Ethical AI, Iteration Velocity, and the Future of Software with Grant Lee (Gamma), Noah MacCallum (Open AI), Malte Ubl (Vercel), Chris Messina (Product Leader), Ash Kumrah (The AI Collective) | Inside Summit 2025 [LIVE]

the most valuable insight today is having access to the problem that no one else is paying attention to. If you have the ability to identify those problems, find distribution, sell, get your first customer, there's a lot of people who can build a lot of software that don't have access to that, would happily partner with you, or you could just build it yourself potentially. So if you have that, lean into it. If you don't, maybe try and do some weird stuff with your career that will get you exposed to non-consensus stuff. That you can double down on.

Welcome back to Season 13 of The Room Podcast. Claudia, can you believe we've been doing this this long?

No, it's been 5 years. So many episodes, so many incredible conversations telling the stories of some of the world's most iconic founders and funders. And whether our listeners are new for the first time today or have been with us from Season 1, we're so grateful for the community we've built.

Completely. We've had the mission to open the door to the room where it happens for for those entire 5 years and have been able to do so over 120 episodes with hard-earned lessons from some of the top technology and consumer founders of our time.

And we've been through a lot personally as well.

We have indeed.

I built, scaled, sold my startup Pryve. You were our first investor, so you've been there through the entire journey. And you've also recently graduated from HBS. You're now building at Vercel. And through it all, we've brought the funder and founder perspective.

Yeah, it's really been an incredible journey, and I'm thrilled that I'm back full-time in San Francisco, and we're probably here to stay. Uh, that means that there's more for us in terms of growth, both inside and outside of the physical room. But what's important to have really out there is these stories. And we've been able to do so over incredible guests from the CEO and founders of Flexport, Perplexity, Particle, Zapier, Zillow, Vanta. I could go on. These guests and their enterprises have generated over $100 billion enterprise value and really, frankly, moved the markets.

And growing, especially in 2025 with everything that's going on in AI. So I'm so excited to share Season 13 stories that are maybe a little bit more AI-focused than past seasons.

Absolutely. And you can find as IRL and URL, especially with our events such as our annual conference, Inside Summit, or upcoming SF Tech Week events and beyond.

And if you want to get information around those events, subscribe to The Room Podcast and subscribe to our newsletter at theroompodcast.com.

Perkins Coie supports the most innovative entrepreneurs and investors in fast-moving and high-growth sectors, addressing their myriad of legal needs. But the firm doesn't just provide end-to-end legal and business counseling to its startup clients. It also facilitates introductions to key advisors and sources of capital. Perkins Coie's interactive website, Startup Percolator, offers access to programs, resources, and rich dynamic content designed to assist entrepreneurs on their startup journey. To learn more, go to startuppercolator.com and perkinscoie.com.

This podcast is brought to you by Mercury. The banking platform businesses like The Room Podcast use to simplify their finances. It's time banking did more than hold your money. Now it can. With Mercury, you can pay bills in seconds, close the books faster, and even send invoices. Not only does Mercury do away with a patchwork of tools, it eliminates guesswork, giving you complete and accurate visibility into your business's finances, all from one account. Apply in minutes at mercury.com. In this Inside Summit episode of The Room Podcast, We're joined by an exceptional group of builders shaping the AI landscape. Moderated by Ash Kumara of AI Collective, this panel brings together Chris Messina, product leader and creator of The Hashtag, Grant Lee, founder and CEO of Gamma, Noah McCallum, product leader at OpenAI, and Malte Uebel, CTO of Vercel. Recorded live at Inside Summit, this conversation explores what it takes to build durable AI products in a rapidly evolving market. The panel digs into how teams move from experimentation to sustained impact, where real product differentiation is emerging as models commoditize, and how leaders think about ownership, leverage, and scale in this next phase of AI. With perspectives spanning frontier models, developer platforms, and AI-native applications, this episode offers a grounded look at how the best operators are navigating velocity, ambiguity, and long-term bets without losing sight of execution. Let's open the door.

