Bản chép lời đầy đủ của Will Bennett & Sivesh Sukumar, Investors @ Balderton, Seedcamp

over time. 72 00:04:05,239 --> 00:04:05,639 Absolutely. 73 00:04:05,639 --> 00:04:09,925 And it seems like that specialism has delved into what we see now as, like, a bit of a a 74 00:04:09,925 --> 00:04:13,625 hypercycle or a race depending on what way you view us in terms of AI. 75 00:04:13,764 --> 00:04:18,004 So I just wanna delve into really why we organize this particular episode and and really 76 00:04:18,004 --> 00:04:20,745 into focus of AI of Alumigate as a whole. 77 00:04:21,089 --> 00:04:24,629 Firstly, incumbents versus, should we say, new startups on a scene. 78 00:04:25,089 --> 00:04:29,490 When I speak with people about this particular subject, they always mention access to training 79 00:04:29,490 --> 00:04:29,810 data. 80 00:04:29,810 --> 00:04:33,089 People say, well, this is where the incumbents really thrive almost. 81 00:04:33,089 --> 00:04:37,274 They have much more customer transaction data, and and they can utilize that. 82 00:04:37,414 --> 00:04:41,735 Others say that there is so much data out there and it is highly that is ready from an access 83 00:04:41,735 --> 00:04:42,555 point of view. 84 00:04:42,694 --> 00:04:46,615 So firstly, how do you guys think about that in relation to incumbents having a bit of a head 85 00:04:46,615 --> 00:04:48,899 start and how much of a head start do they have? 86 00:04:48,899 --> 00:04:52,500 So I guess, yeah, incumbent, the the the broad broad sense of that word now. 87 00:04:52,500 --> 00:04:56,339 I think a lot of the incumbents in the game right now are card based companies that that 88 00:04:56,339 --> 00:04:56,979 can move quickly. 89 00:04:56,979 --> 00:05:02,045 And so I think sort of assuming that incumbents are sort of slow, they're sort of slow slow to 90 00:05:02,045 --> 00:05:03,964 adopting technologies and went and won't be able to keep it. 91 00:05:03,964 --> 00:05:04,764 What was what was going on 92 00:05:04,764 --> 00:05:05,904 now is probably wrong now. 93 00:05:05,964 --> 00:05:08,125 I think Microsoft is is a good example of that. 94 00:05:08,125 --> 00:05:11,084 I think that the the way which they've moved and deployed products has been pretty 95 00:05:11,084 --> 00:05:11,584 incredible. 96 00:05:11,884 --> 00:05:14,685 The way I like to think about especially startups and there's always options to start 97 00:05:14,685 --> 00:05:14,764 ups. 98 00:05:14,764 --> 00:05:16,205 If we didn't believe that, we wouldn't be here. 99 00:05:16,205 --> 00:05:20,629 But I think it's it's sort of where there isn't one incumbent that owns all the data and 100 00:05:20,629 --> 00:05:22,550 distribution, that's where probably where the opportunities could be. 101 00:05:22,550 --> 00:05:26,389 So take the gaming environment where at the end of the day, you really only got Unity and 102 00:05:26,389 --> 00:05:26,889 Unreal. 103 00:05:27,029 --> 00:05:31,529 The opportunities for startups in terms of dev tools might be but take the sales stack where 104 00:05:31,625 --> 00:05:33,705 you've got about 7 or 8 tools that everyone needs. 105 00:05:33,705 --> 00:05:36,205 And these incumbents aren't incentivized to building integrations. 106 00:05:36,585 --> 00:05:38,585 They're trying to capture everyone in their own ecosystem. 107 00:05:38,585 --> 00:05:42,425 So startups have that benefit of being able to integrate with all those different tools with 108 00:05:42,425 --> 00:05:45,965 all the data and distribution that you need to actually build an AI agent that works. 109 00:05:46,149 --> 00:05:49,370 And that that's where I see start to see the the sort of immediate opportunities. 110 00:05:49,750 --> 00:05:51,910 But, yeah, we we've always gotta believe in the underdog set. 111 00:05:51,910 --> 00:05:53,529 That's that's the the name again. 112 00:05:54,550 --> 00:05:57,110 I guess one thing I'd ask is to train an amazing model. 113 00:05:57,110 --> 00:06:01,785 You don't necessarily that much data now, which is something we're seeing emerge quite quickly. 114 00:06:02,165 --> 00:06:07,365 I think businesses where they'd have a lot of invoices or other sorts of payment data, for 115 00:06:07,365 --> 00:06:12,420 example, in construction and agriculture, shipping, logistics, are able to train quite 116 00:06:12,420 --> 00:06:15,480 local forecasting models to really understand their businesses better. 117 00:06:15,699 --> 00:06:17,460 I think there's a huge opportunity there. 118 00:06:17,460 --> 00:06:21,139 But, yeah, as as Suresh said, nimbleness is still the name of the game. 119 00:06:21,139 --> 00:06:25,139 And I think Meta have corrupted move fast and break things, but move fast and don't break too 120 00:06:25,139 --> 00:06:27,745 much can still reap their enormous rewards, I think. 121 00:06:28,044 --> 00:06:32,305 See, and as Suresh mentioned, obviously, Microsoft being, like, a very fast paced 122 00:06:32,365 --> 00:06:34,944 incumbent, which has been shared products very quickly. 123 00:06:35,164 --> 00:06:39,550 Will Deep have any other notable examples on top of that, which you've been nicely impressed 124 00:06:39,550 --> 00:06:39,790 by? 125 00:06:39,790 --> 00:06:41,709 I think Microsoft really is, is the main one. 126 00:06:41,709 --> 00:06:46,589 The thing on the thing on the ad is actually that Microsoft and Facebook and the way they've 127 00:06:46,589 --> 00:06:51,089 built up their GPU staff in a very clever way, like Microsoft through acquiring inflection, 128 00:06:51,389 --> 00:06:56,115 Facebook through a reserving GPU in advance is incredibly smart way to think about building 129 00:06:56,115 --> 00:07:00,055 models and optimizing what the distribution that they already have. 130 00:07:00,435 --> 00:07:04,035 I think, you know, Facebook obviously have a slightly different approach and that people may 131 00:07:04,035 --> 00:07:08,689 never pay for llama, and they're actually building that in more in a more full way. 132 00:07:08,689 --> 00:07:10,470 They're building in CyberSec

evolves. 133 00:07:10,529 --> 00:07:15,569 There's some detection for prompt injection in this thing that could can damage the way people 134 00:07:15,569 --> 00:07:16,129 use models. 135 00:07:16,129 --> 00:07:18,930 So I think they're also thinking really interesting to me about the way their 136 00:07:18,930 --> 00:07:21,785 distribution to really launch this new wave of products. 137 00:07:22,084 --> 00:07:22,564 Absolutely. 138 00:07:22,564 --> 00:07:27,045 And Sudesh pointed towards, like, on cloud brand use cases and being obviously the the 139 00:07:27,045 --> 00:07:28,245 fastest to respond to. 140 00:07:28,245 --> 00:07:31,125 It seems like, obviously, the biggest companies in the world, like the Microsoft side of 141 00:07:31,125 --> 00:07:34,504 matters, etcetera, obviously, the the fastest to to move. 142 00:07:35,259 --> 00:07:40,220 Is is it true when you think about, like, an enterprise stack and, like, on prem use cases 143 00:07:40,220 --> 00:07:40,879 in particular? 144 00:07:40,939 --> 00:07:44,300 I I personally think that the biggest companies in the world will not really allow the majority 145 00:07:44,300 --> 00:07:48,220 of the data to be put through a different solution to be stored in in some cloud that 146 00:07:48,220 --> 00:07:49,360 they'd never heard of. 147 00:07:49,705 --> 00:07:51,545 I personally think it'll be quite sensitive. 148 00:07:51,545 --> 00:07:55,485 It'll be on prem under wraps under a lot of padlocks. 149 00:07:55,705 --> 00:07:56,665 Do you believe that also? 150 00:07:56,665 --> 00:08:01,145 And, obviously, Willie pointed towards data being highly available, and you don't really 151 00:08:01,145 --> 00:08:04,379 need a huge amount of data to actually accurately forecast models these days. 152 00:08:04,620 --> 00:08:06,720 So, how do you guys think about that? 153 00:08:07,339 --> 00:08:08,860 Yeah, I know it's really true. 154 00:08:08,860 --> 00:08:13,020 You think as much as Aon has advanced enterprise productivity, it's enabled an 155 00:08:13,020 --> 00:08:16,540 enormous amount of malicious actors external to the enterprise. 