Sarah Guo, founder of Conviction VC, joins Sam and Shaan to brainstorm AI business ideas ranging from companion apps and video avatars to voice agents and e-commerce automation. They discuss the “AI drop shipping” opportunity where entrepreneurs with strong distribution skills can build million-dollar businesses on top of foundation models, and debate whether these wedge businesses can scale to billions.
Speakers: Sam Parr (host, co-founder of The Hustle), Shaan Puri (host, investor and former Twitch product lead), Sarah Guo (guest, founder of Conviction VC)
Introduction: The AI Drop Shipping Opportunity [00:00:00]
Shaan: There are ways to make a million bucks and then ways to make a million bucks that could turn into a billion bucks, right? This is Sarah, the queen of AI. She’s got a $100 million fund that she’s investing into AI startups. We invited her on to brainstorm AI business ideas.
Sarah: I think the market for this company is very deep because people want a lot of video. I think there’s a billion dollars of video generation revenue.
Shaan: Sarah, how hard is something like this to make?
Sarah: People are generating a million dollars of cash flow for themselves. It’s not because they’re deep AI people.
Sam: And it looks super real. Am I right? Isn’t this wild? This is amazing.
Million-Dollar Businesses vs. Billion-Dollar Businesses [00:01:30]
Shaan: So what do you mean by that? You’re saying there’s ways to make a million bucks and then there’s ways to make a million bucks that could turn into a billion bucks. You have my interest. Let’s go.
Sarah: Okay, let’s go. I think, maybe — and I don’t mean this in any dismissive way — venture capitalists are very often accused of dismissing something as a cash flow lifestyle business or whatever, right?
Shaan: Which, by the way, for anyone who is not in the VC world — you go to a VC, you say “I’ve got this great company, I think it can make $5 million in profit in year eight, and then after that maybe we can grow this for another 50 years and one day it could be a thing.” And they say, “That’s a really nice lifestyle business.” It’s like them saying “that’s cute,” right?
Sam: “That’s cute.” Yeah. I remember my first time doing that.
Sarah: Of course. The thing is, anyone who actually is an entrepreneur, including anyone who’s a VC, they know that you can oftentimes get richer and have a less stressful life if you have a “lifestyle business.” So I’d say there are many types of valid businesses. Also, a lot of things that have become very interesting start very small. I want to recognize that.
But I think a reasonable analogy is: if you can figure out internet distribution and then get super powerful models — getting increasingly powerful — to just do something useful in a niche, those two things together? That’s the new drop shipping.
You know how for maybe seven or eight years — I’m too old to know the exact timeline — but there’s a period of time where people are like, “Oh, I’m an internet kid, I’m gonna figure out some drop shipping thing and make my first $100,000.” I think this is it.
Shaan: Yes. So basically, for people who don’t really know, you could go on Alibaba or AliExpress — that was like the OpenAI in this case, right? It was this thing that exists that you didn’t have to build, but it’s magic. Watch this: you can push a button, you never had to make the product, you never had to warehouse the product, you never have to ship the product. It will just magically appear at your customer’s door somewhere between one and three weeks later. All you have to do is the marketing bit.
And kind of what you’re saying is OpenAI and the other AI companies have built this magic that basically will take a piece of text and turn it into a video or a song or whatever. And if you just do the marketing bit, you can almost drop ship a product or a service to the customer without having to make it yourself. Is that the idea?
Sarah: That is — thanks for explaining it. I think it’s easier with a few examples. Copy editing is probably a prototypical one. You can do not amazing but reasonable copy generation with these models today. So there’s a series of companies where you just have some templates that make it more obvious to somebody writing marketing copy how to use these models, and then you have a website with decent SEO, and you add some Stripe integration and you’re in business.
Shaan: What’s an example?
Sarah: I think Copy AI and Jasper — these companies started this way. And then I have several friends who have shipped AI companionship apps. Just look at paid apps in the app store by charting. Some of those people are generating a million dollars of cash flow for themselves. It’s not because they’re deep AI people.
AI Companion Apps and Character AI [00:05:30]
Shaan: When you say companion, you mean like a digital girlfriend or boyfriend?
Sarah: Right. I think people think that is skewed toward girlfriend in a way that’s not necessarily true. You can have your own ethical opinions about whether or not that’s good for people, but it’s a pretty basic human need. And people want all sorts of different things in terms of niche companionship and how you might distribute that.
Shaan: Aren’t these quietly very huge? Can we do some ballpark — give people a sense of the size and scale that these have gotten to?
So there’s Replika. I think that’s probably the most well-known one, which is a digital boyfriend or girlfriend. They kind of try to say “friend,” but I think the use case is a little bit more in the relationship side of things. I don’t remember their exact numbers, but I don’t think I would be crazy for saying they’re doing like $50 million a year in revenue. I believe she had bootstrapped it for a while, or raised very little money. Is that right? Am I off base?
Sarah: Eugenia’s built a very cash efficient business.
Shaan: Okay. Is that like a code for something? Are you an investor in Replika?
Sarah: No, I’m not an investor.
Shaan: You just said everything without saying a thing. It was basically like you’re friends with them, you know the number, and they’re killing it.
Are they killing it from your perspective?
Sarah: I think they are making a lot more revenue than most startups. I don’t think it’s fair for me to give the number. It’s not my number.
Shaan: Okay. What are the other ones that are interesting?
Sam: There’s Character AI that has some absurd amount of traffic. I’ve heard some things about it — I don’t know if this is all legit traffic or what. There’s Character AI. What are the other ones that are interesting, or what do you want to touch on?
Shaan: Can we touch on Character for a second? I think when you look at consumer companies, one of the things that I learned was that when the behavior patterns really stand out from all other products in their category or previous categories, that’s when you pay attention. It’s the dumbest metric, but it is really clear when something has special consumer behavior.
The thing that is really interesting to me about Character or the companion apps that work really well is people spend hours with them. In terms of the number of products — how many products do you spend hours with every day? Not a lot.
Sam: Social media, Shaan. That’s my product that I spend hours a day with.
