1 00:00:00,200 --> 00:00:02,520 Clinical trial can take upwards of 10 years. 2 00:00:02,680 --> 00:00:05,520 It costs on average 1 to $2 billion per drug to get to 3 00:00:05,520 --> 00:00:07,240 market. These are huge numbers and huge 4 00:00:07,240 --> 00:00:09,080 timeframes. Writing clinical grade 5 00:00:09,080 --> 00:00:11,200 documentation where the stakes are really high, we're talking 6 00:00:11,200 --> 00:00:13,640 about human lives and human safety and things like that is 7 00:00:13,640 --> 00:00:16,960 incredibly challenging. And it just staggers me that how 8 00:00:16,960 --> 00:00:19,720 many people still do this in Microsoft Word, use Microsoft 9 00:00:19,760 --> 00:00:22,600 Word to essentially model much at the author of model these 10 00:00:22,600 --> 00:00:24,480 trials. Patrick Lung is the CTO of 11 00:00:24,480 --> 00:00:26,680 Pharaoh Health and a former engineering lead at Google 12 00:00:26,680 --> 00:00:28,680 Duplex. So we published a paper with 13 00:00:28,680 --> 00:00:31,080 Merck that sort of. To analyze like what is the 14 00:00:31,080 --> 00:00:34,120 benefit of using the FIRE study designer to optimize clinical 15 00:00:34,120 --> 00:00:36,480 trial design? And we found that the savings 16 00:00:36,480 --> 00:00:38,680 over the course of the lifetime of the trial. 17 00:00:38,680 --> 00:00:41,760 Can be north of $100 million. We have customers using this 18 00:00:41,760 --> 00:00:44,040 document generation system and they're super impressed with 19 00:00:44,120 --> 00:00:47,160 documentation that normally takes potentially months and it 20 00:00:47,160 --> 00:00:49,560 can be generated in a matter of, in some cases, minutes. 21 00:00:49,640 --> 00:00:53,000 We're also in charge on building the Google Duplex, right? 22 00:00:53,200 --> 00:00:56,640 Did you already use something like LLM or is it using a 23 00:00:56,640 --> 00:01:00,600 different technology altogether? It didn't use LLMS at all. 24 00:01:00,600 --> 00:01:02,760 It predated that. People started freaking out when 25 00:01:02,760 --> 00:01:04,879 we launched this product because it sounded so lifelike. 26 00:01:04,959 --> 00:01:07,280 And there were even some people, prominent names of the press, 27 00:01:07,280 --> 00:01:08,400 who thought that we faked the whole thing. 28 00:01:08,520 --> 00:01:12,000 And yet it wasn't using any of this modern LLM architecture at 29 00:01:12,000 --> 00:01:13,160 all. It was using other methods that 30 00:01:13,160 --> 00:01:15,840 were much, much simpler. What do you think people should 31 00:01:15,840 --> 00:01:19,240 do in this kind of era where things moving so fast? 32 00:01:19,640 --> 00:01:22,840 AI won't replace humans per SE, but people who are able to 33 00:01:22,840 --> 00:01:25,280 effectively wield AI will definitely replace people who 34 00:01:25,280 --> 00:01:42,740 aren't. Hello guys. 35 00:01:42,740 --> 00:01:44,780 Welcome back to another new episode of the Technology 36 00:01:44,780 --> 00:01:47,340 General Podcast. Today I have with me the CEO of 37 00:01:47,340 --> 00:01:51,520 Pharaoh Health, Patrick Leong. So we have been hearing a lot 38 00:01:51,520 --> 00:01:54,960 about AI development recently. Viral health itself is trying to 39 00:01:54,960 --> 00:01:58,640 apply AI on, you know, clinical trials and life sciences and 40 00:01:58,640 --> 00:02:00,720 health and all that. I think it's going to be really 41 00:02:00,720 --> 00:02:04,200 interesting to understand how AI can be applied in this type of 42 00:02:04,200 --> 00:02:06,360 work as well. So Patrick, welcome to the show. 43 00:02:07,120 --> 00:02:08,240 Thanks, Henry. Great to be here. 44 00:02:09,039 --> 00:02:11,640 Right, Patrick, I always love to invite my guests to share a 45 00:02:11,640 --> 00:02:14,360 little bit more about yourself, maybe by sharing any career 46 00:02:14,360 --> 00:02:16,640 turning points that you think we all can learn from you. 47 00:02:17,440 --> 00:02:19,440 Yeah. I mean, I guess the beginning of 48 00:02:19,440 --> 00:02:20,920 my career was very entrepreneurial. 49 00:02:20,920 --> 00:02:24,320 Like I joined a company that was in the very early stages and it 50 00:02:24,320 --> 00:02:27,640 went all the way to an IPO and then back down to, you know, 51 00:02:27,640 --> 00:02:30,720 near 0 again duringthe.com boom. And so that was a huge learning 52 00:02:30,720 --> 00:02:33,040 experience. And then sometime after that I 53 00:02:33,040 --> 00:02:35,440 made the big decision to join a big company. 54 00:02:35,440 --> 00:02:38,880 I joined Google back in 2007. And so that was a huge career 55 00:02:38,880 --> 00:02:41,240 turning point because I never thought I'd find myself. 56 00:02:41,240 --> 00:02:44,040 At a big company. But against all odds, I did and 57 00:02:44,040 --> 00:02:45,600 I ended up there for over 10 years. 58 00:02:46,440 --> 00:02:48,080 So I think it's really interesting as well. 59 00:02:48,080 --> 00:02:50,640 Back then when you were at Google, right, you work on some 60 00:02:50,640 --> 00:02:53,160 of the AI technologies like Duplex, right? 61 00:02:53,280 --> 00:02:57,120 So I think since then you have been working a lot on AI related 62 00:02:57,120 --> 00:02:59,360 stuff even before this Gen. AI era, right? 63 00:02:59,640 --> 00:03:02,440 So what do you see in the last, I don't know, maybe 10 years or 64 00:03:02,440 --> 00:03:05,760 maybe 8 years, right. So what are the some 65 00:03:05,920 --> 00:03:09,400 advancements that you can see working on AI things right in 66 00:03:09,400 --> 00:03:11,760 the past compared to now? So what do you see are the 67 00:03:11,760 --> 00:03:15,480 differences? Yeah, I mean, it's easy to 68 00:03:15,480 --> 00:03:18,680 forget now because this whole large language model revolution 69 00:03:18,680 --> 00:03:20,960 has just completely taken over everything. 70 00:03:21,080 --> 00:03:24,200 And so when most people on this planet think about AI, they they 71 00:03:24,200 --> 00:03:26,960 really kind of thinking about large language models as opposed 72 00:03:26,960 --> 00:03:29,640 to AI in general, like data science in general. 73 00:03:30,160 --> 00:03:32,800 And So what I saw when I was back at Google was there was 74 00:03:32,800 --> 00:03:36,520 already an AI revolution that was well in full swing, which 75 00:03:36,520 --> 00:03:38,840 was mostly related to image recognition. 76 00:03:39,240 --> 00:03:41,920 Like because processing power kind of passed a certain 77 00:03:41,920 --> 00:03:45,560 threshold, we became much more able to interpret images. 78 00:03:45,640 --> 00:03:48,240 I'm using AI. And so this resulted in all 79 00:03:48,240 --> 00:03:52,080 sorts of amazing advances that started showing up and things 80 00:03:52,080 --> 00:03:53,800 like Google Photos and other products like that. 81 00:03:54,560 --> 00:03:57,000 But now all that's forgotten because it got completely sort 82 00:03:57,000 --> 00:04:00,320 of superseded or overshadowed by this more recent large language 83 00:04:00,320 --> 00:04:02,200 model revolution. But there have been multiple 84 00:04:02,200 --> 00:04:03,560 waves here. And that's been really, really 85 00:04:03,560 --> 00:04:06,600 interesting to see how. As processing power hasn't 86 00:04:06,640 --> 00:04:10,120 increased, there are these new forms of AI that come out and 87 00:04:10,120 --> 00:04:12,560 start really solving certain problems really, really well. 88 00:04:13,240 --> 00:04:16,360 Yeah, it's interesting that we know LLM is like LLM or 89 00:04:16,360 --> 00:04:19,440 generative AI, right, The bigger categories like the all the 90 00:04:19,440 --> 00:04:21,560 things that people talk about regarding AI. 91 00:04:21,560 --> 00:04:24,240 So in the past I remember people are talking about machine 92 00:04:24,240 --> 00:04:27,360 learning, recommendation system, image recognition, just like 93 00:04:27,360 --> 00:04:29,880 what you said. It seems like all of those kind 94 00:04:29,880 --> 00:04:32,480 of like toned down a little bit. Do you still see advancement in 95 00:04:32,480 --> 00:04:36,600 those area or is generative AI is like the go to AI strategy 96 00:04:36,600 --> 00:04:38,920 that we are going as a human species I guess. 97 00:04:39,640 --> 00:04:41,520 Yeah, I think it's pretty extraordinary. 98 00:04:41,520 --> 00:04:43,960 How? General purpose and powerful. 99 00:04:43,960 --> 00:04:47,080 These models have become and of course, you know, with a lot of 100 00:04:47,080 --> 00:04:49,240 the image generation, there are different methods involved 101 00:04:49,240 --> 00:04:50,960 there. Like it's not necessarily LLMS, 102 00:04:50,960 --> 00:04:54,200 it's more like diffusion models, but just these multi model large 103 00:04:54,200 --> 00:04:57,040 language models are pretty incredible, like going beyond 104 00:04:57,040 --> 00:04:59,480 human language into things like genetic coding. 105 00:04:59,840 --> 00:05:02,840 And of course, you know, computer coding, like actually 106 00:05:02,840 --> 00:05:04,760 programming, has become a huge. Use case for. 107 00:05:04,760 --> 00:05:08,040 These language models, and they originated out of the transform 108 00:05:08,040 --> 00:05:10,880 architecture that Google and others sort of introduced. 109 00:05:11,160 --> 00:05:14,880 A while ago, which originally was applied to the use case of 110 00:05:14,880 --> 00:05:17,080 translation from one human language to another. 111 00:05:17,320 --> 00:05:20,440 So from that initial use case, it's become much more general 112 00:05:20,440 --> 00:05:21,920 purpose. So it's pretty incredible. 113 00:05:22,120 --> 00:05:24,920 But I do think, I continue to think that there's always going 114 00:05:24,920 --> 00:05:28,160 to be many, many use cases out there for which other forms of 115 00:05:28,160 --> 00:05:31,320 machine learning that have been sort of largely maybe overlooked 116 00:05:31,320 --> 00:05:34,440 during this whole LLM revolution are going to continue to be 117 00:05:34,440 --> 00:05:37,480 really useful for solving certain other types of problems 118 00:05:37,480 --> 00:05:40,360 involving particularly involving quantitative information like 119 00:05:40,760 --> 00:05:42,280 predicting numbers and stuff like that. 120 00:05:43,440 --> 00:05:45,520 Right. So back then you were also in 121 00:05:45,520 --> 00:05:48,200 charge on building the Google Duplex, right. 