156 00:08:16,540 --> 00:08:20,875 And, you know, phishing's up 500%, ransomware attacks up enormously as well. 157 00:08:21,014 --> 00:08:25,915 And the biggest companies in the world, their data keeps getting extracted, and encrypted 158 00:08:26,055 --> 00:08:27,194 like Okta and HubSpot. 159 00:08:27,254 --> 00:08:31,810 So I think really thinking about how to keep your data secure and protected, is enormously 160 00:08:32,509 --> 00:08:37,549 important and isn't just about how to defend for easy gains through the use models, but it's 161 00:08:37,549 --> 00:08:42,590 actually external actors as well who are trying to trying to use this amazing technology in a 162 00:08:42,590 --> 00:08:43,309 nefarious way. 163 00:08:43,309 --> 00:08:43,470 Yeah. 164 00:08:43,470 --> 00:08:47,605 So this first time at JPMorgan where sort of the first question we'd ask, any startup we're 165 00:08:47,605 --> 00:08:50,164 sort of looking at was if they could deploy on prem. 166 00:08:50,164 --> 00:08:52,644 And if the answer was no, it was sort of red flag. 167 00:08:52,644 --> 00:08:53,684 We we sort of stopped there. 168 00:08:53,684 --> 00:08:58,860 And it's a hurdle, but I think you we do have to accept that there is a lot shifting to cloud 169 00:08:58,860 --> 00:09:00,139 still, and you can see that in the numbers. 170 00:09:00,139 --> 00:09:03,259 Like, these companies are at 10, 100 of 1,000,000,000 of revenue still growing at 30% 171 00:09:03,259 --> 00:09:03,899 year on year. 172 00:09:03,899 --> 00:09:07,179 It it's a constant question as to as to how much is gonna is gonna shift on cloud. 173 00:09:07,179 --> 00:09:10,620 What can startups do to get a meaningful edge on the large incumbents, do you think? 174 00:09:10,620 --> 00:09:14,294 I mean, you think how are they able to compete on this distribution side? 175 00:09:14,375 --> 00:09:20,134 You look at GPT 4, the code assistant, super fast, like, their increased rate limits in the 176 00:09:20,134 --> 00:09:20,634 API. 177 00:09:20,774 --> 00:09:24,294 I mean, how do you compete the distribution, the new updates, and things like this? 178 00:09:24,294 --> 00:09:26,794 Maybe you can share your point of view first, Will. 179 00:09:27,009 --> 00:09:27,809 Yeah, for sure. 180 00:09:27,809 --> 00:09:31,889 I think on the application layer, maybe to start with, I think there's still incredible 181 00:09:31,889 --> 00:09:35,730 value to unlock by building amazing customer experiences, which might sound like a bit of a 182 00:09:35,730 --> 00:09:40,230 cliche, but you look back at other waves we've had, neo banking being 1. 183 00:09:40,289 --> 00:09:45,045 Like, some of the technologies like are very similar to what to what came before, and a lot 184 00:09:45,045 --> 00:09:49,524 of neo banks are banks, and being able to move incredibly fast, build an amazing experience, I 185 00:09:49,524 --> 00:09:52,825 think, will still apply to applications in AI. 186 00:09:52,980 --> 00:09:56,740 And then in the infrastructure layer where startups, where there's white space, look, I 187 00:09:56,740 --> 00:09:59,559 think close to the hardware, there's loads of opportunities. 188 00:09:59,620 --> 00:10:04,419 I think accessing GPU in an efficient and cheap way is something that one of our portfolio 189 00:10:04,419 --> 00:10:07,815 companies, Fluid Stack, has enabled companies to do excellently well. 190 00:10:08,054 --> 00:10:12,955 And then even the basis where there's like early movers, like lang chain app development, 191 00:10:13,335 --> 00:10:19,495 and base 10 of the Inference Cloud, I still think there's LLMs and AI get as big as people 192 00:10:19,495 --> 00:10:20,220 think they will. 193 00:10:20,299 --> 00:10:25,339 There's enormous value to capture for start ups who provide very, very specific services that 194 00:10:25,339 --> 00:10:31,179 untracked value for for

companies who who'd want to use LLMs, but don't necessarily have 195 00:10:31,179 --> 00:10:33,439 all the picks and shovels to to make it work mutually. 196 00:10:34,554 --> 00:10:34,794 Yeah. 197 00:10:34,794 --> 00:10:36,315 So I think it's it's an interesting question. 198 00:10:36,315 --> 00:10:39,754 I think we sort of a lot a lot of people can sort of fall into into the bucket of, okay. 199 00:10:39,754 --> 00:10:43,274 Number 1, you think all the value is gonna be in the model layer, so you spend all your money 200 00:10:43,274 --> 00:10:43,834 on GPUs. 201 00:10:43,834 --> 00:10:47,514 Or number 2, you're gonna avoid any caps going towards GPUs, and you're gonna build up your 202 00:10:47,514 --> 00:10:48,014 application. 203 00:10:49,080 --> 00:10:53,000 But it's tricky at the moment because, obviously, the sort of negatives of spanning 204 00:10:53,000 --> 00:10:57,000 volume in GPUs as they are becomes very unclear, and the negatives of building up 205 00:10:57,000 --> 00:11:00,600 complete application there is it's hard to differentiate because I think what we what we 206 00:11:00,600 --> 00:11:05,514 need to accept here is that most of the value of the of these of these models, of these AI 207 00:11:05,514 --> 00:11:07,695 driven applications is coming from the model. 208 00:11:07,835 --> 00:11:11,995 So if you think about, say, if you're a team moving 50% quicker than the average other 209 00:11:11,995 --> 00:11:14,335 developer, after a year, you're 50% ahead. 210 00:11:14,394 --> 00:11:18,235 But if the 9% of value is coming from the model, then you're only gonna be 2.5% ahead at 211 00:11:18,235 --> 00:11:19,049 the end of the year. 212 00:11:19,129 --> 00:11:22,809 And so that's why I think there's a lot of value in building at the model and sort of fine 213 00:11:22,809 --> 00:11:27,049 tuning, preference tuning, and not training big models, but training reward models, and try and 214 00:11:27,049 --> 00:11:29,610 sort of enhance what these open source models can do. 215 00:11:29,610 --> 00:11:32,809 And then sort of adjusting the distributions of the model towards your specific use case. 216 00:11:32,809 --> 00:11:37,205 Because at the end of the day, OpenAI, Anthropic, Nostral are training very sort of 217 00:11:37,205 --> 00:11:41,524 uniform horizontal generalized based models that aren't gonna be very good at specific 218 00:11:41,524 --> 00:11:42,024 tasks. 219 00:11:42,085 --> 00:11:44,644 And and that's where I see the sort of immediate opportunity. 220 00:11:44,644 --> 00:11:48,164 But going forward 2, 3 years, I can definitely see some great businesses being building 221 00:11:48,164 --> 00:11:49,205 completely application now 222 00:11:49,205 --> 00:11:49,769 as well. 223 00:11:49,850 --> 00:11:50,250 Makes sense. 224 00:11:50,250 --> 00:11:54,330 A couple weeks ago, we had a friend of mine who was a investor in emerging markets. 225 00:11:54,330 --> 00:11:58,170 He was saying a lot of the value will be at the application layer instead of the infrastructure 226 00:11:58,170 --> 00:11:58,490 layer. 227 00:11:58,490 --> 00:12:01,450 Do you guys agree of that, or do you sort of see the opposite view? 228 00:12:01,450 --> 00:12:05,290 I don't think there's any any sort of obvious answer to to where it will accrue, but I think 229 00:12:05,290 --> 00:12:09,804 we've sort of gotta gotta go to to business 101 and and and leverage your network effects, try 230 00:12:09,804 --> 00:12:13,965 and sort of build some sort of strategy to prevent other companies from taking your market 231 00:12:13,965 --> 00:12:16,445 share and and and value accrue, and that can be done at application layer. 232 00:12:16,445 --> 00:12:18,144 That can be done on infrastructure layer. 233 00:12:18,284 --> 00:12:19,965 And it could be done any anything in between as well. 234 00:12:19,965 --> 00:12:22,339 Do you think 5, 10 years out, same as well? 235 00:12:22,339 --> 00:12:25,779 There'll be value for in the application layer as well as the infra layer as well? 236 00:12:25,779 --> 00:12:29,339 So, you know, we just find pockets of opportunity Yep. 237 00:12:29,620 --> 00:12:30,259 In your opinion? 