Shaan: You were at Greylock, I think, when they invested in Discord. Discord was one of those things that was probably overlooked because it was mostly teenagers who play video games using this thing. It kind of looked like a chat room. You’re like, “Well, how’s it going to make money? It’s not like Slack where you can charge the company.” But the stat was people were spending like seven hours a day on Discord. Something ridiculous. Just living in Discord — it was their social life. And so you’re like, well, there’s definitely something there. And they were able to make a ton of money just even selling emojis at that point, because if you have that much engagement, you can’t fake that.
Sarah: Yes.
Shaan: By the way, I just went to Character AI and there’s an option to chat with Elon Musk. The preloaded question is “Why did you buy Twitter?” So I click it, starts a chat with Elon Musk as a character, and the first response literally goes, “You are wasting my time. I literally ruled the world.” Okay.
So by the way, according to SimilarWeb — which is like you multiply by two or three and then divide by two or three, and that’s the huge range — but according to SimilarWeb, it says Character AI has 310 million monthly uniques. Are you kidding me? That’s more than the Wall Street Journal. It’s more than a bunch of really insane — is this company really that big?
Sarah: I think people want companions. This is what I’m saying — the engagement characteristics around this stuff is real. So for anybody starting a new business, one-person company, shipping an AI companion app to a niche, generating a million dollars of cash flow for themselves.
Shaan: Do you know how these things grow? 300 million monthly visits is no joke. What’s the growth channel for something like this?
Sarah: I think that’s going to be an advantage in the future. One of the weird things about these AI capabilities is they are so novel and unique that they do drive word of mouth. For example, with Character you can make new characters and people share them, so there’s inbuilt virality there.
HeyGen: AI Video Avatars [00:10:30]
Sarah: Maybe I’ll give you two other examples of when I say the capabilities are really new and powerful and people want to talk about them. I don’t think you can engineer that, but it’s just characteristic of these companies.
One example is I am an investor in a company called HeyGen. You can make a video avatar of yourself. You cannot tell the difference. Reaching that bar of quality is new as of this past year. People create content that is unbelievable and they share it. Now HeyGen is in tens of millions of revenue. They’ve never spent a dollar on paid marketing.
Shaan: Sam, have you seen this thing before? HeyGen — this is one of those products that I’m seeing all over the place, but it felt like it was just people younger than me talking about it, so I felt embarrassed.
Sarah: This is not like the Character AI — that’s like teenagers sharing stuff, more like Wattpad or something. This is a corporate use case.
Sam: So this is basically: I make a digital AI of me or of a fake character altogether, and then it could be used in training videos, intro videos with customers, things like that. So you could basically create — you don’t have to actually set up a camera, film a video, have it edited, and then post it in order to send a video to a prospect, or send a video internally in a training system or educational product. That’s what this is, and that’s why they — it says they raised $60 million in funding. But the chart I saw was pretty absurd. They basically got to $20 million in ARR very fast.
Sarah: Holy crap. Some of the usage actually is also — yes, it’s a business use case, but it’s all kinds of businesses. Creators, SMBs, high-end enterprise, advertisers. For you guys — actually, I bet there’s a lot more demand for Shaan and Sam talking than the amount each of you can contribute. If the marginal cost of more time of Sam talking is free, you probably do more with it. I think that’s just what people are discovering.
Shaan: Have you guys used this? The landing page makes it look amazing. Is it actually amazing, or is it still up and coming?
Sarah: No, it’s pretty good. This one crosses the line — I would say, usable in real life versus cool demo. Which is the hard thing with AI. You get a lot of cool demos, then you go in and try to use it for your use case and you’re like, “How come the tweet had such a good output but mine is kind of whack every single time?” Or, “This is good, but it won’t let me change the text on it, which is what I would need to use it in my real thing.” I would say this one is definitely production ready. They wouldn’t have tens of millions in revenue if they weren’t actually usable by customers.
Sam: They just did a public campaign with McDonald’s. An advertising campaign.
Shaan: But there’s some good limits, right? You can’t be moving around — it’s like a face on camera. At least that’s what it used to be when I tried it six months ago.
Sarah: There’s some new stuff you should try. You can be walking around now.
Shaan: Oh wow. I stand corrected.
AI-Generated UGC Ads for E-Commerce [00:15:00]
Sam: All right, we’re back. Can they just — can we just upload our YouTube page, or do we have to stand in front of it and film?
Sarah: You have to stand in front of your web or phone camera for two minutes and film. It’s more of a safety thing than anything else because they don’t want people being able to take your YouTube and make you, if that makes sense. They want you saying specific words about “I, Sam Parr, say it’s okay to make this avatar.”
Sam: You said that you started the podcast by saying there’s a bunch of ways to make a million dollars that could eventually become a billion. Is this one of those companies where it started that way?
Sarah: I think — I mean, I think the market for this company is very deep because people want a lot of video. If you just think about the domain of making video — you guys know much more about this than me — but people want a lot of control. They want quality, they want specific expression and brand and motions. They want one person, two people, three people, person walking around product, whatever it is. There’s actually a lot we still cannot do with the research, and the company wants to continually push the bounds of what you can do. So I think this is a good example. There’s revenue for them or for others, but you actually have to invest in the product pretty deeply. It doesn’t mean your wedge can’t be really powerful across a single use case.
Sam: Sam, have you seen the ones that do this for DTC products? The AI for DTC product ads? Go to Arcads.ai. If you scroll down you can watch the video.
So see the video of the Asian dude who’s holding a collagen peptides thing — that’s an AI-generated video. It’s the product in his hand that’s not actually in his hand, with a script that was written. He never recorded it. And now you have a UGC, very authentic-looking ad for an influencer.
You go to the next one — look, he’s holding a different product. That’s because he didn’t re-shoot it. They just put a different product in his hand. And it looks super real. Am I right? Isn’t this wild? This is amazing.
He’s got another one with ramen. And so what he’s doing is interesting — he’s letting actual popular Instagram people create their digital twin that will be able to do branded content. So a brand can come in, request from an Instagrammer with a million followers and say, “I want you to sponsor this video, here’s the script, here’s my product.” And if I click yes, then it will AI-generate that video. I never needed to open up a package, grab a thing, take 10-20 minutes, set up my tripod, record an ad, send it to the brand, ask them if it’s okay, then they say yes and then I get paid.
