Will Bennett & Sivesh Sukumar, Investors @ Balderton, Seedcamp

Welcome aboard Venture Passports, the podcast delivering the inside view of early stage global markets. Me, Jack Richardson, along with my co host, Rich Armstrong, explore insights from the most innovative entrepreneurs and investors worldwide. All indications, coming in, to the control center at this time indicate we are go, go, go. Lift off.

We have a lift off. Awesome to have you both looking forward, certainly for some Point of view. The core. Like And at the autumn, just for a really quick introduction, a bit of a journey to to where you are to know.

I wish. Should start with Will. Change. Yeah, sure.

So I'm an investor in Seedchem. We are one of Europe's older seed funds. My journey here went via journalism, then BCG, and then a startup called Tesseract, now called Fuse Energy, which Sebastian's team kindly invested in back in 2022. Yeah.

Now now absolute pleasure to be part of the SECAM team. Thanks, Sebastian. Investor at Boltzson. Joined a couple of years ago, 4,000 at JPMorgan and their fintech team.

So looking at stream investments, actions, and partnerships, the bank. And before that, I was an engineer. I specialized machine learning, so cover everything on the applied AI and more tech side of Boltzmann. Yeah.

So, I mean, first off, guys, breaking to VC, being at Seedcap in Boltzmann, what's a culture like that when especially when you're super young. Right? You're learning from the industry greats, such as Sarangar, Rob, Carlos, as you mentioned. I mean, tell us about the day to day.

Yeah. So, I mean, our way of thinking about culture is really paying forward, and I think we try and think about relationships as kind of a multiyear, multi decade journey. So within seed camp, we work in lots of different ways of founders. Those split into a few key areas.

So those are the talent network, which SeedCamp's been doing this for about 20 years almost or 17 years, I should say. So a lot of that time gets recycled through our portfolio. It's through people paying it back, who we've invested in before, when they become our LPs, engine investors alongside us. And many founders who we invested in multiple times.

So people like Gabriel and Stan at Dust, who we backed in their previous journey when they were acquired by Stripe. And the culture is really driven by a lot of the people that we worked with back in the day who now support our founders ad hoc as experts, invest in them, and really help us build the European ecosystem. So, Suresh, anyways, tell tell us about Balderton. I joined Vault a couple of years ago.

I sort of stumbled stumbled into Balderton, applied via a LinkedIn post. It was quite an untritional reason to intervention. And so Balderton for the Amherst program was very keen to sort of avoid the the bias that generally comes from hiring via network, which is what we see a lot in venture. And so, yeah, they really opened the wide in the funnel, and and and I popped out for better or for worse.

But, yeah, I'd I'd I'd loved it. It's been incredible. The way the analyst program worked is essentially they take areas of opportunity where we don't have a strong thesis in, and they let the analysts go go free, work the map, what works, the tailwinds, the headwinds, and try and identify some of the best companies you're up to invest in. And that helps people who saw Tristy Junior build sort of strong conviction and also strong brand in those spaces.

So, like, to the Monde Stack, robotics, AI, and and a range of other sectors. But I think that's it. It's it's about sort of earning earning the right to cover a space is how Bolton operates. Everyone starts to generous, and you specialize over time.

Absolutely. And it seems like that specialism has delved into what we see now as, like, a bit of a a hypercycle or a race depending on what way you view us in terms of AI. So I just wanna delve into really why we organize this particular episode and and really into focus of AI of Alumigate as a whole. Firstly, incumbents versus, should we say, new startups on a scene.

When I speak with people about this particular subject, they always mention access to training data. People say, well, this is where the incumbents really thrive almost. They have much more customer transaction data, and and they can utilize that. Others say that there is so much data out there and it is highly that is ready from an access point of view.

So firstly, how do you guys think about that in relation to incumbents having a bit of a head start and how much of a head start do they have? So I guess, yeah, incumbent, the the the broad broad sense of that word now. I think a lot of the incumbents in the game right now are card based companies that that can move quickly. And so I think sort of assuming that incumbents are sort of slow, they're sort of slow slow to adopting technologies and went and won't be able to keep it.

What was what was going on now is probably wrong now. I think Microsoft is is a good example of that. I think that the the way which they've moved and deployed products has been pretty incredible. The way I like to think about especially startups and there's always options to start ups.