238 00:12:30,259 --> 00:12:30,419 Yep. 239 00:12:30,419 --> 00:12:31,139 Definitely agree with that. 240 00:12:31,139 --> 00:12:31,379 So I 241 00:12:31,379 --> 00:12:36,579 assume you guys agree that the future of AI is, like, dominated by 1 or 2, like, single models. 242 00:12:36,579 --> 00:12:37,825 Do you think that'll be the case? 243 00:12:37,825 --> 00:12:41,044 Or do you think we'll have fairly decentralized, fragmented ecosystem? 244 00:12:41,345 --> 00:12:44,804 In mind, personally, for mine, like, if you look at, I don't know, Jarvis or Longchain. 245 00:12:45,264 --> 00:12:50,225 I mean, the future really is the mediator from an input web additionally across the models. 246 00:12:50,225 --> 00:12:51,424 What's your guys' thoughts on that? 247 00:12:51,424 --> 00:12:52,325 What do you say? 248 00:12:53,319 --> 00:12:58,519 I think there will be different models in lots of spaces where that has to be, for example, 249 00:12:58,519 --> 00:13:04,600 like regulated data or very specific production level data for manufacturing plants, where the 250 00:13:04,600 --> 00:13:11,274 data is time series or to do with other components of industry where you don't need 251 00:13:11,274 --> 00:13:12,654 natural language, for example. 252 00:13:12,955 --> 00:13:18,634 For more general enterprise productivity where natural language is the the atomic unit of the 253 00:13:18,634 --> 00:13:22,634 of model, I think it's possible that lots of big models or sorry, a few a small number of 254 00:13:22,634 --> 00:13:24,014 big models will dominate. 255 00:13:24,710 --> 00:13:30,230 But I guess if you think about other other industries where the atomic atomic unit is 256 00:13:30,230 --> 00:13:34,629 quite similar between businesses, like telecoms, like, there's still space. 257 00:13:34,629 -->

00:13:40,004 A gigabyte is a gigabyte is a gigabyte, and there will still be space for multiple winners. 258 00:13:40,004 --> 00:13:42,404 Just on that point, big believer in in in specialized models. 259 00:13:42,404 --> 00:13:45,845 Because I think if you if you think about what what's a neural network is, it's it's an 260 00:13:45,845 --> 00:13:46,884 approximation of a function. 261 00:13:46,884 --> 00:13:51,045 And if you've got one large model and lots of smaller models, they're essentially the same 262 00:13:51,045 --> 00:13:51,205 thing. 263 00:13:51,205 --> 00:13:55,360 But then if you factor in cost and inference speed and everything in around that realm, 264 00:13:55,500 --> 00:13:58,539 you're gonna want the small specialized models because they're they're achieving the same 265 00:13:58,539 --> 00:14:01,519 goal, but it's doing a lot more efficiently when you're using the small specialized models. 266 00:14:01,820 --> 00:14:05,980 Whether or not those multiple models will come from one company is is unclear right now, but 267 00:14:05,980 --> 00:14:09,095 I'm definitely a believer in in in smaller fine tuned models for specific purposes. 268 00:14:09,095 --> 00:14:12,774 I mean, general consensus is, like, as long as somebody specialized in the output of 269 00:14:12,774 --> 00:14:15,434 verticals, don't have to be a competitive mode, would you say? 270 00:14:15,575 --> 00:14:16,054 Absolutely. 271 00:14:16,054 --> 00:14:19,095 And I think a lot of that momentum will actually be customer first. 272 00:14:19,095 --> 00:14:23,174 I hope there's a temptation to think tech first and how the model works and from a machine 273 00:14:23,174 --> 00:14:24,899 learning, like what a model enables. 274 00:14:25,200 --> 00:14:31,620 Some customers will 100% need specific fine tune models due to regulation, due to internal 275 00:14:31,679 --> 00:14:32,960 compliance reasons, and so on. 276 00:14:32,960 --> 00:14:38,419 So I think the market will actually encourage that dynamic rather than everyone reverting to 277 00:14:38,784 --> 00:14:40,784 a small number of big models. 278 00:14:40,784 --> 00:14:45,585 Will we see any model used now used also within 2 to 3 years? 279 00:14:45,585 --> 00:14:47,024 Like, do they have a lifetime, would you say? 280 00:14:47,024 --> 00:14:48,225 It depends on the task, I guess. 281 00:14:48,225 --> 00:14:48,705 Because if you 282 00:14:48,705 --> 00:14:52,950 think about I mean and most of deep learning and production today is is models that were 283 00:14:53,190 --> 00:14:54,549 invented 5 or 6 years ago. 284 00:14:54,549 --> 00:14:57,509 So sort of a lot of them look look like your sort of traditional BERT transformers. 285 00:14:57,509 --> 00:15:00,549 If they're solving your problem, there's there's no real reason that you need to keep on 286 00:15:00,549 --> 00:15:00,870 shifting. 287 00:15:00,870 --> 00:15:03,910 I think it's for these use cases where best is always better. 288 00:15:03,910 --> 00:15:07,769

That's that's where we're gonna be considered using the sort of state of the art models. 289 00:15:08,865 --> 00:15:12,225 Let's maybe discuss about the regulatory changes, right, which need to be made in the 290 00:15:12,225 --> 00:15:13,105 world of AI. 291 00:15:13,105 --> 00:15:17,105 What do you see as the most urgent change that we need, not just maybe in Europe in general, 292 00:15:17,105 --> 00:15:21,425 but on a global scale so that we can keep on improving the advancement of AI? 293 00:15:21,425 --> 00:15:25,350 I'm actually the biggest issue replace our mission a little bit earlier is the nefarious 294 00:15:25,350 --> 00:15:30,230 use of AI and figuring out how to regulate, ensure that companies defend themselves 295 00:15:30,230 --> 00:15:31,110 sufficiently against that. 296 00:15:31,110 --> 00:15:34,330 Because ultimately, they're, for you, they're at as much risk as they are. 297 00:15:34,550 --> 00:15:40,065 And the UK does a good job of this, GDPR regulation, which to some extent is anti AI 298 00:15:40,065 --> 00:15:44,625 hacking, right, because it's data security and the fines that the UK has been able to give out 299 00:15:44,625 --> 00:15:49,264 for companies that don't protect themselves against AI native attacks, I think are 300 00:15:49,264 --> 00:15:49,764 meaningful. 301 00:15:50,065 --> 00:15:52,644 So I think that will be incredibly important. 302 00:15:52,785 --> 00:15:56,919 And then I think the spirit of your question is like, how can we all use them without messing 303 00:15:56,919 --> 00:15:57,659 anything up? 304 00:15:57,960 --> 00:16:03,720 And regulation encourages everyone to track how data flows in through model and out, making 305 00:16:03,720 --> 00:16:08,539 sure that brand assets don't escape and get reproduced by other companies querying a model. 306 00:16:08,784 --> 00:16:12,004 Things like how can I embed an open source model into the enterprise? 307 00:16:12,144 --> 00:16:17,264 What guard layer guardrail layers do I need to ensure that my data doesn't influence that 308 00:16:17,264 --> 00:16:19,764 model, and so on will be important. 309 00:16:20,225 --> 00:16:24,019 But a lot of that is actually coming in with the model providers themselves. 310 00:16:24,320 --> 00:16:28,480 As often is the case, I think some of the tech companies will actually be fast before the 311 00:16:28,480 --> 00:16:30,500 regulators because it's what the customer wants. 312 00:16:30,639 --> 00:16:35,024 And Llama, as I mentioned, is a good example by building Cypressack Kvaer into the product. 313 00:16:35,345 --> 00:16:39,985 They're actually front running some of the the regulation will no doubt come in to say, please 314 00:16:39,985 --> 00:16:43,445 protect employees against malicious prompts, and they're already doing. 315 00:16:44,384 --> 00:16:45,345 I bet you agree with Will. 316 00:16:45,345 --> 00:16:48,465 And I think copyright infringement is definitely sort of top line for me at the 317 00:16:48,465 --> 00:16:51,830 moment because there's sort of a massive elephant in the room there where everyone's 318 00:16:51,830 --> 00:16:53,269 sort of just training over everything. 