Instead, in this case it’s basically: I just approve the brand, it uses my digital twin to make the ad, if I’m cool with the ad I get paid, and that’s it.
Arcads is the same thing — it’s pretty wild. In their case, these are fake actors. These women that you see promoting stuff, these people do not exist. This is an AI-generated woman who looks like a real person that is promoting some product. You script it and you can get these made.
Shaan: Or it might be a real person but they’ve said, “License it to the company, you can do whatever you want with it.”
Sam: Yeah, exactly. I think in this case they might have started with a couple of those — I think they found one of the girls on Fiverr or something. But the idea would be that people are not going to know what the hell’s real and what’s not. These look like real people in their home giving a genuine endorsement of some product that they like, and it is very simple to create.
Sarah, I think this is the type of idea you’re talking about, where two people can take the existing models, maybe customize them, but then it’s just a wedge — in this case it’s for e-commerce companies. They’re going to try to build a business here that will very quickly get to mid seven figures of revenue without much marketing spend or much of anything, just because the product is such a wow product. And then from there who knows if it can get really enormous or not.
Sarah: Yeah. And I think a piece of it is: what’s the difference between the first million and the next 999 million? It is whether or not the capability exists in the company to make the product deeper and keep expanding scope for what you do for your customer.
But there’s a ton of these wedges. Staying with visual content — you can use this category of models, they’re open source, to be fine-tuned for different use cases that are super commercial. It could be models or creator videos for e-commerce as you described. It could be renderings for interior design or buildings.
I don’t know if you guys have ever looked at a floor plan — maybe I just have terrible visual-spatial reasoning — but I can’t look at a floor plan with a couple blocks and then a fuzzy piece of fabric and be like, “Yes, I see it, that is the room. I’ll put my life savings into this.”
Our friend Pieter Levels has a thing where you take a picture of your home and then it does interior design for you and shows you mockups, which is pretty cool.
Sam: But I’d say those renderings, AI-generated, cost thousands of dollars, right? And now if you can give it to people for very little incremental cost, that’s an interesting wedge.
AI Headshots as a Business [00:20:30]
Sarah: There’s a handful of AI headshot companies making revenue. If you guys have ever gotten a professional headshot taken —
Shaan: Yeah, dude. This is actually kind of an interesting version of the drop shipping idea. I bet this would work.
So there was an ad I saw on Facebook. It was basically a guy and he had a headshot — I think of somebody I recognized, maybe a VC in Silicon Valley — basically like, “If you’re in San Francisco, I take awesome headshots. You should have a great shot for your website, for your LinkedIn, whatever. It’s good for business, good for your career.”
You click his site and it’s a bunch of people in the tech industry and it was like $300 or $400. I went to some warehouse-type place, some little photo shoot studio in San Francisco, stood there awkwardly, got a headshot made, and paid this guy $350. He was running Facebook ads profitably to do that — he was able to put in and acquire a customer for whatever $70, and he’s generating $350 off them.
Now you could run that same funnel just without the San Francisco studio, without the guy taking the picture, without any of the cost. You just say, “Awesome, give me a couple of your photos,” and then boom, here you go.
I’ve seen a couple of these go viral — viral headshot, viral yearbook ideas — but I haven’t seen too many people just running paid on them and making them work. But I’m pretty sure you could create a paid funnel that would print cash for a period of time. Maybe not forever, but an arbitrary period of time.
Sam: What’s an example of one?
Shaan: Yeah. Look at this — Aragon.ai. Dude, look at this landing page. This is genius. They just have a side-scrolling carousel, and it’s the befores — and then there’s a line — and then that same photo becomes the after. That is very well done. Looks good.
How Hard Is It to Build These? [00:23:00]
Shaan: Sarah, how hard is something like this to make?
Sarah: There are a million of these wedges. And I think that means it’s an amazing time to, as you were saying, be good at distribution, understand how to make a funnel and how to market something, and to be an idea person. Fundamentally, if you run into problems all the time, you see the right basic capabilities, you’re like, “Oh, I can think of five other use cases for this.” And by the way, you know the distribution thing.
A good example — I invested a little bit in Jasper. Jasper was started by guys who were internet marketers first, not AI researchers, not AI engineers, not even frankly very good engineers probably. They were just internet guys. They were doing something before this that wasn’t really working very well, but they had spent a lot of time building internet marketing funnels.
When they got access to GPT-3 — back before ChatGPT, just when it was GPT-3 — they got access to the API and they built Jasper. It took that same capability but now made it useful for marketers. If you’re a marketer, you needed a blog post written or an email, or you needed copy written for an ad, whatever it was — they just made a standalone tool that would do that. Under the hood, the OpenAI model is doing 80-90% of the work. They maybe customized the last mile of it.
But they were so good at internet marketing that they started running Facebook ads on this thing. It’s the fastest company I’ve ever seen get to $50 million in ARR. They got to $50 million ARR in one year, which is just absurd. Zero to $50 million in revenue in one year.
The reason they were able to do that is because their background as internet marketers — as soon as they have anything that works, they will just plow the maximum amount of cash into Facebook ads and keep optimizing the ads until they get a dollar in equals $1.50 out or a dollar in equals $2 out. And that’s why they were so successful early on — they had a different skill set than most Silicon Valley people. Most people in Silicon Valley don’t ever run paid ads. That’s just a pretty crazy thing.
Sarah: I think if we just go to the difference — the challenge for any one of these companies that gets this wedge, and it’s rare to see zero to 50 in one year, that’s pretty special — but even if you get a product to hit in terms of initial adoption, I think the next 999 million of revenue has to be more traditional.
The problem is, if you were just first with an idea and you hit it on Reddit because it’s a novel capability, I think then you need to get to traditional reasons companies get really big: product velocity, depth of product, ability to serve the customer, social engagement, or something.