If we didn't believe that, we wouldn't be here. But I think it's it's sort of where there isn't one incumbent that owns all the data and distribution, that's where probably where the opportunities could be. So take the gaming environment where at the end of the day, you really only got Unity and Unreal. The opportunities for startups in terms of dev tools might be but take the sales stack where you've got about 7 or 8 tools that everyone needs.

And these incumbents aren't incentivized to building integrations. They're trying to capture everyone in their own ecosystem. So startups have that benefit of being able to integrate with all those different tools with all the data and distribution that you need to actually build an AI agent that works. And that that's where I see start to see the the sort of immediate opportunities.

But, yeah, we we've always gotta believe in the underdog set. That's that's the the name again. I guess one thing I'd ask is to train an amazing model. You don't necessarily that much data now, which is something we're seeing emerge quite quickly.

I think businesses where they'd have a lot of invoices or other sorts of payment data, for example, in construction and agriculture, shipping, logistics, are able to train quite local forecasting models to really understand their businesses better. I think there's a huge opportunity there. But, yeah, as as Suresh said, nimbleness is still the name of the game. And I think Meta have corrupted move fast and break things, but move fast and don't break too much can still reap their enormous rewards, I think.

See, and as Suresh mentioned, obviously, Microsoft being, like, a very fast paced incumbent, which has been shared products very quickly. Will Deep have any other notable examples on top of that, which you've been nicely impressed by? I think Microsoft really is, is the main one. The thing on the thing on the ad is actually that Microsoft and Facebook and the way they've built up their GPU staff in a very clever way, like Microsoft through acquiring inflection, Facebook through a reserving GPU in advance is incredibly smart way to think about building models and optimizing what the distribution that they already have.

I think, you know, Facebook obviously have a slightly different approach and that people may never pay for llama, and they're actually building that in more in a more full way. They're building in CyberSec evolves. There's some detection for prompt injection in this thing that could can damage the way people use models. So I think they're also thinking really interesting to me about the way their distribution to really launch this new wave of products.

Absolutely. And Sudesh pointed towards, like, on cloud brand use cases and being obviously the the fastest to respond to. It seems like, obviously, the biggest companies in the world, like the Microsoft side of matters, etcetera, obviously, the the fastest to to move. Is is it true when you think about, like, an enterprise stack and, like, on prem use cases in particular?

I I personally think that the biggest companies in the world will not really allow the majority of the data to be put through a different solution to be stored in in some cloud that they'd never heard of. I personally think it'll be quite sensitive. It'll be on prem under wraps under a lot of padlocks. Do you believe that also?

And, obviously, Willie pointed towards data being highly available, and you don't really need a huge amount of data to actually accurately forecast models these days. So, how do you guys think about that? Yeah, I know it's really true. You think as much as Aon has advanced enterprise productivity, it's enabled an enormous amount of malicious actors external to the enterprise.

And, you know, phishing's up 500%, ransomware attacks up enormously as well. And the biggest companies in the world, their data keeps getting extracted, and encrypted like Okta and HubSpot. So I think really thinking about how to keep your data secure and protected, is enormously important and isn't just about how to defend for easy gains through the use models, but it's actually external actors as well who are trying to trying to use this amazing technology in a nefarious way.

Yeah. So this first time at JPMorgan where sort of the first question we'd ask, any startup we're sort of looking at was if they could deploy on prem. And if the answer was no, it was sort of red flag. We we sort of stopped there.

And it's a hurdle, but I think you we do have to accept that there is a lot shifting to cloud still, and you can see that in the numbers. Like, these companies are at 10, 100 of 1,000,000,000 of revenue still growing at 30% year on year. It it's a constant question as to as to how much is gonna is gonna shift on cloud. What can startups do to get a meaningful edge on the large incumbents, do you think?

I mean, you think how are they able to compete on this distribution side? You look at GPT 4, the code assistant, super fast, like, their increased rate limits in the API. I mean, how do you compete the distribution, the new updates, and things like this? Maybe you can share your point of view first, Will.

Yeah, for sure. I think on the application layer, maybe to start with, I think there's still incredible value to unlock by building amazing customer experiences, which might sound like a bit of a cliche, but you look back at other waves we've had, neo banking being 1. Like, some of the technologies like are very similar to what to what came before, and a lot of neo banks are banks, and being able to move incredibly fast, build an amazing experience, I think, will still apply to applications in AI.