122 00:05:48,200 --> 00:05:51,400 So when I remember back then when it was introduced in Google 123 00:05:51,400 --> 00:05:54,480 IO or something like that, that was really, you know, like 124 00:05:54,480 --> 00:05:57,680 incredible to me. Did you already use something 125 00:05:57,680 --> 00:06:00,960 like LLM or is it using a different technology altogether? 126 00:06:01,520 --> 00:06:05,080 Yeah, I mean, it's really interesting because it didn't, 127 00:06:05,200 --> 00:06:08,280 it didn't use LLMS at all. It predated that. 128 00:06:08,680 --> 00:06:11,840 And so in some ways, that was sort of like AI wouldn't say a 129 00:06:11,840 --> 00:06:13,520 prototype. It was almost like sort of a 130 00:06:13,840 --> 00:06:16,320 preview of what was to come because people started freaking 131 00:06:16,320 --> 00:06:18,400 out when we launched this product because it sounded so 132 00:06:18,400 --> 00:06:21,000 lifelike. And some people like the press 133 00:06:21,000 --> 00:06:22,960 had this huge media frenzy around this, right? 134 00:06:22,960 --> 00:06:26,120 Because Google introduces some kind of AI that sounds like a 135 00:06:26,120 --> 00:06:27,160 human being. And it's like. 136 00:06:27,320 --> 00:06:31,000 Oh, it's unethical or oh, the game over, you know, the bots 137 00:06:31,000 --> 00:06:32,200 are going to take over the world. 138 00:06:32,440 --> 00:06:35,040 And there were even some people, some, you know, prominent 139 00:06:35,120 --> 00:06:37,160 members of the press who thought that we faked the whole thing. 140 00:06:37,520 --> 00:06:40,200 So it was super interesting to be involved in sort of like a 141 00:06:40,200 --> 00:06:41,920 media frenzy situation like that. 142 00:06:42,040 --> 00:06:44,240 It was on the Good Morning Show, it was on, you know, Colbert. 143 00:06:44,240 --> 00:06:48,080 It was like really hit this very, very big nerve in popular 144 00:06:48,080 --> 00:06:49,560 culture. And yet it. 145 00:06:49,560 --> 00:06:53,360 Wasn't using any of this modern LLM architecture at all. 146 00:06:53,640 --> 00:06:55,560 It was using other methods that were much, much simpler. 147 00:06:55,960 --> 00:06:59,560 And so since then it's just gotten way crazier in terms of 148 00:06:59,560 --> 00:07:00,760 what these systems are capable of. 149 00:07:00,760 --> 00:07:03,880 But it was really a privilege and just so much fun to be 150 00:07:03,880 --> 00:07:06,440 involved in a system that was very much the forefront of the 151 00:07:06,440 --> 00:07:08,200 time in doing this kind of thing. 152 00:07:08,960 --> 00:07:12,560 Yeah, I think Google itself has to rebrand it, you know, those 153 00:07:12,560 --> 00:07:14,760 kind of things like becoming like Google Assistant now Google 154 00:07:14,760 --> 00:07:16,760 Assistants, becoming like Gemini, right. 155 00:07:17,000 --> 00:07:20,480 I think all those speakers that they introduced probably is like 156 00:07:20,480 --> 00:07:23,000 using some kind of things that you use in Google Duplex as well 157 00:07:23,000 --> 00:07:26,760 because the human tone are so natural when I chat with like 158 00:07:26,760 --> 00:07:29,680 Gemini using voice, right. So I think it's really cool 159 00:07:29,680 --> 00:07:32,760 technologies, you know, that comes out of those innovations. 160 00:07:33,080 --> 00:07:37,000 And today specifically we are going to talk how to apply AI in 161 00:07:37,000 --> 00:07:39,240 health or life science and clinical development. 162 00:07:39,560 --> 00:07:42,960 So your company, Vero Health, if you can share a little bit more, 163 00:07:43,160 --> 00:07:47,240 how does Vero Health use AI in terms of, you know, like solving 164 00:07:47,240 --> 00:07:50,560 your problem? First of all, just in brief, 165 00:07:50,760 --> 00:07:54,320 Firo has developed the SAS platform for designing clinical 166 00:07:54,320 --> 00:07:57,080 trials. So this is a really intensely 167 00:07:57,080 --> 00:08:01,080 complicated process that can take, you know, like a clinical 168 00:08:01,080 --> 00:08:03,120 trial can take. Upwards of 10 years. 169 00:08:03,520 --> 00:08:06,440 It costs on average 1 to $2 billion per drug to get to 170 00:08:06,440 --> 00:08:08,680 market. Like these are huge numbers and 171 00:08:08,680 --> 00:08:11,960 huge time frames. And so we developed a platform 172 00:08:11,960 --> 00:08:15,200 that allows essentially biotech companies and pharma companies 173 00:08:15,520 --> 00:08:17,040 to design better clinical trials. 174 00:08:17,640 --> 00:08:19,520 So we're applying. AI in a couple of different 175 00:08:19,520 --> 00:08:21,000 ways. First of all. 176 00:08:21,280 --> 00:08:24,280 One of the big time and cost thinks in the whole development 177 00:08:24,280 --> 00:08:29,000 process is clinical writing. So writing really complex 178 00:08:29,000 --> 00:08:31,880 clinical document like the clinical protocol that 179 00:08:32,039 --> 00:08:34,679 essentially defines the drug trial like so it lays out the 180 00:08:34,679 --> 00:08:37,080 whole design and the schedule and all the things that need to 181 00:08:37,080 --> 00:08:40,600 happen, all the details. So this is a really complex 182 00:08:40,600 --> 00:08:44,840 document that can be north of 100 and 5200 pages and typically 183 00:08:44,840 --> 00:08:46,720 takes an entire team of people to write. 184 00:08:46,720 --> 00:08:49,240 It takes months and. We. 185 00:08:49,720 --> 00:08:52,680 Starting last year, we started applying large language models 186 00:08:52,800 --> 00:08:56,840 to essentially automatically generate various sections in the 187 00:08:56,840 --> 00:09:00,480 clinical protocol document and. You might think, well, what's 188 00:09:00,480 --> 00:09:03,280 the big deal? You know, it's easy to ask GPT 189 00:09:03,280 --> 00:09:04,680 to go and generate a document, right? 190 00:09:04,680 --> 00:09:07,040 I would do a little time. It's a write better emails or 191 00:09:07,040 --> 00:09:09,400 they're, you know, do my homework or this particular. 192 00:09:09,400 --> 00:09:11,560 Use cases, right? But it turns out. 193 00:09:11,560 --> 00:09:15,760 That writing clinical grade documentation that passes master 194 00:09:15,760 --> 00:09:18,880 with regulatory groups like the FDA and where the stakes are 195 00:09:18,880 --> 00:09:21,400 really high, we're talking about human lives and human safety and 196 00:09:21,400 --> 00:09:24,120 things like that is incredibly challenging. 197 00:09:24,600 --> 00:09:27,040 And so we went through many iterations when we started 198 00:09:27,040 --> 00:09:29,800 designing this system to try to figure out the best way to 199 00:09:29,800 --> 00:09:33,200 actually produce really high quality clinical output that 200 00:09:33,360 --> 00:09:36,360 meets all the needs and requirements of a clinical 201 00:09:36,360 --> 00:09:39,120 protocol document. Yeah, this resulted in a lot of 202 00:09:39,200 --> 00:09:42,280 new architecture and new processes and all sorts of 203 00:09:42,280 --> 00:09:44,360 interesting things in order to make that happen. 204 00:09:45,120 --> 00:09:46,480 Right. It's very interesting that you 205 00:09:46,480 --> 00:09:49,960 mentioned applying like generative AI or LLM to solve 206 00:09:49,960 --> 00:09:52,440 something that is, you know, highly regulated. 207 00:09:52,720 --> 00:09:55,280 And it is also kind of like mission critical, right? 208 00:09:55,280 --> 00:09:57,680 So to speak, Right, Because it involves people's lives. 209 00:09:58,080 --> 00:10:01,600 So people, you know, these days know about the danger of 210 00:10:01,600 --> 00:10:04,840 hallucination and all that. How do you actually tackle these 211 00:10:04,840 --> 00:10:07,920 kind of things, applying this to kind of like this highly 212 00:10:07,920 --> 00:10:10,440 regulated and mission critical kind of a problem? 213 00:10:11,200 --> 00:10:12,800 Yeah. I mean, at first we sort of 214 00:10:13,080 --> 00:10:16,160 naively assumed that we could just adjust our prompts. 215 00:10:16,440 --> 00:10:18,280 You know, every time we saw a hallucination, we're like, oh, 216 00:10:18,280 --> 00:10:19,560 we need to go back and change the prompt. 217 00:10:19,960 --> 00:10:23,600 But it turned out that what we really needed was to evaluate 218 00:10:23,600 --> 00:10:25,960 comprehensively the output systematically, right. 219 00:10:25,960 --> 00:10:31,600 And so when you generate. Say a PK sampling section or ECG 220 00:10:31,600 --> 00:10:33,480 section. Like some highly typical part of 221 00:10:33,480 --> 00:10:37,520 the protocol dock, you need to actually evaluate the output 222 00:10:37,520 --> 00:10:39,280 according to a whole bunch of criteria. 