Before we get started, since we have 4 amazing people here, could you just briefly introduce yourself and tell us what you do currently, just so everyone has context? Why don't we start with you, Noah?

Hi everyone.

I'm Noah McCallum. I'm at OpenAI. We're actually starting a new applied evals team internally. So I'm a founding member of that. We're focused on identifying gaps in API capabilities and figuring out how to improve models so that you can all build more effective products on top.

Thank you.

Hi everyone. My name's Grant. I'm the co-founder and CEO of a company called Gamma. You can think of us as building the anti-PowerPoint. If anybody has stayed up late at night working on a deck, trying to format it, find the right layout, we're going to try to save you some time. So we use AI to help you eliminate all that, minimize the manual tedious work. And try to help you with visual storytelling.

Sweet. I'm Chris Messina. As Ash suggested, I'm the inventor of the hashtag, which was like 8,000 years ago. More recently, I help a lot of founders go to market in terms of telling their story, mostly on a platform called Product Hunt. But before that, I worked at Google, I worked at Uber, I've been on developer platform teams, I worked on a lot of open source and open protocols for the social web, and was also a founder of a conversational AI company back in 2017. So I've done a bunch of random things because I'm not good at anything. And so I'm here to be the gadfly on this panel.

Hey, I'm Malte. I'm the CTO of Vercel, Madison's employer. Yeah. Thanks for having me. What we do is if you have an app, front end, back end, AI agent, you throw it over the fence. We're standing on the other side of the fence and we run it for you. And when it goes down at night, we get woken up first and only call you if it's really bad. The other thing is we're making a tool called v0, which lets you type in what you want to build, and it's coming out in really nice source code that you can then deploy to yourself or anywhere else.

Thank you very much. There's a theme that I've been hearing about in the AI landscape, and it doesn't matter what size of your company, doesn't matter what industry you're in. It's even in general news, like mainstream news. It's called ethical AI. Is anyone here familiar with this term? Is anyone like me, like, what is this term? Because I don't know about you, but I keep hearing more and more about ethical AI, and I'm just curious, and we'll start with you, Noah. What do you define as ethical AI, and how do you feel your company is helping address it?

Of course, ethical AI is super important and something that we care about a lot internally. There's a lot of dimensions to this, and it's important to try and pull it apart. And it ranges from inherent model bias, and is it providing outputs that are harmful to individual users and maybe more of a per-interaction way, all the way through how do we make sure that there's not any big scary sci-fi risks for the models becoming autonomous or, uh, recursively self-improving and so on. And we're paying attention at every level of the stack. We're now a global company, hundreds of millions of weekly actives, and the decisions we make have huge impacts on people's daily lives. In particular, we were reporting on sycophancy being a problem where the models are being a little too agreeable for users. So if you say, hey, is this a good idea? You don't want the model to say that everything is a good idea. There's many ways that we can imagine that failing. I think it's an example of, I think, how we showed up to identify that based on feedback in the wild, providing a postmortem on what went wrong and what we're gonna do to improve. It's kind of part of our broader stance of iterative deployment where we know that the way that these things fail in the wild are impossible to predict. So we'll just kind of ship on a very bounded but fast cadence and try to fix things as rapidly as possible. Thank you.

For us, one of the most important things is just have like an open dialogue. So strong partnerships where we're sharing data, what we're seeing. One other avenue at the application layer that we think a lot about early on is around like trust and safety and the things that people are using our product for. And at the stage of startup that we are today, I think we're investing in that much more than I think prior startups have had to. So trying to maintain this sort of ecosystem where everyone's very collaborative and hopefully steering the sort of future together is definitely top of mind.

Thank you.