319 00:16:53,269 --> 00:16:57,029 And I think that those questions will be answered, but I think they really need to be 320 00:16:57,029 --> 00:17:01,830 answered to sort of set up how we how we really advance, these models getting bigger and bigger 321 00:17:01,830 --> 00:17:02,355 and bigger. 322 00:17:02,355 --> 00:17:05,075 Like, it's not just gonna be a few subsets that that that can solve it. 323 00:17:05,075 --> 00:17:09,154 I think we need some proper infrastructures to how we distribute IP and how people can 324 00:17:09,154 --> 00:17:13,015 monetize IP going forward rather than it being the wild west out here at the moment. 325 00:17:14,275 --> 00:17:14,775 Computes. 326 00:17:14,835 --> 00:17:16,930 Like, how much of a problem is that, would you say? 327 00:17:16,930 --> 00:17:21,490 And, like, how optimistic are you that the capacity will increase in the future to host, I 328 00:17:21,490 --> 00:17:23,650 guess, the more good advancement of in this particular sector? 329 00:17:23,650 --> 00:17:25,490 It's always a function of the data you've got available. 330 00:17:25,490 --> 00:17:29,650 Like, there's a there's an optimal ratio of of of how much compute you really should use for 331 00:17:29,650 --> 00:17:32,375 for how much data you've got, and that changed with architectures. 332 00:17:33,075 --> 00:17:35,634 But I think data's currently the the finite thing at the moment. 333 00:17:35,634 --> 00:17:39,634 We can still scale up compute, and I think that the sort of GPU crunch seems to be easing from 334 00:17:39,634 --> 00:17:40,695 from what we've seen. 335 00:17:40,835 --> 00:17:43,575 There's less people training models for the sake of training models. 336 00:17:44,009 --> 00:17:46,970 I think one thing that we haven't scaled up yet, which would be really interesting to see, 337 00:17:46,970 --> 00:17:48,509 is is inference time. 338 00:17:48,650 --> 00:17:49,690 We've scaled up data. 339 00:17:49,690 --> 00:17:50,570 We've scaled compute. 340 00:17:50,570 --> 00:17:53,710 But this at the end of the day, we're still querying these models in in a millisecond. 341 00:17:53,849 --> 00:17:58,554 If we could sort of if we think about how the human brain works, we solve complex tasks over 342 00:17:58,554 --> 00:17:59,434 a long period of time. 343 00:17:59,434 --> 00:18:02,875 If we can allow models to do that, that could increase the accuracy of the model as well. 344 00:18:02,875 --> 00:18:07,115 So I think that's that's one sort of idea of long inference is is something that that I'd 345 00:18:07,115 --> 00:18:08,335 regularly love to see as well. 346 00:18:08,809 --> 00:18:12,269 I think GPU can, in terms of being, I think limiting the right amount of inference. 347 00:18:12,730 --> 00:18:15,929 But I think also in the future, they're going to say the lead

factor. 348 00:18:15,929 --> 00:18:20,650 I think there's other things like chip design may actually limit our rate of development or 349 00:18:20,650 --> 00:18:22,250 Catmull's law more quickly. 350 00:18:22,250 --> 00:18:26,855 And to say, Sebastian's point about inference, freeing out ways to step chips or reuse the 351 00:18:26,855 --> 00:18:32,134 energy produced by chips to avoid like hitting a thermal wall where we can't do inference any 352 00:18:32,134 --> 00:18:37,335 more quickly or more efficiently, I think could well be, a more immediate impact than just pure 353 00:18:37,335 --> 00:18:39,035 production volume of GPU. 354 00:18:40,269 --> 00:18:45,630 In in terms of, like, Mistral and the pricing now of Reliance, like, the pricing of Mistral 355 00:18:45,630 --> 00:18:46,670 particularly has has come out. 356 00:18:46,670 --> 00:18:50,990 So obviously, the media model fairly recently suggests, obviously, perhaps they're they're 357 00:18:50,990 --> 00:18:54,404 reasonably more confident that their models are essentially better than GPTs. 358 00:18:54,865 --> 00:18:56,144 Do you agree or or not? 359 00:18:56,144 --> 00:18:58,224 I think there's there's definitely a land grab at the moment. 360 00:18:58,224 --> 00:19:01,365 And if you think about how how GPUs work, it's it's back to parallelism. 361 00:19:01,585 --> 00:19:04,065 So the more users you have, the cheaper it gets. 362 00:19:04,065 --> 00:19:07,984 And so it does make sense to start offering models with sort of negative unit unit 363 00:19:07,984 --> 00:19:12,450 economics to to capture those users and and to capture that brand, which is something that I 364 00:19:12,450 --> 00:19:14,369 don't think gets discussed enough at the moment. 365 00:19:15,329 --> 00:19:19,029 Because the way these generative models works quite and the output's quite subjective. 366 00:19:19,329 --> 00:19:20,470 Brand really matters. 367 00:19:20,529 --> 00:19:25,474 Like, it's hard to say if OpenAI is really best than Claude or Mistral, but if you use OpenAI 368 00:19:25,474 --> 00:19:28,275 and you enjoy using it, then you're still gonna use it going forward. 369 00:19:28,275 --> 00:19:31,795 And so if you can capture users and they enjoy your brand and they enjoy using their models, 370 00:19:31,795 --> 00:19:33,714 then that that's gonna definitely pay off going forward. 371 00:19:33,714 --> 00:19:36,295 So I think we can see a lot of companies pricing their models. 372 00:19:36,355 --> 00:19:36,855 Yeah. 373 00:19:37,200 --> 00:19:38,159 Very interesting indeed. 374 00:19:38,159 --> 00:19:40,240 Maybe let's discuss about which was more important. 375 00:19:40,240 --> 00:19:40,559 Right? 376 00:19:40,559 --> 00:19:44,659 Do you think it's the model size or the data size, in your opinion? 377 00:19:44,720 --> 00:19:45,839 Maybe this one's for you, Will. 378 00:19:45,839 --> 00:19:51,654 I mean, on the model size, first of all, I think GPU is still enhancing performance. 379 00:19:51,714 --> 00:19:56,134 Like, more GPUs are enabling big models, which I think is like no brainer. 380 00:19:56,275 --> 00:20:00,214 The question is maybe more, how many enterprises really need a 40,000,000,000 381 00:20:00,595 --> 00:20:01,559 parameter model? 382 00:20:01,799 --> 00:20:03,960 That I think is not necessarily solved. 383 00:20:03,960 --> 00:20:07,960 I think there's a bunch of you use cases that are quite clear and people hide to keep 384 00:20:08,039 --> 00:20:08,539 occasions. 385 00:20:09,080 --> 00:20:13,400 But if you think about the way most LMs are being put into production for enterprises right 386 00:20:13,400 --> 00:20:17,975 now, Whether we need all that GPU and compute, I think probably not. 387 00:20:17,975 --> 00:20:20,315 I'm querying, company knowledge. 388 00:20:20,695 --> 00:20:22,535 I'm accessing information really quickly. 389 00:20:22,535 --> 00:20:27,095 I'm trying to share information I know with other people, which are all great use cases. 390 00:20:27,095 --> 00:20:32,200 But whether a huge amount of compute you really require for that, I'm I'm not necessarily sure. 391 00:20:32,820 --> 00:20:38,900 And on the data side is and maybe maybe to add a different ability to Suresh is, I don't think 392 00:20:38,900 --> 00:20:42,974 we've always need loads of data to to get to, ready useful models. 393 00:20:43,214 --> 00:20:48,194 I think a bunch of things can be done with quite a small data center, which I think is 394 00:20:48,654 --> 00:20:51,694 increasingly emerging and especially to fine tune a big model. 395 00:20:51,694 --> 00:20:54,755 Like, you don't actually need that much different data. 396 00:20:54,894 --> 00:20:59,549 Well, if we think about pretraining of base models independently, I mean, the chinchilla 397 00:20:59,609 --> 00:21:04,330 paper stated that there is an optimal ratio for a given parameter count, how many tokens you 398 00:21:04,330 --> 00:21:04,809 should use there. 399 00:21:04,809 --> 00:21:09,130 And so I think as as we can find more data or as the we've got the data out there. 400 00:21:09,130 --> 00:21:11,630 It's cleaning that data and making it appropriate for training. 401 00:21:11,904 --> 00:21:13,984 As we can do that, we can scale up models further. 402 00:21:13,984 --> 00:21:17,744 But as we touched on earlier, I think what what's really important and what we haven't 403 00:21:17,744 --> 00:21:22,224 scaled is is inference and and and and the idea of, like, how can we if if we're scaling up the 404 00:21:22,224 --> 00:21:25,109 training time to months and maybe even years, who knows? 405 00:21:25,190 --> 00:21:29,349 Why can't we scale up the inference time from a millisecond to minutes? 406 00:21:29,349 --> 00:21:31,210 Because that's how the human brain works. 407 00:21:31,269 --> 00:21:33,750 We don't just solve a complex problem in a millisecond. 408 00:21:33,750 --> 00:21:34,789 We take time to do that. 409 00:21:34,789 --> 00:21:38,630 And so we can allow models to sort of