223 00:10:39,800 --> 00:10:42,320 And do this. Per subsection, because 224 00:10:42,320 --> 00:10:46,240 essentially each subsection is so different and has different 225 00:10:46,480 --> 00:10:49,200 sort of technical details that are fed in through our fire 226 00:10:49,200 --> 00:10:51,600 study designer tool. So we have this whole staff 227 00:10:51,600 --> 00:10:54,520 platform that's modelling the trial in great detail and we 228 00:10:54,520 --> 00:10:57,560 pass that data into the LLM. To generate the protocol 229 00:10:57,560 --> 00:11:00,200 document sections. And so we found that we had to 230 00:11:00,200 --> 00:11:04,680 really rigorously sort of query the results and say, well, did 231 00:11:04,680 --> 00:11:06,080 you include this? Did you have this? 232 00:11:06,080 --> 00:11:09,040 Tone, did you sort of like phrase this in this way? 233 00:11:09,200 --> 00:11:11,400 And we have this very specific checklist for every single 234 00:11:11,400 --> 00:11:14,440 subsection that we run to make sure that the quality is high. 235 00:11:15,040 --> 00:11:17,800 And we've also designed this in such a way that customers can 236 00:11:17,800 --> 00:11:20,600 actually make their own checks. So if they have additional 237 00:11:20,760 --> 00:11:24,320 requirements that are particular to them, they can augment or 238 00:11:24,320 --> 00:11:26,720 change the ones that we ship our product with. 239 00:11:26,720 --> 00:11:29,560 So we've made it very highly configurable to match the 240 00:11:29,560 --> 00:11:32,760 customer's needs. So this is what we found was 241 00:11:32,760 --> 00:11:34,960 needed. And over time, that initial 242 00:11:35,360 --> 00:11:37,960 evaluation architecture has evolved into more of an agentic 243 00:11:37,960 --> 00:11:40,560 architecture where there's all sorts of different agents that 244 00:11:40,560 --> 00:11:44,680 are examining the content of the generated clinical protocol 245 00:11:44,680 --> 00:11:48,120 section and looking at it for stylistic consistency with the 246 00:11:48,120 --> 00:11:51,200 rest of the document and tone and deep level of detail and 247 00:11:51,200 --> 00:11:53,840 formatting and stuff like that. In addition to the factual 248 00:11:53,840 --> 00:11:56,160 content to prevent hallucinations and also. 249 00:11:56,960 --> 00:11:58,280 In addition to. Hallucinations. 250 00:11:58,280 --> 00:12:01,800 We also found that sometimes the LLM just misses things, so it 251 00:12:01,800 --> 00:12:03,480 doesn't make things up, it just misses things. 252 00:12:03,800 --> 00:12:05,680 And so we have to sort of check for that as well. 253 00:12:06,160 --> 00:12:09,360 So there's a whole bunch of examination that occurs once you 254 00:12:09,360 --> 00:12:13,000 generate the output, and then you feed that feedback from the 255 00:12:13,160 --> 00:12:15,600 evaluation models back into another iteration of the 256 00:12:15,600 --> 00:12:17,920 generation cycle. O we developed this whole 257 00:12:17,920 --> 00:12:20,880 architecture that's way more complex than just sort of 258 00:12:20,880 --> 00:12:24,480 sending queries to chat GT in order to get the system actually 259 00:12:24,480 --> 00:12:26,480 working. Yeah, you mentioned about 260 00:12:26,480 --> 00:12:29,800 missing, you know, information and also again, like the danger 261 00:12:29,800 --> 00:12:32,560 of hallucinating, right. So how do you actually do these 262 00:12:32,560 --> 00:12:35,080 kind of checks? Is it like still human involved, 263 00:12:35,080 --> 00:12:39,240 like at the end someone who is, you know, maybe capable enough 264 00:12:39,240 --> 00:12:41,360 or expert enough, like reviewing the output? 265 00:12:41,360 --> 00:12:44,320 Because I would imagine, right, these kind of documents is 266 00:12:44,320 --> 00:12:46,960 highly technical, right? Sometimes you probably need a 267 00:12:46,960 --> 00:12:50,120 little bit more time to research whether it's factual or accurate 268 00:12:50,120 --> 00:12:52,600 or not. Or do you completely rely on AI 269 00:12:52,600 --> 00:12:55,840 agents to do that for you? And if so, like how do you to 270 00:12:55,840 --> 00:12:58,680 build this kind of like evaluation using AI? 271 00:12:58,800 --> 00:13:01,200 Is it like something like eval model that people are using to 272 00:13:01,200 --> 00:13:03,600 solve, you know, certain kind of a programming challenge and 273 00:13:03,600 --> 00:13:05,800 things like that? Yeah, I mean in this. 274 00:13:05,800 --> 00:13:08,120 Respect. What we've built with clinical 275 00:13:08,200 --> 00:13:11,040 protocol writing is kind of similar to other AI systems I've 276 00:13:11,040 --> 00:13:13,840 worked on in the past, where in the early stages when you're 277 00:13:13,840 --> 00:13:17,200 first getting the AI model to kind of start producing 278 00:13:17,200 --> 00:13:20,240 interesting content, you need to be very, very hands on and have 279 00:13:20,240 --> 00:13:23,080 a human examine everything. With a fine tooth comb. 280 00:13:23,080 --> 00:13:27,080 And then after a while you can codify that feedback into these 281 00:13:27,080 --> 00:13:30,280 checks and into queries that you can then use to sort of automate 282 00:13:30,280 --> 00:13:32,480 the process. And after a while, you build 283 00:13:32,480 --> 00:13:35,480 enough confidence against a range of different inputs, you 284 00:13:35,480 --> 00:13:38,840 know, looking at different types of therapeutic areas to make 285 00:13:38,840 --> 00:13:41,960 sure that the content is right. Because looking at a cancer 286 00:13:42,240 --> 00:13:45,560 trial like an oncology trial versus say an immunology trial 287 00:13:45,560 --> 00:13:47,560 or infectious. Disease trial, it's just kind of 288 00:13:47,560 --> 00:13:50,040 really different. And so we want to make sure that 289 00:13:50,200 --> 00:13:52,240 the models we build are applicable across a range of 290 00:13:52,240 --> 00:13:54,000 different types of clinical trials. 291 00:13:54,400 --> 00:13:57,040 And so once we reach a level of confidence that, oh, it looks 292 00:13:57,040 --> 00:14:00,240 like, you know, in the 80%. Case the content is pretty good 293 00:14:00,560 --> 00:14:03,120 then we're confident enough to share this with our customers, 294 00:14:03,120 --> 00:14:05,440 our early adopted customers and sort of get them to evaluate 295 00:14:05,440 --> 00:14:07,840 themselves and they probably come up with some issues as 296 00:14:07,840 --> 00:14:09,800 well. And then we sort of gradually 297 00:14:09,800 --> 00:14:12,640 improve the quality of the model using that process. 298 00:14:13,000 --> 00:14:16,320 But we have a quite an in depth process involving engineers, you 299 00:14:16,320 --> 00:14:19,920 know, data scientists, clinical writers, QA, all sorts of 300 00:14:19,920 --> 00:14:22,920 different people involved in this to sort of bootstrap the 301 00:14:22,920 --> 00:14:24,360 process and get everything running. 302 00:14:24,560 --> 00:14:26,920 And then after a while it becomes more automated as we 303 00:14:27,160 --> 00:14:29,680 refine the model. Thanks for sharing that. 304 00:14:29,960 --> 00:14:32,400 So apart from like, content writing, right, is there 305 00:14:32,400 --> 00:14:35,400 anything else where you apply AI as part of the work? 306 00:14:36,200 --> 00:14:38,280 Oh, yeah. So we're also looking at 307 00:14:38,280 --> 00:14:41,240 actually developing what we're calling an AI research 308 00:14:41,240 --> 00:14:43,880 assistant. And so the idea is that, you 309 00:14:43,880 --> 00:14:46,480 know, right now actually designing the clinical trial is 310 00:14:46,480 --> 00:14:50,120 quite laborious, like choosing what kind of activities you want 311 00:14:50,120 --> 00:14:53,400 to include in the schedule, designing the population schema, 312 00:14:53,400 --> 00:14:55,440 all these kinds of things, objectives and endpoints. 313 00:14:55,720 --> 00:14:57,800 These will require a lot of care and in many cases a lot of 314 00:14:57,800 --> 00:15:00,840 research like going out there on the Internet, you know, checking 315 00:15:00,840 --> 00:15:04,120 out the latest clinical research that being published, taking a 316 00:15:04,120 --> 00:15:07,160 look at identifying and taking a look at comparable, the trials 317 00:15:07,160 --> 00:15:09,520 that are out there, learning from them, reading through 318 00:15:09,520 --> 00:15:12,800 written hundreds of pages. This is super laborious and we 319 00:15:12,800 --> 00:15:13,960 can actually automate a lot of that. 320 00:15:14,440 --> 00:15:17,480 So we are applying AI to essentially help to 321 00:15:18,200 --> 00:15:20,640 automatically identify trials that are similar to the one that 322 00:15:20,640 --> 00:15:23,640 you are working on. And then you know, aggregate 323 00:15:23,720 --> 00:15:26,200 metrics and stats do comparisons. 324 00:15:26,360 --> 00:15:27,520 Look at what happened with these trials. 325 00:15:27,520 --> 00:15:30,000 Did they require amendments? Like why were they able to 326 00:15:30,000 --> 00:15:32,040 enroll patients? Were they able to, you know, 327 00:15:32,080 --> 00:15:34,640 stay on track? Did they result in a successful 328 00:15:35,000 --> 00:15:38,320 outcome or not? And so clinical developers who 329 00:15:38,320 --> 00:15:41,040 are or clinical scientists who are working on these trials can 330 00:15:41,040 --> 00:15:44,200 essentially really accelerate the process by which they come 331 00:15:44,200 --> 00:15:47,080 up with an optimal model. And of course, you know, we can 332 00:15:47,080 --> 00:15:50,000 provide insights, we can say, oh, here's your patient burden 333 00:15:50,000 --> 00:15:53,120 of this proposed trial. Here is your, you know, the site 334 00:15:53,120 --> 00:15:56,240 complexity and cost lots of different metrics that help 335 00:15:56,240 --> 00:16:00,360 guide what you're doing and help you come to the right balance of 336 00:16:00,560 --> 00:16:04,040 patient experience, site feasibility, and overall cost 337 00:16:04,040 --> 00:16:06,840 and time to market. So this is a really sort of 338 00:16:06,920 --> 00:16:09,560 multivariate kind of thing that people are trying to optimize 339 00:16:09,560 --> 00:16:12,800 when they put together these trials and it just staggers me 340 00:16:12,800 --> 00:16:15,120 that how many people, many groups still do this in 341 00:16:15,120 --> 00:16:17,520 Microsoft Word. It's like sort of my sort of 342 00:16:17,520 --> 00:16:18,880 catch cry is like not even Excel. 343 00:16:19,440 --> 00:16:22,600 You've used Microsoft Word to essentially model much author of 344 00:16:22,600 --> 00:16:25,960 model these trials and I think that you know, people deserve a 345 00:16:25,960 --> 00:16:27,840 better tool than that. So that's what we're developing 346 00:16:27,840 --> 00:16:31,880 with the help of AI is like really modernizing and shifting 347 00:16:31,880 --> 00:16:34,800 the way that people develop these and design these trials in 348 00:16:34,800 --> 00:16:39,040 such a way that it makes it much, much quicker and with much 349 00:16:39,040 --> 00:16:40,800 better outcomes. Yeah. 350 00:16:40,800 --> 00:16:44,720 So since this is like a very highly niche, you know, specific 351 00:16:44,720 --> 00:16:47,880 kind of a problem domain, right. So I'm just assuming like people 352 00:16:48,240 --> 00:16:50,840 know things like Chachi, PT and Gemini and all that, right. 