Sarah Guo on AI investing: capabilities are still underhyped, but most companies are 'wrongly hyped'

story? I think that was it. I think it was about the backstory. Yeah. And she's here now. Welcome to the show, Sarah. Good to have you. Hey. Hey. Hi. What's going on? It's great to have you. How's it going? Thanks for having me. Uh what's uh what what's new in your world? How did you process the Studio Gibli moment?

Uh how are you feeling? I mean, you've been, you know, bullish on AF for years now. Um what's uh how is it like updated your world model and what's exciting going forward from here? Um well, it's super cute, right? Start with that.

I have a lot of and generated a ton of images of me and my kids and my family, my friends. So not be um ignored like how much people actually do want to create. I think it updates I think it should update a lot of people's world models in terms of how early we still are in quality.

I think the fact that like quality is a very broad term. I think the fact that you had image generators make things that were beautiful or realistic.

I feel like people kind of thought that you know to some degree like images were solved like other um like other domains were uh already more well solved but like what I think the GPT40 and studio Gibli moment taught us is first of all it's like a different technology for image generation but it allows a lot of controllability and it there's more knowledge of like why the image is the way it is and so I think we're going to keep seeing a ton of improvements in our ability to create different things that like surprise people.

Have you gotten a chance to dig into how much of the I mean it's clear that like this new model, whatever they're doing is much better. How much of it is just the design, the algorithm, the strategy versus like scale?

And I I really hope that we have almost like a deep-seek moment where we're like, oh, now it now these now it's a filter on Instagram and it's just a button and it takes one second and it runs on your phone. We've compressed it down. That seems to be a trend in all of these models.

something comes out, it's really cool, it's really big, it's really expensive, then it gets really cheap and quick. Um, what are you expecting this year as this evolves? And do you think any do you think any of those trends will like unlock new uh use cases?

Uh, I I mean I I think this system can do things that uh were really hard Yeah. to get right consistently before, right? Um, and it is it is different. I mean, I don't work at OpenAI, but lots of people believe that GPT40, it's not it's not a diffusion based image generator.

Instead of starting with noise and you like gradually cleaning it up the way like midjourney um models supposedly work, you you're building images piece by piece like sequentially like you have tokens in a language model, right?

And um if the image parts are more like words in a sentence um and you're creating the image one element at a time. So it's like technology brothers in a cyberpunk city with a blue dragon. It's understanding the request with language.

It's connecting the words to visual concepts it knows and then it's doing it like step by step. And this is great because it's more controllable and it's really high quality but requires a lot of compute which is why Sam is talking about servers melting and you know it's it's still slow.

Um, do I would take I would take like every bet that in general all types of AI capabilities due to competition and continued investment um, and technical progress and distillate like it all gets cheaper and faster, right?

So, anybody who's building with this stuff or using it should expect that curve to continue the way it has for the last 18 months. And I think a lot of it will be open. Um, I mean opening I just said they we're committed to this too. Yeah.

And and so like do I I definitely think that the um the the data would would suggest that we should continue to see like new capabilities as the models continue to like invest in or the the foundation model companies invest in multimodality and continued scale. Mhm.

How do you uh what advice would you give to somebody that's doing anything around image generation at the app layer?

I think that prior to last week, if you were building like something that generates Facebook ads, you felt pretty good cuz you were like, "Well, we have great UI and like customers care about workflows and like, you know, we help with this sort of like effectively like, you know, baking in prompt engineering to get great outputs and then OpenAI releases something that like consistently oneshots, you know, prompts with like three words and it's like a fantastic output.

" Yeah.

Uh do you think that um you know do you think if you're working on like sort of image generation or or ad generation SAS that you need to like pivot to just having like new ideas and like kind of thinking deeper into the stack or do you think that there's you know uh do you think that you can still kind of like stay ahead by focusing on workflows?

Um I I definitely think there's a lesson here of like watch what the labs are doing. watch what anybody who is doing like core research is working on and you don't want to fly too close to the sun, right?