have that aspect of system 2 thinking. 410 00:21:38,630 --> 00:21:40,250 I think that that could be quite powerful. 411 00:21:40,955 --> 00:21:41,195 Yeah. 412 00:21:41,195 --> 00:21:44,714 And maybe you guys can touch upon your experiences within your portfolio on this 413 00:21:44,714 --> 00:21:45,535 particular question. 414 00:21:45,914 --> 00:21:51,515 How difficult really is it to educate and really get to grips with what enterprise need 415 00:21:51,515 --> 00:21:54,634 AI and want AI for from a good market point of view? 416 00:21:54,634 --> 00:21:58,690 Does that seem to be a huge frustration from portfolio companies that are trying to sell to 417 00:21:58,690 --> 00:21:59,190 enterprise? 418 00:21:59,970 --> 00:22:03,730 I think there's a lot of frustration right now because there there's a lot of well, freight is 419 00:22:03,730 --> 00:22:07,730 probably the wrong word, but there's a lot of app type enterprises, but I think none of them 420 00:22:07,730 --> 00:22:08,849 really know what they want. 421 00:22:08,849 --> 00:22:09,455 And I think you can 422 00:22:09,455 --> 00:22:13,494 see that in the BCG numbers and the extension numbers, and these consultants are are sort of 423 00:22:13,494 --> 00:22:16,775 the real winners today because they're helping and they're doing the handholding for for 424 00:22:16,855 --> 00:22:19,335 through the enterprises to to help them get to value with AI. 425 00:22:19,335 --> 00:22:23,095 And we've seen that with one of our portfolio companies, writer, who really focused on that 426 00:22:23,095 --> 00:22:24,394 solutions engineering piece. 427 00:22:24,559 --> 00:22:27,039 So at the end of the day, most of the enterprises don't even know what prompt 428 00:22:27,039 --> 00:22:27,759 engineering is. 429 00:22:27,759 --> 00:22:31,440 And if you can help that and and and help them do that in the sales cycle, that's gonna help 430 00:22:31,440 --> 00:22:35,460 them get to get to value, and people are happy to pay a premium for that time to value. 431 00:22:35,919 --> 00:22:37,839 I don't have really strong opinion on this, actually. 432 00:22:37,839 --> 00:22:42,634 And I think that the the consensus is different on this at different times. 433 00:22:42,775 --> 00:22:49,174 And whereas 2 different pillars, there's one that's solving one GTM solving a specific pain 434 00:22:49,174 --> 00:22:53,275 point that people already have that LMs, like, happen to resolve. 435 00:22:53,590 --> 00:22:58,470 Then there's the other component, which is like all the stuff that LMs can do, that people are 436 00:22:58,470 --> 00:23:01,750 currently getting to grips with that need a little bit more education. 437 00:23:01,750 --> 00:23:05,830 And that's, I'm sure, where business development and GTN teams are really figuring 438 00:23:05,830 --> 00:23:10,715 out what that looks like and enabling companies, whether that's product led or really 439 00:23:10,715 --> 00:23:13,134 sales led, to make the most of these tools. 440 00:23:13,914 --> 00:23:17,674 I would say that, like, there are amazing companies out there that aren't necessarily 441 00:23:17,674 --> 00:23:20,815 building the application or the infrastructure, but are actually thinking about, 442 00:23:20,899 --> 00:23:21,179 okay, how 443 00:23:21,179 --> 00:23:26,899 do we embed AI in workflows and really going deep into companies and understanding, okay, in 444 00:23:26,899 --> 00:23:32,179 healthcare billing and in all sorts of different, very specific verticals, look, the 445 00:23:32,179 --> 00:23:36,974 way people work looks like this and therefore, how do we enable amazing existing technology to 446 00:23:36,974 --> 00:23:42,335 come in here and, and be the people to unlock that value, where, where it is at the moment 447 00:23:42,335 --> 00:23:42,835 trapped. 448 00:23:43,295 --> 00:23:48,409 Moving on to like an HR type of alignment questioning, will we see, in your personal 449 00:23:48,409 --> 00:23:53,529 opinion, like, 10, 20 people, AI companies, like, unicorns the next 5, 6, even 10 years? 450 00:23:53,529 --> 00:23:54,909 We've got it in our portfolio. 451 00:23:54,970 --> 00:23:56,190 We've got VotoRoom. 452 00:23:56,490 --> 00:23:57,210 That's happening. 453 00:23:57,210 --> 00:23:58,649 I think that the trends are all there. 454 00:23:58,649 --> 00:24:02,234 The sort of barriers to entry to building software is is sort of tending towards 0. 455 00:24:02,535 --> 00:24:02,855 Yeah. 456 00:24:02,855 --> 00:24:03,255 I agree. 457 00:24:03,255 --> 00:24:09,494 I think especially in areas where SMEs are being served in a product led way for things 458 00:24:09,494 --> 00:24:14,450 that they've really struggled in the past, like accounts receivable, marketing, the CFO 459 00:24:14,450 --> 00:24:14,929 function. 460 00:24:14,929 --> 00:24:20,289 I think amazing AI companies that really fix a specific problem there can make an enormous 461 00:24:20,289 --> 00:24:20,690 difference. 462 00:24:20,690 --> 00:24:24,549 You don't necessarily need huge teams to scale a business like that. 463 00:24:24,609 --> 00:24:29,525 I think where you do need more personnel is where you're looking at a Sandija type company 464 00:24:29,525 --> 00:24:35,125 where you're not only doing amazing services for big clients, but you're also trading on 465 00:24:35,125 --> 00:24:39,299 stable diffusion models and really building that research component in house. 466 00:24:39,299 --> 00:24:42,180 But I think that's not necessarily a mop out for a huge company. 467 00:24:42,180 --> 00:24:46,600 I think there's many other routes where you can remain very lean and build something big. 468 00:24:46,820 --> 00:24:47,140 Yeah. 469 00:24:47,140 --> 00:24:47,299 No. 470 00:24:47,299 --> 00:24:47,860 It makes sense. 471 00:24:47,860 --> 00:24:51,380 I mean, to get to these size companies, you're gonna need to have the best AI talent. 472 00:24:51,380 --> 00:24:51,619 Right? 473 00:24:51,619 --> 00:24:56,765 So where do you think

this will concentrate from, and what would companies have to do to 474 00:24:57,224 --> 00:24:58,105 win this talent? 475 00:24:58,105 --> 00:24:58,345 Yeah. 476 00:24:58,345 --> 00:25:02,184 So I think there'll be, like, a natural concentration over the next few years just due 477 00:25:02,184 --> 00:25:02,664 to the market. 478 00:25:02,664 --> 00:25:07,065 I think talent will be recycled as some companies shrink in size or even go out of 479 00:25:07,065 --> 00:25:09,759 business, and amazing talent will rejoin. 480 00:25:09,819 --> 00:25:12,079 What, what are the new emerging winners? 481 00:25:12,139 --> 00:25:13,759 And that is a natural phenomenon. 482 00:25:13,819 --> 00:25:15,740 That's the reason why you should build in tough times. 483 00:25:15,740 --> 00:25:16,139 Right? 484 00:25:16,299 --> 00:25:18,779 Because you get better access to, to amazing talent. 485 00:25:18,779 --> 00:25:21,659 The other thing is like, a lot of these fields are still really early. 486 00:25:21,659 --> 00:25:27,335 Like, if you think about, a classical MLM is might use supervised learning, unsupervised 487 00:25:27,474 --> 00:25:29,095 learning, reinforcement learning. 488 00:25:29,234 --> 00:25:34,115 Like, they reinforcement learning particularly now really haven't been along for, much time. 489 00:25:34,115 --> 00:25:35,575 It's not usually well understood. 490 00:25:35,875 --> 00:25:39,900 So I think that in itself will just attract people who want to be on the cutting edge of 491 00:25:39,900 --> 00:25:42,000 developing their machine learning technology. 