This is Episode 8 of The OpenVC Podcast. In this episode, John Spindler shares his journey from founder to investor, overseeing £85M+ in investments across 150 startups. With a 60x angel investment win, he breaks down AI startup evaluation, frontier AI opportunities, and key investment strategies
Harrison Faull (00:00.108)
Hi, John. Thank you for joining us on the first season of the Open VC podcast. It's amazing to have you on the show. You're extremely accomplished as an angel investor, as a VC, and also as a founder. So I'd like to take you all the way back to your earliest entrepreneurial experience. How soon did you get started? What was your first entrepreneurial venture?John spindler (00:11.63)
Thank you Harrison.
Well, I didn't, to be honest. I mean, it's an interesting thought for basically for my generation, getting brought up in the 80s and 90s. What do we think a founder was? And I think it's to some extent not changed that much in a mass population. I was brought up in Yorkshire as a founder was a person that built a lifestyle business, basically a trader effectively buying low, selling high. If there's any differentiation on the service.
And it was almost, you know, and the type of entrepreneur was a person who bought low, sold high. And that actually didn't really appeal to my mother, did that. She ran shops and things. so I didn't come in, the thing that can spark my kind of entrepreneur journey was actually the dot -com boom. Because what the dot -com boom got me to think about is you could be an entrepreneur and also to be creative and create something that was impactful.
And I think I still, that doesn't come across a lot for tech founders. Basically, we're trying to ride a new wave based on a new technological breakthrough. A very different from your standard entrepreneur that are looking to optimize the present system. And I think it doesn't come across a lot in media or in basically, especially mass media. You still have this...
John spindler (01:52.907)
with popular programs like Apprentice and others. an entrepreneur is someone who basically buys lemonade for $1 and sells it for $4. Yeah, yeah, yeah. so that was my kind of opening. And I suppose I was in my mid to late 20s when the dot -com boom started and
Harrison Faull (02:06.859)
can get a lot more resizing than that, especially in the digital space.
John spindler (02:21.099)
I got involved in it towards what I thought was the peak, end of 98, 99, and obviously it all blew up in the first quarter of 2020. But at that time, looked like web.1.0, with the development of basically websites, basically, we've just been talking about things like Chrome and browser technologies, et cetera. It all seemed to be a possible.
And my first business was really a software business. It wasn't that exciting. I tried to get into the more exciting, we were pitching like hell, but we ended up doing SQL databases, building them for order. And we did reasonably well. And then that business got sold in 2003, 2004. After the dot -com boom, partners wanted to go off and do other things because of myself. And we sold it for book value. So, you know, never raised any money.
solid for book value, but it was okay actually, conditional, kind of et cetera. of, I did a consultancy business, which was actually quite successful and made a decent amount of money. Sometimes I call that venture building, because we did not do venture building as we described at the time, but what's what we were effectively now looking back doing, we're getting paid to put together and set up businesses by corporates and sometimes governments, et cetera.
And then I went into, I got bored of that, got frustrated. And I think you do if you're gonna be building something for someone else. I did know software and I did know the web. And stupidly I met a friend who was building a hardware business and I went full in and invested life savings at the time that I'd made from previous exits, et cetera, into a hardware business, assuming that they had similar.
projectories from media and basically software. Obviously it doesn't. The company was called Ferguson Hill. We were building audio equipment. Initially this was a day of audio equipment where you just had basically the emergence of iPhone just come on basically. We were building across the board and we did okay.
Harrison Faull (04:24.176)
What was the product?
John spindler (04:47.943)
The product was never perfect. We had many different iterations of it. We sold through the likes of Apple, But the 2008 crash, as well as some massive mistakes we made, kind of really killed the business as a really going concern by around 2011 -12. And I suppose that's when my story is into tech and what we mean by tech now started.
One of the things I never came across when I was in the 90s and 00s was the techniques and methodologies that were developed after the dot -com crash. But basically meant that starting a business was a very different proposition to what it was seen to be before. So you had the developments of people like Steve Blank and Eric Reese and the emergence of accelerators, cetera, et cetera, taught those techniques, techniques like customer discovery and all of those.
Back in the 90s and early noughties, we would write enormous 120 page business plans saying everything would just be like that. And often your first person you recruited would be your CFO. Kind of like to, you know, et cetera. And at a different age and obviously Steve Blank's book, Four Steps to the Epiphanies showed that a startup is actually not a scaled down corporate, cetera. It's a very different beast. And that those methodologies came across. And actually the...