And then in the infrastructure layer where startups, where there's white space, look, I think close to the hardware, there's loads of opportunities. I think accessing GPU in an efficient and cheap way is something that one of our portfolio companies, Fluid Stack, has enabled companies to do excellently well. And then even the basis where there's like early movers, like lang chain app development, and base 10 of the Inference Cloud, I still think there's LLMs and AI get as big as people think they will.

There's enormous value to capture for start ups who provide very, very specific services that untracked value for for companies who who'd want to use LLMs, but don't necessarily have all the picks and shovels to to make it work mutually. Yeah. So I think it's it's an interesting question. I think we sort of a lot a lot of people can sort of fall into into the bucket of, okay.

Number 1, you think all the value is gonna be in the model layer, so you spend all your money on GPUs. Or number 2, you're gonna avoid any caps going towards GPUs, and you're gonna build up your application. But it's tricky at the moment because, obviously, the sort of negatives of spanning volume in GPUs as they are becomes very unclear, and the negatives of building up complete application there is it's hard to differentiate because I think what we what we need to accept here is that most of the value of the of these of these models, of these AI driven applications is coming from the model.

So if you think about, say, if you're a team moving 50% quicker than the average other developer, after a year, you're 50% ahead. But if the 9% of value is coming from the model, then you're only gonna be 2.5% ahead at the end of the year. And so that's why I think there's a lot of value in building at the model and sort of fine tuning, preference tuning, and not training big models, but training reward models, and try and sort of enhance what these open source models can do.

And then sort of adjusting the distributions of the model towards your specific use case. Because at the end of the day, OpenAI, Anthropic, Nostral are training very sort of uniform horizontal generalized based models that aren't gonna be very good at specific tasks. And and that's where I see the sort of immediate opportunity. But going forward 2, 3 years, I can definitely see some great businesses being building completely application now as well.

Makes sense. A couple weeks ago, we had a friend of mine who was a investor in emerging markets. He was saying a lot of the value will be at the application layer instead of the infrastructure layer. Do you guys agree of that, or do you sort of see the opposite view?

I don't think there's any any sort of obvious answer to to where it will accrue, but I think we've sort of gotta gotta go to to business 101 and and and leverage your network effects, try and sort of build some sort of strategy to prevent other companies from taking your market share and and and value accrue, and that can be done at application layer. That can be done on infrastructure layer. And it could be done any anything in between as well. Do you think 5, 10 years out, same as well?

There'll be value for in the application layer as well as the infra layer as well? So, you know, we just find pockets of opportunity Yep. In your opinion? Yep.

Definitely agree with that. So I assume you guys agree that the future of AI is, like, dominated by 1 or 2, like, single models. Do you think that'll be the case? Or do you think we'll have fairly decentralized, fragmented ecosystem?

In mind, personally, for mine, like, if you look at, I don't know, Jarvis or Longchain. I mean, the future really is the mediator from an input web additionally across the models. What's your guys' thoughts on that? What do you say?

I think there will be different models in lots of spaces where that has to be, for example, like regulated data or very specific production level data for manufacturing plants, where the data is time series or to do with other components of industry where you don't need natural language, for example. For more general enterprise productivity where natural language is the the atomic unit of the of model, I think it's possible that lots of big models or sorry, a few a small number of big models will dominate.

But I guess if you think about other other industries where the atomic atomic unit is quite similar between businesses, like telecoms, like, there's still space. A gigabyte is a gigabyte is a gigabyte, and there will still be space for multiple winners. Just on that point, big believer in in in specialized models. Because I think if you if you think about what what's a neural network is, it's it's an approximation of a function.

And if you've got one large model and lots of smaller models, they're essentially the same thing. But then if you factor in cost and inference speed and everything in around that realm, you're gonna want the small specialized models because they're they're achieving the same goal, but it's doing a lot more efficiently when you're using the small specialized models. Whether or not those multiple models will come from one company is is unclear right now, but I'm definitely a believer in in in smaller fine tuned models for specific purposes.

I mean, general consensus is, like, as long as somebody specialized in the output of verticals, don't have to be a competitive mode, would you say? Absolutely. And I think a lot of that momentum will actually be customer first. I hope there's a temptation to think tech first and how the model works and from a machine learning, like what a model enables.