353 00:16:50,840 --> 00:16:52,720 It's more like a general purpose AI models. 354 00:16:52,960 --> 00:16:55,680 Do you actually have to fine tune or even build your own 355 00:16:55,680 --> 00:16:57,800 model to work on this specific niche? 356 00:16:57,800 --> 00:17:00,680 Because like one of the most important things when using 357 00:17:00,680 --> 00:17:03,400 general purpose AI, it's like, what kind of data are they 358 00:17:03,400 --> 00:17:06,000 trained on, right? For example, if you've never 359 00:17:06,079 --> 00:17:08,280 seen this kind of problem before, I'm sure AI will 360 00:17:08,280 --> 00:17:11,920 hallucinate even much worse than, you know, if you train it 361 00:17:11,920 --> 00:17:14,920 with specific model and data. So do you actually have to tweak 362 00:17:14,920 --> 00:17:17,119 or fun tune the model or even come up with your own model? 363 00:17:17,760 --> 00:17:19,839 Yeah, we haven't found the need to do that yet. 364 00:17:19,920 --> 00:17:23,560 So the good news is that a lot of the major LLM AI vendors like 365 00:17:23,560 --> 00:17:26,440 Google and Anthropic are actually focusing quite a bit on 366 00:17:26,440 --> 00:17:28,800 the life sciences space and producing models that are kind 367 00:17:28,800 --> 00:17:32,200 of specific to that domain. And so we've designed our system 368 00:17:32,200 --> 00:17:35,160 in such a way that it's kind of agnostic to the model, like 369 00:17:35,160 --> 00:17:38,800 we're not fixed on GBT or Gemini or Anthropic or anything like 370 00:17:38,800 --> 00:17:40,760 that. And so because we have this 371 00:17:40,760 --> 00:17:45,440 evaluation system, we can pretty easily evaluate new LLMS and 372 00:17:45,440 --> 00:17:48,160 figure out which ones are going to produce the best clinical 373 00:17:48,160 --> 00:17:50,680 output. I would say it's quite possible 374 00:17:50,680 --> 00:17:54,240 that we might in the future want to potentially have our own 375 00:17:54,240 --> 00:17:57,680 model with respect to maybe vectorization like which is a 376 00:17:57,680 --> 00:18:01,280 process that all the text kind of goes through before it gets 377 00:18:01,280 --> 00:18:02,840 indexed. So we built this kind of 378 00:18:03,160 --> 00:18:06,640 retrieval augmented generation RAG system that allows customers 379 00:18:06,640 --> 00:18:10,480 to sort of upload documents that we can then parse and use to 380 00:18:10,600 --> 00:18:14,200 generate certain pieces of the protocol document that require 381 00:18:14,320 --> 00:18:17,360 really technical information. And so that's an area where we 382 00:18:17,360 --> 00:18:20,560 might want to develop our own vectorization scheme that's 383 00:18:20,560 --> 00:18:23,360 really oriented around clinical terminology. 384 00:18:23,760 --> 00:18:26,400 But so far we haven't. Seen the need to even fine tune 385 00:18:26,400 --> 00:18:28,760 an existing model. Even fine tuning these days can 386 00:18:28,760 --> 00:18:30,840 be quite an expensive and laborious process. 387 00:18:31,240 --> 00:18:33,120 And so we haven't really seen the need to do that yet. 388 00:18:33,440 --> 00:18:36,040 What we've really seen is that this evaluation model can 389 00:18:36,200 --> 00:18:39,080 combined with the ability to actually sort of fine tune 390 00:18:39,080 --> 00:18:42,320 prompts is enough to be able to produce clinical grade 391 00:18:42,320 --> 00:18:44,760 documentation. It might change in the future as 392 00:18:44,760 --> 00:18:47,440 we get into different documents and and domains and so on, but 393 00:18:47,600 --> 00:18:49,680 so far that's been really good for us. 394 00:18:50,560 --> 00:18:51,960 Yeah. So what I know when you do 395 00:18:51,960 --> 00:18:54,920 clinical trials, right, like you mentioned, it could be, yes, 396 00:18:54,920 --> 00:18:56,360 right. It could be a decade or even 397 00:18:56,360 --> 00:18:58,160 more. Not to mention, you know, like 398 00:18:58,160 --> 00:19:00,880 having to do the testing and wait, you know, for the effect 399 00:19:01,240 --> 00:19:03,160 in multi years. So using this kind of 400 00:19:03,160 --> 00:19:06,960 technologies, what do you see potentials, time savings or 401 00:19:06,960 --> 00:19:10,440 maybe even effort, right. You mentioned like $2 billion 402 00:19:10,440 --> 00:19:12,320 just to finish the whole clinical trial. 403 00:19:12,320 --> 00:19:14,960 So maybe what are some of the things that you think can be 404 00:19:15,200 --> 00:19:18,440 reduced or optimized by having this kind of AI tools? 405 00:19:19,440 --> 00:19:21,840 Yeah, I mean, the interesting thing is that when you're 406 00:19:21,840 --> 00:19:24,960 talking about a project that takes, you know, upwards of 10 407 00:19:24,960 --> 00:19:28,680 years in many cases or more to execute, it means that small 408 00:19:28,680 --> 00:19:32,280 changes in the design upfront can actually result in huge cost 409 00:19:32,280 --> 00:19:34,520 savings. So we actually published a paper 410 00:19:34,520 --> 00:19:36,600 with Merck. Maybe we can include a link to 411 00:19:36,600 --> 00:19:37,160 it. It's on my. 412 00:19:37,160 --> 00:19:39,800 It's somewhere on my LinkedIn profile that. 413 00:19:40,000 --> 00:19:42,920 Sort of analyzed like what is the benefit of using the fire 414 00:19:42,920 --> 00:19:45,400 study designer to optimize clinical trial design? 415 00:19:45,840 --> 00:19:48,600 And we found that the savings over the course of the lifetime 416 00:19:49,000 --> 00:19:51,320 of the trial can be north of $100 million. 417 00:19:51,800 --> 00:19:55,640 So obviously this was a really big Merck is a huge, you know, 418 00:19:55,640 --> 00:19:58,560 top ten, top 20 pharma company and so on. 419 00:19:58,560 --> 00:20:01,880 And so that they do these really significant trials, but it just 420 00:20:01,880 --> 00:20:03,840 gives you an idea of what's possible, right? 421 00:20:03,840 --> 00:20:06,960 And this was even before some of these AI features that we have 422 00:20:06,960 --> 00:20:09,120 coming up like clinical writing and so on. 423 00:20:09,120 --> 00:20:11,800 So the true savings actually could be even more than that 424 00:20:11,960 --> 00:20:15,440 when you factor in the time savings as well of accelerating 425 00:20:15,440 --> 00:20:18,080 the clinical writing process and making the output better. 426 00:20:18,640 --> 00:20:23,000 Because getting to market faster is hugely, hugely valuable to a 427 00:20:23,000 --> 00:20:25,760 pharma company, right? Like every day is like hundreds 428 00:20:25,760 --> 00:20:27,840 and hundreds of thousands of dollars of additional revenue 429 00:20:27,840 --> 00:20:30,600 that you can realize sooner when you divorce the drug faster. 430 00:20:31,040 --> 00:20:33,200 So it's not only cost savings, it's also getting to market 431 00:20:33,200 --> 00:20:36,600 faster with the new treatment. So we're really excited about 432 00:20:36,600 --> 00:20:40,000 that because there's also the potential to really over time, 433 00:20:40,080 --> 00:20:42,840 as this technology becomes widespread to really sort of 434 00:20:42,840 --> 00:20:46,840 like for a smaller players to come in here and really try 435 00:20:46,840 --> 00:20:49,920 things out. And it's going to kind of enable 436 00:20:50,440 --> 00:20:53,840 a wider range of different treatments and trials to become 437 00:20:53,840 --> 00:20:56,520 more feasible. And we're really excited by that 438 00:20:56,520 --> 00:20:59,120 because there's so much innovation going on in the very 439 00:20:59,120 --> 00:21:02,400 early stage drug discovery realm, but it gets sort of 440 00:21:02,560 --> 00:21:05,160 blocked or delayed when it comes to actually going through 441 00:21:05,160 --> 00:21:07,360 trials. And so we want to really widen 442 00:21:07,360 --> 00:21:10,000 the pipeline there. And so hopefully there'll be a 443 00:21:10,000 --> 00:21:13,000 lot more treatments for more conditions and lives saved and 444 00:21:13,000 --> 00:21:14,640 suffering alleviated all the good things. 445 00:21:15,440 --> 00:21:17,320 Yeah. And hopefully reduce the cost of 446 00:21:17,320 --> 00:21:20,720 the kind of the medicine that is produced out of the clinical 447 00:21:20,720 --> 00:21:23,480 trials as well, so that people can get much more options in 448 00:21:23,480 --> 00:21:26,600 terms of availabilities and also like the cost is not less 449 00:21:26,600 --> 00:21:29,120 expensive. So you mentioned about 450 00:21:29,120 --> 00:21:31,840 regulations. So I think in some parts of the 451 00:21:31,840 --> 00:21:34,800 world, they are still kind of like cautious about applying AI 452 00:21:34,800 --> 00:21:37,440 technologies again, especially if it affects a lot of 453 00:21:37,440 --> 00:21:40,040 population. So what do you see some of the 454 00:21:40,040 --> 00:21:43,320 stance from regulations or do you even have to convince them 455 00:21:43,320 --> 00:21:47,000 hard to actually ensure that the process is robust enough? 456 00:21:47,920 --> 00:21:51,520 I think that by their very nature, regulatory bodies are 457 00:21:51,560 --> 00:21:55,640 very, very oriented around evaluating the quality of what's 458 00:21:55,640 --> 00:21:58,880 presented to them. And so at least in my mind, like 459 00:21:59,360 --> 00:22:02,680 if the quality of the output is good, at least as good as human 460 00:22:02,680 --> 00:22:05,120 generated protocols and we strongly believe that it will 461 00:22:05,120 --> 00:22:08,320 be, then the regulatory bodies will be OK. 462 00:22:08,840 --> 00:22:11,880 I think there are various risks involved in using AI. 463 00:22:12,120 --> 00:22:13,680 There's many, many different types of risks. 