Jill Chase of CapitalG joins Nick to discuss Where AI Value will Accrue in the Tech Stack, How Far we are from AGI, The Role of Adversarial AI, and Enterprise Adoption of AI in 2024. In this episode we cover:
- AI Technology and Investment, LTM Nets and Potential for Fully Autonomous Software Engineering
- AI Technology, Growth Strategies, and Outlier Metrics in the Industry
- AI Adoption in Enterprise, Challenges in Measuring ROI
- Ethical Considerations Around Data Usage for Model Training
- AI Research and Its Potential Impact on Society
Guest Links:
The hosts of The Full Ratchet are Nick Moran and Nate Pierotti of New Stack Ventures, a venture capital firm committed to investing in founders outside of the Bay Area.
Want to keep up to date with The Full Ratchet? Follow us on social.
You can learn more about New Stack Ventures by visiting our LinkedIn and Twitter.
Are you a founder looking for your next investor? Visit our free tool VC-Rank and we’ll send a list of potential investors right to your inbox!
0:18
Jill Chase joins us today from San Francisco. She’s a Partner at CapitalG, a growth-stage venture firm with $7B AUM where she focuses on AI/ML, data infrastructure, and enterprise technology. .Prior to CapitalG, she was the CEO of PE-backed, Interlaced, and prior to that was Founder of Y Combinator-backed Nimble. Jill, welcome to the show!
0:41
Thank you so much for having me. I’m excited to
0:44
be here. Yeah, it’s a pleasure to connect. So you were an operator before becoming an investor. Tell us a bit about your journey prior to capital. G,
0:51
yeah, for sure, I always like to say there are kind of four things you can do in business. There’s sort of consulting or advising, there’s investing, there’s founding, and then there’s operating. And I like to say I touched them all, and kind of saw which I liked the most and was best at, and then sort of landed there. So started my career in consulting right out of undergrad in Boston, at a group called the Parthenon group, which was and out of Bain really enjoyed that, but decided after four years. You know, all I could do was make slides, which is, turns out, a very valuable skill. But I wanted more out of life than than slide making. And so I decided to go to Stanford Business School, where I wanted to sort of take two years, put my head down, try to found a company and see, you know, if I was any good at that, if I liked it. I had an idea that I that I wanted to pursue. So spent two years at Stanford doing just that, and ended up going through Y Combinator and the summer between. Between my two years, founded the company called nimble, which was actually an AI based applicant tracking system. So a little bit early on that on the AI wave after business school, I decided to take a role at Alpine investors, which is a mid market private equity fund in San Francisco that hires folks out of GSB and HBs to run their portfolio companies. So the concept is a lot of PE back companies are sort of founder led, and when they have an exit to a private equity shop, maybe they don’t want to continue to run the company, but you have folks fresh out of business school who sort of do have that toolkit and are excited to take over. And so got the chance to run interlaced, which was an IT services firm selling to startups. So we actually sold a lot of the technology and some of the companies that I look at now, I was sort of selling into and deploying into startups at the time. And then after that, my husband actually was and is a partner at red point, Patrick Chase, and I sort of saw what he was doing, and I was wrapping up my time at interlaced. We restructured the business, and sort of got it back to growth, and then landed it on another platform. And I was sort of thinking, Okay, do I want to go back into operating or do something like investing? And I saw what Patrick was doing, and I was like, All right, do you sort of like get coffee and learn a lot about interesting things and kind of chop it up with really smart, motivated people? Because I feel like that’s something I would love to do. And after seeing that, I started to take the path towards venture and landed at capital G, amazing.
3:05
So capital G not to be confused with Google Ventures, right? Tell us a bit about capital G, you know, what is the relationship to Google and alphabet you know? What are you guys up to at capital G?