The challenge with talking about ethical AI is that you have to start talking about like ethics and that concerns human values and what is important between people and cultures and societies. What I have found having worked on social media for so long is that a lot of the technologies that we build are downstream of our own idiosyncrasies and our own unaddressed traumas and the things that we don't know how to talk about or don't have language for. Or if we do have language for them, the language then essentially with a large language model is somewhat deterministic of the outcomes that might happen. There's been obviously some terrible instances lately where people have gone down rabbit holes where talking to a chatbot that gives them a sense that everything that they think is a good idea or something that is appropriate or is acceptable or where self-harm is okay. Is something that in normal culture, with normal humans, typically, you don't have a great deal of structure to actually unpack that conversation and to support in a way that's helpful. If someone were to come to me and ask about suicide, I would want to be there for them, but that doesn't mean that they're not going to be pursuing that direction some other way. And so when it comes to ethical AI, I think we have to be pragmatic and realistic to look at our own culture that is informing these technologies and not put the entire onus on the technology producers themselves. In other words, like, we have some responsibility to become more clear about our behavior and how it informs these things, as opposed to assuming that creating a censorship regime is going to prevent all negative applications of these technologies.

Thank you.

Maybe to give a slightly different perspective, I think one thing that's important is often you talk about AI and immediately the ethics question gets asked. And I'm actually most excited about the very boring apps and boring kind of use cases, like, expense report classification, which used to be like very difficult and now you can like build a prototype in an afternoon. That's very good.

Yeah.

Hold this to expense social drinks though, or?

Amen.

Exactly.

Right.

But like then you're actually at the level where like you always have to answer that question. And so I think there's like, obviously it's worth having the discussion, but also worth disassociating it as a stigma for every possible use case.

Okay.

I really appreciate the diverse opinions we have here. And along those lines, Speaking of diverse initiatives, how many people here are building companies right now? There's a reason why I'm asking.

Okay.

And how many people here are looking to raise money? Raise your hand. And last question, it's not a trick question. How many people here, regardless of your stage, are really obsessed right now with how do I take what I have and go from like 0 to 100? It's already working. I now have lit the match and I'm really focused on growth. And deployment. Perfect. So this question applies to all three. What I love about this panel, or I call the Avengers, is you all have had amazing experience building products, investing in products, and scaling products. If you were an entrepreneur right now, given what you know about AI in the world that you reside in, what advice would you give to these entrepreneurs who are looking to really scale their products? Because we have a lot of builders here, as you can see when I asked them to raise their hand, and they're all hungry and they all seem like they really want to go. They're not at the ideation stage. They're like, no, we're here because we want to build something. We want to be in your seat maybe in a year or two, if not sooner. What would you say?

I have this conversation a lot with founders that I'm working one-on-one with, and I think it's a good distinction going zero to one and one to 100, where if you just have a demo, sometimes that can be enough to raise money. But the one to 100 is where You start to scale, robustness starts to really matter, and you are trying to figure out how to make this actually work in the real world. This is where so many AI products will end up failing. And there's kind of two main loops that I really advise people to nail. One is the sort of vibes-based improvement. There's a narrative shift around evals right now. Some people say that it's the only way, you have to have evals, and it's softening a little bit. And I think this is right, that the first 70% can come from vibes. If you have someone who's very opinionated, who has a lot of taste, and is in a very tight feedback, uh, improvement loop with the product. So they see how it works, see a fail, they update the system prompt, they tweak some UI. You can get actually quite far. You'd be surprised some of the startups that we work with are quite biased in this direction. But as things start to fail, then build targeted evals so that you can start to systematically improve. On those really tricky, really inconsistent capabilities that are impossible to squish out with the vibes-based loop. And then in order to build those evals, you need to have really good tooling and do that quickly and make sure that they're grounded and useful so you're not wasting time. And there's major failure modes in both directions that I see a lot.