John spindler (06:16.645)
I do remember a friend mine called Sal Varani, basically who had been working in Silicon Valley and came to London in 2009 -10, invited me to an event in a part of London called Camden on a Saturday morning where there was Eric Rees talking about lean startup. And I remember, this is my third business by then, kind of like, and I'd never heard of it and stuff like.
had you really done any customer, had you really spoken to your customers and worked out what they wanted? And particularly in hardware businesses where they're engineering -led, often, even today I see other hardware businesses don't do enough customer discovery. And the problem with hardware is interesting because people are saying now why doesn't investors invest in hard tech like hardware? Obviously the margins are very different. You have to have stock, have to lot of upfront tooling. But the big problem with hardware compared to software.
social platforms, etc. is the time it takes you to learn and the cost of learning. So when you iterate and produce a product in hardware, just like in software or in social or whatever, it's probably likely to be wrong. You know, you probably want to iterate and adapt it and develop it, but in hardware that's incredibly hard because you've committed a production run of thousands and thousands of a product coming down production runs sitting in warehouses, which you know isn't great.
John spindler (07:40.91)
probably over engineered, probably overpriced and you can't sell it. And that was some of the problem we had, et cetera, on that. that's really hard. And obviously this was also an interesting time because what happened in the Northeast was China entered the world market and China initially owns the world market as contract manufacturers. So you can see this today with Foxcom still doing that for people like Apple is that you've got the Chinese turn up and say,
the cost of putting together a factory production run in Europe, enormous amounts of money. And there wasn't much VC money around at those times. But if you partner with us, our government will cover that cost. And you don't have to pay a fraction of the cost to do it, some of the tooling. And also they would fund things such as your cash flow. So they would fund the production run. So you wouldn't have to pay for the production run until you actually sold.
John spindler (08:38.786)
And therefore that means drove down the prices of hardware considerably, massively. Because previously you would have to basically sell a hardware product, make enough margin so you could build the next one. But now they were funding, the Chinese were funding everything upfront. You could drop your prices and if you made a 10 or 20 % margin, it was really profit to fund growth because the cost of the production was
Harrison Faull (09:11.348)
It goes against conventional angel investing wisdom today where people are told to run away from hardware startups because it's so risky. And actually being a contrarian investor, maybe there are opportunities in that space.
John spindler (09:23.847)
I think it is and it's still people do, know, early stage companies still do contract manufacturing. So you rely on someone else with all the manufacturing facilities, expertise, logistics and cash flow to cover your initial costs. It's, you know, the Chinese, as far as we understand it, have stopped doing the Wolf & Jarl cash flow, which has kind of, I think they need to do it as much.
But yes, it still has that of issue. people don't realize as well when you're selling through platforms or into retail shops is they take such large margins. You the sales distribution. You actually don't get paid until you've sold. So you spend all the money of production, marketing, et cetera, to get it into say, well -known shop or well -known online platform.
John spindler (10:22.534)
and they will pay you 60 days after the first item has been sold.
And so it's very difficult, the cash flow, the amount of money required. People say, well, why do small VC funds fund hardware? It's difficult because we don't have enough money. If we were, know, roughly a million dollar investment in a software company, you'd probably need around five to 10 times that for a hardware.
Harrison Faull (10:50.134)
And there's probably less upside as well. So there's a lot of reasons to not go into that space. And actually after all these lessons that you've learned in the hardware space, you became an investor. started as an angel investor, had a very successful initial investment out of one of the first eight investments that you made that was up 60 times on your original investment.
John spindler (10:52.909)
It could be less upside.
John spindler (11:11.332)
Yeah, I mean, that was a complete accident to be honest. I went into angel investing because I had some cash. And because I was working at Capital Enterprise, I CEO, and we were supporting accelerators. And I was seeing all these interesting startups coming about in 2011, 2012. And yes, most of my peers were saying, why don't you put your spare cash in a bank or let property like everyone else is doing? And I said, no, I'll go and do. And I quite frankly didn't know what I was doing.
And look, and actually, you know, my first eight investments, if you came back to me two years later, I think virtually all of them were dead or underwater. But one of them persisted. And it looked like it was going nowhere. For about five, years, I forgot about it. And then the founder got back in touch and basically said, this time I've been doing X and Y, we've reinvented it. The market has changed, blah, blah, blah.
And now I can say, it's success. But I think how I learned to invest was the experience I had setting up and running London Co Investment Fund. And that gave me a much more deeper knowledge of how to make investments at the precedency level.
Harrison Faull (12:31.894)
So for those that don't know, the London Co -Investment Fund, it deployed around 23 million pounds, but then there other co -investors that came in and I think they deployed 80 or so. So the collective group made over 100 million pounds worth of investments into...
John spindler (12:40.161)
Yeah, I want them.
John spindler (12:45.246)
Yeah, I think it was actually about 121 million, I think, but it's roughly in that figure overall. then obviously we did some follow -ons as well. it's probably another five or 10 million on top of that.