464 00:22:14,000 --> 00:22:16,360 And so we can mitigate a lot of them. 465 00:22:16,360 --> 00:22:19,640 Like for instance we are running our AI models in a sort of a 466 00:22:19,640 --> 00:22:21,680 private. Environment where data is not 467 00:22:21,680 --> 00:22:23,680 shared with any other groups and so on. 468 00:22:23,680 --> 00:22:25,760 So we don't have any of those risks of leakage and people 469 00:22:25,760 --> 00:22:28,720 being able to kind of discover what other companies are doing 470 00:22:28,720 --> 00:22:30,600 or submitting. So there's that risk that we can 471 00:22:30,600 --> 00:22:32,680 mitigate. I think there's a bunch of 472 00:22:32,680 --> 00:22:35,840 others involving potentially bad actors coming in and trying to 473 00:22:36,120 --> 00:22:39,240 sort of discover things they shouldn't be discovering by 474 00:22:39,240 --> 00:22:40,880 manipulating queries and stuff like that. 475 00:22:41,040 --> 00:22:43,200 That's mitigated by the fact this is an enterprise tool. 476 00:22:43,200 --> 00:22:45,840 It's not a consumer tool. And so only people who have 477 00:22:46,160 --> 00:22:49,160 signed up as Faro, you know, customers will be able to use 478 00:22:49,160 --> 00:22:51,280 the. System so we just have to go 479 00:22:51,320 --> 00:22:54,040 case by case and look at all the different risks involved in AI 480 00:22:54,440 --> 00:22:58,480 and mitigate them and we have AAI governance policy that we've 481 00:22:58,480 --> 00:23:00,720 been working really hard on to sort of like really. 482 00:23:01,080 --> 00:23:03,680 Alleviate our own customers concerns because however 483 00:23:03,680 --> 00:23:06,360 concerned the FDA is, you know, our customers are even more 484 00:23:06,360 --> 00:23:08,640 concerned about all these things because it's their crown jewels, 485 00:23:08,640 --> 00:23:11,240 it's their intellectual property that they're, you know 486 00:23:11,240 --> 00:23:14,880 essentially developing using our platform and so we take this 487 00:23:14,880 --> 00:23:17,280 very seriously and. We want to actually at some 488 00:23:17,280 --> 00:23:20,640 point probably publish our AI governance policy so that other 489 00:23:20,640 --> 00:23:23,600 people can learn from it and we can help raise the bar across 490 00:23:23,600 --> 00:23:26,120 the whole board in our industry because we just think this is 491 00:23:26,120 --> 00:23:27,680 really important to. Everybody to get right. 492 00:23:28,520 --> 00:23:31,760 Thanks for sharing that. So any kind of a cool success 493 00:23:31,760 --> 00:23:35,120 stories that you have done so far, you know, with Farrow or is 494 00:23:35,120 --> 00:23:38,160 there anything that is already applied in the, you know, real 495 00:23:38,160 --> 00:23:39,880 world or in the life science industry? 496 00:23:40,800 --> 00:23:43,480 Yeah, I mean, I think that again, the MOC study is sort of 497 00:23:43,800 --> 00:23:47,600 what we have as far as publicly available case studies of just 498 00:23:47,600 --> 00:23:50,080 how impactful this technology can be. 499 00:23:50,520 --> 00:23:53,760 I would say just anecdotally, like we have customers using 500 00:23:53,760 --> 00:23:56,600 this document generation system and they're super impressive how 501 00:23:56,600 --> 00:24:00,040 quickly it's able to come up with documentation that normally 502 00:24:00,040 --> 00:24:03,320 takes potentially months and it can be generated in a matter of, 503 00:24:03,560 --> 00:24:07,000 in some cases minutes. That's a startling extraordinary 504 00:24:07,000 --> 00:24:09,160 kind of multiplier in terms of productivity. 505 00:24:09,760 --> 00:24:11,040 We're just really excited about that. 506 00:24:11,040 --> 00:24:13,080 You know, like, I think that in some ways the industry is 507 00:24:13,080 --> 00:24:18,240 waiting for a really widespread and impactful application of AI 508 00:24:18,240 --> 00:24:20,680 because often times when you get past the surface level of like, 509 00:24:20,680 --> 00:24:23,880 oh, I play with GBT and, you know, it's really cool and 510 00:24:23,880 --> 00:24:25,720 interesting. Like once you really start to 511 00:24:25,720 --> 00:24:28,720 get serious about applying AI, then that can rapidly turn into 512 00:24:28,720 --> 00:24:31,840 disappointment or sort of like, oh, you know, it doesn't go deep 513 00:24:31,840 --> 00:24:33,560 enough. And there's all these stories of 514 00:24:33,560 --> 00:24:36,400 people who start building with AI, but then they kind of get 515 00:24:36,400 --> 00:24:39,080 disillusioned because it turns out to be way harder than they 516 00:24:39,080 --> 00:24:41,360 think. And so that's where I guess, you 517 00:24:41,360 --> 00:24:44,200 know, companies like Faro come in, is that we'll take it to the 518 00:24:44,200 --> 00:24:46,560 next level. You know, we will figure out all 519 00:24:46,560 --> 00:24:50,040 the hurdle and the challenges and the risks involved in 520 00:24:50,040 --> 00:24:53,400 actually applying AI to do something really impactful, like 521 00:24:53,560 --> 00:24:57,000 want to make clinical, you know, protocol documentation or like 522 00:24:57,040 --> 00:25:00,040 optimize a trial design. And so that our customers don't 523 00:25:00,040 --> 00:25:03,040 have to do that themselves because that requires a lot of 524 00:25:03,520 --> 00:25:06,920 data science expertise, software engineering expertise, clinical 525 00:25:07,520 --> 00:25:10,560 design expertise, and so on that they may not want to sort of 526 00:25:10,560 --> 00:25:12,880 invest in becoming a tech company themselves. 527 00:25:13,400 --> 00:25:15,640 And so we're really excited about that. 528 00:25:16,720 --> 00:25:18,640 Yeah. So I think these days some 529 00:25:18,640 --> 00:25:21,440 people also kind of like envision the AGI thing, right? 530 00:25:21,440 --> 00:25:24,240 That could happen in some years. I don't know whether it's short 531 00:25:24,240 --> 00:25:27,760 term, medium term, long term. So for someone who has worked on 532 00:25:27,760 --> 00:25:31,600 this field, AI fields for quite some time now, how do you feel 533 00:25:31,600 --> 00:25:33,880 about this AGI? Is it going to happen? 534 00:25:33,880 --> 00:25:36,640 Are you bullish about it or do you have some kind of 535 00:25:36,640 --> 00:25:38,920 scepticism? So maybe share a little bit of 536 00:25:38,920 --> 00:25:41,880 your thoughts here. Yeah, it seems like, you know, 537 00:25:41,880 --> 00:25:45,040 in the past, everybody thought AGI was really, really far away. 538 00:25:45,040 --> 00:25:48,120 Like the joke is that for the last 50 years of, you know, AGI 539 00:25:48,120 --> 00:25:51,240 has been 20 years away. So so, you know, sort of like 540 00:25:51,240 --> 00:25:54,320 always in the future. But now there's a lot of very 541 00:25:54,320 --> 00:25:57,160 prominent people in this field who believe that it's actually 542 00:25:57,160 --> 00:25:59,840 only a matter of like three to five years away, maybe even 543 00:25:59,840 --> 00:26:00,640 shorter. Who knows? 544 00:26:01,120 --> 00:26:05,720 And I am not in that camp. Like, I just think that there's 545 00:26:05,720 --> 00:26:10,200 a test out there called Arc AGIARCAGI and it's sort of 546 00:26:10,840 --> 00:26:13,880 presents itself as being like, oh, this is a test, you know, 547 00:26:13,880 --> 00:26:15,760 for AGI. And you look at what this test 548 00:26:15,760 --> 00:26:18,320 actually involved. And it's the kind of thing that 549 00:26:18,520 --> 00:26:21,200 many of the questions are sort of like a bright kid could 550 00:26:21,200 --> 00:26:23,960 answer these questions, you know, like moving blocks around 551 00:26:23,960 --> 00:26:26,400 on a grid, you know, in such a way that the block sort of fills 552 00:26:26,400 --> 00:26:29,160 in a space, stuff that smart kids should be able to do. 553 00:26:29,600 --> 00:26:33,280 And I sort of to me, and you know, there's systems that is, 554 00:26:33,280 --> 00:26:35,680 that is gradually getting better at this, but it's not yet human 555 00:26:35,680 --> 00:26:37,280 level. And I just to me, that just 556 00:26:37,280 --> 00:26:39,920 means that we're so far away from having something that's 557 00:26:39,920 --> 00:26:41,400 genuinely creative and intelligent. 558 00:26:41,680 --> 00:26:44,640 Now, I will say that there have been really brilliant moments in 559 00:26:44,640 --> 00:26:47,440 the history of AI, like if you remember, you know, Alphago, 560 00:26:47,680 --> 00:26:51,200 which was this non LLM based system was using reinforcement 561 00:26:51,200 --> 00:26:53,680 learning, was using other methods of AI that are now 562 00:26:53,680 --> 00:26:56,680 completely sort of forgotten in the wake of the LLM revolution. 563 00:26:57,320 --> 00:26:59,680 But Alpha Go there was at the moment when it was playing 564 00:26:59,800 --> 00:27:02,680 against the world human champion and it made this move that 565 00:27:03,000 --> 00:27:06,560 nobody in the 3000 year history of this game had ever seen 566 00:27:06,560 --> 00:27:08,120 before. And people were just all the 567 00:27:08,120 --> 00:27:10,920 experts were like gasping like, Oh my God, this is incredible. 568 00:27:11,040 --> 00:27:12,640 What a brilliant move. Like how did I come up with 569 00:27:12,640 --> 00:27:14,920 that? And so, you know, AI can, under 570 00:27:14,920 --> 00:27:19,400 the right conditions, be capable of behaving like a genius. 571 00:27:19,640 --> 00:27:22,280 But it's such a special that was such a specialized example. 572 00:27:22,800 --> 00:27:25,760 And so I'm sure as time goes on, there'll be more examples of 573 00:27:25,760 --> 00:27:28,120 that where, oh, there's a flash of brilliance and AI came up 574 00:27:28,120 --> 00:27:31,760 with some mathematical proof that was beyond the, you know, 575 00:27:31,800 --> 00:27:33,960 like what any scientist has come up with. 576 00:27:34,280 --> 00:27:36,680 But I think in terms of just having an intelligent 577 00:27:36,680 --> 00:27:40,440 conversation and really being inspired and looking at a piece 578 00:27:40,440 --> 00:27:44,120 of art and being truly like this belongs amongst the classics, I 579 00:27:44,120 --> 00:27:45,280 think we're a ways away from that. 580 00:27:45,280 --> 00:27:48,560 But I could be wrong. I mean, this definitely the 581 00:27:48,560 --> 00:27:50,960 progress even since these live language models are first 582 00:27:50,960 --> 00:27:54,480 introduced has been really nothing short of amazing in 583 00:27:54,480 --> 00:27:56,600 terms of the increasing capabilities and so on. 584 00:27:57,040 --> 00:28:01,040 But I think AGI in terms of like this is a peer to humanity. 