It's a really big question. I think I have a lot to say on this. Oftentimes I'll talk to founders and it's really exciting right now to be building with AI and you might start seeing some initial traction. Throwing a lot of these ideas at the wall, and maybe the temptation is as soon as you see something start working, you're going to go all in. I think the downside is that the reality is you're going to be working on this thing for many years. This is a multi-year journey, and if you're not really committed to that problem or that problem space, you might be married to a bad idea forever because you saw that little bit inkling of, oh, there's something here. When we started Gamma, this was pre-AI. We focused on really what we cared about, which is We want to help change how people communicate ideas. This was a concept we were passionate about. I started my career in consulting, investment banking, lived in slide decks, and so was always frustrated with that format. That medium has been around for almost 40 years. And so we started that because of that. We were handed a great gift of like, obviously AI is moving really fast. We can incorporate that into how we build Gamma. But that wasn't the reason we were pursuing this problem space. And so I feel like for those that end up going like 1 or 2 years in chasing the wrong dream, you're already worrying about scaling when you're not even sure if this is the mountain you want to climb, that's going to be something that is almost irreversible. And then you're going to lose some of your best years working on something that maybe isn't something you're passionate about.

Thank you.

The way that I tend to think about this, I spent 2 years investing in vertical AI applications. It seems to me that we get confused between what is technology and what is a medium. And this distinction has become more valuable and useful over time. So rather than thinking about something just as a technology, which can be a tool that's delivered at scale, If you think about mediums allow people to express new ideas in ways that previously weren't economically viable. To draw an analogy, once we came up with agriculture, suddenly we had lots of different ingredients that we could use to then create different types of foods and dishes and different cuisines. Once you had those cuisines, then you could make different types of restaurants and different types of expressions from the lowest execution to the very highest. To me, what AI is doing is it's changing the ability for people to create different types of software experiences. That are incredibly bespoke, are very personalized, and used to be uneconomic to pursue. So to your point about picking a, a problem area or domain that you feel is richly expressive of something that you wanna say to the world, that is actually gonna be much more likely to work out over the next 5 to 7 years than to try to solve a deep technical problem unless you're one of these like AI labs or something where you're building something foundational. So be very clear in the thing that you want to create in the world. I guess another way to think about this is we are at the beginning stages of what I would think of as the record labelification of software, meaning that by applying taste, you can create a record label of different types of software applications that look like something certain, or that have a certain typography or typeface or whatever, but they're quite commoditized. To your point, Gamma is its own specific expression and experience. It needs to get the basics right of creating slides, but then from there you can do jazz and it becomes very exciting. So you might be creating more of a jazz label than actually like a piece of software.

I've been making this one point a lot lately, which is that the way I think about AI is that I use this Venn diagram where it's like the whole thing is all the software that would be awesome to have. And then there is the software that we used to be able to economically viably write. Maybe with $100 million, we could maybe do something, but like practically we wouldn't, right? And now what AI helped us is like fill out the Venn diagram. There's now more software that you can actually do. And that's also why we're seeing like this emergence of startups because suddenly there's this entire like white area on the flag, which is obviously incredible. More tactically, I think my personal mantra is that iteration velocity solves all known problems. And that really just means that you look at everything you do and see like, can I cut down on making that a little bit faster? Might be evals, might be vibes, right? It's really, that's the same decision, right? Where is my iteration velocity bound by someone doing evals or not? And the trade-off of that changes over time. And similarly, like in the past, you might have said, well, do I get an experiment running? And then I need like 4 weeks to get statistical significance where I honestly knew on day 1 that one side was shit. That type of thing, like is really fruitful. If you look at every step of the way, like how fast could I go from I have an idea to like, it's something I can actually try and then make sure it's good.

Thank you.

You had said something earlier about boring industries and paying attention to them. And I was thinking about this. Along those lines of maybe not a boring industry, but is there an industry or industries you feel that AI has not really tapped into yet? Yes, we know the obvious. We know about the consumer side, obviously the LLM side and various vertical AI, but is there like some industries that if you had the time to spend more time on, or maybe you're working on publicly, you're allowed to share, you're like, you know what, watch out for this. 'Cause there might be someone in this room who might be working on that solution or might have thought the same thing and they might get some insights from you. We'll start with you, please.