Harrison Faull (12:58.889)
It must have been one of the first government funded VC programs in the UK. So why was the London mayor incentivized to create a London Co -Investment Fund?
John spindler (13:13.224)
We had a pot of money that wasn't theirs. It was given to by the then British government. So was given to all the regions, not just the London mayor, others as well. And it was there to stimulate private sector investment. And you're interested now because we have a new government that's saying it's all about stimulating private sector investment. And at the time, there was a big shortage of capital.
in the scene and the idea that we came up with is people were finding it possible to raise a hundred, two hundred thousand pound equivalent now would be somewhat like half a mil three quarters but they couldn't get much further and it was difficult and so we came up with a program where we would top that up for every pack you know and what the genius of the scheme ended up becoming because we end up realizing very quickly that the best thing and I'll say this to others
co -investment funds are popular around the world now is source your own deal flow because I think that's so important all the best deals we did was sourcing our own deal flow and not relying on our co -investment partners because there's a general tendency not always and it's not kind of thing but if the deal is really really strong and you have the available money and there's more money available now you want to take all that round it's only when the deal is maybe a little bit
John spindler (14:38.078)
that you would then go to others. Now that's not always the case and I think deep attack deep deep attack.
Harrison Faull (14:41.505)
So you think it's a negative signal if multiple investors go into the round because there isn't enough conviction for one fund to take the whole round?
John spindler (14:49.18)
Now what?
I think in many ways it's the only way you can do it because there's not enough funds to take the kind of thing, especially the bigger funds that are coming early. I think if you do them now, you're probably very much stricken to pre -seed. We did seed for instance, rounds up to two mil and two mil ends, going to about four five now. But I think what I'm actually kind of saying is that if you source your own deal flow, the co -investment gets into the good deals and we set up a whole operation.
John spindler (15:20.703)
to see around 5 ,000 startups across London. And then we would take them to our co -investment partners and they would basically look at them and say, I will basically complete the due diligence, do that deal. And we were already in. that, well, we were very lucky because we had a whole, sorry, I was running a capital enterprise at the time. So capital enterprise in 2013 had managed to find a way of taking unspent EU money.
Harrison Faull (15:33.944)
So how are you finding these 5000? How are you building that deal flow?
John spindler (15:50.686)
that was allocated to the UK regions and using it to fund and set up and fund accelerators. So we were the first funders of Techstars in London, first funders of Seed Camp of Entrepreneur First et cetera. And we kind of hacked a kind of funding that was supposed to be like X. We said, well, you could do it like Y. And that basically meant that the London scene in 2012, 2013, 2014, 2015.
John spindler (16:19.111)
was very much accelerator led. And it was partly accelerator led because it was dragging, it was sucking in talent from all over Europe in particular, as well as the country. And accelerator is a great way of basically taking founders from one area and embedding them in an ecosystem. so Seedcamp had already been doing that, but we enabled them over five years, given sufficient money that they could go from
a very small team to quite an extensive team and basically do much more basically investments and make their business model work. And entrepreneur first again, we helped them for two, three years. We set up new itself, it's like Space Camp, which basically still going kind of. And so that gave us all this deal flow. And we also had lots of people coming in because we ran on the back of that, we ran a program of investment readiness. So for people that didn't get into accelerators.
who were still kind of coming into the system, but didn't want to wait three, six months because they were mainly episodic, know, kind of. So, so that gave us lots and lots of deal flow. we, you know, we saw lots of great companies and invested in some of them and unfortunately didn't invest in all of them. One that we, to this day, very apt that we decided we didn't do what you should have done. there was, know, Queen Cup and Lip was a company at the time called Revolut.
Harrison Faull (17:25.017)
very
John spindler (17:47.526)
But we kind of, those are now co -investments, I won't mention the co -investor, kind of somehow managed to knock it in for mistakes. Actually, we actually wanted, it was a whole mess. But anyway, kind of like, you know.
Harrison Faull (18:00.212)
You can't leave with your anti -portfolio. Give us some of the winners.
John spindler (18:04.859)
Well, we did companies that have kind of pretty well. We've had about 13, 14 exits on that and companies that doing well for us today and people like Curve and Cognizism and Ably and Wealth, Kernel and kind of, we also rode a lot of the kind FinTech wave of that period. And others as well, we Medical.
Harrison Faull (18:31.867)
It was the PayPal -Mathio effect, wasn't it? Silicon Valley had the big PayPal exit and then the early team there went and created loads of other companies. London, being such a financial hub, had a lot of financial experts. And the earliest success of the startup scene was the Fintechs. Those are the ones that we're now seeing that have done incredibly well.