585 00:28:01,040 --> 00:28:03,840 This is a, you know, like a system that we can have a really 586 00:28:03,840 --> 00:28:07,520 truly intelligent conversation about life and philosophy and 587 00:28:07,520 --> 00:28:09,440 things like that. I'm not so sure. 588 00:28:09,800 --> 00:28:12,320 And I sort of, I guess I come down on the young Likun side of 589 00:28:12,320 --> 00:28:14,560 things where just knowing a little bit about how these 590 00:28:14,560 --> 00:28:18,600 systems work, it is sort of like a super glorified, super, super 591 00:28:18,600 --> 00:28:21,320 well read autocomplete. And I know that there are 592 00:28:21,760 --> 00:28:25,320 efforts out there to introduce more real world knowledge and 593 00:28:25,320 --> 00:28:28,080 factual information into these LLMS to give them some grounding 594 00:28:28,080 --> 00:28:30,480 in reality. But I think that's going to be a 595 00:28:30,480 --> 00:28:33,600 long path to make sure that those that grounding is actually 596 00:28:33,600 --> 00:28:35,160 sort of super accurate and meaningful. 597 00:28:35,480 --> 00:28:38,680 But we'll see, Yeah. And I mean, those kind of the 598 00:28:38,680 --> 00:28:41,360 LLM Gen. AI leaders, thought leaders, 599 00:28:41,360 --> 00:28:44,280 right, think that maybe the path to AGI is using LLM. 600 00:28:44,360 --> 00:28:47,640 I'm personally also not very convinced since simply because 601 00:28:47,920 --> 00:28:50,800 of the way it works right at so many probabilistic thing that 602 00:28:50,800 --> 00:28:54,040 could go wrong as well. So do you think LLM or 603 00:28:54,200 --> 00:28:57,920 generative AI kind of things is going to be the path to AGI? 604 00:28:57,920 --> 00:29:01,400 Or is it more like combining all different, you know, AI, machine 605 00:29:01,400 --> 00:29:03,960 learning, kind of like capabilities that we have? 606 00:29:04,280 --> 00:29:06,960 And I haven't really heard about people, you know, trying to 607 00:29:06,960 --> 00:29:09,680 combine, combine different models from, you know, LLM or 608 00:29:09,680 --> 00:29:12,520 maybe other things, neural networks and all that being 609 00:29:12,520 --> 00:29:14,840 combined to kind of like solve a particular problem. 610 00:29:14,840 --> 00:29:18,920 So do you still see some rooms where we can apply, you know, 611 00:29:18,920 --> 00:29:21,920 more AI advancements to the part of AGI? 612 00:29:22,760 --> 00:29:24,640 Yeah. I mean, I find it hard to 613 00:29:24,640 --> 00:29:27,920 believe that an LLM based architecture alone would be 614 00:29:27,920 --> 00:29:31,360 capable of producing AGI. However, as I said before, you 615 00:29:31,360 --> 00:29:34,200 know, there are groups that are looking to merge in knowledge 616 00:29:34,200 --> 00:29:37,440 systems that are aware of facts in the world and how they relate 617 00:29:37,440 --> 00:29:39,240 together. Kind of going back to some of 618 00:29:39,240 --> 00:29:42,640 the really much older architectures for AI where they 619 00:29:42,640 --> 00:29:45,000 were attempting to like CYC, where they were attempting to 620 00:29:45,000 --> 00:29:48,000 actually model reality, like create this big huge knowledge 621 00:29:48,000 --> 00:29:50,720 graph that links together all these different concepts that 622 00:29:50,880 --> 00:29:53,800 collectively model reality. This was sort of a path that AI 623 00:29:53,800 --> 00:29:57,600 took a long, long time ago that was kind of abandoned when deep 624 00:29:57,600 --> 00:30:00,880 learning came along in favor of a more statistical probabilistic 625 00:30:01,200 --> 00:30:03,560 methods, including LLMS. Ultimately. 626 00:30:03,800 --> 00:30:06,440 So I think, you know, I can't help thinking and I think I'm 627 00:30:06,440 --> 00:30:09,520 not alone in this. I'm pretty sure well others 628 00:30:09,520 --> 00:30:12,360 might agree, like some kind of Harvard where we have this 629 00:30:12,800 --> 00:30:16,320 fantastic language generation capability of the LLMS combined 630 00:30:16,320 --> 00:30:18,000 with an actual. Sort of factual. 631 00:30:18,000 --> 00:30:22,240 System that helps to guide the LLM to not really say stupid 632 00:30:22,240 --> 00:30:23,800 things sometimes that ignore reality. 633 00:30:24,440 --> 00:30:25,880 And we'll just have to see how that goes. 634 00:30:25,920 --> 00:30:29,040 You know, like it's one of those things where the more computing 635 00:30:29,040 --> 00:30:31,760 power and the more efficient these algorithms and 636 00:30:31,760 --> 00:30:34,240 architectures become, the more things we can try out. 637 00:30:34,600 --> 00:30:36,800 And of course, there's a whole wild card of quantum comuting, 638 00:30:36,800 --> 00:30:39,000 like once quantum comuting reaches a level of maturity 639 00:30:39,000 --> 00:30:41,600 where it's able to be practically applied to this and 640 00:30:41,600 --> 00:30:43,280 all bets are off as far as what's possible. 641 00:30:44,120 --> 00:30:47,280 Yeah. So at one time, I feel excited, 642 00:30:47,280 --> 00:30:50,480 but at the other time, I feel a little bit scared as well. 643 00:30:50,480 --> 00:30:53,200 You know, so many people are thinking about jobs being 644 00:30:53,200 --> 00:30:54,840 changed. You know, people's roles are 645 00:30:54,840 --> 00:30:57,080 going to be diminished and all that. 646 00:30:57,320 --> 00:31:00,360 And obviously, you also have dealt with this kind of worries 647 00:31:00,360 --> 00:31:03,080 as well. What do you think people should 648 00:31:03,080 --> 00:31:07,240 do in this kind of era where things moving so fast every 649 00:31:07,240 --> 00:31:08,920 other day? I think you might have heard 650 00:31:08,920 --> 00:31:11,640 about cool things being applied using AI to improve 651 00:31:11,640 --> 00:31:14,040 productivity, remove redundancies and all that. 652 00:31:14,600 --> 00:31:16,560 So, yeah. What do you think people should 653 00:31:16,560 --> 00:31:19,360 do in this era? Yeah, I think first of all, 654 00:31:19,360 --> 00:31:21,240 maybe a little bit of an inspirational story. 655 00:31:21,240 --> 00:31:23,880 Like I love to give this example of radiologist. 656 00:31:24,040 --> 00:31:27,000 So, you know, seven or eight years ago, Geoffrey Hinton, 657 00:31:27,000 --> 00:31:30,720 who's a brilliant guy, like one of the forefathers of deep 658 00:31:30,720 --> 00:31:33,200 learning, one of the people who really helped make this whole 659 00:31:33,280 --> 00:31:35,120 AI, the first AI revolution really happened. 660 00:31:35,520 --> 00:31:37,680 He made this prediction that, you know, there would be no more 661 00:31:37,680 --> 00:31:40,800 radiologists within a few years, that we should just, you know, 662 00:31:40,800 --> 00:31:42,440 forget about radiologists because they're going to be 663 00:31:42,440 --> 00:31:44,520 replaced by computer vision based systems. 664 00:31:44,960 --> 00:31:48,600 And Fast forward to today and there's more radiologists than 665 00:31:48,600 --> 00:31:50,240 ever. So how could that be? 666 00:31:50,640 --> 00:31:54,360 And the answer is, well, you know, radiology is a high stakes 667 00:31:54,800 --> 00:31:57,720 kind of field in the sense that there are human lives at risk 668 00:31:57,720 --> 00:31:59,040 here. And so there is a need for a 669 00:31:59,040 --> 00:32:02,080 human in the loop. And the fact that we automated 670 00:32:02,080 --> 00:32:05,240 so much using machine learning means that essentially there's 671 00:32:05,240 --> 00:32:07,720 just more and more people who are out there looking to 672 00:32:07,720 --> 00:32:10,640 actually have radiology based procedures performed. 673 00:32:11,040 --> 00:32:13,280 Like these days, if you get an injury, it's like, Oh yeah, just 674 00:32:13,280 --> 00:32:15,160 get an MRI. You know, like just get a 675 00:32:15,160 --> 00:32:18,720 diagnostic MRI to find out if you might have some sort of 676 00:32:18,720 --> 00:32:19,320 problem. Like it's. 677 00:32:19,320 --> 00:32:22,640 Become so commonplace that could not have happened, I don't 678 00:32:22,640 --> 00:32:26,240 think, without these computer vision sort of advances. 679 00:32:26,280 --> 00:32:28,680 And it's resulted in actually more radiology jobs because 680 00:32:28,680 --> 00:32:30,000 there's still needs to be a human in the loop. 681 00:32:30,520 --> 00:32:33,840 And so it's kind of a volume thing where even though a lot of 682 00:32:33,960 --> 00:32:35,960 tasks would have made it, the volume increased. 683 00:32:36,160 --> 00:32:39,640 So that's point #1 and I think point #2 is there's this quote 684 00:32:39,640 --> 00:32:44,160 by, I think a former CEO of IBM saying, you know, AI won't 685 00:32:44,160 --> 00:32:47,880 replace humans per SE, but people who are able to 686 00:32:47,880 --> 00:32:50,840 effectively wield AI will definitely replace people who 687 00:32:50,840 --> 00:32:53,240 aren't. And so if you're listening to 688 00:32:53,240 --> 00:32:55,440 this, if you're going to take anything away from this whole 689 00:32:55,680 --> 00:32:58,840 conversation, that would be really immerse yourself, like 690 00:32:59,040 --> 00:33:02,240 look into how you might be able to use AI to make your own job 691 00:33:02,480 --> 00:33:04,720 better, to make the way you do your job better. 692 00:33:04,960 --> 00:33:07,640 There was a study, actually a very recent study published by 693 00:33:07,640 --> 00:33:10,720 the Harvard Business Review called the Cybernetic Work 694 00:33:10,720 --> 00:33:13,960 Companion, something like that. If you Google have a business 695 00:33:13,960 --> 00:33:15,600 review and cybernetic, you'll find it. 696 00:33:16,000 --> 00:33:17,920 And it's super interesting. Like they actually quantified 697 00:33:17,920 --> 00:33:20,920 this and said, oh, you know, somebody who's with the help of 698 00:33:21,000 --> 00:33:24,000 AI, suitably trained, can actually perform the work of two 699 00:33:24,040 --> 00:33:26,240 people. And I think that actually might 700 00:33:26,240 --> 00:33:28,440 be sort of higher than that, but it just goes. 701 00:33:28,440 --> 00:33:30,520 To show like wow, you could be at least twice as productive 702 00:33:30,520 --> 00:33:33,040 potentially using AI so. Figure out how to do that, 703 00:33:33,360 --> 00:33:35,800 because if you don't, someone else you know might. 