and also the most privacy focused like they're fundamentally incompatible. So I think there's still all kinds of space around what these these companies do. But I you know you're you're living this. What's your take?

Boris Berenberg: So when I was in the Atlassian ecosystem, my last company, one of the things we did was build apps on top of Jiren and Confluence. And periodically Atlassian would sherlock one of these app, not not just ours, but you know, everyone in the ecosystem, they would shock one of these. And whoever was getting Sherlock would invariably freak out. They're like, "Oh no, my business is dead." And like the reality is it's the opposite, right? What would happen is that Atlassian would then do a ton of marketing around whatever thing they had built. That would get consumed by the lowest denominator of users who didn't care about product depth, quality, etc. Right? Because Atlassian just like Anthropic can't do everything, right?

And so if you specialized in one area, what happened was they drastically increased buyer awareness that this was a that there was a problem here that there are solutions. Customers would try that baseline solution, get a baseline level of awareness and then would go and buy kind of the more specialized solution. You know, I was talking with with somebody who runs like a specialized private credit thing where they're very sensitive around privacy because I guess their customers are contentious with the US government as it is today. And they said they cannot use cloud co-work because claude coowork can write code and execute code on the device and that code can exfiltrate data and I was like well yeah that's true it kind of can like and they're like okay so it can't we can't have that right so you know is Anthropic going to come in and come in with better sandboxing at some point for this yes is it going to meet every compliance requirement no right to your point about trade-offs

Like if you go on Twitter, you'll see like Daytona and all the other sandbox people. They're always like getting in a dick measuring contest essentially about like whose sandbox spins up the fastest. I can tell you ours are not very fast. They're about 2 seconds. They're all at about like I think 500 400 milliseconds. So we're like three times slower than them. But we cache a lot less. So, we know that like it's much less likely that something sneaks through between different invocations. Yeah. Right. So, there there's always going to be trade-offs around these kinds of things. And the question is will customers value that?

Yeah. I mean, you know, it's not that revolutionary because Devon did this at their launch, right? Like Devon was a Slack only launch. And that's like two years ago now, right? So, I think the idea of at Claude, honestly, I didn't realize you couldn't do that until they announced you could. I think the really big thing that they announced today was it actually they built an ambient feature. So, you can now have it it can proactively do things for you without you telling it. I think that's by far the most interesting thing that they've announced. I have no idea how well it works. You know, could be wonderful, could be terrible. Our testing on this because we've tried to do the same thing is that it is very very difficult to tune in a way that's not annoying.

Anna Kirk: How how do you think about that from like an enterprise buyer perspective? And Boris, I'm sure especially because you're prioritizing that top down motion, this comes up a lot where there's this like, you know, maybe it's a kind of CIO visionary type who says, I want this to be the easiest thing to use ever, and I also want it to be highly governant and highly secure. Like to Rob's earlier point, like you're not going to change the laws of physics. There will always be an inherent trade-off between like ease of use and governance. So, is this like an adoption cycle thing where we need to let people try and fail the easy route or do you think that you know people will self- select to the more secure option if it's a real need for them right away?

Boris Berenberg: Well, we believe that on our end that going after organizations that don't have a choice, right? The choice is made for them by the regulations under which they operate. I mean, look, they can break the law. That's on them, right? Like that's not like our business model is not to find customers who are interested in breaking the law. Right. Right. And so for us we went in this direction from day one specifically because I mean Rob you kind of brought up that like you know you have these conversations with your investors. Every single investor that we have and every investor that we don't have asked like what are you going to do when you know Claude etc do this type of thing. And you know I think that if you are talking to executives that understand that they need to adopt AI but that they cannot place the rest of their bit like you cannot buy upside by selling your company on risk right and so being positioned to sell to that specific segment gives us kind of a structural advantage where we don't really have to have those conversations. Right? The conversations are more around like, "Hey, you've met our security thresholds now. Are we actually getting enough value out of it?" Right? It's actually like like we're never skipping the security conversation with anybody. In fact, we typically have to have that before we can have anything else. Yeah.

Ryan Eppley: Cool. Well, before we wrap up, we always like to end on some New York City trivia. So I have a question for you guys today and I'll we'll give you credit if you even get within 50 of the answer of this, but is how many subway stations are in New York City proper? That's take a guess.

Boris Berenberg: Yeah. Why you got to throw me under the bus like that, Ryan? Why do I have to go? All right.

Ryan Eppley: What do you got for?

Boris Berenberg: I said 470. 470.

Rob May: 70. I'm going to go lower. I'm going to say 300.

Anna Kirk: Yeah, let's let's do like 672.

Ryan Eppley: Boris was within two. It's 472. [laughter] Very nice. Very nice.

Boris Berenberg: I can I can shout out something called MaximumN NYC. Yeah. It's a guy here. I'm blanking on his name right now, but he does classes on the history and politics of New York City. It's quite quite affordable if you are especially a tech person. I think it's like two grand. Is very very worth your time to spend a couple months learning about the cool

Rob May: Boris, thanks for being on today and everybody out there, thank you guys for watching. Please rate us, like us. We have a newsletter now. Please subscribe. And then check out Zo Computer our sponsor. And so again, Zo lets you build anything that you want. They host it. Super easy to use. Please check them out and we will see you next week. Great. The next