Yeah, part of this is if it's on my radar, it's probably one of the larger segments. So I'd flip the question where the most valuable insight today is having access to the problem that no one else is paying attention to. If you have the ability to identify those problems, find distribution, sell, get your first customer, there's a lot of people who can build a lot of software that don't have access to that, would happily partner with you, or you could just build it yourself potentially. So if you have that, lean into it. If you don't, maybe try and do some weird stuff with your career that will get you exposed to non-consensus stuff that you can double down on. Be at a very high level. I'm personally very excited for the potential in healthcare to be realized. It's very good for the world. It's very interesting. It's very hard from a sort of reasoning and intelligence perspective, which is something that from the research side we really care about. If you look at the history of providing leverage for doctors, there's some small attempts. You have a human scribe or you have a, nurse practitioner that are trying to make the doctor do less of the rote work and more of the differentiated sort of top of their discipline type work. But that is only the tip of the iceberg. Every doctor's complaining about doing too much clicking around in EHRs and just not doing the interesting work, which is probably some combination of like hard diagnosis, empathetic conversations with patients. Like, how do we get them doing more of that and none of the busy work? Healthcare is super complex, but AI is really well suited to dealing with that complexity. I'm excited to see that realized.

Is there anyone here working in healthcare at all?

Healthcare?

You are? You should definitely talk to him after.

That's a cool topic.

Thank you for sharing.

There's a lot of different AI applications. In your mind, you can ask, is it good enough for the AI to be an intern or does it need to be an expert? We're trying to build Gamma such that our AI is going to be your expert design partner. So every step of the way, AI can help you. That doesn't need to be the case for all applications. I think when you look at certain applications today are very successful, like customer success or customer support, certainly coding. There's almost this failover where even if AI doesn't get it right, human can still be in the loop and help pick things up where it left off. So I think it's helpful just to understand different applications will have different needs. Maybe your skill sets will be well suited for one versus the other and just give that some thought as you're exploring ideas.

Can I ask you a follow-up on that? Just focusing on Gamma for a second, because I realize this question might be a little out of the focus of what you're working on. Is there an industry or type of vertical that you were surprised to see a lot of customers customers penetrating with Gamma?

Early on, where we saw the sort of more— most usage and stickiness was around small teams that are doing a lot of external-facing communication. So sales and marketing teams, client people that are pitching their clients, customers, prospects, partners. And what we thought was that, okay, as we get bigger and bigger, even large organizations, large enterprise, you know, it would be maybe the sales and marketing teams. What we're seeing today is actually there's a demand for a much more wall-to-wall deployment, which is Everyone's using slides, as they call, like the language of business. Large organizations use it as a way to spread information and make decisions, make really tough decisions and important decisions. And so Gamma could be a replacement for that language. And that's what's been surprising to us is that if you can actually create a new creative language, you can earn the right to actually be replacing the incumbent in this case.

Thank you.

My response to that is going to be challenging if you guys are looking for pragmatic answers.

I wouldn't expect anything less from the inventor of the hashtags. More about it.