704 00:33:36,320 --> 00:33:38,080 And so that's what I have to say about that. 705 00:33:38,920 --> 00:33:40,840 Yeah. So definitely you have to give 706 00:33:40,840 --> 00:33:42,680 it a try, right? And know what are the 707 00:33:42,680 --> 00:33:45,160 possibilities? Because sometimes like even out 708 00:33:45,160 --> 00:33:47,920 of our imagination, our current imagination, right, you wouldn't 709 00:33:47,920 --> 00:33:50,720 think that these kind of things can be solved by AI or some kind 710 00:33:50,720 --> 00:33:54,000 of a tools that can actually shortcut, you know, the time and 711 00:33:54,000 --> 00:33:56,400 effort required, right? So knowing just that 712 00:33:56,400 --> 00:33:59,480 possibilities, I think it's also really important and especially 713 00:33:59,480 --> 00:34:01,040 if you have used the system right. 714 00:34:01,720 --> 00:34:04,120 I mean, as you can imagine, we've been really looking into 715 00:34:04,120 --> 00:34:06,480 the use of AI internally, not just in our product, but 716 00:34:06,480 --> 00:34:09,280 internally the way we developed our technology and our software. 717 00:34:09,520 --> 00:34:13,760 And at first I was sort of a bit of AI would say a sceptic, you 718 00:34:13,760 --> 00:34:16,600 know, thinking like, hey, it's one thing to create a proof of 719 00:34:16,600 --> 00:34:20,320 concept or an early stage prototype using AI, but actual 720 00:34:20,320 --> 00:34:22,719 production code, no way. You know, we're still going to 721 00:34:22,719 --> 00:34:24,800 need a whole bunch of software engineers doing things pretty 722 00:34:24,800 --> 00:34:26,920 much the way they've always done them to do that. 723 00:34:27,239 --> 00:34:29,760 But now I'm not so. Sure, like we've had engineers 724 00:34:29,760 --> 00:34:32,440 do extraordinary things. Using the latest version of 725 00:34:32,440 --> 00:34:36,040 these code generation tools. And I really am starting to 726 00:34:36,040 --> 00:34:38,880 fully believe that the Harvard Business Review articleized, I 727 00:34:38,880 --> 00:34:41,280 have mentioned before about how people can be twice as 728 00:34:41,280 --> 00:34:43,639 productive. It absolutely holds for software 729 00:34:43,639 --> 00:34:46,199 engineering. And so, yeah, I mean, we're 730 00:34:46,280 --> 00:34:49,560 practicing what we preach. We're using AI to develop 731 00:34:49,560 --> 00:34:52,320 software a lot more, you know, productively and faster than 732 00:34:52,320 --> 00:34:54,120 before. And it's a pretty exciting time 733 00:34:54,120 --> 00:34:57,360 because the underlying tools also are evolving so fast. 734 00:34:57,680 --> 00:34:59,720 We're finding that the latest versions of the code. 735 00:35:00,000 --> 00:35:02,560 Generation tools are just way more capable than the ones even 736 00:35:02,560 --> 00:35:06,840 just from a few months ago, so that's also really exciting to 737 00:35:06,840 --> 00:35:09,600 be surfing that wave of advancement in AI in that way as 738 00:35:09,600 --> 00:35:12,000 well. Yeah, personally, as someone who 739 00:35:12,000 --> 00:35:14,560 have been around in the industry for quite some time, right, I 740 00:35:14,560 --> 00:35:17,680 find myself during this transition also a little bit 741 00:35:17,680 --> 00:35:20,920 skeptic at once, but actually seeing maybe the junior people 742 00:35:20,920 --> 00:35:23,080 doing stuff that I couldn't imagine. 743 00:35:23,200 --> 00:35:25,040 Sometimes I feel like, wow, amazing. 744 00:35:25,080 --> 00:35:28,800 Like sometimes I'm having to, you know, switch my perspectives 745 00:35:29,200 --> 00:35:31,360 to be more optimistic rather than skeptic, right? 746 00:35:31,640 --> 00:35:33,960 I think this is also maybe some challenge for those of you who 747 00:35:33,960 --> 00:35:36,720 have been around in industry because we used to think this is 748 00:35:36,720 --> 00:35:39,440 how we solve things. But now I think the 749 00:35:39,440 --> 00:35:42,160 possibilities are there, right? So as long as you can imagine, 750 00:35:42,160 --> 00:35:45,000 you give it a try and maybe I could help you in some parts of 751 00:35:45,000 --> 00:35:46,840 the job. And I think your productivity 752 00:35:46,840 --> 00:35:50,360 can be proof as well. So you mentioned about 753 00:35:50,360 --> 00:35:52,240 engineers, right? So let's switch a little bit to 754 00:35:52,240 --> 00:35:54,480 that side. So is there any difference 755 00:35:54,480 --> 00:35:58,440 managing or leading AI engineers versus typical software 756 00:35:58,440 --> 00:36:01,240 engineers, You know, like things like back end, front end and 757 00:36:01,240 --> 00:36:03,240 those kind of things? Yeah. 758 00:36:03,240 --> 00:36:07,840 I mean, even before the LLM revolution came along, it's 759 00:36:07,840 --> 00:36:11,360 always been quite different managing data scientists or ML 760 00:36:11,360 --> 00:36:13,240 engineers as they sort of used to be called, right? 761 00:36:13,760 --> 00:36:16,840 Simply because a lot of what you're doing is exploration. 762 00:36:17,240 --> 00:36:20,440 Like, I love what you're doing is you had this idea like, oh, 763 00:36:20,520 --> 00:36:22,880 you know, we can use data science to solve some problem, 764 00:36:23,320 --> 00:36:26,800 but it's not like traditional software engineering where you 765 00:36:27,080 --> 00:36:30,800 have this idea, you design it, you specify it, and then you 766 00:36:30,800 --> 00:36:33,440 sort of build it all and according to spec. 767 00:36:33,520 --> 00:36:35,320 And it's a very deterministic process. 768 00:36:35,760 --> 00:36:38,240 With data science in general, you know, you're just going to 769 00:36:38,240 --> 00:36:39,520 find out, like, is this feasible? 770 00:36:39,960 --> 00:36:42,280 OK, I have some idea. I can predict, you know, let's 771 00:36:42,280 --> 00:36:45,520 just say I can look at the whole bunch of historical trials and I 772 00:36:45,520 --> 00:36:49,280 can start to predict things like amendments or outcomes of any 773 00:36:49,280 --> 00:36:51,400 kind. That's a thesis that you need to 774 00:36:51,400 --> 00:36:54,000 go out there and just test kind of like science, right? 775 00:36:54,000 --> 00:36:55,320 That's why they're called data science. 776 00:36:55,440 --> 00:36:58,080 And so that inherently is quite different because you have to 777 00:36:58,080 --> 00:37:04,200 embrace more uncertainty. Now I think maybe people also 778 00:37:04,200 --> 00:37:07,960 want to sort of know what about all this AI vibe coding or AI 779 00:37:07,960 --> 00:37:09,640 enabled coding? Like how does that change? 780 00:37:09,640 --> 00:37:14,320 And I think that what it does is essentially it's a force 781 00:37:14,320 --> 00:37:18,000 multiplier for people. So maybe to build a prototype 782 00:37:18,280 --> 00:37:20,040 these days, you don't necessarily have to be a 783 00:37:20,040 --> 00:37:22,040 software engineer anymore. You can get to a certain point 784 00:37:22,040 --> 00:37:24,000 without any software engineering skills whatsoever. 785 00:37:24,320 --> 00:37:27,040 So that's kind of different. What we're observing out there 786 00:37:27,040 --> 00:37:30,240 actually is that there's a lot of entrepreneurs and even sort 787 00:37:30,240 --> 00:37:33,680 of like groups that are out there selling who are basically 788 00:37:33,680 --> 00:37:35,800 building prototypes where before they would have been presenting 789 00:37:35,800 --> 00:37:39,560 marks or, you know, conceptual sort of descriptions of what 790 00:37:39,560 --> 00:37:41,400 they intend to build. And now it's like, no, we 791 00:37:41,400 --> 00:37:43,920 actually just went ahead and built it because these coding 792 00:37:43,920 --> 00:37:46,680 systems, these automated coding systems make things so much 793 00:37:46,680 --> 00:37:48,960 easier. So there's this huge shift 794 00:37:48,960 --> 00:37:53,480 towards essentially what a lot of the people who are really 795 00:37:53,480 --> 00:37:56,960 behind the agile process we're already advocating, which is you 796 00:37:56,960 --> 00:37:59,720 just got to prototype stuff, try it out and test it with real 797 00:37:59,720 --> 00:38:01,920 customers. And I think in the past, a lot 798 00:38:01,920 --> 00:38:04,360 of people didn't do that because it was just too much work. 799 00:38:04,600 --> 00:38:06,880 It was like, well, you know, prototyping, it just takes so 800 00:38:06,880 --> 00:38:10,080 much work to try to make all these ideas real, like, and then 801 00:38:10,200 --> 00:38:11,880 potentially throw them away. It's just like we don't like 802 00:38:11,880 --> 00:38:14,560 throwing things away. Whereas now the bar is so much 803 00:38:14,560 --> 00:38:18,560 lower as far as what the amount of effort it takes to produce a 804 00:38:18,560 --> 00:38:20,440 working prototype. That it's much more. 805 00:38:20,440 --> 00:38:23,360 Feasible to do what people like Marty Kagan, you know, out there 806 00:38:23,360 --> 00:38:26,040 say, which is just pretty like a whole bunch of stuff, you know, 807 00:38:26,040 --> 00:38:28,680 within two weeks or less. Before you fall in love with the 808 00:38:28,680 --> 00:38:30,640 idea, just test it out and throw it away if it doesn't work. 809 00:38:30,960 --> 00:38:33,880 So that philosophy and that process will become way more 810 00:38:33,880 --> 00:38:36,120 feasible with AI, which is really exciting too. 811 00:38:37,160 --> 00:38:38,920 Yeah, definitely makes sense, right? 812 00:38:38,920 --> 00:38:42,160 So the process, the iteration now can be much faster and even 813 00:38:42,160 --> 00:38:45,120 people are not skilled enough in development now can do software 814 00:38:45,120 --> 00:38:47,120 development in some sort of form, right? 