Some customers will 100% need specific fine tune models due to regulation, due to internal compliance reasons, and so on. So I think the market will actually encourage that dynamic rather than everyone reverting to a small number of big models. Will we see any model used now used also within 2 to 3 years? Like, do they have a lifetime, would you say?

It depends on the task, I guess. Because if you think about I mean and most of deep learning and production today is is models that were invented 5 or 6 years ago. So sort of a lot of them look look like your sort of traditional BERT transformers. If they're solving your problem, there's there's no real reason that you need to keep on shifting.

I think it's for these use cases where best is always better. That's that's where we're gonna be considered using the sort of state of the art models. Let's maybe discuss about the regulatory changes, right, which need to be made in the world of AI. What do you see as the most urgent change that we need, not just maybe in Europe in general, but on a global scale so that we can keep on improving the advancement of AI?

I'm actually the biggest issue replace our mission a little bit earlier is the nefarious use of AI and figuring out how to regulate, ensure that companies defend themselves sufficiently against that. Because ultimately, they're, for you, they're at as much risk as they are. And the UK does a good job of this, GDPR regulation, which to some extent is anti AI hacking, right, because it's data security and the fines that the UK has been able to give out for companies that don't protect themselves against AI native attacks, I think are meaningful.

So I think that will be incredibly important. And then I think the spirit of your question is like, how can we all use them without messing anything up? And regulation encourages everyone to track how data flows in through model and out, making sure that brand assets don't escape and get reproduced by other companies querying a model. Things like how can I embed an open source model into the enterprise?

What guard layer guardrail layers do I need to ensure that my data doesn't influence that model, and so on will be important. But a lot of that is actually coming in with the model providers themselves. As often is the case, I think some of the tech companies will actually be fast before the regulators because it's what the customer wants. And Llama, as I mentioned, is a good example by building Cypressack Kvaer into the product.

They're actually front running some of the the regulation will no doubt come in to say, please protect employees against malicious prompts, and they're already doing. I bet you agree with Will. And I think copyright infringement is definitely sort of top line for me at the moment because there's sort of a massive elephant in the room there where everyone's sort of just training over everything. And I think that those questions will be answered, but I think they really need to be answered to sort of set up how we how we really advance, these models getting bigger and bigger and bigger.

Like, it's not just gonna be a few subsets that that that can solve it. I think we need some proper infrastructures to how we distribute IP and how people can monetize IP going forward rather than it being the wild west out here at the moment. Computes. Like, how much of a problem is that, would you say?

And, like, how optimistic are you that the capacity will increase in the future to host, I guess, the more good advancement of in this particular sector? It's always a function of the data you've got available. Like, there's a there's an optimal ratio of of of how much compute you really should use for for how much data you've got, and that changed with architectures. But I think data's currently the the finite thing at the moment.

We can still scale up compute, and I think that the sort of GPU crunch seems to be easing from from what we've seen. There's less people training models for the sake of training models. I think one thing that we haven't scaled up yet, which would be really interesting to see, is is inference time. We've scaled up data.

We've scaled compute. But this at the end of the day, we're still querying these models in in a millisecond. If we could sort of if we think about how the human brain works, we solve complex tasks over a long period of time. If we can allow models to do that, that could increase the accuracy of the model as well.

So I think that's that's one sort of idea of long inference is is something that that I'd regularly love to see as well. I think GPU can, in terms of being, I think limiting the right amount of inference. But I think also in the future, they're going to say the lead factor. I think there's other things like chip design may actually limit our rate of development or Catmull's law more quickly.

And to say, Sebastian's point about inference, freeing out ways to step chips or reuse the energy produced by chips to avoid like hitting a thermal wall where we can't do inference any more quickly or more efficiently, I think could well be, a more immediate impact than just pure production volume of GPU. In in terms of, like, Mistral and the pricing now of Reliance, like, the pricing of Mistral particularly has has come out. So obviously, the media model fairly recently suggests, obviously, perhaps they're they're reasonably more confident that their models are essentially better than GPTs.

Do you agree or or not? I think there's there's definitely a land grab at the moment. And if you think about how how GPUs work, it's it's back to parallelism. So the more users you have, the cheaper it gets.