I try on the one hand not to like over-fixate on my own lived experience, but it's hard not to because I lived it. And when we were early building social media and I came up with the idea for the hashtag, we were solving a bunch of different problems to try to decentralize the social web. And the thought was that the web had been built mostly for academics and for the military and was a way to distribute documents and we knew, many of us being 24 in San Francisco 20 years ago, that people needed to be represented on this network as well. And so we went to work solving that problem. The reason why I'm not going to give you a pragmatic answer is because essentially I and many of my friends had a vision for the future about how people would behave, and they would probably behave more like us. And as a result, we needed to make those technologies easier to use, more transparent, and more engaging. Many people at the time wrote or thought that the things that we were doing were stupid or crazy, or that no one would ever do them. Who's going to post pictures of their breakfast to social media. And now social media is this crazy force on the world. Therefore, when you think about what you're going to do over the next 5, 10, and 20 years, imagine as though you actually succeed. And if you actually succeed, what is it that you achieved? I'm all about finding ways for organizations to have better internal communications, but it might not be a matter of actually creating better slideware to do that. It's more about I don't want to put words in your mouth, but I imagine it's about clarity of thought and then communicating those right messages to the right people, perhaps with depersonalization based on the audience and what's relevant to them. That doesn't exist today. You create a slide deck, you send it around, it's the same for everybody. It doesn't have to be that way in the future. So you asked about opportunities and things that I'm looking at. I'm still trying to solve the same social media problems and social networking problems that I was working on 20 years ago. And I still think that they're very interesting. And I think the future is going to be made up more of like digital twins and personal agents and custom agents that are interacting with and on behalf of ourselves. And the ways in which we control those things are completely unknown right now. So that's what I'm excited about.

Just a follow-up question, if I may. Along those lines though, similar to what I asked with Grant here, was there a certain behavior or certain type of audience that you have seen accelerate that you didn't expect when you were in those initial heydays with your colleagues building the hashtag and building at the time Twitter? And was it called Twitter then?

It was definitely, yeah, Twitter.

Definitely Twitter.

Yeah.

I guess I'll answer that differently because I was building a conversational AI before we had the LLM, it was incredibly brittle and it was stupid. You'd talk to it and then like you'd go a couple turns and it would just like fall off and you'd be like, like what happened? Once I saw ChatGPT come out, I posted it to Product Hunt and then it became product of the year. That is what caused me to become an investor because I knew that was the linchpin of the thing that I was missing. So the degree to which there's a younger generation coming up where they are able to have conversations with artificial intelligences as though it's normal. All the old folks have no idea what's coming. They have no idea how to operate in that space. They have no idea how to feel like it's okay to have an artificial friend that you build some sort of connection or intimacy with. They have a hard enough time having intimacy with each other. If you imagine like the Electron generation, like having Electron intimacy as being normal, that's going to set a whole bunch of new behaviors for the next 5 to 15 years. So I look at that and the adoption of ChatGPT We were right about what we were trying to build before, but it was the LLM that unlocked that new possibility. And if you imagine language being the original software, meaning that you can change it, right? Like the hashtag is a piece of software that is linguistic software. I changed language so that we could communicate better through groups. That is what we are all giving to each other through the LLM.

Thank you.

I think one mistake we all are making all the time, because that's what we, the world we live in, So we still have a scarcity mindset on what software should exist. It was very expensive to make software. And so there would be a small amount of software. There would be lots of winner-take-all market where something that's like, has a relatively broad applicability, kind of wins that entire market. And it's just the best product.

Can I just add to what you're saying? The reason why software is hard is because we're trying to communicate with an electron system that is quite brittle. And so when you have an idea for what you want to have happen, you have to communicate that in such a way that the computer can understand it and then perform that function over and over again. And the more that code builds up over time, you get more and more complexity such that any different thing that happens could break. So LLMs allow the system to become much more dynamic and comes closer to the way that we actually communicate and think ideas in the sort of like glossy patternistic way that's resilient relative to the way in which we used to build more brittle systems. So to your point, like the reason why we can have so much more software is is because we can communicate in a way that's natural to us, as opposed to having a very small number of wizards that could write the incantations that computers could use effectively.

And so like what that means is that almost none of the software that will be ever rewritten has been written. You are asking like, are there still any opportunities left? Yes. Essentially all of them are still left. We are like building a platform for deploying agents. Okay, we have to have some demos and you're like, oh, I have, I can't think of any. And so you go to like your finance person. And you're asking them, what do you really hate about your job? And they're like, bam, bam, bam, bam. You're like, holy crap, every single one of these I can just build, right? Like this used to be like very difficult to build software. And I think I've been having this experience basically in every field. So every field has just abundance of opportunity and almost all of it is implementable.

