It's a rapid evolution in an old paradigm. Like this concept of websites is going to disappear very quickly before we know it.
Nvidia is taking no balance sheet risk on this. The people that are being stuck with the risk I'm pointing, but it's like my mother, your parents, it's the pensioners. The liability outlives the asset.
This isn't a toll road. You know, how much money can you spit off versus the duration that these are useful for?
No one's ever securitized compute. We have not answered the post labor existential question. We just haven't. You can't vibe code a home health nurse to show up in your living room to take care of you post surgery.
Your financing risk is pretty low. Your talent risk is pretty low and your upside is going to be way higher. Are you kidding me? That person just raised how much and what.
I would say AI is a lot more akin to something like, you know, GPS on a phone or a database where it is a technology.
Fake stuff is really helpful for society right now because it inoculates us and makes us realize we can't trust what we see anymore.
Hello and welcome to a ridiculous episode of More or Less where my co-hosts Dave Min, Britt Morin, and Jessica did the terrible, terrible judgment error of letting me host without them. They're all actually on vacation, and I am actually also on vacation, but pretending like I'm not on vacation. So, I'm thrilled to be here. We have assembled a world-class community here. I think this group represents every flavor and every area of venture capital investing known to man as well as like every product experience you could imagine. So I'm very, very excited to have a great group here to run through.
First we'll start with we have Peter Deng here. Um Peter actually has the distinction Peter is the one who brought me into Facebook in a lot of ways. He was running profiles right before I was. And so when I showed up on day one, he said, 'Great, someone else can run this' and taught me everything I need to know and then was very happy to peace off into his next project, which ended up being Instagram and then Uber and then Airtable and then big role running product at OpenAI and now is a dirty venture capitalist like the rest of us. So Peter, welcome. It's good to have you. You're GP at Felicity.
Rachel Holt. I've actually known Holt longer than anyone on this call. I first met Rachel on our first day of Bain and Company orientation in New York in 2005 when we sat down at orientation about seven of us in Bain, New York. As soon as there was a break in our seven person onboarding,
You cannot tell this story, Sam.
Why not? It's a great story.
I know exactly where you're going. I knew that this was going to come up.
It's my favorite story. She basically we have our first break at Bain and Company in a training. This girl who like is with six people she doesn't know, her new co-workers, takes the speaker phone in the middle of the conference room, dials Comcast and starts yelling at them just to set the tone. I was like, 'This is an amazing, I love this.' So, but Holt from there has done amazing things. She ran Uber North America.
First of all, I got my cable attached on day one.
Did you get it that day? That was amazing. No one had their cable attached.
Yeah. August 1st, 2005. Time Warner was there.
Boom. Boom. So, she took those operational chops to then run Uber North America, has started Construct Capital, represents the rare DC venture capitalist, but a breed that is growing and get to do a lot of business with her now. So, Rachel, great to have you on.
Scott Belsky, I mean, look, Scott's done everything. I knew him originally when he was running Behance and founded Behance, which he sold. He's now a partner at A24 and running A24 Labs, founded it and is thinking more about kind of the future of creative than anyone I know. So, I'm thrilled to have him here.
Scott Stanford, I mean, you started out an even dirtier job than venture capital running Goldman or a big piece of Goldman for a long time on the TMT side. You can correct me on the details, but you somehow did that forever and then have had an illustrious venture career now is the co-founding partner of Acme.
And then Josh Wolf, who does do you even need an introduction at this point?
I need the longest intro.
Of course. So Josh, co-founder, managing partner at Lux, which has done basically did the contrarian thing, which has now become consensus around defense and deep tech and hard tech and is kind of one of the men of the moment for calling things like Anduril early and backing up the truck when that wasn't cool, which is now very cool, which means hard to seed invest in anymore because these numbers have gotten so big. So, we have an amazing gang here and I have my topic list, which obviously I had my bot write of what's going on to run through about what's going on with Nvidia, this crazy balance sheet stuff.
Peter, we're going to have to talk a little bit of Airtable. Um, given the fact you ran product there for years and I was an investor and things like that, but Scott, you started out when we were BSing before hitting record.