If you think about companions — what are the arguments for why somebody gets to dominate that business? There’s a version of a companion business or any business with paid spend that is just a treadmill, right? I make money but I have to keep putting money in. It’s the opposite of compounding. If I stop working hard or other people compete with me, the treadmill gets steeper or I fall off.
I think one simple answer is — did you guys ever play The Sims growing up?
Sam: Of course.
Shaan: Sure.
Sarah: It’s very hard for me to not imagine The Sims being better if the characters are smarter and richer in interaction, and have what looks like realistic video and voice. Instead of it just being “I’m talking to a person,” it could be that person has some combination of memory of me, other interactions, goals, and the media experience of them is richer. We haven’t gone there yet, but I think there’s a version of that company that’s somewhere between a companion and a game world that will be very big.
Sam: It’s kind of an interesting exercise. Well, if I could just get to a million, then I’ve increased my likelihood, and then maybe I can get that to 10 and then 100 and then a billion.
I actually firmly believe that if something can scale to $10 million — it may take a while — but if it can get to 10, it can almost always get to 100. There’s enough people in the world to make that work. But getting it to a billion — that’s been hard for me to figure out how to do. But it’s a fun exercise to think: if I can just get to a million, I bet I can get to 10, and if I get to 10, I know for a fact I can get to 100.
Shaan: By the way, The Sims lifetime sales: $5 billion. Without AI, The Sims was able to get to $5 billion in sales. If you made it more engaging by AI-powering all those characters, that’s going to be even stickier. It’s going to be a big business.
Why Not Just Build It Yourself? [00:28:00]
Sam: Hey Sarah, dude, you’re pretty in the know — why don’t you just go do this? This sounds way more fun than investing in it.
Sarah: I get to — I really like doing the zero-to-one thing repeatedly. I think you just have to figure out what you’re motivated by. I am really motivated by working with people that are entrepreneurs that I like and respect and think are super special. I do not like working with people that I don’t have as much enthusiasm about. That’s a very specific personality trait. Law of large numbers — as soon as you manage very large teams, not everybody is going to be at the same level.
So doing investing and being able to contribute to other people being successful that are really special, and then the competitive nature of being right with skin in the game and knowing what is happening — I like all of that. But never say never. We incubate companies where it’s essentially like, “Oh, I see it, I see it, I see it,” and then there’s frustration that the right set of people you’re really excited to back just hasn’t come together around a certain idea.
Shaan: Sean, you are more technical than me, but you’re still not technical, I would say. But you’re more than me.
Sam: Classic compliment. Thank you very much. “You’re not technical. You’re more technical than me, but you’re not technical.” But you’re also not technical. You’re almost good-looking.
Shaan: Yeah. You’re hotter than me, but I’m a one. You’re a three.
Sam: Did you — I know you’ve been studying this stuff. This seems like a really fun weekend thing just to play with. Are these actually — would it be really hard for me to learn how to do this? Would it be hard to build one of these, just a really simple project? Because Sarah’s getting me all hyped on this. I’m like, this looks really fun to mess around with.
Shaan: Yeah, I mean, I think it’s like anything else. You’d have to have a partner to speed you up. You learning to code to be able to do these things would be the slow way. The easy way is you find an engineer who’s excited about this and doesn’t have clarity of vision around it, maybe doesn’t want to run the business side of things. And you say, “Great, let’s build X together. I have a clear idea that X will work and I’ll handle the marketing side. You’ve got to make this product do this.” And that’s not so hard. That’s pretty easy.
Conviction’s Requests for Startups [00:30:30]
Sam: This is exciting. You get to see a lot of cool stuff. Let’s do some of your specific theses. So you have this website, Conviction — good website, by the way. How’d you get that domain?
Sarah: I’m an internet person.
Sam: Yeah, okay. All right. Did you see her website? She has a website for her — I think it’s the incubator — where you’ve got to code in order to get access to it. You don’t really code, but the menu is set up like that. It’s a little CLI type.
Shaan: What’s that URL?
Sarah: I think it was called Commit — it was our program for hackathons, college students, etc.
Shaan: It’s commit.conviction.com. Sean, it’s a pretty cool website actually. You open it up, it’s a terminal.
Oh God. Let me see, let me try to do this. Slash is like a folder. I don’t know how to do this.
Sam: All right, so you have a website with a bunch of basically requests for startups — things that you think are going to be built in AI. So let’s run through some of these.
Idea: Your Personal Seller [00:32:00]
Sam: Let’s do one that you call “Your Personal Seller.” Do you remember this? You might have wrote this a while back.
Sarah: It might have been one from my partner Prateek Reddy or something, but we can certainly talk about it.
Sam: Okay. I’ll give you the summary. The idea here is that there’s a bunch of places online to sell stuff — Etsy and eBay and Amazon, a bunch of different places to sell things. But actually doing that is a bunch of work: creating the store listings, changing prices, writing the copy, all of that. And I think what you’re saying is somebody should be able to just have a product and then the AI should be able to do the actual e-commerce management — setting up the shop and running it. Is that what that means?
Sarah: Yeah, I think it matches a larger theme that I really think is exciting about AI, which is: because all of these skills — and it could be running a basic social marketing campaign, or sending email to your customers that are likely to be repeat customers, or improving your website for indexing — there are a bunch of things that are probably not related to the core entrepreneurial decision.
Let’s say it’s a Shopify drop shipping store for a particular type of sock, and you love socks as an entrepreneur. It’s not related to the merchandising decision or the design decision — like, what is the sock I want to give the world? That’s the essence of why people become entrepreneurs. So can you take a bunch of these tasks that require skills in all these different domains and just automate them, at least at a basic level? I think you can now.
The platforms — Shopify and Square, etc. — they now have native assistant products that help you use the platforms better. But I think across the spectrum of how to be a good internet entrepreneur in the e-commerce sense, I think there’s more opportunity there.
Shaan: How do companies do that now? Let’s just say you’re a company with 10,000 SKUs. How do you get accurate descriptions for all of them?
Sam: Well, usually if you have 10,000 SKUs, it’s not 10,000 unique totally variant things — it’s color variant, size variant, things like that.