What is something you are working on right now? And please focus on a specific product. I would rather you talk about something very specific since we have builders here. Is it a product? Is it a feature? Is it an upgrade? Something that's on your mind that you're working on in your spare time outside of your day-to-day. Why don't we start with you?

Currently, I'm obsessed with coding agents and how to be the best model for coding agents. And in service of that, how to accelerate the flywheel that I was just describing. Can't share much more than that, but if there's anyone that has complaints about coding agent performance with OpenAI models, please come talk to me.

Yeah, we like to dogfood a ton at Gamma. So one of the things we're just releasing now is our own API. And I oversee all of our marketing efforts. So one of the things we're really excited about is now that we're going more into B2B from our prosumer roots, starting to actually expand into selling to organizations, is all of a sudden I can automate all of our deck creation. So any customer or client I need to pitch, I can take my notes from Granola. As soon as the meeting's done, it can send that to Gamma. I can give an immediate recap of everything we discussed. All the deliverables, all the action items, and before the meeting's even done, I can have that email to the client or email to the customer. So we're really excited about that. I think that allows us to kind of not only fulfill the vision of like true personalization, but automation at scale. And I can imagine now that once we actually have a sales team to arm them with that, it's going to be a really powerful unlock.

Thank you.

You mentioned Granola. I'm a big Granola fan. I do a lot of coaching for founders. And so I still use a very manual process in Airtable to manage all my calls. And so I'm just like, I need to make this fully AI where something comes in, it goes to Airtable, like the notes come in from Granola, like the email is sent, like all that's done. And I'm just so embarrassed that I haven't done it yet. The other thing that I want to work on, and I have been working on, is creating a decentralized social network for both people and for agents. It's something that I've been working on for many years. It's something that the protocols that we started in 2008 or 9 are now pretty mature. And I feel like now more than ever, the opportunity exists to do that. So that's something I'm doing.

Thank you.

To end the panel on a low note, what I'm actually spending a lot of time on is like agent security. It's actually not that long ago, let's say 15 years ago, where security was like in the duff. We were not using HTTPS yet. The NSA was reading everything you wrote. Everything was getting hacked all the time. If you were on LinkedIn already, all your data is in the cache of some hacker and it has gotten really good. And now we are getting back into the LLM phase, which has no security model at all. Every MCP that you put into your agent gets to prompt ejecute without any hindrance, right? And so we're like in this Wild West phase and we have to figure it out. And so that's something I really care about. One thing that's really fascinating is that we have this everybody-can-cook world where like our like SDRs are shipping apps to make their job easier, which is amazing. So you're in this world now where the security model for apps changes. Where you go from, I have these like ultra competent engineers who are like engineers, who like do things the right way, to like people who have no business judging whether the thing is good or not. And so like what I'm personally working on is kind of saying like, can we extract certain things from the apps so that you can't go wrong? Let's say auth, like it's not the app's problem that auth works so that even if the app is like completely broken, they cannot give access to the wrong person. Or maybe the app doesn't control how you talk to the backend. So again, if the app is broken, your security model doesn't fall over. Basically, there's a Wild West situation and we're trying to establish security principles in that world.

Madison had mentioned the AI Collective. We're a nonprofit. We are here to like humanize the AI experience through in-person meetups and events. So if anyone's interested in learning more about that, you could check out aicollective.com. But on this note, anyone have a question? Don't be nervous now. Yes, you, sir.