Um uh I I do think one surprise might be uh what anthropic and open AI explicitly care about is like AGI or ASI whatever you want to call it, right? Um, and one might have argued that like image generation, voice, like these things are not on the path to AGI, right? They're um they're products you kind of get for free.

Yeah. Um I I think there's like maybe two lessons here for entrepreneurs. One is well these companies they do care about making like enduser progress and commercial progress in part because they need to build a business that requires consumer engagement to keep justifying the training cost and the research cost, right?

And so, um, I think if they feel like there are capabilities that fit into a broader consumer engagement or, um, like productivity platform, they're probably going to pursue them.

Um, but I I also think that like the at least my experience with creative tools is that people want a huge amount of control and like not one tool will serve everyone and and so like the things that OpenAI and the labs have produced are very democratizing and yet like that is lowering the floor for if you want something that looks like a decent ad.

I also think there's a lot of room to like continue working on the last mile and raise the ceiling and we're gonna have just like much better content, much richer content, more diversity, higher quality from people spending a lot less money and there is opportunity in that.

Yeah, I think last mile is just like an interesting framework where like you know the base models are going to get you to like, you know, pretty great and then how do you get to excellent and there's like probably a lot of value that you can capture by delivering excellent and I think you can make like reasonable predictions about what the labs care about, right?

Like um uh let's just use some examples like do they care about code generation? Yes, they care about code generation. They think it's on the path to better reasoning.

But do they want to build every workflow for a certain type of code base or migration every workflow and serve it like in customer service or law or finance or accounting or even um I don't know video capture and editing like absolutely not they neither can nor want to do that. Yeah.

It seems like there's been this uh trend over the last few years of maybe don't train a foundation model, stick to the application layer and then go operate on highly protected data sets like legal or finance or investment banking, uh medical records, etc.

And I'm wondering if there will be a version of that that relates to image generation.

Um, but do you think we're kind of at the end of that trend or do you think there's more niches that people can discover if they just think really hard about where the labs aren't building towards and then go find something really cool? I actually think we're still like pretty early in quality. Mhm.

Um like in your like artists and advertisers and communicators like the the specificity that people have for what they want in an image like you know the concept here is controllability. Yeah. Um like I think the ceiling is really high and to the point at the beginning like I think people thought we were kind of done.

We're not done. And so um I actually think there's still opportunity for large companies here. And then relatedly like video is much harder than images and um requires like different architectural change.

And so if if images are this expensive and slow to generate today like I do think also efficiency is going to matter here and workflow is going to matter here and collaboration is. And so I um I still think there's opportunity uh for other companies that are not necessarily um just training pure textto image models.

Yeah. Oh, sorry. I was going to take it in a different direction, but I I I was going to ask um it it seems like there's been this evolution of the value will acrue to the foundation model layer. Now, it will acrue at least in the short term to the application layer.

uh every foundation model company kind of has a dance partner on the distribution side whether it's uh Google obviously has their own distribution through their apps and meta has all the different uh social media apps X and XAI are now part partnered and together uh even openAI has gotten so many people to install the OpenAI app and the chat GPT app uh do you think there'll be an increasing trend of hey yeah it's not enough to just be a foundation model company you got to find a partner on the distribution side if you really want to accelerate over the long term and remain relevant.

Uh I I certainly don't think it hurts but the um like to the best example would still be open and catch. Yeah, they didn't actually like their distribution is their own. Yeah, right. But it depends on how novel the capability is.

And so I think there's a deeper question in like maybe that you're implying that in what you just asked of like well are people going to build new experiences that are so different that they can create their own distribution. It is fascinating that they're have this deep partnership with Microsoft.

I've used Outlook for one email inbox for the last two years.

Haven't seen any open AI products in I don't know if I just haven't clicked the right button, but doesn't seem like Microsoft is really pushing like, hey, we're going to build these models in, although they've talked a lot about co-pilots in different office products. But, uh, Jordy, what's your question?

Uh, do you think that it it almost feels like there's there's never been more hype for AI, but I almost feel like it's like underhyped. Like you shared an article yesterday or a study that shows like AI therapy like actually seems to work really well.

