Panel: How to Get VC Investment for Your AI Startup by Nick Black


Nick Black is our moderator for today's panel, and he will introduce everyone else sitting in front of you today. Nick Black is a deep tech expert, founder, a mentor, he's a board advisor, and really an all-round tech enthusiast, so we're really excited for him to be here today. He's put together this panel.

of VCs here to answer your questions on building an AI startup and how that works in the VC world. So thank you very much, Nick. Thank you. Great to be here. So amazing to be here to host tonight's panel.

Nick Black's Personal Journey with AI Startups

This topic is very personal to me because between about 2011 and 2016, I spent five years taking my then location-based services startup into a machine learning startup. So we completely pivoted. We moved away from our existing technology. We found a new customer base.

And over five years, between 11 and 16, we built the foundational layer for machine learning, the data layer, and the application layer. And that was fun. We got to do some pretty fun stuff. But it was eye-wateringly expensive, complex, and took a long time.

The Evolution of AI Technology

One of my aha moments in the current AI wave was last winter when I started to dig into what you could do with this generation of technology. And I was just blown away by the fact that things that would take a team of my developers three months to prototype, I could prototype in a few hours. So I believe there has never been a better time to be a founder with some technical experience or a technical person who is inspired to start a company.

Introducing the Panel of Investors

And with that, I want to hand over to, if you're one of those people, one of your future partners here, we have tonight three of London's top investors all actively investing in this space. So I'm going to hand over one-on-one. You can give a quick 30-second intro, and then we'll dive in.

Evgenia Plotnikova

All right, I'll start. Hi, everyone. Nice to meet you all tonight. My name is Evgenia Plotnikova.

I'm a general partner at Dawn Capital. We are a B2B software-focused fund of which data, analytics, and AI are one of the bigger verticals. We made quite a few investments in the space. where I focus is a series A and series B, which is when you found your product market fit and you're looking to accelerate the go-to-market machine and truly scale and create a global champion, and we're all based in London.

We are investing out of our fifth fund, which is 700 million, and we still have an odd sort of 20 investments to make, so we're absolutely active in the market and looking for more.

Ben Prard

Hello, everyone. I'm Ben Prard. I'm a partner at GP Bullhound. I'm on the IC.

For those of you that don't know GP Bullhound, we're an interesting organization. We are an advisory house, so M&A and fundraising. The mission has always been to support tech entrepreneurs, and people kind of know us very well as an advisory house, but over the last 10 years, we've been developing our investment business as well. We now have close to a billion under management.

We invest... maybe slightly later than these guys, Dawn, I would say more Series B companies that have revenues, but we do have a little slice for fun stuff, so we will do earlier AI stage investments, and we have done AI at the very early stages as well. We are software predominantly B2B, but unfortunately or fortunately, AI is permeating everything we do right now. So whether we're looking at an AI company or not, one of the first questions is, how are you dealing with AI?

And for founders that don't have a good answer for that, well, from an investment point of view, that's a problem.

Adam from Episode One Ventures

Hi everyone, I'm Adam. I'm a partner at Episode One Ventures where I lead our AI investing. We're an early stage investor at the seed and pre-seed stage, mainly focused on software businesses. I also build our AI internally that we use to predict which founders to back and why.

So we have models trained on sort of 20, 25,000 founders, probably 50 or 60 variables for each one. Prior to Episode One, I did my PhD in machine learning for credit default risk at Cambridge.

Wonderful. Thanks everybody. So I'm going to dive straight in.

The Importance of Data Advantage

One of the things that we hear a lot as founders when fundraising is the need to have a data advantage. Can you explain what that means and give some good examples of data advantages that you're seeing? Jump in, whoever wants to go. Ben, you look keen.

Well, look, I'm an investor. I've been an investor for a long time. I'm a bit of an old hat, but AI is clearly a new thing. It's hard to steer away from the old kind of acronyms around investing.

But when I think about AI, I think a little bit like a gold rush, where you have, you know, people going out there trying to get rich. And often in gold rushes, the saying in investing is, you know, be the person that's selling the spade. And so when I think about AI, there's three elements to it. Clearly data is the key element to the whole thing.

For me, the data is the land. Who owns the land? That's a unique asset. The shovel is the compute.

The shovel is the computing power, the chips, which seem a little bit kind of taken right now by Nvidia. And then in the middle, there is the founders, the miners, the people out there doing the hard work. That seems very well bid already. That seems like the multiples are very high.

That's good for you as founders, not so good for us as investors. So definitely the data element and companies that have data or proprietary data is very much what we're focused on right now. So from our side, when I think of data, it really means you have some sort of differentiated product. It means your model, your product can give insights that non-fine-tuned models are not able to give.