This is like the security kind of conversation around the bytecode discussions on Twitter. People saying, oh, like, it's the end of libraries because you can just bytecode anything now. Putting things in the platform level for like security, like secure by default versus like teaching AI how to do everything. What do you got to think about the juxtaposition? There's two ends of the spectrum, I guess you could say. One is like just bytecode everything. One might be like putting in a library that you know is tried and true, interoperable, modular, but now people are saying, oh, my agent, they just code that. Why would I use this in a bit PFR? Yeah. I mean, yeah, it's kind of directly what I was referring to. I think we're in the earliest innings of a maturation phase. There was always this notion of shadow IT where like someone would just upgrade from Excel or like do something in their organization that accesses some backend and maybe that wasn't good already, right? And so we just put like fuel onto this phenomenon, but it's not necessarily new, which does mean that like in large enterprise settings, you have to find a way to make it work. I think there, there can be this world where you end up with an architecture where you have this like professionally developed services and then you have vibe-coded frontends that access these services, right? But you're retaining all the like core variants of your system in those backends. I could definitely see like one of the like biggest running jokes of the industry is like how horrible Salesforce is as an interface. Like you can Viacode a better interface for whoever in your company is using Salesforce in an afternoon. Doesn't mean you need to stop using Salesforce, but it's just an API now and it can maintain your business rules, but everyone in your company will be happier.

Great question.

You, sir. Curious what each of your biggest hot takes are. What I mean is something that's truly not in the Zeitgeist?

It's a good question. One off the top is maybe coding agents aren't just for coding. You were alluding to this a little bit, but, and also to your question, agents work best when they have really good tools and really good abstractions. And all a coding agent is, is an agent that's really good at using tools and interacting with environments, doing research, making changes to things. And maybe for the N+1 domain, it's all kind of a variation of like, oh, it's actually just another tool and it's all kind of code.

Anyone else want to answer that?

As a VC, like, I guess I can say that conventional VC might be dead, which makes it very strange for me to sort of be in this position, but I don't know how capitalism is going to change or evolve. And this is also going to sound totally messed up. And so you asked for a hot take, so here you go. Like sometimes you come up with ideas and it's just not the right time. So communism and socialism. Might be actually like the right idea, but just the wrong time. We didn't have the right means of production. We didn't have abundance, we didn't have scale. And so there were a lot of externalities as a result of going down that path that really worked very poorly for humans at the time. But in the future, where you have a lot more services or agents managing things and having a deeper sense of what's going on and the metrics to know like what's working and what's not, maybe some of those concepts are worth reinvestigating. Now, I'm not proposing that. I'm not saying we go directly there. But I think it's worth thinking about the ways in which we've tried to solve problems in the past and recognize that the timing maybe just wasn't right.

To specifically ask about MCP, definitely overrated technology. You either own the server or you own the agent, but very often if you build a startup, you own both the tool and the agent, and you definitely don't need MCP. It's only for when you want to be in ChatGPT and eventually they'll maybe allow non-developers to use MTP. That's cool, right? That'll work because you don't own the agent, it's ChatGPT's. Or you are ChatGPT and you want someone to give you some stuff. But if you're building an agent, you don't need it.

I want to thank everyone, especially our panelists. Thank you all for attending. This is such a great experience. It's truly an honor to be your moderator for this. Thank you.

Thank you so much for joining us at The Room Podcast. If you want more from The Room every week, subscribe to our newsletter at theroompodcast.com/newsletter. We'll be back next week with a new episode, an inspirational guest, Tuesday, 10:00 AM Eastern, 7:00 AM Pacific. See you in the room. Perkins Coie supports the most innovative entrepreneurs and investors in fast-moving and high-growth sectors. Addressing their myriad of legal needs. But the firm doesn't just provide end-to-end legal and business counseling to its startup clients. It also facilitates introductions to key advisors and sources of capital. Perkins Coie's interactive website, Startup Percolator, offers access to programs, resources, and rich dynamic content designed to assist entrepreneurs on their startup journey. To learn more, go to startuppercolator.com. .com and perkinscoie.com.

This podcast is brought to you by Mercury, the banking platform businesses like The Room Podcast use to simplify their finances. It's time banking did more than hold your money. Now it can. With Mercury, you can pay bills in seconds, close the books faster, and even send invoices. Not only does Mercury do away with a patchwork of tools, it eliminates guesswork giving you complete and accurate visibility into your business's finances, all from one account. Apply in minutes at mercury.com.