We want to start with Grok bot.
Well, yeah. I mean, I think that the agents have been a topic that we've all been thinking about and talking about for quite some time, but we're like just in this moment, it seems where they're becoming like really useful on a consumer level. And I don't know about a lot of you, but I've been playing with some of these. They're always a little scary, these startup agents, because they ask you to like log in via Google and give it access to everything. And you get those like crazy Google warnings of this is not an authorized secure blah blah blah, but you do it anyways. But I have to say like they're kind of blowing my mind and seeing some of the demos that people have shared in the last 24 hours of their Grok bot and I think it's a combination of obviously the agent capabilities but also the memory and this notion of like having the bot have a computer instance of its own and even seeing their computer set up. It's truly this notion of someone with a screen and then you get this like consumer click of oh anything you can do on a screen they can do on a screen is like a really fascinating turning point it seems.
It's a shift from help me do this to just do it.
I'd like to point out that like I'm so down the rabbit hole on this that none of this stuff impresses me. I have a persistent cloud instance running in Digital Ocean. I had it do its own analysis and math right now. I'm not talking about the AI. My bot is running about three grand a month of infrastructure, database, app servers, etc. Whenever I want something, it builds a new app for me. It like runs its own databases, etc. And I actually interact all with it through email. So I like basically start email threads and I'll like ask it 50 things and then just it dispatches them, build an app, analyze my health, send me this cron job, whatever. So, like I've given my bot so much infrastructure that I'm not that impressed with it has a computer, but I'm kind of curious if any of you guys I mean Scott, you built some crazy stuff with AI.
You're not normal, Sam. So, it's like I think that
Yeah, I'm curious what others think, but it does feel like there's like a crossing the Rubicon moment for consumer accessibility and just the ease of onboarding and giving it authority for certain websites and just it's something seems to be clicking to me. But it's a rapid evolution in an old paradigm. This concept of websites is going to disappear very quickly before we know it. And this concept of like help me do a bunch of stuff versus just do it. Predict I need it, get it done. And you I love the Comcast example, Rachel, of like you having to call Comcast. Forget that. Like there's no humans involved anymore. It simply I don't even need to tell it that my cable's out. It knows my cable's out. It will reach out to the agent at Comcast. Will sort it out.
No, no, but I mean, where were we last week, right? So, we look at this with such a microscopic lens. This is going to happen before we know it. And then we're all going to be sitting around looking at each other saying, 'What are we supposed to do?'
But Scott Comcast doesn't give a. Like, they're not going to like solve that.
The reason why they're going to give a is when they get bombarded by a bunch of automated agents that are hitting them 24/7 saying, 'Fix this.' And they have to respond. And so then they rev up their agents and then it's agent to agent and we fish or we ride on Boston whalers or we do whatever.
So to me that question because when you said it's inevitable directional arrow of progress, every one of us is going to have many, many bots doing many, many things from the trivial to the substantive. Which bots? Because then it comes down to one key factor which is trust. And we're screwing around with lots of different ones. And at some point I do believe that the distinction between them will be who do you trust or entrust with your data? Who are you giving your APIs? Who are you giving your inbox to? Who are you giving your calendar? Who are you I mean clearly Grok
And whose answers do you trust like is the upshot right? It's like whose matrix multiplying are you willing to rely upon in what scenarios? But Peter you worked at OpenAI. Give us the scoop.
I think people's perception of AI is just going to change. We know it's continuing to change and the idea of trust is actually getting people are just starting to trust it a little bit more. In the beginning everyone was so like kind of skeptical of like oh it's going to steal all our jobs. Yes, it's taking some jobs but the fear is going to be there until people really experience it and I said this when I was at OpenAI one of the best things that OpenAI did for AI was just make it just free and usable by anyone to be like oh yeah you could try it.
Well Peter can I push you on that? That is maybe you are in my narrative. But there was a study I saw recently that said that AI is so hated. It's actually hated more than Jeffrey Epstein by the general public, right? It's like it is absolute it is the worst brand like you could possibly imagine. No one trusts it.