3:16
Yeah, for sure. So capital G is alphabets independent Growth Fund. So two parts of that independent meaning, we invest purely for financial returns, so not strategic to Google, and then growth meaning, sort of post product market fit. I think what’s more relevant and maybe unique about us, and sort of what got me excited about joining here, after after my time at interlaced, is we are highly thematic investors. We have teams that are fully oriented towards specific categories. We cover those categories for years and years and years. We spend a lot of time and happy to talk more about this and what we did with AI, but we spend quite a bit of time with folks within the alphabet and Google ecosystem learning about things sort of technical or otherwise to help inform our thesis areas, and then we spend a lot of time meeting companies in those target thesis areas, and then we make an investment, and we make very concentrated bets. So we’ll start with, you know, a $25 million check, get a series B, and then grow that over time. And that’s sort of a paradigm that we’ve done with UiPath, with CrowdStrike, with stripe, with Databricks, all companies that we sort of are big backers of and have supported for a long time.
4:23
And what are some of those major thematic areas that
4:26
you guys, yeah, for sure. So I focus on AIS, as you mentioned. And I think actually, in the way I got into AI was when I started at investing at capital G, it was 2020, and I really had zero finance background, as I described, like I didn’t come from investment banking. I didn’t really focus on that at GSB and capital G was really taking making a bet on me. Everyone else, sort of in the growth investment field typically has some type of finance background, and so when I came in, I’ll be honest, I was a little insecure, and I thought, like, I’m never gonna be better at the finance stuff in this job than other people, because I just it’s not my passion. Area. I haven’t done a lot of it, and there are some really brilliant people at capital G and in the growth investment world more broadly. And so I thought, what is one area like? How can I be differentiated? And to me, the answer was, find something extremely technical and go spend six months learning everything about it. And ideally it would be something extremely technical that Google also is uniquely well positioned to help that for portfolio companies, and well positioned to help me get up the learning curve on and so it very quickly became clear that AI, which at that time in 2020 was actually not that cool of an investment area. There wasn’t really anyone covering it at capital G, but it seemed like an interesting thing for me to really dig my teeth into leverage the resources at Google to learn and go from there. And so as part of that, I actually ended up spending quite a bit of time with Jeff Dean and the Google Brain team to really understand, like, Okay, what was some of the cool stuff that they were doing through that I saw a lot of the work that Noam shazier was doing, actually on transformer models. And at this point it was like, early 2021 so way earlier, I think, than than most folks sort of caught on to this, which I was around, sort of November 2022 but I sort of saw the technology within Google, and I was like, damn, this is gonna change the world. And I think, you know, there are obviously, whereas a lot of that technology floating around a lot of different big tech labs, but just sort of had privy to the access through Google. And so started building out our AI thesis then. And at that time, I was like, All right, there’s basically two categories, machine learning as a service, or the old mlops stack. That was what everyone was talking about in sort of 2021, or or, sorry, that was byoml, which was the old ml op stack, or machine learning as a service, which is what I had labeled, this sort of like LLM category at the time it was lambda, was the big model out of Google, and I went sort of deep down the path of machine learning as a service, because it just seemed like such cool technology would have such interesting applications. And it led to quite a bit of interesting investment opportunity, and sort of continuing to build out that thesis
6:55
area. And what are the other thematic areas at Capital? G,
6:59
yeah. So I do AI, my colleague James does cyber security and healthcare. My colleague Kelly does fintech. We have another colleague who does sort of horizontal application software. Those are the big ones that we focus on. But we are people will sort of ebb and flow out of big thematic areas every six months. I’d say amazing.
7:18
Jill, tell me a bit about magic dot Dev, right? An AI startup who’s series A you led in February of 23 they’ve become quite buzzy. Tell us a bit about what they do. Why are they creating so much attention, and why investors like yourself and Eli Gill and Nat Friedman and Daniel gross act them?