John spindler (18:50.532)
Yeah, yeah, yeah, I think so. I think it was partly the talent that was here. Partly the money as well. know, money in London to this day, to Britain is still kind of, you know, disposable surplus money, especially at the angel level, it's still in the financial services and people tend to invest in what they understand and what they can see. And I had to say that the regulation is very good. You know, it was open.
It was keen, there was people in government pushing for new entrants. The Financial Conduct Authority was much more encouraging than they are now of new entrants. And I think we built a platform of good companies. But today, our work is quite a bit of money. I think we also attracted a lot of Europeans to...
to come here. yes, mean, mean, London Co Investment Fund, we were doing 30 to 40 investments a year. And we followed quite a lot of the trends, FinTech, we did a lot of B2C platforms, know, marketplaces, which I must admit, hasn't worked out as well. We did some e -commerce subscription, which was popular at some of those stages. we did quite a, it was quite broad and we did...
And actually one of reasons I ended up doing AI Seed is that we had these AI companies coming out of our accelerates and our universities and they didn't fit at the stage the popularity for things that could get into revenue very quickly and that were not technically that difficult. And so we found that these companies were finding very difficult to kind of raise money. I think we had only really like 10 partners, maybe one or two partners that would touch AI, but there was Nathan Bennett, Playfair.
He was at PlayFair down in Scott. Yeah, it wasn't very hot yet. I want to do AI. didn't understand. It's hard. What they want to do is AI when it stopped being tech when there was no technical risk. When it becomes when you could apply SaaS metrics to it and say, kind of a but
Harrison Faull (20:44.421)
because of investor appetites. Investors didn't, they didn't, wow.
Harrison Faull (21:01.531)
I suppose it is a bit alien, right? The first wave, seeing that trend before, it's never obvious until it's a bit too late and the wave is popular and everyone knows about it. So you broke that trend early with AI Suite.
John spindler (21:12.391)
And yeah, that was definitely the case. There was not as much appetite for AI. And lot of AI companies, and you get this almost to, you've got companies that were AI, anything like it, would call themselves AI because they thought it was trendy. And they therefore basically kind of made, kind of you'll see their pictures in their decks. And the last thing they wanted to talk about was actually their tech staff.
because I didn't have much kind of like, and I remember we saw things like, yeah.
Harrison Faull (21:43.385)
I love the story. I love the story of when you went to a pitch event, but you were the VCs and people being asked, what's the USP of your fund? And you get up and you speak and you say, actually, we have a technical due diligence expertise. We know about technology and we can help founders. And then at the end, when founders could come and approach funds and ask future questions, no one came to your stand because they were a bit intimidated, but you could actually uncover the fact that they had nothing under the hood.
John spindler (22:10.367)
Yeah, and that was kind of okay. And it was a kind of element which looked, and there still is, AI, you know, let's get it to market. Let's get kind of traction and we'll find the AI later. So people would tell us, not AI today, but yeah, we'll probably put some AI in the future. And some of that can work. know, there is different ways to skin cats, but we thought that particularly from tech, we had a...
John spindler (22:38.717)
I did a deal called Blues B.A.I. with SeaCamp and it was a difficult deal to get across and we went round and round the kind of things and we ended up basically investing and that actually got bought by Meta. And that was one of the kind of cases that showed you how difficult it was in a great team blah, blah, blah to get anyone interested. And so that's why we, me and Tom Stone and a base at the time.
John spindler (23:04.372)
decided what we wanted to do, up a thing called AIC. Tom had just sold his company, founded a partner, got his PhD from UCL in a company called Prediction IO that was bought by Salesforce. He'd moved back to the UK. And so we put together AIC, then we ended up doing what's 46 investments over three years. Our original partner has also now taken AIC on, has basically restarted AIC.
seed and so investors may see start to see investing again but not with me not with Tom but with basically a new team and I think a slightly different focus but I'm not quite sure yet because we're not involved but you know so yeah at that time and we invested in say 4 -6 companies we were writing checks about 100 to 200k kind of and we did very we did very well that funds had five exits now we turned all its caps
to all its investors and it's got some nice kind of companies at series A, series B doing very, very well. And overall, think we'll stonking return to all our investors in the next three to four or five years. that was it. when I stopped investing in that and we stopped in 2020 just before COVID.
The next two two and a half years, I then went back to angel investment. So I did quite a lot of angel, I did about a dozen, maybe 13, 14 angel investments, kind of writing roughly 20K checks into kind of companies, of, you know, kind of some of them doing very, very well.
Harrison Faull (24:57.341)
What kind of, so these 20 angel investment that you made after AI seed, what kind of things were you doing differently to when you started? What are some of the lessons, what are some of the disciplines that you now impose upon yourself that you think are good principles that other ancient investors can learn from?