And the idea that like society broadly, everybody would have free access, infinite access to high quality therapy like should probably be like the top story in the world because of how transformative it could be if every individual could understand themselves better and sort of, you know, mature and and whatever they're getting out of therapy.

It it feels to me like just AI broadly is still underhyped even by the cohort of humans that is like you know like fully AGI pill. Um I think it's I I think it's wrongly hyped. Maybe that's the the the answer.

Um uh like it is still underhyped and the things that people are really excited about are not necessarily the right ones. Okay. Um maybe I'll go back to like the Microsoft thing for a second.

This is not like I mean clearly SA has done an amazing job with Microsoft but it's still a very large beast and I think it is like kind of impossible to overstate how hard it is to get existing products in existing businesses to move quickly, right? And all of this [ __ ] has happened so quickly.

Um and like should there be intelligence in your email? Absolutely. Right. like that technical capability exists today.

But what does anybody who is a traditional email vendor Microsoft or Google like navigating here they're like oh what does our user agreement say about if we can process our d that data what is like the safety consideration around processing that data what does the user promise around privacy right like there's um how much is it going to cost does like do we even have our infrastructure set up to share that data with model inference um I I think it's it's complicated to go deploy this stuff especially at scale with organizations that by virtue of actually having users and customers are more conservative and um and so I I think they're like people can go challenge it by yeah I mean staying on email I was I was setting up a new Gmail inbox and it's there's Gemini buttons everywhere but they don't seem to do anything that I want to do.

It's just like, "Oh, give me a twoline summary of this email that I could already kind of skim. " It feels like, you know, during the mobile era, there was this whole drum beat of like Mailbox. I don't know if you remember that company that got acquired by by uh Dropbox. It like went viral.

Everyone was using it in it introduced a couple different UI patterns that were really cool and then eventually adopted. And it was never like a power law outcome, this massive trillion dollar company, but it was like a really cool product. People enjoyed it.

It was probably a good outcome for the investors, the founders, the users, everyone had a good time. And I feel like there's a little bit of a meme right now that like, oh, you're you're just a chat chat GPT rapper. You shouldn't even go and try and build that.

And and I've always wondered like like should we be telling founders like just go build and maybe don't even worry if it's aligned with venture capital necessarily. It's okay to go build something because there's there's so much opportunity with these new tools.

You might become really big, but you also might just build like an okay subscription business and people are paying 20 bucks a month for it for a while and then maybe you get steamrolled, but maybe you land somewhere great and you get acquired somewhere and it's all worth and at least we get the product.

It just feels like again and again I think of a product and then I can reason my way out of that's not a venture bet, but we don't get the product in the interim and I'm like I just want the product right now as a consumer. Yeah.

Um, I don't think like you can give entrepreneurs advice on things that are really about like personal motivation. Sure. Right. If they want to just build the product, like go build the product.

If they want to build a hundred billion dollar company, they should think a lot about like what's going to make a hundred billion dollar company. Yeah. Um, we are explicitly not going to back anything that's going to like be a small outcome that is a tuck in at least that we believe is, right?

But I I also think that um there are spaces that people think of as like very red ocean that people have creative approaches to. What does red ocean mean? Red ocean like just overly complicate competition. Blue ocean the opposite of blue ocean. Yeah, exactly. Like a lot of people are fighting over it.

It's not like an unknown market like um you know I think I'm pretty excited for notion mail. Lots of other mail services exist as you Yeah. Uh, have you I'm sure you've looked at a bunch or gotten pitched a bunch. Have you done any AI rollups or do you have a broad thesis there?

When I when I look at some of these strategies where where sort of you know uh startup people like basically are trying to reinvent the private equity playbook being like we're going to buy businesses and make them more efficient.

Like it's sort of the same thing as like you can make them more efficient with software and now the narrative is is AI. Uh how have you thought about opportunities like that? There's obviously a bunch of them and I'm sure you've Yeah. probably seen them all. Yeah.

Um, we have something still in stealth in the portfolio that you might characterize this way.