7:39
Yeah, so magic dot Dev is building an autonomous software engineer. So if you think about the landscape for AI coding, you are no doubt familiar with copilot and Microsoft GitHub copilot, which is akin to your writing an email, and the next word, or several words pop up, and you click tab, and it fills out your sentence, and you’re delighted, and it’s helpful, and you sort of save a little bit of time, but not a ton of time, but certainly like really great tool apply that to coding. And that’s sort of GitHub copilot. What magic is doing is sort of the exact opposite of that spectrum, which is, instead of just giving you, you know, the next snippet of code and allowing you to sort of click tab and have it auto fill, it’s actually building a true AI pair programmer, so a real autonomous software engineer. You can say, Hey, can you code the front end of this application and, you know, make it in Christmas colors and make it, you know, do X, Y and Z, and it will automatically spin something up like that. Now that’s obviously incredible, if you can make it happen. And so when I first led the investment in magic. It was actually right before chatgpt came out, early November 2022 and at the time, I had been looking thematically, as I described on, I had seen the technology for llms from within Google. I knew that it would be a very obvious application for code generation, and so I was looking at all the startups in the code generation space, and I felt like I was seeing a lot of the sort of Code Complete, or versions of Code Complete. And I also felt like, as this technology evolved there with like it is, I am 100% certain that at some point in the next two years there will be an autonomous, fully autonomous software engineer that is like the same as a mid level software engineer. And so if you believe that fact, then it’s very hard to invest in any code complete company, because it sort of has uncanny valley vibes, where you’re like, Okay, these are gonna be interesting for a couple of years, but once this exists, they are not interesting at all anymore. And so when I met Eric, and he was raising the series A he sort of described exactly the vision that I had just described. And we were like, I remember so clearly being on the call, and we were like, mind meld, like, this is exactly what I was looking for. And he was excited, because I sort of got the thesis right away. And of course, you know, the rubber meets the road to believe that he could actually do that and build it, and that’s where I got a ton of conviction. I think that’s also where a lot and Nat and Daniel have all sort of gained conviction as well. The way he is building it is through a new type of model architecture, so not a transformer based model architecture, but rather something called LTM net, which I won’t get into the detail, but basically the way in which it’s architected allows for long context windows. These long context windows are incredibly valuable if you are building an AI software engineer, because it means you, the model itself, can ingest entire code repositories, which makes it way more powerful as a product, and allows you to get to the sort of like fully autonomous future, versus just if you imagine a code complete product like the reason it can only predict the next little snippet of code is because it’s only ingesting a few several lines, 1000 lines of code before that. And so it makes sense that it could only sort of give you the next very small piece. If it can ingest an entire code repository, or several repositories, then suddenly you can understand how it could code an entire application. And so we spent quite a bit of time validating the technology. We were able to get, the help of some Google Brain engineers, made the investment, and then went from
11:09
there, are there modalities, like you described, where LTM nets can thrive? Are there modalities where they fall down and they’re weaker than their transformer counterparts?
11:20
For sure. I think, you know, the transformer counterparts obviously have so many obvious strengths, particularly around sort of like next token generation, which is a huge component to a lot of what we’re a lot of the technology that we’re seeing today. I think that what you could believe is that the LTM nets are a lot stronger, or have the potential to be a lot stronger on, sort of reasoning capabilities, because, as you can imagine, with the human brain, like it’s way easier to reason when you have context on something, or are delivered full context on something. It’s way harder to reason when you’re getting a snippet of context and then you make a decision, and then you get another snippet of context, and then you make a decision, and the game of telephones, sort of like gets you to a point where the conclusion you draw maybe isn’t exactly what you had hoped. So. Again, all of this technology is is very early and evolving, but I think that’s sort of the bet you can make on the modality question. We’re focused fully on code and text right now, so not doing anything in image, obviously. Got it. Got it. Jill,
12:17
as you think about the AI stack, you know, from cloud compute to llms, to applications and everywhere in between. Lots of layers here, in what layers will the most value accrue and why? Yeah,
12:30