And so it does make sense to start offering models with sort of negative unit unit economics to to capture those users and and to capture that brand, which is something that I don't think gets discussed enough at the moment. Because the way these generative models works quite and the output's quite subjective. Brand really matters. Like, it's hard to say if OpenAI is really best than Claude or Mistral, but if you use OpenAI and you enjoy using it, then you're still gonna use it going forward.

And so if you can capture users and they enjoy your brand and they enjoy using their models, then that that's gonna definitely pay off going forward. So I think we can see a lot of companies pricing their models. Yeah. Very interesting indeed.

Maybe let's discuss about which was more important. Right? Do you think it's the model size or the data size, in your opinion? Maybe this one's for you, Will.

I mean, on the model size, first of all, I think GPU is still enhancing performance. Like, more GPUs are enabling big models, which I think is like no brainer. The question is maybe more, how many enterprises really need a 40,000,000,000 parameter model? That I think is not necessarily solved.

I think there's a bunch of you use cases that are quite clear and people hide to keep occasions. But if you think about the way most LMs are being put into production for enterprises right now, Whether we need all that GPU and compute, I think probably not. I'm querying, company knowledge. I'm accessing information really quickly.

I'm trying to share information I know with other people, which are all great use cases. But whether a huge amount of compute you really require for that, I'm I'm not necessarily sure. And on the data side is and maybe maybe to add a different ability to Suresh is, I don't think we've always need loads of data to to get to, ready useful models. I think a bunch of things can be done with quite a small data center, which I think is increasingly emerging and especially to fine tune a big model.

Like, you don't actually need that much different data. Well, if we think about pretraining of base models independently, I mean, the chinchilla paper stated that there is an optimal ratio for a given parameter count, how many tokens you should use there. And so I think as as we can find more data or as the we've got the data out there. It's cleaning that data and making it appropriate for training.

As we can do that, we can scale up models further. But as we touched on earlier, I think what what's really important and what we haven't scaled is is inference and and and and the idea of, like, how can we if if we're scaling up the training time to months and maybe even years, who knows? Why can't we scale up the inference time from a millisecond to minutes? Because that's how the human brain works.

We don't just solve a complex problem in a millisecond. We take time to do that. And so we can allow models to sort of have that aspect of system 2 thinking. I think that that could be quite powerful.

Yeah. And maybe you guys can touch upon your experiences within your portfolio on this particular question. How difficult really is it to educate and really get to grips with what enterprise need AI and want AI for from a good market point of view? Does that seem to be a huge frustration from portfolio companies that are trying to sell to enterprise?

I think there's a lot of frustration right now because there there's a lot of well, freight is probably the wrong word, but there's a lot of app type enterprises, but I think none of them really know what they want. And I think you can see that in the BCG numbers and the extension numbers, and these consultants are are sort of the real winners today because they're helping and they're doing the handholding for for through the enterprises to to help them get to value with AI. And we've seen that with one of our portfolio companies, writer, who really focused on that solutions engineering piece.

So at the end of the day, most of the enterprises don't even know what prompt engineering is. And if you can help that and and and help them do that in the sales cycle, that's gonna help them get to get to value, and people are happy to pay a premium for that time to value. I don't have really strong opinion on this, actually. And I think that the the consensus is different on this at different times.

And whereas 2 different pillars, there's one that's solving one GTM solving a specific pain point that people already have that LMs, like, happen to resolve. Then there's the other component, which is like all the stuff that LMs can do, that people are currently getting to grips with that need a little bit more education. And that's, I'm sure, where business development and GTN teams are really figuring out what that looks like and enabling companies, whether that's product led or really sales led, to make the most of these tools.

I would say that, like, there are amazing companies out there that aren't necessarily building the application or the infrastructure, but are actually thinking about, okay, how do we embed AI in workflows and really going deep into companies and understanding, okay, in healthcare billing and in all sorts of different, very specific verticals, look, the way people work looks like this and therefore, how do we enable amazing existing technology to come in here and, and be the people to unlock that value, where, where it is at the moment trapped.

Moving on to like an HR type of alignment questioning, will we see, in your personal opinion, like, 10, 20 people, AI companies, like, unicorns the next 5, 6, even 10 years? We've got it in our portfolio. We've got VotoRoom. That's happening.