Um, I uh I'll I'll like talk about what I think the component parts of this thesis um are and uh probably piss a bunch of people off by starting with um I think that there are asset classes where finance people are pretty good at finance, right?

private equity is a mature asset class and like a bunch of engineers being like you know what we're going to be really good at underwriting private equity. I'm like like why why would you have any alpha in this?

Like I think you seem really smart like a very good engineer but but like the component pieces of I think it's really interesting when people just like discover other sectors that have existed for a long time and don't think about what the basis of competition.

Well, that's a classic Silicon Valley approach being like, I'm an outsider. Like, I know how to do this better than you do.

And then it's like private equity is like the most optimized asset class ever where there's people out there that were like, I will buy this knowing that I will only get a, you know, 7% return, but that's okay because I'm going to do it enough times throughout the fund life cycle that it's going to, you know, work or whatever.

So, I I I uh I'm glad we're on the same page here or I'm like I'm I'm pretty skeptical of most of these things that start with like we're just better at finance. Yeah.

But being, you know, I could imagine uh hiring or actually being good at the under like the asset selection and underwriting piece of this as being like a baseline for a rollup type play working.

And then the the other premises here are like uh we think that there are a bunch of sectors in the world that would benefit a great deal from technology or AI and benefit being like it shows up in the growth rate or in the EBIDA right pretty quickly but they're not um by by uh virtue of the industry structure they're going to like adopt these technologies very slowly they're fragmented they don't invest in technology the um the user base is really far away from Silicon Valley and like sales is hard, right?

Like that's that's kind of the premise and and like it and maybe there's some broader technical thesis about why working across these companies is interesting.

Um I just said there there is something in our portfolio that you would characterize as a play on this and um I would just want to believe that these component pieces are true, right?

um that the team can be world class at finance, the team can be world class at the technology piece and the technology piece is going to translate to something that is not like incremental but fundamental to the quality of the business and your um scalability in an industry or ability to compete.

You heard it here first. It's an HVAC roll up for sure. HVAC plus docuign printing money.

Uh, how do you think about as a GP wanting to constantly learn and learn about new industries and and get enough conviction to make bets in areas where maybe you didn't have like a ton of context 5 years ago while at the same time sort of like knowing your lane and like having areas that you're super confident in and having a deep bench of sort of like experts because like the thing that I think is you know every every VC became an AI investor over the last you know year and a half two years but also a bunch of people became defense tech investors and a bunch of and like for us like we we angel invest it's low stakes if there's a founder we like we'll we'll put money in and we want to support them but then you're running a firm you kind of have to know your your edge and your lane and I'm curious how you sort of like sort of balance that like what and think about like your sweet spot yeah it's a pretty core question I was actually having this discussion uh uh with my friend Eric Vishria who's at benchmark and he you know is like an infrastructure enterprise software person by background or like me and we're talking about companies that are out of domain and his point which I think is like pretty aligned with my view is if you are too familiar with the status now I'm going to like you know make fun of myself because I just said all this stuff about the private equity rollup but if you are too familiar with the status quo um you uh the the experts in the field do not necessarily see the future Yeah.

Um and especially as you like I I genuinely I mean believe that uh large models can extend to many domains that were not software before. Like I was not particularly interested in vertical SAS or the legal industry before Harvey. But I think Harvey is a really interesting company.

I still don't know anything about the legal industry. I know like 1% of what like I should as an investor here. But like the the the point is um the the TAM has massively expanded because we are not just doing some incremental little productivity thing.

We're trying to take huge amounts of the profession and democratize it and raise the bar of expectation for what you get from your lawyers.

Um and and so like if that opportunity is large enough and I think the hammer is a really good hammer like I am so excited to learn about the legal industry and work on it for 20 years. Yeah. Um, we just got a legal bill like a day or two ago and John was like, "Oh, maybe we should have chat GBTED some of that stuff.

" And I was like, I think we're like in the last like maybe like few months before we're going to have good enough, you know, uh, legal AI, you know. I think we're going back to handshakes and uh, and smoky back rooms and threats of violence because we can't handle these legal bills anymore.