you know, if I knew the answer to that question with certainty, I would, I wouldn’t be on this podcast, but I’ll give you my you know, here’s what I think. I think that there’s really two layers I’m really excited about right now that I think will accrue the most value as this evolves over the next 10 years. The first is infrastructure. And I think the infrastructure layer, I mean, infrastructure has always been an interesting and good place to bet, whether it’s sort of cloud or mobile or, you know, AI like it’s it’s a great place to put capital to work because it’s an, I’ll give the AI example, like it’s agnostic to the hardware beneath. So you don’t need to bet on it being TPUs or GPUs or whatever the neck or grok or what you know. You don’t need to bet on the chips themselves. And you don’t need to bet on the end applications, and which of the end applications are going to get it right. All you have to bet on is the fact that you believe that more and more companies will build their own models and build their own use cases leveraging that leverage, AI. And if you believe that, which I think most people do, and I think it’s becoming even more clear as these open source models have come out and been, you know, effectively equal to state of the art models from APIs off the shelf, like, if you have an open source model that is cheaper and allows more customizability, customization, customization, than model APIs, I think it’s fairly intuitive that you would expect most enterprises and startups alike to start using more open source models to power their AI applications. If that is true, they’re going to need some element of infrastructure that sort of abstracts away some of the like puts and takes of exactly how you set up a cluster and manage the cluster and run the cluster, and so it’s a really exciting category that I think will end up capturing quite a bit of value. And the company I’m most excited about in that space is modal. And then the second big category that I think is going to likely to accrue a lot of value is end applications that are selling into enterprises that have thick layers on top of whatever model. So the examples I’ll give there are something like a Sierra or a decagon or a Maven, any of these, in the sort of customer support space where it’s like that there is so much profit pool and customer support, it is such an obvious application for AI in both sort of like cost efficiencies and revenue generation over time. And it’s actually like quite hard in enterprises to deploy this type of software, because the customer support software has to be integrated with your CRM. It has to be integrated with your system of record. It has to be integrated with whatever like sales platform you’re using, etc. And that’s actually quite hard when you have enterprises that have have all different systems, and have set up those systems very differently. And so even if there’s no differentiation on the model, like, even if all these companies are using the same model to power that, the actual AI features of their product, if you can build out and land in these enterprises and sort of get your teeth sunk in to like, getting integrated into all the systems and making that more of a productized journey, like you can actually capture quite a bit of value. So that’s one. And then the other is a company, something like a bridge and healthcare space, where I’m sure you’ve heard of this one, but it’s sort of, you know, doctors can have an app on their phone that listens to your interaction with the patient, and it automatically transcribes it, throws it into all the right fields in epic, etc. Again, like on its surface, you’re like, Okay, if that’s just using an off the shelf model, how is it differentiated? But it’s like, dang, that’s actually quite hard. There are, like, a lot of nuances, particularly in healthcare, from a distribution side and from a regulatory and compliance side, like you and getting integrated with epic in the right ways, and that’s actually quite hard. And so those are two examples of like. Those are thick layers selling into big enterprise categories with big profit pools that have very clear value add.
16:32
What does the thick layer mean? Yeah,
16:36
so I’ll give you, in order to sort of describe what it means, I’ll give you a non thick layer. So an example of a thin layer would be something like, and I think these, these are all amazing companies that have grown incredibly quickly, but something like a jasper or a writer or a copy like the, I think the the challenge with those is there’s not a ton like you’re basically building a UI on top of a model. And you could argue, or what you’d have to argue is, like the UI is so compelling that it out competes the other folks that are building a UI on top of the model. But there’s not a lot that’s really hard to do on top of that, like anyone can sort of build a compelling UI. And I think with the two I just described, like getting those integrations set up, figuring out the distribution angle, like, that’s hard. And so even if the AI is the same and everyone’s using GPT four, it’s still pretty hard to build the product. And so once you’ve built it, and get signed contracts in healthcare systems, in the case of a bridge, or in large enterprises, in the case of Sierra, you kind of got a flywheel effect going. That’s pretty compelling. I don’t mean like, I think Jasper and copy and writer are all like, they actually have seen growth faster than any other business. Like their growth trajectory is incredible, because I think the nail on the head, on a on a customer need, but it’s just you constantly worry a little bit about that differentiation piece.
18:05
So it sounds like it’s a function of like, how challenging the technology is, the differentiation, the number of stakeholders. Like, that’s exactly right, got it? Joe, many AI models are only as strong as the underlying data set that they’re based or trained on. What characteristics do you look for in a data set that signals a high value, long term, sustainable competitive advantage?