I'm not disputing that. What I'm saying is that that barometer is going to shift over time. I was just on a call with a bunch of creatives and maybe Scott I'd love to hear your thoughts on this because you're in the thick of it, writers who just started saying like how awesome it's been for them to build portfolios for their friends with Claude code that they were unable to do before. That's like not a conversation I would have had with them six months ago. I'm not saying where the trust is right now Sam. I'm saying that that shifts as people get more comfortable with it and it's just a matter of time. Do you trust all models equally right now or are there some brands where you're like, I'm not giving it my data?
No, I have an irrational distinction between what I use Claude for and what I use ChatGPT for.
But is that on trust or on performance, Peter? Because I think you're basing it on something different than a lot of other people.
It's actually based on what has my history to be honest. It's like, oh yeah, like it has all my work context on Claude, so I'm just going to It's easier. It's like back in the day when we got iPhones and we had too many apps on our app screen, we started just choosing icons to like to just compartmentalize things, right?
You got to in terms of where this actually plays out. One, this idea that AI companies be able to lock in your data makes no sense to me. You're going to like self-sovereign your identity in some way, shape, or form. I do it by owning all my own infrastructure and then I just point at different models for different types of tasks. But this idea that like with a consumer app, you're going to like lock people in because of their data, I just don't believe it. It's one of the only things the Europeans did well with GDPR. GDPR is a piece of.
Well, wait, wait, wait, but but some significant portion of the average population which again today is probably just on ChatGPT is not anywhere near sophisticated. Sort of like what Scott said about you. You are special. You are unique and you are advanced user. And the vast majority of people do get habituated and will just use what they've always been used.
Sure. And that's why lots of people still use yahoo.com for search.
Yeah. So it's not about I never said that there's lock-in Sam. I didn't say there's lock-in. I just said that there is like there's an irrational preference.
Can you take a step back though and ask like what's the ideal customer experience that would be most likely to be successful that is like we're from a tech and see trust lens is one thing but for a consumer they like get the new version of their iOS software and they get a magic text message that's like hey I'm the Apple agent. I've got all access to everything on your phone. Don't worry it will never be in the cloud. It will never be exploited by a third party. What do you want me to do for you now? And you just like go through and give it connectors and access as you need certain things. It saves your login as you need it. Why wouldn't they just win this?
Well, I think Apple might win it, but I think the interesting thing is I don't think you want to be the one who takes on that liability. If anything, I think the goal is to like radiate out risk to other people and be like, well, we're not like to keep yourself safe and be the one that's aligned because like the reality is you're like, Peter, you're saying that you think the ban is going to go up. The reality is the hacks are only going to continue to grow and scare people more and more and more, right? Not less and less and less. Everyone's going to have them. And so I just wonder if the real thing is going to be if you really want to win in like the new economy in this AI thing, you have to somehow figure out how to hoard the data and the trust and build the trust, but then when things go badly, always have a third party to point at to blame.
I'm with Scott on this. He's got the name advantage, number one. But number two, I mean, Gemini just hit a billion users. It's pretty obvious. It's friction. You get rid of friction and most consumers are going to gravitate in that direction. People are lazy.
I would question that metric. I think they probably counting like the people who are just typing the search box and like maybe clicking expand. There's many ways to game the metric, but I think distribution is a good part of it, but I do think consumers' taste and their bar goes up in terms of what they want to use. They have to really like using. I mean, there's many examples of this in the past. Why certain apps win in the app store and why others, you know, why Instagram took off and a bunch of the other ones didn't. Consumers have a discerning eye.
Wait, sorry, just consumers are also really cheap. Consumers want free and they're lazy and they're locking. I mean, it's kind of one of the things like is it really better experience or is it cheap, accessible, free, and habituated? Okay, so let me ask you this. Like Siri available, when's the last time you trusted it with a question?
That's not a thing. That's not even the same class. That's why they hired Google to build a real AI agent, which will be here yesterday.