But but yeah, going back to your question you said that I think is an interesting thing to pull out. You said like experts can't see the future, but they absolutely can see today.

It's like sometimes you like, you know, look and you hear a pitch and you're excited about it and then you talk to a friend who like actually knows the industry really well and they just point out like here's like the one reason like why this is not going to work today.

Is it just you have to like step back and like almost kind of view things at a macro level and then ultimately make a founder bet? Um I I mean like why do we end up working in areas where we don't have deep domain expertise?

It tends to start with like we love the founders and there's some first principles view of like it makes sense why like we should be able to restructure the industry or the technology is going to be massively important. Yeah. Like the thesis just makes sense, right? Yeah.

Um, and so at least my personal approach is like I'm going to try to understand what every expert in the status quo believes. And usually if it's a good thesis, like somebody thinks it's viable, right? You'll just get mixed opinions back.

But um I don't want to discover that there were obvious reasons it was not going to work. I would like to understand why everybody else believes it is not going to work and then like, you know, still.

It seems like there's a lot of I mean obviously there's a lot of money slloshing around in the AI deal making space right now but at the same time it's it's easier than ever to test an idea and see oh well we got to 100 million ARR in three months like maybe there's something here.

Uh can you talk about just like the pace of play in AI both on the deal making side and then on the actual revenue making side because it feels very different than the dotcom boom where companies were going public with like $2 million in ARR and they're trading at a billion dollar valuation.

Uh some of these companies, they're at high valuations, but they have a lot of money coming in. Yeah. I mean, if any of you guys are experimenting your way to 100 million on run rate, please call meiction. com. Um I still think it's pretty hard, but uh I maybe to your point like people are doing it very efficiently.

Uh and I think it just speaks to like the technology, the magic box is very powerful. Um and you don't necessarily need it's not like if you just think about the SAS world, right? Like I need an army of engineers to implement the business logic. Yeah. Of like how does payroll work in Slovenia? Like what?

Like I'm I'm turning the law into code. Yeah. Yeah.

Um uh and uh I think that um the experiment like the iterative nature of like I'm trying to make a model produce outputs that are great in these specific workflows but the workflow is so valuable enables some of these companies to grow in an unhinged way like the ROI is there.

I think healthcare is a really interesting example of this because for you know for like a 10-year period I was like oh man like I would be really careful about you know with love to my friends who are healthcare venture capitalists like massochist not it's not a super fastm moving area huge part of the economy very important but it's just like you know it's not um it's not like an early adopter industry and yeah you have companies that um like a bridge and others that like have massive revenue new pacing including in the enterprise.

Um and I think the basic answer is like now people are making things that are worth buying quickly. So that the fundamental piece is like pretty exciting here. I think the pacing for entrepreneurs and for investors is either also very exciting or very stressful depending on your personality. Yeah.

Uh do you think do you think some of the the sort of bad activity in the space like is there so much pressure now on a like the bar in AI has now been set that like if you're a general access product and you're not like 2xing every single month or like you know not not every single month but like you know growing like extremely quickly.

Uh everybody's seen the charts of like you know whiz and cursor and ramps and and and deal and all these companies.

It just feels like the pressure is now so intense that people are like counting CR as like you know uh like trials demos like do do you find yourself giving advice to founders of being like you know ignore these sort of like uh ignore the charts and just like focus on your on your customers because if you're too focused on like how do we just grow like bestin-class then you maybe end up like not building you know building stuff that's not reliable or, you know, robust.

Yeah. I I I I think people focus more on uh what's out there instead of what's possible for their business way too much, right?

Like I think it's a pretty um so there is something fundamental in what you said which is like the world has become faster and the internet gave us the ability to distribute software products much faster than previously.

Like you were not going to go to like a hundred million of run rate in a couple months before you could discover things on the internet, right? Like what are you going to do? Like call somebody phone and be like come here with buy the CD ROM. That's how soft bank got started.

I mean or it was it was just like structurally a bit slower. Yeah. Yeah. Um uh so so I I think the the ambition for the number of connected people in the world, the number you you can serve, how quickly things can grow should structurally increase for entrepreneurs.