18:30
Yeah, I think you want, you want a few things. Obviously, volume is the number one thing that matters. Having a high volume of data is incredibly useful. You also want that data to be diverse and have as much context, kind of what I was talking about before, with the sort of context window piece. If you have a data set where all of the data is sort of steeped in context in its own right, it becomes really valuable because it’s just when the model is training like each individual token is more powerful than sort of a random token just thrown at it in a data set, if that makes sense. And then you also want it to be highly accurate. I mean, this is why things like scale or snorkel have built in massive businesses around this data piece, because they can get highly accurate high volumes of diverse sets of data,
19:20
is it? Is that? I mean, in theory, could you get better context out of audio data and and image based data? Then you can out of just pure text?
19:28
I Yes, the it’s interesting, because audio and image data is, it’s a double edged sword, because there is just way less of it, obviously, and it’s harder to generate, but it’s so rich, and it’s there is actually quite a bit of differentiation on how you use the data that you can generate. So the example I’ll give is like, if you’re if you’re training a text based LM, or if I’d say true, it’s large language model, everyone has access to basically the same amount of data, which is like the corpus of data on. Internet, there are very obvious ways in which you can also generate more text data to train your models. And like, a text data is text data. There’s not a lot that you’re doing with it to extract, like, different types of value, or you’re not being creative with, like, how you use a word or a token to make it more interesting. With image and audio data, there’s not a lot of it. It’s pretty hard to generate, but the metadata aspect of it is way more interesting. And so you have companies like world Labs, which is Fei Fei Li’s company, you know, or Black Forest Labs, which is one of the guys from stability company. Like they their claim is that they are using image and video data in a really creative way, like they are leveraging the metadata aspect in a way that’s differentiated from one another and from others in the space, that actually makes it, like, quite interesting. And so it’s like it’s a double edged sword, like you have less of it, but you can be more creative with it. I guess it’s kind of
20:55
interesting to think about, like, who might be the big players in each of these spaces. Like, I have to think I signed a waiver somewhere with Zoom, where they’re collecting, you know, every video I’ve ever participated in. And there’s, there’s a lot of virtual, you know, recordings out there that somebody is is processing, no but it’s funny like you. We recently had my old friend Mike drosh on the program at Bessemer, and we were talking about a bridge specifically, and how the models have shifted from, like, collecting audio and transcribing that and then analyzing a transcription, to actually listening to the audio for semantics, you know, for sarcasm, for, like, all this contextual data which actually changes the tone and the tenor of the conversation. It’s not just the words being spoken
21:41
totally and I think that the the concept you just described as sort of like empathetic AI on the audio side, and the speech modality in particular, is something that even on the customer support side, folks are spending quite a bit of time on. So like even Sierra Carlos, another one that’s doing this like they’re spending a lot of time. If you are representing a brand and you are going to have, like, a fully AI agent, instead of just AI assisting human agents, you need that that AI agent to really have like, sarcasm or make jokes or, you know, pause at the appropriate time, or be able to react when it sounds like someone’s really upset on the phone, you know, or be able to, like, be bubbly when someone’s really, you know, having a great time. And so I do think, yeah, there’s a ton you can do there, and I think that it will make a meaningful difference in the value of these applications.
22:34
It’s funny how, like, for many years, we’ve been talking about technology enabled, you know, or technology assisted, but it almost feels like a paradigm shift, where, in the future, it’s going to be, you know, human assisted, AI, instead of the
22:50
reverse. I think that’s, you know, I think that’s exactly right. I think that’s a huge debate. I think people have, reasonable people have differing views on that exact point, which is, and we’re, you know, we’re thinking about that a lot right now within customer support, like if you believe that technology will get to a point where having a fully autonomous virtual agent is possible in the near term, then, like, the tech stack for customer support and enterprises, looks very different than if you believe that it’s not possible To have a fully virtual agent in the near term, and instead, we should be leveraging this technology to assist humans in contact centers. Then you believe, a whole different technology stack. And so as investors, it’s kind of fun, because as growth state investors, you know, you don’t always get the chance to do as much of the sort of like creative dreaming about how the world is going to evolve, because a lot of our job, oftentimes, is a little bit of that, but a lot of like, okay, this company is already doing well. They’re already scaling. Well, how quickly can they scale? And can we make sure that they’re going to continue to be better than competitors, et cetera? But in AI, as a growth investor, you get to do that job, but you also get to think like, dang, how is the world going to evolve, and what is the right place for this technology in that incredible
24:02
Jill you’ve cited that you think you’ve cited that you think a lot about outlier metrics that set a company ahead of the competition. Can you give us a real example of these that you’ve observed?