492 00:25:42,220 --> 00:25:45,900 And then I think there's a suite of other things that in the future will be incredibly 493 00:25:45,900 --> 00:25:46,220 exciting. 494 00:25:46,220 --> 00:25:50,779 Like, federated learning will attract huge amount of talent whereby a central model can be 495 00:25:50,779 --> 00:25:55,075 trained by decentralized private data, which some banks in China are already experimenting 496 00:25:55,075 --> 00:25:55,555 with. 497 00:25:55,555 --> 00:25:59,315 You know, neuromorphic computing, reflecting the brain's qualities where inference and 498 00:25:59,315 --> 00:26:01,095 training can happen at the same time. 499 00:26:01,234 --> 00:26:06,434 And many other kind of disciplines, which are not yet underlying in LLM or machine learning 500 00:26:06,434 --> 00:26:07,494 is perhaps in general. 501 00:26:07,690 --> 00:26:11,710 But I think in the future, it will no doubt, like, enable huge progress and attract 502 00:26:12,009 --> 00:26:15,690 I'm very bullish on on open source frameworks helping to to flatten that curve 503 00:26:15,690 --> 00:26:16,009 of talent. 504 00:26:16,009 --> 00:26:16,410 Because if you 505 00:26:16,410 --> 00:26:21,255 think about 10 years ago, the number of people that could train a a neural network would be in 506 00:26:21,255 --> 00:26:24,075 sort of a 100 people or a convolutional that would be a 100 people. 507 00:26:24,134 --> 00:26:27,575 And

now if you think about the amount of people that could train a language model last year, 508 00:26:27,575 --> 00:26:29,515 they've probably been sort of same order of magnitude. 509 00:26:29,894 --> 00:26:34,134 Whereas if you think about the number of people that can train a CNN now, it's it's anyone with 510 00:26:34,134 --> 00:26:35,195 7 lines of code. 511 00:26:35,380 --> 00:26:40,259 And so very bullish as to as to how sort of the likes of PyTorch, TensorFlow can can sort of 512 00:26:40,259 --> 00:26:43,859 really flatten that curve as to who can do what, enabling a sort of a a broader 513 00:26:43,859 --> 00:26:45,059 distribution of of what talent 514 00:26:45,059 --> 00:26:45,859 can open or close. 515 00:26:45,859 --> 00:26:46,740 That's the big question. 516 00:26:46,740 --> 00:26:47,220 Right? 517 00:26:47,220 --> 00:26:49,035 I'm I'm I'm bullish on on open source. 518 00:26:49,035 --> 00:26:53,035 I think as as we touched earlier, I think that there's a lot that can be done within the 519 00:26:53,035 --> 00:26:56,555 models and and sort of adjusting the premises of the models, adjusting the distribution to 520 00:26:56,555 --> 00:27:00,394 what the models can do, to to help really get to any of the value. 521 00:27:00,394 --> 00:27:02,815 And I think you can only really do that with open source models. 522 00:27:03,140 --> 00:27:07,700 I think open source helps solve a lot of concerns in how we think about and it's not 523 00:27:07,700 --> 00:27:09,140 just the there's a broad sense of open. 524 00:27:09,140 --> 00:27:13,220 I think we we sort of really should shift properly open source models where we sort of 525 00:27:13,220 --> 00:27:16,820 really know what's going on in the data going into these models, the training pipelines, and 526 00:27:16,820 --> 00:27:20,035 that that helps enterprise bring comfort in the end output as well. 527 00:27:20,035 --> 00:27:23,335 And that that's what we've seen from from writers, which are now portfolio companies who, 528 00:27:23,714 --> 00:27:26,994 are very transparent about what what's going into their model, and then that helps bring the 529 00:27:26,994 --> 00:27:29,815 comfort to enterprises that are that are pointing these going forward as well. 530 00:27:30,250 --> 00:27:30,490 No. 531 00:27:30,490 --> 00:27:31,309 It's really interesting. 532 00:27:31,450 --> 00:27:36,329 More generally, would you say from a macro point of view, what the biggest or most common 533 00:27:36,329 --> 00:27:39,789 misconception that public have about AI, which you wish you would change? 534 00:27:40,970 --> 00:27:41,450 It's probably 535 00:27:41,450 --> 00:27:42,404 that AI doesn't work. 536 00:27:42,565 --> 00:27:47,205 I think at the end of the day, I know a lot of people that sort of try Chatbt once and say, 537 00:27:47,205 --> 00:27:48,244 oh, it didn't didn't do this. 538 00:27:48,244 --> 00:27:49,285 It didn't do that. 539 00:27:49,285 --> 00:27:52,805 But at the end of the day, the only real way

you you sort of get get max value of these 540 00:27:52,805 --> 00:27:54,164 models is is by trial and error. 541 00:27:54,164 --> 00:27:58,200 And so I think keen for for more and more consumers to really get to grip through these 542 00:27:58,200 --> 00:28:00,539 models and understand what they can do, what they can't do. 543 00:28:00,840 --> 00:28:05,160 And and that should should solve a lot of concerns to to sort of model market fit today. 544 00:28:05,160 --> 00:28:06,119 I think it's a good question. 545 00:28:06,119 --> 00:28:08,619 The one that really frustrates me is that it's magic. 546 00:28:08,920 --> 00:28:13,055 I think models are basically using especially in national language models. 547 00:28:13,055 --> 00:28:16,494 They're listening at like, what are other common sentences or common sentence 548 00:28:16,494 --> 00:28:18,255 construction, dah, dah, dah, dah. 549 00:28:18,255 --> 00:28:21,775 What might this next sentence be or next word, next token, so on? 550 00:28:21,775 --> 00:28:22,835 Like, not magic. 551 00:28:23,134 --> 00:28:29,690 Enabling that education component will will make conversations around AI AGI much more 552 00:28:29,910 --> 00:28:32,490 effective, where people can share some of the same premises. 553 00:28:32,549 --> 00:28:36,230 That this isn't anything like weird and spooky and impossible to understand. 554 00:28:36,230 --> 00:28:40,570 It's actually explainable in the traditional sense, if not machine learning explainable. 555 00:28:42,254 --> 00:28:46,494 You guys are much more calm, obviously, than the mainstream media, which capturing the 556 00:28:46,494 --> 00:28:48,335 headlines is obviously the end of the world. 557 00:28:48,335 --> 00:28:50,174 As we know, it's gonna replace jobs. 558 00:28:50,174 --> 00:28:54,595 For instance, being the lawyer from a services point of view or an accountant could become 559 00:28:54,734 --> 00:28:57,380 extinct in the next 5 to 10 years. 560 00:28:57,380 --> 00:28:59,240 Do you guys agree with that or not? 561 00:28:59,539 --> 00:29:00,980 I think it's really hard to say. 562 00:29:00,980 --> 00:29:08,019 I think in areas where LMs are doing amazing jobs at automating repetitive knowledge 563 00:29:08,019 --> 00:29:13,835 intensive information oriented tasks like law and procurement when there's a lot of forms and 564 00:29:13,835 --> 00:29:16,095 heavily natural, natural language based tasks. 565 00:29:16,634 --> 00:29:19,855 I think they'll accelerate the work completed over time. 566 00:29:20,075 --> 00:29:24,394 Now, I suspect it will actually mean we all get more lawyers and there's more litigation in 567 00:29:24,394 --> 00:29:24,894 general. 568 00:29:25,150 --> 00:29:29,710 It doesn't necessarily mean that lawyers will disappear because I think lawyers as an 569 00:29:29,710 --> 00:29:36,029 interface with clients, lawyers as the ultimate arbiters in many cases, will be continue to be 570 00:29:36,029 --> 00:29:37,009 very, very important. 571 00:29:37,294 --> 00:29:40,174 So I don't necessarily agree that it will it will remove those jobs. 572 00:29:40,174 --> 00:29:45,375 I think it will actually if the promise future comes through where, like, a lot of this stuff 573 00:29:45,375 --> 00:29:48,894 comes cheaper and accessible, I think it will actually enable people to access great 574 00:29:48,894 --> 00:29:50,595 services, more easily. 575 00:29:51,190 --> 00:29:51,670 Definitely agree with that. 576 00:29:51,670 --> 00:29:55,450 And I think big big big fan of sort of the the the augmenting thesis for copilots. 577 00:29:55,590 --> 00:29:59,750 And if we look at what AR is really good at, it's preventing anyone from doing the same 578 00:29:59,750 --> 00:30:00,490 thing twice. 