John spindler (25:12.709)
I think one of key understandings is that if you're at pre-seed and seed you're investing in signals, not data. And that's the big distinction. And I've done some work in later stage funds as well, series A, series B funds. And that's very much where you can build you can invest in data and you can build your own models. You can build your own financial model, market model, et cetera. Cause you have the data, not just on the company, but the markets that watch us. You've got data on competitors, et cetera, et cetera. So you can basically start to do some modeling.
to predict whether my investment investing at this valuation will be worth X later. You're not able to do that kind of interest. So you've got to kind of understand that a lot of it basically is around signals and what signals you can gain kind of are important. We still do every to know that there is a systematic way to assess still basically is useful.
but to know that you don't have the data to kind of fully put that model there. So I think there's still the classic six risks apply pre seed and seed and some of them more so when they do it series B, series C, you can use your things you get taught in MBAs, et cetera, how to do kind of valuations, pricing, predictions, et cetera. So I think it's still the most important one is the team and what we call the execution risk. Do we think this team
can meet the challenge. And it's a technical challenge, which majority of AI is, and they must be able to basically have within their team, the credibility, the experience, the know how to be able to build AI products, AI systems. And then you've got obviously the market risk, are there customers? Is there a market? Is there a problem that people care about? Blah, blah, blah. You've got obviously the third risk, which is often
quite important in especially companies like AI or anything deep tech is the technical risk. Can they build it? Tech product risk and can they make it work? And can they kind of, and this is really important. A lot of people dismiss it. go, do, know, yes, it will be, when it works, it will generate this value. But when it works, how it's going to work is the reason why there isn't thousands of them already doing it. Financial, yeah, yeah.
Harrison Faull (27:31.901)
I'd love to dive into that because I feel like, I think I read somewhere that founders apply to TwinPath. They are asked to submit a 55 -part questionnaire about technical risk.
John spindler (27:44.974)
It's actually gone up to 58 questions actually, I keep adding questions.
What kind of things are you trying to extract? You don't have to give us all 58, but what are some of the bigger ones that you're trying to uncover?
John spindler (27:56.557)
It's based on the five P's which I originally got from a guy called Matt Turk who was a well -known investor in New York, although we define the P's a little bit different. the five P's are people, so obviously you're looking at experience credentials, not just the team now, but who do they need to hire and also their advisors and etc, the gaps and do they understand it and blah, blah, blah.
You, and obviously this is holistic, every sector, you're looking at the purpose, which is a strong, why is AI required? Why does it add value? Why does it allow you to do something that others can't do? Where does it kind of give that massive uplifting value to your clients? And is it really worth it? AI is expensive, is difficult, probably won't work. You better have a strong kind of purpose. The bulk of it is around what we call the product, which is we take, look.
for them to give us data diagram actually and for the data from on the product from data ingestion to data outcome. We want to know how they extract and transform and load data across that architecture, what models they're using, what evaluation frameworks they're using, what basically how they assess performance today and how does it what's it need to do to get to be best in class in the future. Everything in AI is always a trade off.
That's the truth. And therefore we're looking at what their view is the trade off given the purpose of what the client thinks they need to do. And then we look at the 4P is slightly embarrassing. It stands for petabytes because we couldn't find another. That's what Mark Tert uses it for. I couldn't find another P. So it's actually data. what is the data? Do they have enough data? What's their data strategy?
John spindler (29:52.158)
Defensibility, moat in the data, they have a data kind of flywheel? And, you know, is that where the IP fundamentally is? And then we look at positioning, the final piece position, which is what most investors, professional investors are looking at, which is, you know, who's the ideal client? How many clients do you originally have? What's your go -to -market? What's your, you know, basic business model? Who you're competing head -to -head with, with anyone?
basically what are the critical success factors and why you win. And we put all those in. And we only give that kind of questionnaire out once we've had an initial meeting. Because we say, you know, it's a lot of work. Do you want to do it? Is it worth it? Are we a fit? We're keen to kind of know that the innovation and the novelty and the long term defensively is in your app development and application of AI. If it isn't brilliant, go and do it. You know, you don't need us.
go off and kind of like, it's a business model innovation or it's just a go -to market, you know, et cetera, et cetera. And we have a lot of people, we've had a lot of teams, some are very successful ones, we've gone on to raise quite a bit of money. Obviously, I don't know they'll be successful in longer term. Who would just say to us, well, we don't really have a kind of, for instance, a North Star performance, we don't judge the engineering, it's just, does a client like it?
Harrison Faull (31:09.661)
Mm -hmm. Yeah.