So from our side, we see it usually in two different ways. One way is in the raw training data. So we have a company called CircuitMind that helps you optimize the positioning of components on a circuit board. How do they do that?

They have tens of thousands of pictures of circuit boards. Very hard to find those online, so the only way is to set up a camera in front of 10,000 circuit boards. It takes a long time, but it gives you a unique advantage. A different way is to have human feedback as your unique data.

We backed a company called Pandas AI, which some of you may or may not have used, where it uses a fine-tuned language model to help with Pandas code. What users can do is submit thumbs up or thumbs down. you can use that data to help with reinforcement learning via human feedback. And even with a small number of responses, as few as 10 or 15,000, you can have an order of magnitude improvement in your performance.

So that's really the way we see it. You have the raw data, and then you have the fine-tuned human data for RLHF. I guess I agree with all of those. I think what we have noticed is that there are a lot of startups that will come and pitch to us and kind of use AI for the sake of AI.

Data as a Mode of Investment

Being an AI business doesn't mean you're a great investment. You still have to create a mode of which data could be one out of many modes. And I think the way that would summarize it is absolutely having maybe some specific data that is unique to you, where sometimes incumbents, large corporates, may actually have an advantage because they have the data, they may have captive distribution, something to think about when you choose your vertical. But also perhaps the way that you access the data, i.e.

data collection and data labeling, have to be reliable, automated, and that will also sort of help create the feedback loop. And that's how we think about it.

Challenges at the Application Layer

Okay, so if we don't have this data advantage, right, if we're not able to fine-tune a base model based on our own proprietary data, should we bother? Are there opportunities for value creation at the application layer of the stack? So we talked about this at something we were at recently.

The thing with the application level is you've got these massive companies out there, these dominant probably seven companies in the US. They have bags of cash. They have all the compute power. They have the cloud to access. It's very hard for you as founders to compete with what they're doing already.

that does give you an opportunity of course because if you can come up with something they haven't done they're probably interested in looking at buying your products and adding that to their network very quickly but this is definitely an issue i mean microsoft has 120 billion in cash sitting on its balance sheet it can just take up anything that it sees out there all the roads do lead back towards these guys but 1But I think within that, having said that, where you're building models, language models that are very specific to a particular subject that go deeper. If I use the analogy, if I'm going to go and buy an outfit for the Oscars, I'm not going to buy it from Amazon. You know, I'm going to go on to a really cool fashion retailer. So if I think about that from a point of view, if I'm going to do something very precise and important, I'm probably not going to get it from a general model.

Yeah, so from our side, I think when I break it down into the three things that I would look for, you have the unique data edge. If you don't have that, the two things are unique founder market fit or traction. So even if it's a very generic product, if 2000 people are signed up for your waiting list, clearly there's something exciting happening there. And even if the technology is not super differentiated, if you have enough traction, that's often enough to get the interest of investors.

Founder Market Fit and Traction

The other element is the unique founder market fit, which I think we'll probably get to a little bit later as well in more detail. But essentially, that means when I look at you, when I look at what you've done, when I look at your industrial experience, your educational experience, are you one of 500 or 1,000 people on the planet that is uniquely positioned and capable of solving this particular problem? Or are you someone who heard about generative AI and decided to go into that business or that industry, which is completely fine, but it's a very different industry. individual or founder to someone who did their PhD in NLP in 2005 and was working on these things years or even decades before GPT was a thing.

All I want to know is how I get the Oscars invitation now.

So I guess all then and wholeheartedly agree I think for us actually we like application layer. We are actually like vertically integrated. The reason being is that I think what sometimes people forget when they come and pitch us is We will represent different stages. We also represent different sizes of the fund. I mean, obviously our responsibility vis-à-vis our investors is to generate returns.

And so sometimes the size of our fund dictates the size of the outcome as well as the size of the investment that we have to make. And so for me, everything that is hardware specific like NVIDIA or competitors to NVIDIA, anything that's cloud distribution specific or foundational models can be extremely capital intensive, require billions of dollars. I think foundational models raise close to $15 billion to date. That's hard for someone who does series A and series B, I'll be crushed by that preference stack. And so that's something that we need to think about.