I think that the trends are all there. The sort of barriers to entry to building software is is sort of tending towards 0. Yeah. I agree.

I think especially in areas where SMEs are being served in a product led way for things that they've really struggled in the past, like accounts receivable, marketing, the CFO function. I think amazing AI companies that really fix a specific problem there can make an enormous difference. You don't necessarily need huge teams to scale a business like that. I think where you do need more personnel is where you're looking at a Sandija type company where you're not only doing amazing services for big clients, but you're also trading on stable diffusion models and really building that research component in house.

But I think that's not necessarily a mop out for a huge company. I think there's many other routes where you can remain very lean and build something big. Yeah. No.

It makes sense. I mean, to get to these size companies, you're gonna need to have the best AI talent. Right? So where do you think this will concentrate from, and what would companies have to do to win this talent?

Yeah. So I think there'll be, like, a natural concentration over the next few years just due to the market. I think talent will be recycled as some companies shrink in size or even go out of business, and amazing talent will rejoin. What, what are the new emerging winners?

And that is a natural phenomenon. That's the reason why you should build in tough times. Right? Because you get better access to, to amazing talent.

The other thing is like, a lot of these fields are still really early. Like, if you think about, a classical MLM is might use supervised learning, unsupervised learning, reinforcement learning. Like, they reinforcement learning particularly now really haven't been along for, much time. It's not usually well understood.

So I think that in itself will just attract people who want to be on the cutting edge of developing their machine learning technology. And then I think there's a suite of other things that in the future will be incredibly exciting. Like, federated learning will attract huge amount of talent whereby a central model can be trained by decentralized private data, which some banks in China are already experimenting with. You know, neuromorphic computing, reflecting the brain's qualities where inference and training can happen at the same time.

And many other kind of disciplines, which are not yet underlying in LLM or machine learning is perhaps in general. But I think in the future, it will no doubt, like, enable huge progress and attract I'm very bullish on on open source frameworks helping to to flatten that curve of talent. Because if you think about 10 years ago, the number of people that could train a a neural network would be in sort of a 100 people or a convolutional that would be a 100 people. And now if you think about the amount of people that could train a language model last year, they've probably been sort of same order of magnitude.

Whereas if you think about the number of people that can train a CNN now, it's it's anyone with 7 lines of code. And so very bullish as to as to how sort of the likes of PyTorch, TensorFlow can can sort of really flatten that curve as to who can do what, enabling a sort of a a broader distribution of of what talent can open or close. That's the big question. Right?

I'm I'm I'm bullish on on open source. I think as as we touched earlier, I think that there's a lot that can be done within the models and and sort of adjusting the premises of the models, adjusting the distribution to what the models can do, to to help really get to any of the value. And I think you can only really do that with open source models. I think open source helps solve a lot of concerns in how we think about and it's not just the there's a broad sense of open.

I think we we sort of really should shift properly open source models where we sort of really know what's going on in the data going into these models, the training pipelines, and that that helps enterprise bring comfort in the end output as well. And that that's what we've seen from from writers, which are now portfolio companies who, are very transparent about what what's going into their model, and then that helps bring the comfort to enterprises that are that are pointing these going forward as well.

No. It's really interesting. More generally, would you say from a macro point of view, what the biggest or most common misconception that public have about AI, which you wish you would change? It's probably that AI doesn't work.

I think at the end of the day, I know a lot of people that sort of try Chatbt once and say, oh, it didn't didn't do this. It didn't do that. But at the end of the day, the only real way you you sort of get get max value of these models is is by trial and error. And so I think keen for for more and more consumers to really get to grip through these models and understand what they can do, what they can't do.

And and that should should solve a lot of concerns to to sort of model market fit today. I think it's a good question. The one that really frustrates me is that it's magic. I think models are basically using especially in national language models.

They're listening at like, what are other common sentences or common sentence construction, dah, dah, dah, dah. What might this next sentence be or next word, next token, so on? Like, not magic. Enabling that education component will will make conversations around AI AGI much more effective, where people can share some of the same premises.

That this isn't anything like weird and spooky and impossible to understand. It's actually explainable in the traditional sense, if not machine learning explainable. You guys are much more calm, obviously, than the mainstream media, which capturing the headlines is obviously the end of the world. As we know, it's gonna replace jobs.