John spindler (31:11.272)
If the client likes it, they can have this, they can have that, they can have this, it doesn't matter. And we said, well, you
Harrison Faull (31:17.309)
which is quite telling. It's quite telling if you extrapolate that sort of attitude, that approach into the next 10 years, that's probably not going to turn into an outlier investment for you.
John spindler (31:25.626)
It's hard, my view is, obviously, well I name the one, but that's probably raised more finance in the last two years from what I've done this season. Still today, he lives near me, can sometimes bump into him, he still kind of smiles at me and says, still kind of, we'll kind of turn around whatever the client wants.
John spindler (31:51.423)
So it's different ways to kind of do it. There's never a set way. This is where the innovation comes and actually to do this. And people are always looking for formulas and I don't think they're necessary there. And coming back to what should angels look at, signals. And actually one of the things that think we need to understand a little bit, business model, how do they make the money? Do they make enough money to basically to get to a stage where they're gonna generate this time?
as cash flow that can fund their own growth without always diluting to investors. And this can be very important to angels who are fearful of always getting kind of diluted or especially liquidity preferences getting kind of taken out of a deal, even if it exits. And also finance risk, very big one, which is kind of, do they have enough money to get to the milestones? And if they don't,
Who's going to helicopter money to them when they haven't hit those milestones? Who's going to give them water when they're going across the kind of valley of death, the desert, to keep them alive another six months, a year, et cetera? And there is always a problem with that bridge route to come to kind of see, is it just a recognition of failure? Or is it a recognition they were under -capitalized and they were just hopeful?
that their investors would find some money for them sometime down the future. And it's very important for angels. That's what I learned from angels. You just can't be the person to keep going back to. You get to quite good money going after bad, et cetera, et cetera. And obviously the difference between VCs and angels is VCs have, to make money, they have to have big exits and they have to have businesses working in big markets.
to get those big exits. So if your market is really niche, they can't see the big exits to fund the kind of business models.
Harrison Faull (33:57.113)
In terms, so Twin Path now, it's AI focused. And I would say when I'm looking at Deal Flow, I'm probably most confused about the positioning side of things when it comes to AI products now, with the big releases of LLM tools and it being open sourced and then there being open AI and so many others. How do you evaluate the market at the moment? And how do think it's going to move and
John spindler (34:23.332)
It is moving very quickly. One of the aspects of large language models, especially large context windows, the emergence of incredible performance on zero -shot prompts, et cetera, is that sometimes the old verities do get challenged. Like how important is data if all the data is contained in the large language models and the context ones are so large?
John spindler (34:52.918)
So you need specialized data to get the same performance. And that's the small language model compared to the big language model. It's difficult in many cases to see whether a small language model can compete apart from in some very specific niches. So what do we...
Harrison Faull (35:08.859)
Yeah, one good one I've seen has been like a breast cancer screening product. They've got access to unique data source of breast cancer screens from an NHS trust. And they've been able to refine their model on that very specific data, which an open -sourced LLM doesn't have access to because it hasn't got all the permissions.
John spindler (35:23.795)
At the moment, at the moment, that's the issue. At the moment, they're soaking up all data sources they can fit. But you're quite right. If you're building a supervised model, trying to detect, say, a known disease with a known biomarker, then basically the old thing about quality of that data,
the ability of that data to be correctly annotated, labeled, et cetera. The models themselves aren't really that difficult. You're obviously changing the weights and nodes all the time to improve the specificity or sensitivity or whichever is more important in that detection or that classification. But you're quite right in those cases, if the data is unique and out and basically, and I think the problem with
with that data is that it has to be very correctly annotated and labeled. But people are, even in those cases, they're bringing LLMs in to basically, yeah, so it's not kind of one or other. And I think we're looking at, you know, with classic applied areas, you're looking at what else needs to be done. I mean, in classic applied, say, medical imagery,
Harrison Faull (36:32.103)
Yeah, to identify things properly, yeah.
John spindler (36:51.477)
the accuracy has to be so high. The regulator is so difficult that you're probably still going down that old pathway. that's probably not as effective. LLMs will be a part of the tool to help speed that up a bit, but they will replace kind of on that. And you can see the regulators in particular often are the barrier because they want higher proof points and higher transparency, et cetera, et cetera.
Harrison Faull (37:19.453)
which is sensible and good for society. So, no.
John spindler (37:21.12)
Yeah, I mean, I was having sometimes it may be too much, but yes, generally overall,
Harrison Faull (37:28.007)
So what excites you? At Twin Path, you must have made some recent investments. What's standing out? What do you think is defensible? Yeah. Could you tell us maybe a recent investment that's been publicly disclosed?
John spindler (37:38.367)
Well, I'll tell you a little bit. We invest in three categories. We of AI, we invest in frontier. So this is when you're still challenging what can be done and you're using the latest research, preferably your own, to be able to kind of push that further and further back and have a product that can do something that no one else can do at present.