I was just I want to go back to Scott's thing because there's a sort of a profound declaration that Apple wins and how you define that, but they have not spent the capex that everybody else has, which is an advantage. They have waited. You could argue that what they did in search, which is effectively get paid to let Google, you know, be on the search bar and Safari and ecosystem, they could do the same thing in AI. Now, there are two facets of my phone that are probably the most used that today. And Sam, you can enlighten me if there's something that you found for MCP plugins or others that across your agents are working that are the highest frequency utility and to me the least surface area exposed to AI of productivity which are my WhatsApp chats and my text messages. It can ingest my Gmail. It can ingest my calendar. It can do search. It can query all of our internal docs. It can use my Google Drive.
Location's pretty key, too. No one's exploiting it the way it will be exploited, but location's key.
So add those three things though because each of those are today's sandbox. Meta with WhatsApp, Apple with messages, although Claude can sort of access if you're using it through your laptop. And then you can use which a lot of people don't trust screen pipe which basically does like constant DVR recording of your screen and if your WhatsApp is open on your laptop, which is probably 10% of my utility, 90% on my phone, then it'll record your WhatsApp messages. But those two things to me, to Scott's point, feel like
Apple could win on those.
Add two more. Payments. So, Apple Pay, so they're seeing where you're spending your money. And then the second thing is location. I don't want to underestimate that. I created an app called Dippity, which I wanted to be like snap maps for old people. So, you can't see where people are, but you know if they're near you or you know if they're in town. So, I built this thing and it was it's pretty good. But, as I dug in, Apple and Google or Android have both locked Wi-Fi location. You cannot access it as a developer. And that's why Find My is so freaking accurate. But me, the independent developer, I can't get near it. I'm looking at pinging towers and it's just stuff like that that I think Apple keeps a lock on. I don't want them to win. Let's be clear.
Like to be clear, I think there are certain things that Elon has said that I actually really grudgingly completely agree with. And like one thing I think he said several times, I'm not sure if it's his or someone else's, but I think it's true, is that the most hilarious outcome is usually what's going to happen. And Apple winning is the most hilarious outcome, right? So, I think that's like a very good chip in the factor of like Apple wins by doing nothing. And actually, it's not even because they tried to do nothing. From everything I've heard, Apple internally is extremely confused organizationally on AI. There's a lot of reasons they're having trouble moving forward on it. So, in some ways, like organizational dysfunction creates the buffer for them to be highly successful by doing nothing, which is hilarious. That's the Apple thing.
Here's the thing I want to ask about is like are we all just arguing about what has APIs and what doesn't? Like ironically the reason that iMessage and WhatsApp have some defensibility or lock-in to them is they don't have APIs. Like forget MCP. MCP is stupid. Like you can just use API. I don't really understand why MCP is a big deal when everything has an API. You can just rip through it with Fable, right? So like there are things you can access programmatically which have no barriers and therefore go into the hive mind. And then there's this interesting exception to it is like Apple has a platform order which can enforce API rules, right? And they can have effectively APIs to things that no one else can have APIs to, right? So is the real story just it turns out that openness and API is bad, right? And that what you really want is let's close the stack with valuable data you can have uniquely.
I mean this is literally what's going on right now in like the industrial and the physical AI world. This data is not readily accessible. You can't commoditize it nearly as quickly. And so it gives startups like a chance to actually gain distribution in like a more methodical way before the mega platforms can swallow it up. Like this is exactly what we're seeing play out.
Will that continue more so or you think that will
Well, I think it creates choice around who gets the data in the continuum and the like the race between that exists between startups and mega platforms. Like in some ways you think it actually gives startups a chance, right? Because if they can gain distribution faster than like a mega platform can innovate or a hyperscaler could commoditize it, those factors actually favor a startup maybe more so than what you're seeing in the digital world.
It definitely evens the playing field. There's just a lot of data that's like it's not even in a PDF somewhere.
No, the factories weren't connected.
And that becomes their defensibility.
Yeah. A good example is like we have a machinist population that's retiring and a lot of the knowledge is actually in their head and I think that you know companies that can go and figure out what does it sound like when a tool bit's going to break and how do you what is the path that you take. I mean those are bits of data that I think you know in some ways Sam I think you and I saw this at Facebook it's like what did Facebook do it just digitize information that was in our head who's your friend.