And so in the applicative layer, wholeheartedly agree with the team, agree with the points made as well. But I think for us, we also really want to see real business problem being solved. I go back to the sort of AI for the sake of AI. I think we want to see the workflows be reimagined, right? So it's not just saying, all right, well, I'm an incumbent with thousands, millions of customer data points and customer service calls, for example. Well, that's great. But am I just plugging and chat GPT API on top to be a little bit more effective? Well, that's probably not imaginative. That's not revolutionary. We want revolution rather than evolution of that workflow. And so what we'd look for is, well, how do I completely rethink customer service from day one using AI as just another tool, another set of picks and shovels to make a true difference?

And I think that's kind of what we are after in addition to the data asset and the team. you I mean, just to add to that, we love application layer. The problem is since ChatGPT 4 came out, the valuations have doubled, tripled. So it's very hard for us as an investor. We've been investing in AI for the last seven years. We have five or six companies in before ChatGPT hit the public imagination. But the moment that happened, you went from a revenue multiple of maybe 10 times revenue to 40, 50 times revenue for a really cool application. which for us as an investor is too expensive. I agree. Gravity doesn't exist when it comes to generative AI evaluations.

Predictions and Disruptions in AI

But one thing I would note is that I know we're talking about mostly generative AI today, but I think AI is a much broader topic. 1It's not just about systems of creation. It's also about systems of prediction. We have done a lot of predictive AI investing in our fund as well. And so I'd encourage everyone to think more broadly and sort of beyond the hype, effectively. Okay.

So if you're not one of those thousand people who know how to build your own foundational model, and if you're not one of the seven companies who has access to all of the NVIDIA GPU space, what we're saying is a good way to go into this is to either demonstrate traction to investors, as in 200,000 people signed up on your wait list, or to show a combination of some traction along with a fine-tuned model. Is that a good summary of where we're at? I would add that just from what we've seen here tonight, there are still issues with the interaction with the consumer. It's still not easy to use. We were talking about earlier, I think there are still opportunities to build applications that make it much easier for my mum, for instance, to be able to use something like this. That's way beyond her right now. So a simplification, I think, is something that could be done. And I think for me, again, it goes back to real pain points, real business problems. I think actually I might get a bit of a challenge on the traction part. We have seen a lot of businesses have a bit of a hungover effect, so they'll go they'll grow very fast and then they'll have a very rapid decline as well. I think Jasper is actually a perfect example of a company that grew really fast and then faced challenges. And so we, at our stage, a sort of A's and B's stage, we'll look at the calorific density of your revenue, i.e. how quality driven is it? Is it overly dependent on chat GPT? And the minute that you're squeezed, your business busts. Okay.

Ecosystems and Foundational Models

So at the moment, if we move things on a bit, go to some crazy predictions, right? At the moment in the US and Europe, you've got Facebook with their alliance. You've got OpenAI, Google, and Microsoft all fighting it out to create an ecosystem on top of that. What is your prediction within the next 12 to 24 months? How many ecosystems are we going to see? And how many foundational models?

So I think the most controversial point I would make is that I think in 12 to 18 months, the hegemony that OpenAI have will be disrupted. I think if you look at GPT-1 versus 2 versus 3 versus 3.5 versus 4, you have an exponential increase in the cost per token, but you also have a logarithmic increase in the actual performance. So effectively, in plain English, that means you're paying more and more for less and less benefit. you're going to hit an asymptote.

And we're already seeing that where Mistral and their mixture of experts models are catching up very quickly. Give it a year, and the difference is going to be minuscule between these. I think the key thing with open AI is they don't have the ability really to break into the assistance model or market for phones. If you look at Google, they have Pixie, which they're already working on, which will go straight into Google phones. Apple have been really secretive somehow, but they're working on this behind the scenes. And before you know it, that will be on the iPhone. OpenAI, I think, will really, really struggle to break into that. So I think the hegemony that they have will be disrupted.

Stole my thunder on Apple. I agree. I was going to mention that, but I was going to say also, I believe in open source, actually, when it comes to foundation models. So I think we'll see more of that. I think we'll also see perhaps, I operate in the world of enterprise, so perhaps more verticalized models will become true. I think in enterprise, we're still in the very, very early innings. And what we sometimes see, you know, is six-figure POCs being signed, and it seems very, very exciting because it's so different, but it is because at the very top of the pyramid, at the board level, everyone is like, oh, do we have an AI strategy, right? And so it finally made it to the boardrooms, and so it's not that those are insignificant, but it's still very experimental, And so I think this year is perhaps the year where enterprise truly starts adopting AI. And what I'm looking for is companies that will help enterprise do that in a way that is ethical, in the way that is safe, in the way that is governed.