And one area of that is aegenic AI, especially what we call long horizon aegenic, where you're training things to do task over a long period where the reinforcement correction, you you write or run is three or four, five, 10, 20 steps in advance. And that means you as a human are able to keep on doing that. kind of these bots often struggle to go beyond one, two, three, four kind of verifications.
But there's been great developments there. You're looking at teams coming out of leading universities. We backed a company, for instance, called Wiko out of UCL. And they're building a long horizon agent model to do data science coding. And to get that to work, have to break, they have to produce so many innovations to get that kind of, I've backed a company called Cosine, basically, but has managed to get to a level.
which is two times better than Devin in the last nine months. so, know, but there's still only at the, there's still only at roughly about 30 % equivalent of what a human software developer can do. The 30 % they do is better than any human, but there's over 70 % of tasks you still learn. And we've got one that's in stealth at the moment we think is a lot better than that, or could be a lot better than that, looks like, in what we've seen, it will be when it's launched in September.
And that, that, okay, so the, a, a, a, Agenic side is still really hard. It's really hard once you go beyond simple systems and to do more open.
Harrison Faull (39:38.823)
In terms of the exit opportunity, is that going to be a trade sale or are they going to own that space? How do you feel about that from the financial exit position?
John spindler (39:44.957)
Yeah, the exits are those either way. They can either be bought by Big Tech or they go into basic, go big themselves.
Harrison Faull (39:56.201)
I I had an advisor who sold to Google and he said, I think you had about 120 engineers for him. And Google's metric of acquiring the company was, I will give you a million for every engineer that you bring.
John spindler (40:07.74)
I think when companies buy you for the tech, not the cash flow, they're looking to buy capability. And so there has to be a capability threshold. So you're just like two people in a kind of kitchen. The capability for you to then to become that, that aogenic AI basically kind of unit for whatever is probably more limited, that's why. But it's not as, actually, because I've sold to quite a lot of companies, big tech companies, it's not.
that it is important, but it's not everything on that. So we like HNIT, we like Frontier, we've seen a lot, we're investing a lot in picks and shovels at the moment. So these are, what's happened with large language models is that enterprise clients and big banks, big pharma, government, you name it, have always had this thing, do we build or do we buy? And large language models seems to give a capability to build internally.
using your own data, using your own expertise, especially building a genetic workflows as well to kind of cover your own tasks. And to do that, they're going to need tooling. They're going to need orchestration layers. They're going to need long tail data. They're going to need evaluation kind of tooling, all of those facts. we've seen kind of that and they need things that help them integrate with their own. So we did, we invested in a company called REAPERCH.
They
We want to basically innovate. want to bring more code into our platforms at a quicker space because we have these LLMs, but we're frightened of the impact. So we need tooling to do that. So this, the picks and shovels, the tooling, the orchestration, the data layers, the things that enable people to employ LLMs in business, I think are very exciting still. I think other things that really exciting is still...
John spindler (42:32.173)
We're still seeing, and I'm quite like, there's areas where LLMs don't really play much of a role, such as embedded AI, where the AI comes back, which is embedded in basic hardware or sensor technology, where it's capturing real -time data and using AI to basically process, classify, predict what that data means. And I think that is really interesting and fascinating, because you're not, you're affecting
quite frankly, established old industries outside the tech area and you're bringing in expertise and knowledge to do that. And that's still really interesting. My biggest angel exit, my biggest exit from AIC to date was a company called Odin Vision, who basically is focused on basically computer vision data, classification computer vision data from endoscopes, know, the cameras that go down your throat to detect precancerous polyps.
So the signs of cancer. So a supervised learning model. You've got to be able to take the data, store it accurately, basically reconstitute it, basically, and then go to very high levels of sensitivity and spasability. And that company that got built last year by, last January, so about 18 months ago now, by Olympus. You know, the Japanese, you know, kind of, I knew them as Cammer Monkeys, but actually they make all their money out of medical imaging.
And that enables them to have a kind software AI product to fit in with their hardware and make their hardware products, et cetera, much more useful, efficacious, et cetera, et cetera. So I think that embedded is still there. the final one I think is really interesting, especially on a day where we learned that Essentials has been built, the AI drug discovery company out of Dundee and Oxford has been bought by an American basically kind of biotech pharma company.
John spindler (44:30.134)
is people are using AI to generate de novo, to discover something new, where actually what they're selling is not the platform, not the software, but actually what they find or what they develop. So the innovation, a drug compound or a new nanotechnology basic material or whatever it is. The AI is effectively a process, but you're not selling the process to others to replicate your actually
John spindler (45:00.617)
generating something and then taking that and then developing that and selling that. And I think that's an interesting area. And therefore you're getting in areas where AI companies in areas such as drug discovery are going from being software to full drug companies like Accenture did and others like LabGenius is doing at the moment and others to do that. And that's fascinating.