For instance, being the lawyer from a services point of view or an accountant could become extinct in the next 5 to 10 years. Do you guys agree with that or not? I think it's really hard to say. I think in areas where LMs are doing amazing jobs at automating repetitive knowledge intensive information oriented tasks like law and procurement when there's a lot of forms and heavily natural, natural language based tasks.

I think they'll accelerate the work completed over time. Now, I suspect it will actually mean we all get more lawyers and there's more litigation in general. It doesn't necessarily mean that lawyers will disappear because I think lawyers as an interface with clients, lawyers as the ultimate arbiters in many cases, will be continue to be very, very important. So I don't necessarily agree that it will it will remove those jobs.

I think it will actually if the promise future comes through where, like, a lot of this stuff comes cheaper and accessible, I think it will actually enable people to access great services, more easily. Definitely agree with that. And I think big big big fan of sort of the the the augmenting thesis for copilots. And if we look at what AR is really good at, it's preventing anyone from doing the same thing twice.

And so, yeah, I think it it's rather than replacing work, I think it's gonna help people work on things that they actually want to, enabling more productivity across the enterprise. What do you think about economies that have large segments that rely heavily on, like, freelance work? Right? How much of a impact do you think AI will have?

Yeah. There there are some savings that, like, will to some extent be limited in scale. Like, some of their classical consulting use cases where young folk, like, find data, organize it, and then communicate it graphically. I think of all pretty good AI use cases.

And frankly, enterprises were paying too much for freelancers to do that in 1st place. So I think that's like one interesting way to think about it. And I mean, a lot of those were very smart people anyway. I feel like they found interesting new ways to contribute to society.

I guess other freelancing will just be immensely enhanced, like things like graphic design, website design. We're already seeing a lot of AI take place in websites like Upwork and Fiverr. I think everyone who who earns via those channels will, their workflows will be enhanced by AI. And now I think their earnings will probably also be enhanced because they can complete more high quality work in a specific time period.

Yeah. I completely agree. I I would love to get your take on, we spoke about before, specializing in a particular vertical or 2. And last time I saw Suresh, it was mainly a discussion around particularly the health care space.

If you had to pick one vertical that you think is most promising or most excited about in the current context of AI, What would you say? I think I think we'll touch on this earlier, but I think we're seeing and then the way we're jam try and think about the application there with AI is is the notion of model market fit. It's like, as these models get better, what can they really do? And we've seen them with sort of SDR space enable a lot of value, and recently, we've seen these models really nailed, some sort of, lower value lee legal workflows.

And that's that's driving a genuine shift in in how a lot of legal firms operate. They're they're they're using legal legal AI, and they're now shifting from billable hours to value based work. And whenever you start to see those shift in business models, that's that's a lot of reason to get excited. We're we're seeing, I think, a a complete rising tide across the legal tech sector at the moment, which is a notoriously hard set to crack because lawyers hate using software.

But there's there's there's a huge amount of market demand for it. Right? Yeah. I mean, I I understand that health care, which isn't necessarily based on, Ellen Moran to develop.

It's actually a lot of it is based on previous iterations of AI technology and 2 companies in our portfolio, VIZAI and Ezra. I think they're just creating enormous benefits in terms of access, affordability for pretty heavy lift healthcare. Ezra does MRI scans, For example, they're very easy to to have some demand. And I think AI's potential to bring enormous benefits to society and healthcare, is fantastic.

And I think you have this like enormous field of operational stuff that was in, hospitals and care homes and staffing and so on where there will be amazing companies too. But I've been really on that, like, survey people application layer application is almost the wrong word because it's like, they're not necessarily users. Like the doctor maybe is the more of the user, but they're certainly the customer. I think that kind of thing, will be amazing.

Yeah, absolutely. I think especially like the back office solutions that are using Copilot, I think are particularly interesting even in healthcare. As Richard Lee said, we want a bit of a quick fire and I guess I probably already know the answer to this particular question. So I'm looking forward to seeing how strong willed you guys are.

But, yeah, like, super short concise answers to these. Is now, like, a fundamental shift in AI, or is it a result of potentially just hype cycles doing their work? Let's maybe start with Suresh on this front. Yeah.

I I think it is as as a January friends, we also shift. If you think about what AI engineering is today, it's it's very different to what it was 2 years ago. I think 2 years ago, everyone was stuck in PyTorch and TensorFlow, dealing with their loss rates and and everything in between that. But now people are just interfacing with a track interface, and and that's really what AI engineering is today.