Yeah, 100%. I mean I used to we used to talk about I mean like talk about the composer and Facebook era it was a data pump right it was an efficient free data pump to bring data online that couldn't exist anywhere else.
And the reality is if you go way back in the Facebook history, the real story was on day zero, internet's scary, so no one wants to post on it. You're never going to put a real photo on it. Facebook created the trust and security for people to take a bunch of real world data, photos, relationships, messages, etc., and bring it online, right, for the first time and created an enormous amount of value doing that, right? Okay. And so now we're in this weird multiverse where it's like the internet has all been scraped to hell, right? Like good luck New York Times with their lawsuit, right? But like that has been copied a thousand times and distilled a thousand times.
You have personal private digital information, what's in Gmail, what's in where people have some say in theory as opposed to it is all being like dumpable effectively. And so that's where we were talking about the trust things like that.
There's then like the whole class of companies that are out there trying to in some ways build new data sets, the Handshakes, the Scale AIs where you just pay a ton of money and you turn financial capital in some form into like unique data and then there's all the offline stuff. And so I guess the question for me is like assuming that multiplying big numbers is a commodity and will be, right? Like is the story you just like really you go back to day one which is like if your data isn't unique, you're not going to be valuable long term.
I think there's also a class of data that you didn't talk about which is something that Periodic Labs is doing which is like science data and material science data that doesn't even exist yet. You have to actually do a lab experiment to get that data. So absolutely like finding those bits of data has always been I think the name of the game it's going all the way back to Facebook and you and I used to talk about this in those small little rooms about how we're just trying to extract a bunch of data and how you obsessed you were about buying the DMV.
Not just trying, obsessed. I didn't. No, that is true. Peter, we used to have meetings for what we should buy and my pitch was always that Facebook should buy the DMV.
The DMV.
Okay. Can I pivot the conversation a little bit unless people want to go further on? There's another tack I'm curious people's take on which is open router. Like talk a little bit about enterprise and things like that. If we're saying that your data is so valuable, right? And but at the same time enterprise really cares. Oh my goodness. A guest visitor, Jessica Le. Wow.
How's it going, guys? How's Sam doing?
I am so jealous. Continue on. I'm just here. I'm just, you know, in a taxi listening.
Jess, just to get you up to speed, we've got two Sams and we've got two Scots, but there's uppercase and lowercase Scott. So,
Hi, guys. Guys, I am so happy to see all of you. I just came back from the big city of Manhattan, so I'm on route, but dive in. Dive in.
Are you Wait, are you in I just want for dinner planner. Are you in Massachusetts or in New York?
I'm in Massachusetts. I made it back.
Oh, you are. Oh, great. I'll see you soon.
I was in Midtown at 3:30. It's really a miracle. I'm here.
Open router routing. Everyone wants cheaper stuff. AI the matrix multiplication commodity. There's no reason to pay a PhD to fold your laundry. But you also don't want to give open router all your data. Or do you or do you not care? Like how does that play out?
I do think that the bigger question here which is going to also bifurcate it's not going to be one takes all is open versus closed and I am saying this obviously with a bias as I've said in the past that where you stand on the issue depends on where you sit in the cap table and we're large investors in Hugging Face which made a lot of news over the past few weeks but I think that the future is for the more sophisticated users for the enterprise it's going to be your longitudinal proprietary repository silo of data that is yours that is exactly as we were just talking about not ingested by the machines. And if you have that and can run open source models on that time series of data that is exclusively yours, whether you're a pharma company, a finance company, insurance company, a retail company, a venture firm, increasingly you are realizing that to get the benefit, you're giving a benefit. And you're going to say, I'm not going to give that data, especially if I'm not going to give that data and watch as one of the closed labs competes me away. And you saw that first with Figma, which was sort of scary. And I think the push now from Anthropic and OpenAI to go into pharma and pharma companies saying, 'Whoa, whoa, do we want to do this? Do we want to give them our data?' And then potentially open up competitors. And I think there's a big move right now for them to try to ingest as much information from as many non-in sources as they can. And I think that there's going to be a movement underway, whether it is industrial, manufacturing, defense, pharma, biotech, and all of our personal data that we're going to say, you know what, the models are good enough. They are performative enough and I'm going to retain the sovereignty of that in the same way that Sam does and I'm not giving it over. So I do think that that's where the value is going to occur.