579 00:30:00,869 --> 00:30:05,369 And so, yeah, I think it it's rather than replacing work, I think it's gonna help people 580 00:30:05,744 --> 00:30:09,345 work on things that they actually want to, enabling more productivity across the 581 00:30:09,345 --> 00:30:09,744 enterprise. 582 00:30:09,744 --> 00:30:13,744 What do you think about economies that have large segments that rely heavily on, like, 583 00:30:13,744 --> 00:30:14,464 freelance work? 584 00:30:14,464 --> 00:30:14,944 Right? 585 00:30:14,944 --> 00:30:17,585 How much of a impact do you think AI will have? 586 00:30:17,585 --> 00:30:18,085 Yeah. 587 00:30:18,144 --> 00:30:22,720 There there are some savings that, like, will to some extent be limited in scale. 588 00:30:22,720 --> 00:30:28,820 Like, some of their classical consulting use cases where young folk, like, find data, 589 00:30:29,039 --> 00:30:31,940 organize it, and then communicate it graphically. 590 00:30:32,515 --> 00:30:34,994 I think of all pretty good AI use cases. 591 00:30:34,994 --> 00:30:39,234 And frankly, enterprises were paying too much for freelancers to do that in 1st place. 592 00:30:39,234 --> 00:30:42,434 So I think that's like one interesting way to think about it. 593 00:30:42,434 --> 00:30:45,075 And I mean, a lot of those were very smart people anyway. 594 00:30:45,075 --> 00:30:47,654 I feel like they found interesting new ways to contribute to society. 595 00:30:47,869 --> 00:30:52,029 I guess other freelancing will just be immensely enhanced, like things like graphic 596 00:30:52,029 --> 00:30:54,450 design, website design. 597 00:30:54,670 --> 00:30:59,970 We're already seeing a lot of AI take place in websites like Upwork and Fiverr. 598 00:31:00,190 --> 00:31:04,904 I think everyone who who earns via those channels will, their workflows will be enhanced 599 00:31:04,904 --> 00:31:05,464 by AI. 600 00:31:05,464 --> 00:31:10,265 And now I think their earnings will probably also be enhanced because they can complete more 601 00:31:10,265 --> 00:31:12,605 high quality work in a specific time period. 602 00:31:15,144 --> 00:31:15,304 Yeah. 603 00:31:15,304 --> 00:31:16,444 I completely agree. 604 00:31:16,984 --> 00:31:21,690 I I would love to get your take on,

we spoke about before, specializing in a particular 605 00:31:21,690 --> 00:31:22,970 vertical or 2. 606 00:31:22,970 --> 00:31:26,730 And last time I saw Suresh, it was mainly a discussion around particularly the health care 607 00:31:26,730 --> 00:31:27,210 space. 608 00:31:27,210 --> 00:31:27,610 If you had 609 00:31:27,610 --> 00:31:28,009 to pick 610 00:31:28,009 --> 00:31:33,450 one vertical that you think is most promising or most excited about in the current context of 611 00:31:33,450 --> 00:31:35,644 AI, What would you say? 612 00:31:36,585 --> 00:31:40,264 I think I think we'll touch on this earlier, but I think we're seeing and then the way we're 613 00:31:40,264 --> 00:31:44,184 jam try and think about the application there with AI is is the notion of model market fit. 614 00:31:44,184 --> 00:31:47,144 It's like, as these models get better, what can they really do? 615 00:31:47,144 --> 00:31:52,419 And we've seen them with sort of SDR space enable a lot of value, and recently, we've seen 616 00:31:52,419 --> 00:31:57,000 these models really nailed, some sort of, lower value lee legal workflows. 617 00:31:57,299 --> 00:32:00,644 And that's that's driving a genuine shift in in how a lot of legal firms operate. 618 00:32:01,045 --> 00:32:05,125 They're they're they're using legal legal AI, and they're now shifting from billable hours to 619 00:32:05,125 --> 00:32:06,025 value based work. 620 00:32:06,565 --> 00:32:09,285 And whenever you start to see those shift in business models, that's that's a lot of reason 621 00:32:09,285 --> 00:32:10,105 to get excited. 622 00:32:10,644 --> 00:32:14,325 We're we're seeing, I think, a a complete rising tide across the legal tech sector at the 623 00:32:14,325 --> 00:32:18,700 moment, which is a notoriously hard set to crack because lawyers hate using software. 624 00:32:19,000 --> 00:32:20,039 But there's there's there's a 625 00:32:20,039 --> 00:32:21,319 huge amount of market demand for it. 626 00:32:21,319 --> 00:32:21,819 Right? 627 00:32:22,519 --> 00:32:22,839 Yeah. 628 00:32:22,839 --> 00:32:27,960 I mean, I I understand that health care, which isn't necessarily based on, Ellen Moran to 629 00:32:27,960 --> 00:32:28,460 develop. 630 00:32:28,835 --> 00:32:33,394 It's actually a lot of it is based on previous iterations of AI technology and 2 companies in 631 00:32:33,394 --> 00:32:35,335 our portfolio, VIZAI and Ezra. 632 00:32:35,555 --> 00:32:40,674 I think they're just creating enormous benefits in terms of access, affordability for pretty 633 00:32:40,674 --> 00:32:41,575 heavy lift healthcare. 634 00:32:42,035 --> 00:32:46,430 Ezra does MRI scans, For example, they're very easy to to have some demand. 635 00:32:46,570 --> 00:32:53,529 And I think AI's potential to bring enormous benefits to society and healthcare, is 636 00:32:53,529 --> 00:32:54,009 fantastic. 637 00:32:54,009 --> 00:33:00,345 And I think you have this like enormous field of operational stuff that was in, hospitals and 638 00:33:00,345 --> 00:33:05,305 care homes and staffing and so on where there will be amazing companies too. 639 00:33:05,305 --> 00:33:11,740 But I've been really on that, like, survey people application layer application is almost 640 00:33:11,740 --> 00:33:14,460 the wrong word because it's like, they're not necessarily users. 641 00:33:14,460 --> 00:33:17,819 Like the doctor maybe is the more of the user, but they're certainly the customer. 642 00:33:17,819 --> 00:33:20,399 I think that kind of thing, will be amazing. 643 00:33:21,115 --> 00:33:21,755 Yeah, absolutely. 644 00:33:21,994 --> 00:33:25,515 I think especially like the back office solutions that are using Copilot, I think are 645 00:33:25,515 --> 00:33:26,954 particularly interesting even in healthcare. 646 00:33:26,954 --> 00:33:30,474 As Richard Lee said, we want a bit of a quick fire and I guess I probably already know the 647 00:33:30,474 --> 00:33:31,755 answer to this particular question. 648 00:33:31,755 --> 00:33:34,670 So I'm looking forward to seeing how strong willed you guys are. 649 00:33:34,670 --> 00:33:37,330 But, yeah, like, super short concise answers to these. 650 00:33:37,390 --> 00:33:42,029 Is now, like, a fundamental shift in AI, or is it a result of potentially just hype cycles 651 00:33:42,029 --> 00:33:42,670 doing their work? 652 00:33:42,670 --> 00:33:45,089 Let's maybe start with Suresh on this front. 653 00:33:45,544 --> 00:33:45,704 Yeah. 654 00:33:45,704 --> 00:33:48,105 I I think it is as as a January friends, we also shift. 655 00:33:48,105 --> 00:33:52,265 If you think about what AI engineering is today, it's it's very different to what it was 656 00:33:52,265 --> 00:33:52,904 2 years ago. 657 00:33:52,904 --> 00:33:58,184 I think 2 years ago, everyone was stuck in PyTorch and TensorFlow, dealing with their loss 658 00:33:58,184 --> 00:33:59,944 rates and and everything in between that. 659 00:33:59,944 --> 00:34:05,039 But now people are just interfacing with a track interface, and and that's really what AI 660 00:34:05,039 --> 00:34:06,099 engineering is today. 661 00:34:06,159 --> 00:34:10,019 And so, yeah, I'm bullish that this is a this is a genuine shift in in in what we're doing. 662 00:34:10,559 --> 00:34:11,859 It's both jacked, both. 663 00:34:12,079 --> 00:34:14,875 We are on the high cycle and valuations high, and everyone's I 664 00:34:14,875 --> 00:34:15,994 was waiting for that answer. 665 00:34:15,994 --> 00:34:16,394 It did. 666 00:34:16,394 --> 00:34:20,635 As Samir said, it's there's also like amazing things being built and a thing. 667 00:34:20,635 --> 00:34:25,275 In all the categories we've mentioned today, like almost all of them are big enough for like 668 00:34:25,275 --> 00:34:30,500 several amazing companies who are guessing all the competitive dynamics and things we we we 669 00:34:30,500 --> 00:34:31,719 discussed about commoditization. 670 00:34:32,340 --> 00:34:35,960 Like, there is amazing opportunity, and I