24:13
Yes, I will give you an an example from my portfolio for investment that I made so I led our secondary investment in rippling about a year and a half ago, and the net expansion, net dollar retention, or sort of expansion within customer contracts in that business is it’s incredible. It’s like nothing you’ve ever seen. And I’m sure you’re familiar with rippling, but in case you’re not, it’s the HR Payroll Software company that sells to sort of like small and mid market companies, and the knock on on businesses selling to sort of non enterprise or SMBs, is that, like, while there are so many SMBs, it’s very hard to get to an ACV or an average contract value, or, like, basically price that’s high enough to build a big business. Because. Because SMBs, by their nature, don’t have a lot of budget, and so they’re not spending a lot on software, and it’s just hard to create build a big business that way. The thing that’s exceptional about Parker and rippling is they were able to land with sort of the HR and payroll module, which in its own right is like a reasonable ACV, but not incredible. But then they were able to build out product modules time and time again and cross sell them into their same customer base based off of the same underlying product. And that part’s important,
25:35
because what were the big early modules that they added? Yeah, they have
25:39
glow. I mean, they have a bunch of modules early that are sort of similar to HR and payroll. So like the PEO module, they have Global Payroll, which is akin to companies like deal or papaya or remote. They have financing, defined contribution, maybe, yep, exactly. They have finance, which is doing really well as sort of like a ramp competitor. And so those are the types of things that like in their own right. You’re kind of like, Huh? What? What is like rippling right to win in finance against ramp or brex when you’re selling into these SMBs. But the insight that Parker had that was incredible, that led to these expansion metrics that were, again, just like, truly, like nothing you’ve ever seen, is that he spent a ton of time upfront investing in a shared sort of data layer, or said what he has a data primitive, which is the sort of like permissions element and the like data graph of how everyone, every employee in a company, is interconnected. And if you invest in building that out, then suddenly, like, that’s helpful for HR, obviously it’s helpful for payroll, obviously it’s helpful for something like finance, because you can very quickly, in one sentence, be like, Hey, can you let everyone on our sales team have a stipend of $200 because they’re going to X, Y and Z, which would be very hard if you were just in brex or rammed, but would be quite easy if you’re in rippling, because you have that sort of shared knowledge base. So anyways, the the actual outlier metric, what this showed up in when I saw the numbers, was NDR, but it it their NDR represents just like incredible insight, on behalf of Parker, that it was really cool to see.
27:15
I have to imagine all that helps with gross dollar retention too. I mean, that’s
27:19
exactly right. And I mean, the interesting thing about about GDR and SMB businesses, again, is because you’re targeting SMBs, and a lot of small businesses go out of business. Not only usually GDR is really low, but because rippling is able to land and expand across modules like you sort of have this dynamic where the you obviously have logo churn like anyone else, but it’s very rare that someone entirely wants to, unless they’re going out of business, entirely wants to churn off of every single one of the modules. And so, exactly to your point, it helps out with GDR, for sure.