Let me tell you about our case. I have an e-commerce store. We spend probably $5,000 or $6,000 a month on just Shopify Plus, the enterprise Shopify thing. That’s just the Shopify cost. On top of that, I’d say we probably have another $3,000 to $5,000 a month on Shopify apps. You need an app for search, you need an app for bundles, you need an app for this and that. There’s a ton of things that Shopify doesn’t provide. So my all-in software cost is at least $10,000 a month, probably a little bit more, on top of the fees they take of every transaction.
Then I have an e-commerce store manager. His job is just to run the store — we have new products coming up, make sure those launches go well, move things around, “oh, this is broken, there’s a bug,” whatever.
We then have a merchandiser. The merchandiser goes every day, looks at the collections and says, “This thing is sold out, it shouldn’t be at the top anymore. We don’t have sizes for this or we don’t have colors for this, so let me move this other thing to the top.” Or, “Hey, the season just ended, these need to be rearranged.” So there’s a human being that does that — there’s also apps that do that, but you kind of need the app plus the human today because the app’s not quite good enough to do it by itself.
We then have VAs that go in and do all the product pages — the descriptions, the templates, the tagging — so that our inventory data is correct. Because we need to be able to analyze our inventory, and to do that you need to tag every product accurately.
So there are four or five people just making sure the store runs, in addition to five apps that make the store run. That today shouldn’t be the future state of things. That’s just the current state.
Sarah: And Shaan, I think the future state is: for entrepreneurs who cannot recruit, manage, and pay the five people it takes to run your store — what do they do? As Sam said, I think it will be easier in the future.
Voice AI Agents for Every Business [00:37:00]
Shaan: Yeah, I think this is also similar to another area that we are — and I’m personally really interested in — the voice automation market. A lot of your listeners will have seen the GPT-4o demo where it’s a voice that may or may not sound like Scarlett Johansson, talking in real time.
Sam: Well, we played with ElevenLabs —
Shaan: No, but that’s dubbing. She’s talking about just being able to talk to it and it just talks back and sounds like Scarlett Johansson. Just like ChatGPT but you don’t have to type.
Sarah: Yeah, both of these things — either in your voice or some spokesperson for a brand or a company — but the ability to give reasoned, knowledge-based responses in a human voice, I think it’s just really powerful. I don’t think people are thinking enough about the opportunities here.
You mentioned ElevenLabs — there’s exactly one independent voice API business in tens of millions of revenue, and that’s ElevenLabs. They’re great, that’s amazing. I think there are other opportunities. There’s a company called Cartesia that does more real-time voice, for example.
Shaan: You think ElevenLabs is at tens of millions in revenue?
Sarah: They are. They’re definitely at a large number in the tens of millions of revenue. Hope I’m not surprising the market with that.
But a lot of developers will immediately gravitate toward an API business. That is not how the rest of the world works. The world is full of niches and people running businesses that don’t think about APIs and won’t use them.
Just like your personal seller, I think there are going to be a bunch of interesting voice services for everything from restaurants to HVAC companies to dental reception that are just: answer the phone.
Shaan: I think that’s one of the ideas we had. It could be informational — “we are open from 8:00 a.m. to 6:00 p.m.” — or a lead generation business where it’s like, “Well, my plumbing broke. Are you available tomorrow at 3 p.m.?”
Sam: Huge believer in this. Huge.
And Shaan, this simple idea — when the internet came out, one of the obvious things was: every restaurant just kind of needs their menu online. You should put that your restaurant exists, where it’s located, and then put your menu up there, even as a PDF. That’s still value-add for you. It became where every business needed a website.
And now what I think is going to happen is that every business needs an agent. So what’s the agent for most small businesses?
I called pest control because we always get a bunch of mice jumping in our pool for whatever reason. Try calling pest control — nobody ever picks up the damn phone. Because it’s usually run by, like, “Mike’s Pest Control,” and Mike’s out in the field doing things all day, actually controlling the pest, doing work. So he doesn’t pick up the phone. Then you leave a message and then you call 10 of them because you’re not sure if Mike’s gonna get back to you. So whoever gets back to you first — Mike loses business because Mike doesn’t pick up the phone. Mike also is not going to hire somebody to just sit there and wait for the three phone calls a day that he’s gonna get. It just wouldn’t make sense.
But now, I built one of these in our AI weekly tutoring session. I was like, “I want to build one of these.”
We have the same problem for our offshore recruiting business. We own an offshore recruiting business called Somewhere — you can find amazing talent, they’re just somewhere out in the world, and Somewhere finds you elite talent.
Now, the big problem: if you go to Somewhere.com, you say, “Okay, I’m looking for a designer,” or “I need somebody who could get me leads for my marketing business or my real estate business,” or “I need somebody to do data entry.” You have all these jobs. The button on the site is basically: “Fill out this form.” You fill out the form and then it’s like, “Awesome, we will get back to you soon,” or “Schedule a call — here’s the call tomorrow or two days from now.”
No matter how many sales agents we have, a call tomorrow is not as good as “talk to me right now” about what I need. Because right now is when I’m interested, right now is when I’m on your website, right now is when I’m not thinking about other variations of how I might solve this problem, and you have an opportunity to sell me.
Shaan, I don’t know if you’ve seen this — check out Bland.ai. This is one I built on. It lets you build a phone agent for yourself. So I went on here and I built a phone agent — a guy who could answer the phone so that when somebody goes to Somewhere and they want to hire somebody, it’ll be like:
“Awesome, what are you hiring for? Have you ever hired overseas?”
“Yeah, I have.”
“Cool. Tell me what you’re looking for in a couple sentences.”
“Oh great, it sounds like what you’re looking for is somebody who could be a developer for your Shopify store. Our normal budget for that is $2,000 a month. Would that work for you, or are you looking for something a little bit more, a little bit less?”
And then it answers — it basically does the intake, the initial sales call for you. “No problem, we’ve hired this month for 85 other Shopify brands who are looking for Shopify developers. You’re in good hands. We do this all the time. I’m going to start looking for candidates now. I’m going to email you tomorrow with three candidates. How does that sound?”