do think it will take good lives. 671 00:34:36,500 --> 00:34:38,019 I would like to move on to the next one. 672 00:34:38,019 --> 00:34:40,739 We talked about employment and having effect on AI. 673 00:34:40,739 --> 00:34:45,285 Do you think citizens nowadays should be worried that AI will firstly, will they be 674 00:34:45,285 --> 00:34:45,605 worried? 675 00:34:45,605 --> 00:34:50,164 And secondly, do you think you'll destroy more jobs than create more jobs, in your opinion? 676 00:34:50,164 --> 00:34:54,210 I think when technologies like this, they're like, they often, in the short term, create 677 00:34:54,210 --> 00:34:57,969 amazing jobs and we discussed how like there'll be a boom in law, graphic design, all these 678 00:34:57,969 --> 00:34:59,570 things that LLMs are already doing. 679 00:34:59,570 --> 00:35:03,409 In the medium term, there will be some reallocation of jobs. 680 00:35:03,809 --> 00:35:09,094 In consulting, as I mentioned, we organize data, format it and sell a piece of a 681 00:35:09,094 --> 00:35:10,474 PowerPoint presentation, typically. 682 00:35:11,414 --> 00:35:13,434 Some of that time might well be reorganized. 683 00:35:14,614 --> 00:35:18,614 We're talking about people who a lot of them have the skills to reeducate themselves. 684 00:35:18,614 --> 00:35:23,460 And for the last 50, 100 years, like, the economy has been driven by people reeducate 685 00:35:23,519 --> 00:35:25,299 themselves as as technology emerges. 686 00:35:25,760 --> 00:35:29,599 And, I definitely think this will be the tide that lifts all boats as opposed to a lot of 687 00:35:29,599 --> 00:35:33,599 benefits flowing to specific people in specific roles while other people are displaced, which I 688 00:35:33,599 --> 00:35:35,059 think is a exciting phenomenon. 689 00:35:35,474 --> 00:35:40,114 So you think, in your opinion, that a lot of certain industries, right, will be a lot of 690 00:35:40,114 --> 00:35:42,934 reskilling needed and a sort of upskilling in certain sectors? 691 00:35:43,235 --> 00:35:44,114 There'll be some. 692 00:35:44,114 --> 00:35:49,930 I think people understanding how Alums can support them while continuing to do quite a 693 00:35:49,930 --> 00:35:50,489 similar job. 694 00:35:50,489 --> 00:35:55,710 There will also be people where the LMS can take over a a big chunk of their fundamental 695 00:35:56,489 --> 00:35:57,469 unit of work. 696 00:35:57,610 --> 00:36:01,530 And some of those people will be able to do more of it as a result, but some of them will 697 00:36:01,530 --> 00:36:02,969 also move on to other roles. 698 00:36:02,969 --> 00:36:08,734 And, yeah, I think one of the good thing about this wave of technology is that it is possible 699 00:36:08,734 --> 00:36:09,695 to reeducate yourself. 700 00:36:09,695 --> 00:36:13,135 Probably using perplexity or to research other things you can do. 701 00:36:13,135 --> 00:36:16,110 I I think, that's a natural thing that will happen a little bit. 702 00:36:16,349 --> 00:36:16,510 Yeah. 703 00:36:16,510 --> 00:36:17,630 I think I think completely agree. 704 00:36:17,630 --> 00:36:20,110 I think, it's a good point in education. 705 00:36:20,110 --> 00:36:24,110 If we look at the sort of recent OpenAI demo from yesterday, it's it's it's a kid learning 706 00:36:24,110 --> 00:36:29,025 how to learn Pythagoras theorem straight from from AI and in real time as well. 707 00:36:29,184 --> 00:36:33,505 And so I think it's the scope of what we can do now and the scope of what possible now has has 708 00:36:33,505 --> 00:36:38,545 increased, and people can can use the these models to to help leverage them going forward. 709 00:36:38,545 --> 00:36:42,670 So very very much in the camp that there'll be a net creation of jobs going forward. 710 00:36:42,670 --> 00:36:44,190 Let let me finish off with the final one. 711 00:36:44,190 --> 00:36:45,469 Maybe this is for the both of you. 712 00:36:45,469 --> 00:36:45,949 Right? 713 00:36:45,949 --> 00:36:49,949 We're all young, but in terms of in your opinion, you guys have done pretty well working 714 00:36:49,949 --> 00:36:51,309 at the top European funds. 715 00:36:51,309 --> 00:36:55,710 I mean, if there was someone to that was looking to break into VC, what piece of advice 716 00:36:55,710 --> 00:36:58,264 would you give them to break into the 717 00:36:58,565 --> 00:36:58,804 industry? 718 00:36:58,804 --> 00:36:59,545 Good question. 719 00:36:59,684 --> 00:37:04,164 I think it's important to to read yourself, and that sounds too shaded and boring, but it is 720 00:37:04,164 --> 00:37:06,724 because everyone has a different style doing this job. 721 00:37:06,724 --> 00:37:10,905 And some people are driven by purely curiosity about software. 722 00:37:11,119 --> 00:37:15,219 Other people are driven by pure hustle of people, like pure commitment to entrepreneurs. 723 00:37:15,519 --> 00:37:16,079 By the way, all 724 00:37:16,079 --> 00:37:17,240 those are really important, but like 725 00:37:17,360 --> 00:37:17,599 Yeah. 726 00:37:17,599 --> 00:37:18,400 You need all those. 727 00:37:18,400 --> 00:37:19,440 You need all of those. 728 00:37:19,440 --> 00:37:21,360 But different ones of them come written actually. 729 00:37:21,360 --> 00:37:26,515 So whether it is like sourcing deals and sending them to VCs or supporting founders in 730 00:37:26,515 --> 00:37:30,675 your spare time, or even investing a £100 on seeders because you are obsessed with that type 731 00:37:30,675 --> 00:37:31,335 of technology. 732 00:37:31,635 --> 00:37:33,554 I think those are all relevant ways in. 733 00:37:33,554 --> 00:37:35,635 I think people often talk about writing as a as a way in. 734 00:37:35,635 --> 00:37:38,309 I think that is important and it is amazing to organize your thoughts. 735 00:37:38,309 --> 00:37:42,309 I think there's a lot of young VC doing, so I'm not not necessarily convinced it's it's the 736 00:37:42,309 --> 00:37:42,809 way. 737 00:37:43,670 --> 00:37:44,710 Or or doing a podcast. 738 00:37:44,710 --> 00:37:45,030 Right? 739 00:37:45,030 --> 00:37:45,750 Or doing podcasts. 740 00:37:45,750 --> 00:37:46,250 Exactly. 741 00:37:46,789 --> 00:37:49,289 One word where that sticks to mind here is proactive. 742 00:37:49,670 --> 00:37:52,784 I think this is is definitely a job where you're out there. 743 00:37:52,784 --> 00:37:54,385 You you've really got to be proactive. 744 00:37:54,385 --> 00:37:57,824 You really got to sort of be independent in sort of what makes sense. 745 00:37:57,824 --> 00:38:01,985 And I think, yeah, if you can be proactive, you can sort of build your network, speak to the 746 00:38:01,985 --> 00:38:05,159 right people, then I think everything should should should fall into place. 747 00:38:05,639 --> 00:38:05,799 Yeah. 748 00:38:05,799 --> 00:38:06,119 Absolutely. 749 00:38:06,119 --> 00:38:07,319 It's been honestly amazing. 750 00:38:07,319 --> 00:38:10,119 It's been really useful for both me and Richie, and hopefully all the listeners. 751 00:38:10,119 --> 00:38:12,199 It's been a blast having you both on. 752 00:38:12,199 --> 00:38:13,960 Hope you really enjoyed it as much as we did. 753 00:38:13,960 --> 00:38:14,679 It's been a pleasure. 754 00:38:14,679 --> 00:38:15,500 Thanks, guys. 755 00:38:15,960 --> 00:38:16,280 Well, I'm 756 00:38:16,280 --> 00:38:18,380 sure you can all agree that was an awesome episode. 757 00:38:18,724 --> 00:38:20,804 Thanks so much for joining us on Bench Passports. 758 00:38:20,804 --> 00:38:24,724 We hope you've discovered new insights and inspirations from today's episode that you can 759 00:38:24,724 --> 00:38:29,045 apply in your own line of work, and make sure to tune in next time as we continue to unveil 760 00:38:29,045 --> 00:38:31,602 the extraordinary in every corner of the global market. 761 00:38:32,002 --> 00:38:35,062 In the meantime, you can follow us on socials at vcpassport. 762 00:38:35,762 --> 00:38:36,582 Safe travels.