The Future of AI Startups

And so looking for the classic B2B software tooling applied to AI and MLOps is, I think, a very exciting area that's going to develop more this year. Yeah, so following on, I mean, so on a WhatsApp today, Dimitri, who may or may not be in the audience, pointed out that there are 100 companies in YC's, Y Combinator's portfolio that mention LLMs in their description. How many of those are feasible? How many will survive in 12 months? So the key issue that we saw with YC this year was the post-money valuation on the safe for a business that pretty much was just an idea or three months out of being an idea was $20 to $30 million. And it just makes absolutely no sense to me. No matter how good your team is, in those YC cohorts, there are five, six, seven, eight other companies doing almost exactly the same thing. I think YC is great to get a barometer for what's happening. So they run, you know, 150 or 100 to 150 twice a year. Of those, half of them are doing something in generative AI. Just go to their website, literally filter for gen AI, which you can do, just see what people are working on. And all the ideas are condensed into a very small subset, I think. So a great, I think, litmus test for whether what you're doing is unique or not is to go on YC

have a look at the latest cohort, and tell yourself, what I'm doing, is this just going to blend into this wash of 70 companies, or is it something that's going to float to the top? How many of you have watched Silicon Valley, the TV show? Yeah, I love. Do you remember there's a part where they're all on stage and like, oh, I'm building something that's social, local, and mobile. Yeah, same, but AI. So I kind of feel that there's been a lot of what we discussed on the call is AI washing. I think I am highly, the minute someone comes in and I'm like, oh, I am AI and I also have a little bit of crypto and I have a little bit of that. I'm immediately skeptical. Again, like for me, I don't know how many will survive, but I'll come down to, is this a team that were technical experts before generative AI were cool? And do they understand the business, the pain point that they're actually solving?

Yeah, I mean, I think that we shouldn't get carried away. We're very early stages of this revolution. It's almost thinking back to the computers in the 1950s, the size of this room to do very basic calculations. I think there's years and years of development. I don't think there's massive jumps in technology. There will be small... evolution, small jumps in efficiency. And I think for the people in this room who have their finger on the button, for those small changes that eventually takes a big computer to a mobile phone in your pocket, that journey will offer lots and lots of opportunities. But I think you have to have the skills, as you say. I don't think you can just show up. you need to have some background and really understand that evolution and where it may be, not next year, but maybe five years down the road and what's the building blocks to that. And honestly, that's what's awesome about it, right? That's why we're all here. Because let's be honest, for the last five years, SaaS has become kind of boring. It's been a bunch of people who know how to execute going in and executing well. And for anyone who likes to be hands-on with technology, there hasn't really been much tech innovation space in SaaS. And that's what's, I think, thrilling the audience in places like this.

Would someone be able to help with the mic?

Verticals Ripe for Disruption and VC Bias

And then just generally, for anyone on the panel, what verticals would you see are ripe for disruption beyond maybe what some of the kind of obvious ones that we kind of hear about every day? In terms of verticals, health is a big one for us. Clearly, there's so much money spent in the health sector. Really, technology hasn't really done a lot in the last few years.

It's a fantastic question and we have 10 years of data and I have gathered gender data at least according to the LinkedIn pronouns for each one of those individuals. And if you control for education, you control for experience, you control for how technical you are, what the data showed is that there is significant bias in VC where the female founders with the same background as the male founders were half as likely to raise money. So clearly that's not representative of your skill. I don't think it's a selection bias, that there's something else going on.

So the first step is to have the model to actually explicitly be able to quantify and show that this bias exists. And that allows you then when you're going into meetings to help and correct that bias.

Health and Mental Health Innovations

This is an amazing opportunity to use data to really help people's lives, both in physical health but also mental health. So that's a big area for us. The overlap there with sports as well. Sports.

Sports, yeah, yeah, yeah. Happy to, by the way, I agree appalling in general in terms of female founders backing. So hopefully that changes.

We tried to host an event for B2B female founders, just finding enough people that were series A and series B stage was so little. So I hope that changes over time to fit into the fabulous data model.

I think for us, utilizing AI to disrupt certain sectors of real economy that weren't touched by innovation before, which health could certainly be one, where in COVID, digitization suddenly touched those sectors. And, you know, say in factory floors, people were forced to use technology in a way that they haven't had before, where population may be aging. or people don't really want to do certain hazardous jobs before, I think could be quite interesting.

Closing Remarks

So anything where in real economy, there's been pen and paper and aversion to tech, I think AI could potentially change. Thank you to Yevgenia, Ben, Adam for sharing all that this evening. We're going to hang around a bit and join us with the pizza. So I hope to see you come up and ask some more questions. Thank you, everybody.

Done? Mark to tidy your library & unlock trivia