So where are we in terms of people waking up versus like door closing on the like are do you think that these guys are going to get away with getting enough data before people wake up that they're going to be competitive or do you think that the game's too obvious?
I think that this past month since the OpenAI Hugging Face moment which I think there's probably three or four factors. That's one companies waking up to the token economics and realizing like how much are people spending and like what's the productivity gains we're getting. The rise of the Chinese models and then a combination of like fear-mongering and regulatory capture attempt thereof against the contrast of people that have been in the White House and been the sort of AI czars that are worrying about this and saying no no no if you use these cutting edge Chinese models but you use it on sovereign hardware and in instantiations that are effectively air-gapped you have nothing to worry about. So where are we? I think that still going back to the first point about habituation, the vast majority of people are just going to continue to use closed labs and trust it more and they're going to try to ingest more and more and more and you have whole teams particularly at Anthropic but increasingly at OpenAI that are trying to make big pushes into every enterprise. I mean Scott, I'm not sure if you can reveal like at A24 but like a lot of proprietary data and information on everything from like every pitch that you've never green lit to the ones that you have that never got made to that's a trove of valuable information.
I don't think anyone wants to like share their company data with a company. I do know a number of startups who were kind of using Claude individual accounts and you know using their Ramp card type stuff and then now have to for whatever reason because of scale go to the enterprise plan and they realize that it's like 10x more expensive and then they're realizing that they're using a lot of tokens for things that are pretty mundane and then someone on the engineering team is like I think we can do this with this open model that we can as Josh said there's no risk it's like local we can mess with the weights and do whatever we need. And so the next question is, okay, what router should we use? And it seems to me like there's a lot of router options out there and it's going to be probably pretty commoditized pretty quickly. And so you're going to all have routers that direct all sorts of spend to look more local, more efficient models. Like this is going to become the default.
My question is actually like what more opportunities do routers have to not become commoditized? Like I thought you were going to go in that direction. Like routers can start to carry logic around which models you should use. They're all going to do that. Like and they're already doing that, right? I think the question is I think was what Peter started with which is like at some point like you do have to expose something to them or you just build your own. But like who do you trust, right?
A lot of companies are building their own. It's not rocket science to build it. And Josh, I'm on the same side as you, board member at Arena. We see every week the models change in who's the leader. And not only that, Arena has a great insight into exactly for each use case, accounting or whatever, what is the actual leaderboard for it. And quite frankly, you don't need to drive a Ferrari to do the task of a certain smaller job. You're going to see a lot more of this like I believe in a multimodal world and I believe there's going to be a lot of customization and tuning and you got companies like Applied Compute that's doing this and many others as well. So I think it's going to be a pretty tough road for something like Uber Router. I just I feel like either roll your own or you got to find another way to make a model your own.
Okay, this is a great bridge. Can I bridge this? Let's talk about the crazy shenanigans Nvidia is pulling with financing these things because the question you got to ask yourself, right, is like, okay, multiplying big numbers, kind of a global commodity. You're not going to have PhDs folding your laundry. You're going to route smartly over time. Like, what percentage of things do you actually need an edge model for? It seems like you have this machines that can compute numbers. You have people who can use numbers being computed and then a lot of stuff in the middle, right? And it seems like Nvidia with this new 500 billion Wall Street crazy alliance with everyone is propping up putting a lot of computers in places like discuss.
Do you guys do you remember Lucent? I know I'm dating myself.
Give us some history lessons, Scott.
So Lucent, sorry old guy. And hey, for the record, Sam, I'm now in my 13th year independent venture capital. I was in Goldman for 13 years. So at least
Well, no, no. And then I was at a startup in the dot bomb trying to beat Google looks smart for four years.
There you go. Why just look when you can look smart? Um so okay, Lucent,
Did you come up with that? Was that your contribution, Scott?
Yeah, that explained the bust right there. 99 captured. Lucent put what? 8 billion, I can't remember what the number was into fiber.