And the person’s like, “Great, I guess I can just wait for that to happen.” Or it’ll pull from our existing database and be like, “Here’s an example resume. This is the type of person we’d be looking for. Would this person fit your needs — yes or no?”
So then the human salesperson will come into work and see a ticket that’s like: “The AI agent did the intake sales call, found the customer’s requirements, kind of already warmed the sale up, told the customer what they needed to know — the things you repeat every time on the phone.” And now you could follow up with a more bespoke answer for that person.
That’s the future that I see. I wasn’t able to fully build that — I did a prototype of it — but that’s what I think websites, even like ours which is an internet business, should have. Which means that every plumbing and pest control and restaurant — they’re going to have their version of that. This is 100% way better than having a call center, or it will be, as long as it works as good. But this is absolutely the way to go.
Idea: NextGen Autocomplete [00:43:30]
Sam: All right, let’s do some more. You have another one on here — I think an easy one that’s cool — NextGen Autocomplete. The idea here is you do a Chrome extension or a browser extension that doesn’t just autocomplete and help you fill in the next word it thinks you’re going to say, or how to spell a word. What you have here is that it starts to learn your voice so it can write your emails or your blog posts in your voice. Which is kind of the next level up from autocomplete, next level up from Grammarly. It doesn’t just correct or spell check your stuff, but it actually writes the way you write because it has watched the way you write. Is that the thesis here?
Sarah: Yeah, absolutely. And I think it can be lots of different types of business communication, but especially email.
This is actually my friend Mike Vernal’s idea. I think he suffers from the same thing I do that might be true for you, which is: I’m an incredibly picky writer. I will use the models today for generation of basic content, or I’ll ask my amazing EA to draft emails for me, and then I will go rewrite the whole thing because I don’t like the tone, because it doesn’t sound like me, or because it’s not tight enough, or because I want to use a certain phrase.
I think the next level of value and impact is definitely going to be fine-tuning to specific voice. Nobody wants to write like ChatGPT. Nobody wants to be the generic AI either. So what everybody wants is the thing in between.
Sam: This shit’s all wild to me. Is there anyone right now doing that that you like? Because I would like to use this today.
Sarah: I mean, Superhuman has really interesting AI features, but I think the unlock is going to be the personalization.
Software 3.0 Thesis [00:46:00]
Shaan: What’s your overarching thesis? You have this thing called Software 3.0, which, by the way — most VC thing to do. To be like, “Oh, Software 3.0, Web 3.0.” You’ve done it. You’ve gone full VC. What is Software 3.0?
Sarah: Okay, so the seed for that phrase — Software 3.0 — comes from actually an essay that Andrej Karpathy wrote years ago about Software 2.0. The base premise is that you had to write a lot of software by hand in a prior generation before machine learning.
Software 2.0 — Andrej worked at Tesla, was working on Autopilot — was really about dataset labeling. You are teaching a machine learning model by the data you choose to put into the pipeline how to do new tasks.
Software 3.0 is the idea that the next generation of software, a lot of it is about manipulating foundation models. They’re called foundation models because they have a lot of capability out of the box. You don’t need to train them from scratch. You just need to give them guidance, reinforcement, the information specific to your business.
An example would be: Shaan was talking about for his lead capture intake form voice — he doesn’t need to go train a model. He doesn’t need to go collect data for that software application. The voice agent is a software application. He just needs to make sure it’s plugged into his schedule system and his database of candidates and be able to retrieve the right information about the business and respond consistently to customers in a certain tone.
That’s more about manipulating a bunch of this base work that Labs have already done for you. And the premise is: that last mile of getting a foundation model to be something that serves all these use cases in the real world — the research labs think of them as niches, but the world is composed of very large niches. I think it’s a really big opportunity for entrepreneurs and for us.
Contrarian Takes and AI in Healthcare [00:49:00]
Shaan: What are some of your hot takes or contrarian takes? Anything that you think might be counter to what most people say, most people do, most people are betting on?
Sarah: I’m going to give a somewhat arrogant answer, which is: I don’t spend a lot of time trying to figure out what the entire market thinks. I don’t know which of these things are contrarian. I can tell you where my opinion has changed dramatically.
Let me give you one example. For many years, including the tenure of my investing at Greylock, I was one of several people who were like, “Okay, we’re going to go understand healthcare and digital health.” And I was like, “Ah, healthcare sucks, right? It’s a quarter of the economy, it’s really important. How could you not want to work on this mission? But it is so slow and the incentives are so screwed up that trying to enter that market with technology or the speed of entrepreneurship that Silicon Valley entrepreneurs are seeking is not a good idea.”
And we just did a healthcare administration automation company. So I’m like, “Oops — changed my mind. Real hypocrite here.”
One of the reasons being: if you think about the mind-numbing work that happens in healthcare administration — billing, authorization, coding, claims processing, even not mind numbing but just expensive and manual, like patient support — it’s actually really fertile for an AI company. We backed something that’s growing really quickly in one of those domains.
Sam: I went to the pediatrician yesterday. My doctor is there with my baby, and I’m sitting there and she’s got her iPad on a table, and there’s a video there. I’m like, “Who the hell is that guy?” And they’re like, “Oh, that’s just my scribe. He’s just listening in and he’s taking notes.”
But she was like, “I used to stay up until 3:00 a.m. taking notes on all of my patients.” Oh wow, they just do it for me. Obviously the wheels are turning in my head — yeah, that job is going to be unnecessary in a few years. But it was amazing to have a medical scribe. I’ve never seen such a thing. She’s like, “Oh, this has been around for a long time.” I was like, “I’ve never seen that.”
Sarah: Yeah. I do think one framework for your listeners thinking about different ideas is: what parts of work have been outsourced to services already? Because it used to be the doctor taking the notes, and they’re like, “Wow, we pay this person a lot. They should see more patients and think about their patients more. Let us outsource that to a cheaper tech in our office. Let us outsource that tech to India or the Philippines.”
And now there are a number of scribe businesses in medicine that are growing really fast — Abridge, Nabla, Freed. It is happening. I think that will happen in a bunch of different areas where basically, if you can create separation of that work already to outsource it, then maybe you can outsource it to a machine as well.