27:56
So back to enterprise. You know, as you speak with enterprise buyers about enterprise AI solutions, what is the biggest difference in their buying behavior in this cycle versus, let’s say, previous tech waves like cloud computing or ASP SaaS or productivity and communication tools? I’m curious how the buying behavior has changed and the tone of those conversation has changed when you’re trying to sell AI in at the enterprise level,
28:22
yeah, I think that there’s a huge difference between cloud and AI in enterprise that drives very different, like, incentives in buying behavior. Cloud is like, the cloud transition is hard. Like, that’s like, a very hard thing for an enterprise to do. It’s very nerve wracking, because you’re moving a lot of your data. You want to make sure there’s, like, it’s a very thoughtful process, and so it just takes a long time, like, we’re still not that penetrated, and the cloud transition is kind of crazy, like many years into it, because it’s quite hard, it’s so important, and it’s pretty business disruptive to do if you if you make a mistake with AI, it’s actually not that hard for enterprises to leverage AI or to sort of build it in. It’s certainly quite easy for someone to trial AI, either via a product or themselves, by just like depending on an AI feature, leveraging an API from whether it’s Google or open AI or whatever, to then validate product market fit to then build it out at more scale. And so the dynamic that you have is this sort of the cloud transitions, like it was a much more it’s a longer sales process, it’s a much more thoughtful sales process, but it’s much bigger contracts. You feel, from an investor perspective, like, you feel really good about the quality of the revenue because, like, you’re not like, oh, maybe I’ll try this thought transition, you know, it’s like, no, no, you’re you’re doing it, and you decide to do it, and it’s going to be multiple years across, you know, a variety of different, different applications with AI, it’s actually a lot harder to understand the quality. The revenue because it’s pretty easy. Like, first of all, it’s pretty easy, like I said, and then second of all, it’s so hyped right now. Like, I think I saw a stat the other day that was, like, earnings calls mentioning aI have gone up something like 5x in the past, like, six months, or something, where everybody’s just wanting to mention AI, to feel like they’re on the cutting edge. So they’re just saying, like, oh, yeah, we have a contract with XYZ. Or, like, yeah, we signed something with open AI. Or, you know, we’re trialing our AI feature. But it doesn’t, it doesn’t not mean anything, but it also doesn’t mean anything really. It’s just, like, still so early, and so I’d say that’s sort of the big difference is how hard it is to do, and what that implies for kind of the the quality of the revenue and the longevity of the contracts. Interesting. I
30:43
have to imagine that boards, Public Market boards, are being pressured to adopt, embrace report on ROI and KPIs for, you know, AI enabled retrofits or new platforms you know, that incorporate AI. So there’s got to be some good motivating tailwinds, at least in the enterprise sales cycle.
31:09
You know, I think the thing that’s exciting to me about AI, and there’s been a lot written about this, this is not a unique perspective, but, like, we’re just still so early in it. And so I think, yes, everybody is motivated to like I genuinely believe that the technology that we’re seeing in AI will change the world. It will actually upend like software markets as we know them. I think there’s a lot that will change as a result. I don’t think that will happen tomorrow, but I think that like the world we will live in, and you know, 1020, years will be very different for the better. And I think we’re just early in that journey, and folks are trying different AI products, but I think the big unlock will be when these sort of agents, and we talked about this earlier, but when these agents are sort of fully autonomous, you know, whether it’s a fully autonomous customer support agent that’s virtual or a fully, you know, I don’t think you’ll have the same thing happening on doctors, but things like that, like you’ll Just, I think it will look a lot different. And we’re just not there yet. The technology is not there yet. So
32:05
So Joe, I’d like to get your take on adversarial AI, we had a listener, Dan C, a friend of mine reached out, and he was telling me about a company called nightshade, which has a tool that turns any image into a data sample that is unsuitable for model training, so more precisely, it transforms images into poison samples, so that models training on them without consent will see their models learn unpredictable behaviors that deviate from expected norms. What you know? What is the future of permissioning for data that is being used to train and also adversarial. Ai,
32:41
yeah. I mean, it’s a huge topic right now. I think undoubtedly, we’ll be in a very different situation in the next several years on how data is leveraged, who’s getting, you know, who’s permissioning it, who’s getting value from it? You know, you’re already seeing a bunch of lawsuits on this exact topic. I don’t know exactly what it will look like, but I do think it will. We will move in the direction of users managing more permissions over their own data. Companies that are generating a lot of data getting compensated fairly for generating it, even if it’s sort of being leveraged in other models. So I think we’ll move in that direction. How that looks. I’m not exactly sure.
33:18
How far do you think we are from AGI?