Um but and like an investor from the outside being like best-in-class is this company which has nothing to do with your company is not very useful.

If it is like if there is a data point in your customer segment of how quickly something can grow and you are not winning that is relevant data right then the bar has actually risen and like you should figure out how to win.

Um but I I don't uh I I think there's also like an important question like we talk about um uh uh there's another investor who calls it chicken nugget revenue but we just talk about revenue durability right like there there's a big difference between like deep usage multi-year contract enterprise revenue with integrations.

Yeah. Um and like people trying something with 30% a month churn. Is that what chicken uh chicken nugget revenue is? Is trial revenue basically? Yeah. I don't know if people like Yeah. trial revenue or just like revenue that you know is um perhaps commoditized. Yeah. Right.

It's the novelty of people trying new things because like AI can do so many cool new things. Yep. That makes sense. Not really changing their dayto-day like it's not like a 3d7 behavior where I'm somehow working better like playing more being happier.

And so, um, I I I think people should figure out what the bottleneck to growth is in their business for durable revenue and then anything else they should treat as distribution. Um, and like I don't know what to tell investors about that. Buyer beware. Buyer beware. Uh, last couple questions.

Uh, what introduction have you made that's generated the most enterprise value? uh Andrew Reid was on the show and and said that like I think he said that you connected him to Pat and and that's you know been uh pretty pretty good intro.

Is there anything else that that comes to mind uh that uh that you're particularly proud of? Um uh I don't I don't know. I make a I make a lot of introductions. I I think Rita is a very good investor. Um, no. I I think the um I think like probably created the most enterprise value.

I think there are companies that I've been a part of, be it Figma or Harvey, where I'm like giving an investor friend a hard nudge and being like, I really think this is going to work. And like this guy or gal's like good, you should work with them.

And um that included like Andrew Reed had the thesis at Figma, but I was also like you should do this, right? Like to to Dylan and Andrew. And I think I think everybody's thrilled about that. Um, I introduced a uh I introduced my husband Pat to Harvey. I think that is making progress.

I introduced a very close friend Sinking Zeb to Harvey um at GV. And so I think I'm like excited about this company, right? Um well, I'm excited. I'm excited like you know 20 years from now for the email screenshots from you of like all you know all the AI winners being like, "Yeah, I made that intro. I made that intro.

" Uh last last question I had and then we'll let you go. I know we're at time. Uh were were did you did you ever in intend on being a solo GP? You hired, you know, Mike. I think that got announced last year. I'm assuming you just had this like 30-year master plan to like build a real firm.

Uh or or or something like that, but but how did you think about when are you IPOing? I've heard uh Andre's going out. When can we expect to buy some shares? I um I don't think we have the ambitions for selling equity carriers scale. Um, okay.

So, I I think I would be much more interested in how high the hit rate can be. That's great. Um, and how high the multiple can be and how we are by our entrepreneurs. So, um, master plan. No, kind of the opposite, right?

I um I I I had a very strong instinct that I like wanted to do I wanted to do early stage investing in a concentrated way and I believe in partnership. And then also like I was in a big hurry to like get going. Um, we launched in October of 2022.

If you'd like seen research previews or yeah, played with AI or like thought about it for a number like it was it was accelerating, right? And and so I thought that there was a window to go figure out what was going on and try to be in great companies.

And so um I'm I just figured it would be better to like start figuring it out and um and like date very slowly, right?

The idea that the perfect partners would be immediately available right when I was leaving Greylock seemed like unrealistic and um it took me a year to convince Mike that he like wanted to work 110% and I'm I'm super thrilled to be working with him. That's awesome. Amazing. Thank you so much for coming on.

Fantastic conversation. We'll have to have you back soon. See you guys. Have a great day. See you. See you. Thanks a lot. Coming in next, we got Nick from Light Matter. gonna go deeper on Nick Light Matter himself in interconnect. It's a very interesting company.

They raised a ton of money uh all focused on how we can improve uh these