Shaan: Yeah, Sam Altman had a good thing. He was like, “Everybody worries about AI taking your job, but that’s not the right way to think about it. It’s: AI will take your tasks.” You have to think about it not at a job level but at a task level. There are certain tasks it can do really well, certain tasks it can’t do really well, certain tasks today it can’t do that in the future it can do.
Eventually, a job becomes a bundle of tasks. But for now, you can’t think of the whole bundle because it can’t replace the whole job, but it can replace specific tasks.
Sam: Which might be just the way it works in the long run — there’s a huge slew of tasks that can be done by AI, and then there’s people that bundle those tasks together to make sure they’re getting done well or at the right time.
Sarah: I think that’s approximately right. But to be intellectually honest — there was a scribe in that outsourced BPO that had that job. So it’s not taking the doctor’s job, but it’s taking the piece of the job — the doctor’s job that already got separated out, the task that they didn’t like. But that became a job of its own.
Shaan: Yeah, the task became a job, and the job goes back to being a task. Basically, in this case, yes.
The AI Data Center Opportunity [00:53:30]
Shaan: Sarah, are you investing exclusively in AI-related businesses?
Sarah: I am a technology investor. I’m not a machine learning researcher. I’ve been working on this stuff for a handful of years and I really believe it. I think it’s the most important thing to happen in technology in a long time. But I’m also here to invest in great tech companies.
Shaan: And you’re also here to get paid.
Sarah: I am also here to work on things that will work, that are important. If an entrepreneur that I think super highly of, or that I’ve worked with before, comes to me and says, “I have a great idea, nothing to do with AI,” I’m still definitely going to be really interested in that. If you ask me what are the ideas that we think about or are hunting, it is all in AI.
I do want to put one more thing out there which is definitely not an idea that just anyone can go after — it’s kind of the opposite of the easy wedge idea. But I do really want to hear from anyone who has a point of view on what happens to the Nvidia compute monopoly, and overall what’s changing in the data center.
I don’t know if this is a hot take, but I think a lot of people in technology intellectually are like, “Oh yes, of course workloads are changing from not-AI to AI.” But they don’t actually think about what that means in terms of scale and market cap. That means chips, memory, bandwidth, networking, energy, storage, optimized system design. That was a lot of technology company market cap before. If that’s true, there’s going to be a bunch of different new specialized solutions, and it’s trillions of dollars of value at stake. And it’s not just a single direct attack on Nvidia — that’s the opportunity.
Shaan: What else would be in that category if it’s not just “hey, our chip is better than Nvidia’s chip?” What’s another shape of a company that could be in that space?
Sarah: An example would be: well, what are other bottlenecks? Memory bandwidth. What if you designed storage to be specific for AI data centers? What if you did cooling systems? If you just reimagine the entire data center around big AI inference, I think you end up with totally different needs.
Sam: The New York Times had this article the other day. I don’t remember the stat entirely, but it was something like the amount of AI capacity or chips currently created right now — we need to create another $4 trillion in market cap in order to satisfy the amount of capacity that we need.
They were writing it in a way of, “I don’t know, are we going to be able to do that?” But then when you think about it the other way around — well, in 1998, if you said how big is the internet going to be, I’m sure it went far beyond virtually 100% of the experts’ opinion as to how big it would get. I was reading this article and I was like, “That’s just absolutely astounding that we’re in one of these moments.”
Shaan: Sequoia came out with a sort of blog post or PDF about this. I think they called it “the $600 billion dollar hole” or something. It was basically saying we’ve invested this much in capex, so if you invest that much in capex, what do you need to get out to make that return? And that’s a VC saying that, which is not just some journalist who doesn’t get tech.
Sarah, what was your reaction to that?
Sarah: I think it is a lot of capex. If you put it in context of how does it compare to other big capex spends in the past — let’s say the broadband buildout. We wanted the internet. We spent about $2 trillion on broadband to date. We’re not there yet, right? That was worth it.
So what I would say is: yes, it’s a totally valid question. We’re spending a lot of money — what are we gonna get out of it? I think we’re gonna get a lot of value.
AI: Doom or Abundance? [00:58:00]
Sam: We had Dharmesh on the pod. Dharmesh founded HubSpot. Dharmesh is amazing. He’s really wise. He tends to be right more than he’s wrong. He said something great when I asked him — I’m like, “Man, I’m a little nervous about a lot of this stuff, where the world is going to go.” And he’s like, “Well, I’m fairly educated, and I think that it’s not going to be as good or bad as you think it’s going to be.”
Do you agree with that sentiment?
Sarah: No. I think it’s actually pretty bimodal. It could be much much better — so great — or it could be bad.
Shaan: It’s the opposite. It’s either going to be much worse or much better is kind of your take.
Sam: What’s the bad? What’s the bad situation look like? For example, I think the bad situation — and I’m fairly educated, so take it with a grain of salt — the very bad situation is that there’s just going to be this massive gap between the haves and the have-nots. If you have money now, that’s going to grow and you’re gonna be awesome.
Sarah: The bad situation is AI kills us all, right? That’s the doom situation.
Sam: That’s a bad situation. But in route to that, there is just this massive separation of the haves and have-nots. You know what I’m saying? That kind of freaks me out. What’s your take — where do you see the bad situation going?
Sarah: I think it is not necessarily that correlated — that your resources or your capital today means that you most take advantage of the AI revolution. I actually think people have a lot of agency in this. They can go start these businesses, make a million dollars.
Sam: That’s such a small group of people. Why does it have to be?
Sarah: Because of human nature?
Sam: How many people know about this stuff? Your parents are tech entrepreneurs, but go ask Shaan’s mom and dad. Go ask my parents. Go ask my brother and sister. People are not entrepreneurial. Even if this widens the number of people who can be successfully entrepreneurial, it’s not going to go from 0.1% or 1% of the population to — I don’t know — not 50, right? It’s not going to go that far.
Sarah: Yeah, I don’t know if it has to express in pure entrepreneurialism versus increased productivity for people in lots of different types of jobs. And it’s not obvious to me that that’s just the people who are already most highly paid today.