33:20
Not far. I don’t think not that far. I really, yeah, I guess, my guess is probably three to five years. I know, yeah, I know it’s a hot take, because I think open AI recently said something closer to 10 to 15. I think, if I’m not mistaken, yeah, it’s a belief that I have having seen a bunch of this technology. I also think that people, I think there are different ways in which you can define AGI, for me, like the point at which you’re getting to virtual agents across a variety of different different fields like that, to me, is AGI, and I think that there are a lot of exciting approaches to accomplishing that, both sort of through, you know, Agent frameworks on top of transformer based models or new model architectures. Or I talked to a company the other day that was had rebuilt a digital brain that could be used in conjunction with transformer models, where transformer models are doing token generation, but the digital brain is doing the reasoning. So the most exciting thing about this field, in my view, is it’s so like, someone could come up with something tomorrow that changes everything, and there’s, like, zero linearity to it, like it’s it is such a, what I always call it sort of a small team problem, like the fact that open AI has, you know, 1000s of researchers, or hundreds of researchers, like, it doesn’t make them more equipped to solve AGI, because, like, the cumulative Power of those researchers is better. It just makes them a little bit more likely, because there is a likelihood assigned to every individual researcher. But like, each researcher could discover something tomorrow, if that makes sense, versus, like, I think, with a lot of other fields and software or cloud, like, yeah, if you have a team working on something, you can continue to build the team and make it more and more powerful. But I feel like AI research is quite different, and it makes it very hard to underwrite like when age something like AGI is gonna gonna happen? What about
35:06
asi? Is that? Is that a real eventuality? And if so, you know, I
35:11
don’t know that. To me, that I have, I’m much less equipped to speculate on my I feel like we’re pretty far from that. But I don’t spend a lot of time thinking about that to be honest. You
35:21
know, while many are excited about AI, some, including Elon, have expressed pretty extreme fears about what the future may hold with AGI and asi. Do we have reason to be fearful? Why or why not?
35:34
I think any reason we would have to be fearful is, sadly, because of human nature, in the same way that, like you’re fearful with any technology on how humans will use it, like, I think it’s sort of a similar, it’s kind of sad answer, but that’s my view. I don’t think that inherently AI will be date like the systems themselves are, will be dangerous. I do worry a little bit about what happens if you have sort of, like, gives that access broadly to people who may, may or may not have good or bad intentions. Jill,
36:08
if we could feature anyone here on the show, who do you think we should interview and what topic would you like to hear them speak about?
36:14
Ooh, I probably should say someone in tech, but I’m gonna say Biden, and I would like to hear him speak about the days leading up to his decision to drop out. Wow. Heavy topic. I mean, it would be interesting, though. I think it would be like I wouldn’t be interested to
36:28
hear, agreed, we can have Nancy on as well. And, yeah, perfect. There we go. Get them all here. Jill, what book, article or video would you recommend to listeners? Ooh,
36:37
I’ll. The one I listen to every day is the AI daily briefing. I think it’s really good. It’s like a 15 minute podcast. It just gives you, here’s the stuff you need to know from today, and it’s very factual. I love that one who does that? Nlw, it’s this guy, Nathan whitting, more. I think is his last name, okay, but it’s a good one. Amazing.
36:55
Jill, do you have any habits, tactics or behaviors that are a force multiplier?
37:00
Gosh, I wish I were more organized to say that I did. I would say my best answer to this is work with exceptional people. Like, if you get a players working with you, it just pays dividends. It’s like that that is definitely a force multiplier. And my team is truly exceptional. So that would be, I guess, my, my, I don’t know if that’s a cheap answer, but that’s what I would
37:20
say. It works. And then finally, here Jill, what is the best way for listeners to connect with you and follow along with capital G?
37:26
You can always shoot me an email at Jill Greenberg at capital g.com so feel free to always shoot me a note, especially if you’re working on something exciting or just want to learn. You know, talk more about getting into investing or otherwise, and yeah, follow us on all our social media. I guess she
37:42
is Jill Chase, the firm is capital G Jill. Thanks so much for the time and all the insights on AI. I look forward to running. Thank
37:48
you so much. Yeah, thank you so much for having me. This was awesome. Thanks. Y’all.
37:57
All right, that’ll wrap up today’s interview. If you enjoyed the episode or a previous one, let the guest know about it. Share your thoughts on social or shoot them an email. Let them know what particularly resonated with you. I can’t tell you how much I appreciate that some of the smartest folks in venture are willing to take the time and share their insights with us. If you feel the same, a compliment goes a long way. Okay, that’s a wrap for today. Until next time, remember to over, prepare, choose carefully and invest confidently. Thanks so much for listening.