Harrison Faull (45:23.404)
We listened to some of your other podcasts, which have all been great appearances. I heard you mentioned that some of these models have gone disastrously wrong at times. And it's not necessarily just the building of the models that's difficult, but also maintaining them and making sure that they're consistent. Could you elaborate on, you don't have to name any names on something that, know, an angel invested today that might not have made an investment. What should they be mindful of in this maintaining of the model?
John spindler (45:46.1)
Yeah. Well, the other thing is basically you're always like the idea in, you know, and you'll see this in AI companies, they don't kind of, you know, they still have to keep on doing R &D. They still have to keep on building, having, you think they've built it. Why have they got even more computer scientists now? Because AI is a probabilistic science and anything probabilistic.
basically changes on conditions. And so the conditions change, the kind of nature of the data changes a little bit, the kind of, you know, other things, noise, instrument changes, calibrations, you can get completely different results. And so, you know, and so you're always having this kind of a kind of a problem, but it's not I paint the wall, I walk away, I have to come back in two years time. It's a constant battle. It's wild software.
as opposed to kind of rules -based tamed software. you know, that's the truth. And that's slightly frustrating if you kind of look in them and say, well, can't they just reduce their engineering costs? Kind of like, why they're always there? Because it takes so much to maintain it and keep consistency of results.
Harrison Faull (47:02.871)
lot of sense. In terms of the future for Twin Path, what are your goals and aspirations and targets for that?
John spindler (47:12.274)
Well, obviously we, you won't say that, maybe one of my LPs will be listening. I want to make my LPs money, serious amounts of money, and I'm fully motivated. We're smallish funds, and so my own commercial return will come actually only on exits, not on management fees. I want to back some really interesting and impactful businesses, because that's what keeps, me happy and interested and curious and kind of to want to do this job.
And, you know, to, you know, and this is a pleasure and, and, basically a privilege. So that's definitely true. And obviously, I, you know, overall, I don't, you know, I want to see AI make the impact I think it can do for the good of the world and society. And I think, you know, maybe you say I've dragged in the Kool -Aid, but I do feel
that is a foundation technology you could do an enormous amount of good and I get very frustrated where basically most the general public even my own family what they know about it is all based on fear and some old Hollywood films.
Harrison Faull (48:22.916)
wow.
Harrison Faull (48:27.16)
Yeah, well, there's several, Elon Musk, isn't he? He's quite outspoken about the dangers of what he thinks could happen with AI. And I think he has now set up a charity to try and prevent it.
John spindler (48:39.406)
Well, AlopenAI was supposed to be that thing. That's his own bugbear, isn't it? I want to say that we get a lot of it is just, you know, a lot of it say, I'm going to say that safety is not important and all the rest, etc. it is. But I think sometimes in the media and across the board is that we're not looking at the how transformational this can be and what it enables to do. It will discover drugs we don't know about. We'll discover new materials we can't do.
John spindler (49:04.77)
It will optimize basic process of reducing the cost and allow the whole world to use something that previously was just used by very wealthy people. So all of these kind of things are kind of It would drive productivity up if done correctly. All of these kind of things, which will create wealth that we can fund our own things that people want to fund. I do think that it's the best opportunity and chance that we have to basically in the world to get this growth that was so wanting.
basic output lies we all want.
Harrison Faull (49:37.507)
With you on that, I'm very optimistic. It definitely needs to happen. And for scaremongers to prevent that change or slow that change down doesn't seem like a rational outcome for all of humanity.
John spindler (49:51.671)
I agree and definitely not a good outcome for the UK as well.
Harrison Faull (49:55.309)
No. Where can founders find you? Where should they approach you?
John spindler (50:00.397)
Okay, well, we've got a website called twinpath.vc, which is a little odd. know we designed it. We may not design like that, but we quite like the fact that it's so yellow. They can contact me directly. I'm very open. Queen Path is free partners. We see everyone. There's no principles or associates between us and the partners that make decisions. So John at twinpath.vc or Nick at twinpath.vc or the amazing Katie at twinpath.vc. It's Katie.
Twimp Up will not amazing, Katie, although sometimes I feel like maybe that would be better. obviously LinkedIn as well, John's been pretty open for LinkedIn requests.
Harrison Faull (50:40.1)
Thank you. This has been a great episode. I really appreciate your time, There's so much insight in there for angels and founders. Yeah, it's been amazing. Thank you.
John spindler (50:48.887)
Thank you so much Harrison, have a great day.