Shaan: You’re somebody who thinks a lot about AI, you spend your time in the AI ecosystem. A lot of very smart people are actually worried about the doom scenario, from Elon Musk to — we had Emad Mostaque on the podcast, and Emad’s a smart, thoughtful guy. He’s like, “The P(doom) — the probability of actual doom here — is pretty scary. It’s not zero.” What do you think about that? What are the odds that AI truly is a critically dangerous thing?
Sarah: I don’t actually spend a lot of time thinking about this problem because it is conjecture about the future of both the objectives of these models and capabilities of these models that are kind of hand-wavy.
When you talk to experts about some of the suggested scenarios — here are two classic ones. “Oh, people are going to use this to design a virus that kills us all.” Bioweapons. Or, “Somebody is going to make the objective for a foundation model that is super powerful to be ‘make the most money’ or ‘generate the most paperclips,’ and it’s going to take over all of the resources in the world and kill us all.”
There’s no linear path from here to there. When people ask me about the doom scenario, I am much more concerned about abuses we actually do understand. For example, what if people don’t understand what information is true or not? People are going to use this stuff for hacking and fraud and lots of bad activities today. We should go understand that and react as quickly as possible to that, and as a country probably want to stay ahead on these capabilities technically.
Real AI Threats: Fraud, Deepfakes, and Misinformation [01:02:00]
Shaan: Have you heard any wild examples of how people use this for hacking or for fraud?
Sarah: For my company, we get emails from me — it’s not really me — and sometimes it will have a link to something that sounds like it’s in my voice.
I think that’s the simplest example: what happens if you can create really authentic-sounding media? Are your parents going to not pick up the phone if it’s a spoofed phone number and it sounds like you and you say you need something? That’s a bad scenario. I think we need more tools to protect against that and general education about it.
So I worry more about that. And I’d say the probability of a bad scenario — I said it is possible. I can’t see exactly how we get there. If you ask me what are the reasons in which broad use of cheap intelligence are going to be great, I can give you so many reasons.
Andrej Karpathy just started a company around education. The fields that have been super resistant to cost improvement — basically healthcare, the government, and education — I think this will actually move the needle on some of the domains that matter a lot to all of us.
When people talk about the doom scenario, it’s really fun and scary to talk about the dystopian doom scenario. But I think the opportunity cost of not exploring the ways in which you can have an economy of abundance — we need to talk about that. And that is really what I focus on.
Sam: Sam, do you know who Andrej Karpathy is?
Shaan: No, but I love his name. It’s a lovely name.
Sam: Andrej is an amazingly well-respected research scientist and educator who’s trying to create an experience that is AI-powered in education, where the most amazing expert in a domain is a personal tutor taking you through the material interactively.
He’s one of the five big thought-leader type guys. He ran Tesla’s AI program in terms of self-driving cars. He was one of the five most known and respected guys about that. He then went to OpenAI — he’s listed as a co-founder of OpenAI. He was the early mind behind it. And at Tesla, he was basically the guy leading their entire self-driving unit.
San Francisco: The AI Capital [01:05:30]
Sam: When I lived in San Francisco — it was a fun period, I lived there from 2012 to basically 2020 or 2022 — back then it was the Airbnbs of the world, and Tesla, and Uber. We had Sidecar back then, where it was like, “Holy crap, we’re going to get in a stranger’s car, and this is so exciting. This is so new.”
You’d go to hackathons and people were working on meal delivery services, and that was really cool. I went recently — this was about a year ago — and I was walking around the Ferry Building and this kid recognized me. He’s like, “Oh, Sam, I like the pod.” I go, “Hey, what’s up, man?” He said, “I’m doing a hackathon right now in the Ferry Building upstairs. Do you and your wife want to come up and see what’s going on?” And I was like, “Hell yeah, let’s do it.”
So we go up there and it was so invigorating. I was like, “Dude, we used to do this exact same thing, but it was around the sharing economy and all this type of stuff.” I was just talking to people about what they were building, and I remember thinking, “This is the Renaissance. There’s something really, really cool going on.” And everyone was doing AI stuff. I just thought it’s magical.
When I moved there, it was mobile. It was like, “Oh, X for mobile — everything, we got to make it work on iPhone and Android.” Then you would see some false flags — FrontBack came out, it’s like, “Oh, this is the next thing, this is the next big social app.” And then it would kind of die. But then Instagram, Snapchat — they actually stuck. I remember the early days of Musical.ly that now became TikTok.
So mobile was the big thing at the time, then it became crypto, and it became the crypto hub. But it started to lose a little bit of the steam for crypto because crypto was a lot more international. But now it looks like, for AI, San Francisco at least is back as the hub.
Shaan: Sarah, are you in San Francisco?
Sarah: Yeah, we’re in the Mission in San Francisco. I think we really believe in the community aspect of it. If you’re thinking about looking for ideas for companies and being inspired to be committed to the grind and have the right ideas, then the right thing to do is not do it alone in your basement.
What you have in San Francisco are people who are optimistic and work-oriented. They believe lots of things are possible. They’re learning about what’s going on at the frontier. We actually do this grant program — embedded.conviction.com — to create that kind of community and a bunch of other stuff. It is around this idea that people want to have the experience that you described, Sam. Not all of this is going to work, but what are smart people trying? That is some version of the future in this area of AI, and that will probably educate and inspire me. And some of it will be really big.
Sam: Yeah, I think that if you’re 22 and you’re young and single and you’re into this stuff, I would just say two words: go west. Go west, young man.
By the way, people always talk about San Francisco — it’s dangerous, it’s dirty, it’s lawless. That’s the appeal, baby. You can say you made it in the war-torn city of San Francisco. You don’t want to be a billionaire who is coddled. You want to be a billionaire who grew up on the mean streets of San Francisco.
Shaan: Not that mean.
Wrap Up [01:09:00]
Sam: All right, I think we have to wrap up. Sarah, thanks for coming on. Where should people find you?
Sarah: You can just Google “Sarah Guo” or conviction.com. And I’m on Twitter.
Sam: All right, that’s it. That’s the pod. Thanks, guys.