And so, yeah, I'm bullish that this is a this is a genuine shift in in in what we're doing. It's both jacked, both. We are on the high cycle and valuations high, and everyone's I was waiting for that answer. It did.

As Samir said, it's there's also like amazing things being built and a thing. In all the categories we've mentioned today, like almost all of them are big enough for like several amazing companies who are guessing all the competitive dynamics and things we we we discussed about commoditization. Like, there is amazing opportunity, and I do think it will take good lives. I would like to move on to the next one.

We talked about employment and having effect on AI. Do you think citizens nowadays should be worried that AI will firstly, will they be worried? And secondly, do you think you'll destroy more jobs than create more jobs, in your opinion? I think when technologies like this, they're like, they often, in the short term, create amazing jobs and we discussed how like there'll be a boom in law, graphic design, all these things that LLMs are already doing.

In the medium term, there will be some reallocation of jobs. In consulting, as I mentioned, we organize data, format it and sell a piece of a PowerPoint presentation, typically. Some of that time might well be reorganized. We're talking about people who a lot of them have the skills to reeducate themselves.

And for the last 50, 100 years, like, the economy has been driven by people reeducate themselves as as technology emerges. And, I definitely think this will be the tide that lifts all boats as opposed to a lot of benefits flowing to specific people in specific roles while other people are displaced, which I think is a exciting phenomenon. So you think, in your opinion, that a lot of certain industries, right, will be a lot of reskilling needed and a sort of upskilling in certain sectors?

There'll be some. I think people understanding how Alums can support them while continuing to do quite a similar job. There will also be people where the LMS can take over a a big chunk of their fundamental unit of work. And some of those people will be able to do more of it as a result, but some of them will also move on to other roles.

And, yeah, I think one of the good thing about this wave of technology is that it is possible to reeducate yourself. Probably using perplexity or to research other things you can do. I I think, that's a natural thing that will happen a little bit. Yeah.

I think I think completely agree. I think, it's a good point in education. If we look at the sort of recent OpenAI demo from yesterday, it's it's it's a kid learning how to learn Pythagoras theorem straight from from AI and in real time as well. And so I think it's the scope of what we can do now and the scope of what possible now has has increased, and people can can use the these models to to help leverage them going forward.

So very very much in the camp that there'll be a net creation of jobs going forward. Let let me finish off with the final one. Maybe this is for the both of you. Right?

We're all young, but in terms of in your opinion, you guys have done pretty well working at the top European funds. I mean, if there was someone to that was looking to break into VC, what piece of advice would you give them to break into the industry? Good question. I think it's important to to read yourself, and that sounds too shaded and boring, but it is because everyone has a different style doing this job.

And some people are driven by purely curiosity about software. Other people are driven by pure hustle of people, like pure commitment to entrepreneurs. By the way, all those are really important, but like Yeah. You need all those.

You need all of those. But different ones of them come written actually. So whether it is like sourcing deals and sending them to VCs or supporting founders in your spare time, or even investing a £100 on seeders because you are obsessed with that type of technology. I think those are all relevant ways in.

I think people often talk about writing as a as a way in. I think that is important and it is amazing to organize your thoughts. I think there's a lot of young VC doing, so I'm not not necessarily convinced it's it's the way. Or or doing a podcast.

Right? Or doing podcasts. Exactly. One word where that sticks to mind here is proactive.

I think this is is definitely a job where you're out there. You you've really got to be proactive. You really got to sort of be independent in sort of what makes sense. And I think, yeah, if you can be proactive, you can sort of build your network, speak to the right people, then I think everything should should should fall into place.

Yeah. Absolutely. It's been honestly amazing. It's been really useful for both me and Richie, and hopefully all the listeners.

It's been a blast having you both on. Hope you really enjoyed it as much as we did. It's been a pleasure. Thanks, guys.

Well, I'm sure you can all agree that was an awesome episode. Thanks so much for joining us on Bench Passports. We hope you've discovered new insights and inspirations from today's episode that you can apply in your own line of work, and make sure to tune in next time as we continue to unveil the extraordinary in every corner of the global market. In the meantime, you can follow us on socials at vcpassport.

Safe travels.