815 00:38:47,560 --> 00:38:50,680 So as Acto yourself, you mentioned about vibe coding. 816 00:38:50,720 --> 00:38:52,440 What is your view about vibe coding? 817 00:38:52,440 --> 00:38:55,400 Do you see your engineers doing some kind of vibe coding? 818 00:38:55,400 --> 00:38:59,360 And if so, right, is there any danger that you foresee and how 819 00:38:59,360 --> 00:39:03,640 do you actually mitigate that? I think vibe coding is great. 820 00:39:03,640 --> 00:39:08,240 Meaning just using these AIF systems to generate code that 821 00:39:08,240 --> 00:39:11,120 you don't even necessarily look at, you just prompted and 822 00:39:11,120 --> 00:39:13,920 prompted and prompted until something works. 823 00:39:14,200 --> 00:39:17,960 I think that's great for doing proof of concept initial 0 to 824 00:39:17,960 --> 00:39:20,840 one kind of prototyping. I don't think that's a viable 825 00:39:20,840 --> 00:39:23,360 path to actually building enterprise software simply 826 00:39:23,360 --> 00:39:27,680 because when things go wrong, as they inevitably will, you know, 827 00:39:27,680 --> 00:39:30,160 with AI generated code. Like I started playing with some 828 00:39:30,160 --> 00:39:33,480 of these vibe coding tools myself and very quickly I dug a 829 00:39:33,480 --> 00:39:35,360 hole. Like once I passed a certain 830 00:39:35,360 --> 00:39:38,440 threshold of complexity, it just became really hard to add new 831 00:39:38,440 --> 00:39:40,440 features. They had this magical button 832 00:39:40,440 --> 00:39:42,840 called like try to fix it. And I was pressing that button a 833 00:39:42,840 --> 00:39:46,400 lot. And so I don't know, like I, I, 834 00:39:46,560 --> 00:39:49,800 I think it's got a place and I think it's wonderful that people 835 00:39:49,800 --> 00:39:52,120 who don't have a computer science background or haven't or 836 00:39:52,120 --> 00:39:54,200 not, you know, experienced software engineers can now 837 00:39:54,200 --> 00:39:56,920 actually build technology like that's magical, that's 838 00:39:56,920 --> 00:39:59,080 incredible. But it only goes so far. 839 00:39:59,240 --> 00:40:02,160 And I think in the future, sure, vibe coding will become more 840 00:40:02,160 --> 00:40:04,040 powerful. It will become better. 841 00:40:04,520 --> 00:40:06,440 But I do think at the end of the day, if you are producing 842 00:40:06,440 --> 00:40:10,120 enterprise grade software that's used by real paying customers, 843 00:40:10,440 --> 00:40:12,760 you're going to want to have people who actually understand 844 00:40:12,760 --> 00:40:15,800 the technology, even if they're using AI to help them code. 845 00:40:16,120 --> 00:40:17,440 To actually fix things when they go. 846 00:40:17,440 --> 00:40:19,240 Wrong and to. Ensure that things are designed 847 00:40:19,240 --> 00:40:21,320 in a good way. At least that's my view. 848 00:40:22,080 --> 00:40:23,880 Yeah. So you mentioned like building 849 00:40:23,880 --> 00:40:26,200 enterprise system, right? So the things that are complex 850 00:40:26,200 --> 00:40:27,840 and maybe need to evolve over time, right? 851 00:40:27,840 --> 00:40:30,560 Because I still don't know whether you know, just simply 852 00:40:30,560 --> 00:40:34,160 doing vibe coding can actually evolve your system, you know, in 853 00:40:34,160 --> 00:40:37,000 a predictable manner, right? So rather than, you know, keep 854 00:40:37,320 --> 00:40:39,080 prompting AI until it gets fixed, right? 855 00:40:39,080 --> 00:40:42,160 I think the code that gets produced also can be quite 856 00:40:42,160 --> 00:40:44,000 terrible if human are looking at it, right? 857 00:40:44,000 --> 00:40:46,440 Because simply they don't maybe adhere to a certain design 858 00:40:46,440 --> 00:40:49,160 pattern. There's no refactoring done in a 859 00:40:49,240 --> 00:40:52,240 concerted effort, I guess. So I think vibe coding is 860 00:40:52,240 --> 00:40:55,000 definitely is wonderful for things like POC or things that 861 00:40:55,000 --> 00:40:58,120 you're not looking at or something that is more siloed. 862 00:40:58,120 --> 00:41:01,360 I guess that you can just change the whole thing once you think 863 00:41:01,360 --> 00:41:04,840 the quality is not there. So I think also another thing 864 00:41:04,880 --> 00:41:08,560 when we hire engineers, right? Do you see any kind of 865 00:41:08,560 --> 00:41:11,480 difference these days in terms of, I don't know, attitude, 866 00:41:11,480 --> 00:41:14,800 skill set or non-technical skills that you see when you 867 00:41:14,800 --> 00:41:17,840 hire AI engineers? Yeah, I mean, I think that we're 868 00:41:17,840 --> 00:41:19,520 still at the early stages of this. 869 00:41:19,880 --> 00:41:22,920 And what I do observe as a sort of a hiring manager is there's a 870 00:41:22,920 --> 00:41:26,040 lot of resumes out there with AI on them because everybody wants 871 00:41:26,040 --> 00:41:29,240 to be as perceived as someone who's versant in this because 872 00:41:29,240 --> 00:41:31,000 they know that there's a lot of demand. 873 00:41:31,520 --> 00:41:33,520 But when we start interviewing, when we start testing people's 874 00:41:33,520 --> 00:41:37,520 actual coding capabilities, because this is still relatively 875 00:41:37,520 --> 00:41:41,760 recent, it's still kind of difficult to find people who 876 00:41:41,760 --> 00:41:45,080 really do have experience with this and who have demonstrable 877 00:41:45,280 --> 00:41:48,360 ability in this area. And so I think it's going to 878 00:41:48,360 --> 00:41:51,000 shift over time, just like in the old days, you know, it was 879 00:41:51,000 --> 00:41:53,960 kind of hard to find like way back in the 90s was hard to find 880 00:41:53,960 --> 00:41:57,480 people who knew HTML and he knew like how to build full and then 881 00:41:57,480 --> 00:41:59,800 later to who really knew how to build full stack applications. 882 00:41:59,800 --> 00:42:02,600 And before too long, within a few years, everybody did. 883 00:42:02,840 --> 00:42:05,040 It was the norm. And so I think this is just a 884 00:42:05,040 --> 00:42:09,320 natural technology adoption curve where we're sort of moving 885 00:42:09,320 --> 00:42:12,680 into a more of a mainstream situation where everybody's seen 886 00:42:12,680 --> 00:42:15,040 the value and. You know the majority of. 887 00:42:15,560 --> 00:42:18,360 Software engineers out there are going to be conversant in this 888 00:42:18,360 --> 00:42:20,440 technology and before too long it'll be ubiquitous. 889 00:42:20,800 --> 00:42:23,000 But we haven't quite seen that in the hiring pool yet. 890 00:42:23,560 --> 00:42:25,680 But I think, as I said, it's sort of inevitable. 891 00:42:26,240 --> 00:42:28,400 Maybe from your existing engineers that you have in the 892 00:42:28,400 --> 00:42:31,520 companies, Do you see someone that stood out simply because 893 00:42:31,520 --> 00:42:34,520 you know they can combine all these possibilities and also 894 00:42:34,520 --> 00:42:37,640 with a certain attitude, do you see some kind of persona of 895 00:42:37,640 --> 00:42:40,000 engineers that stood out that you think you can share with us 896 00:42:40,000 --> 00:42:41,560 here? Like, what makes a better 897 00:42:41,560 --> 00:42:43,720 engineer now that AI technologies are there? 898 00:42:44,720 --> 00:42:47,920 Yeah. I mean, I talk a lot about 899 00:42:48,320 --> 00:42:51,920 career development and about the qualities that I've observed 900 00:42:51,960 --> 00:42:54,200 over the many years that I've been working in this industry 901 00:42:54,520 --> 00:42:57,280 that lead people to become successful and sort of more, 902 00:42:57,280 --> 00:43:01,000 quote, senior in some way. And one of the big ones is kind 903 00:43:01,000 --> 00:43:02,960 of curiosity. There's this thing kind of 904 00:43:02,960 --> 00:43:05,400 inherently excited about learning new things and trying 905 00:43:05,400 --> 00:43:08,160 things out. And that's a quality that I 906 00:43:08,160 --> 00:43:12,640 think really, really serves people well in this new world of 907 00:43:12,680 --> 00:43:14,880 AI, because a lot of it is kind of unknown. 908 00:43:14,880 --> 00:43:17,680 It's like just trying to apply AI in ways that. 909 00:43:17,680 --> 00:43:19,400 Have been done before. Like we're doing here at Faro. 910 00:43:19,800 --> 00:43:24,480 And so I do think that that's an example of a quality that is 911 00:43:24,680 --> 00:43:27,920 really a strong one. And I think also people who can 912 00:43:27,920 --> 00:43:31,120 span multiple domain by people who are able to think in terms 913 00:43:31,120 --> 00:43:34,640 of product, like what makes a good product, what makes a well 914 00:43:34,640 --> 00:43:39,400 designed product like usability wise, along with the ability to 915 00:43:39,400 --> 00:43:43,160 sort of decompose that into commands that you can give to 916 00:43:43,160 --> 00:43:44,840 the AI that it's going to understand. 917 00:43:45,280 --> 00:43:47,600 So this is a combination of different skill sets. 918 00:43:47,600 --> 00:43:50,160 It's not just like you do a computer science degree and 919 00:43:50,160 --> 00:43:52,680 you're good. It's more like, hey, there's a 920 00:43:52,680 --> 00:43:55,320 part of this which is customer facing, like knowing what 921 00:43:55,320 --> 00:43:57,120 customers really would respond well to this. 922 00:43:57,120 --> 00:43:59,640 Part of this, which is communication, being able to 923 00:43:59,640 --> 00:44:01,920 kind of express those requirements well, like a 924 00:44:01,920 --> 00:44:04,160 product manager does. And it's part of this which is 925 00:44:04,160 --> 00:44:07,640 like UX designer and of course part of it, which is straight up 926 00:44:07,640 --> 00:44:09,920 programmer. So I think the people who can 927 00:44:09,920 --> 00:44:12,960 think cross functionally also will do really well in this new 928 00:44:12,960 --> 00:44:16,440 world because they now have the force multiplier, you know, in 929 00:44:16,440 --> 00:44:18,640 the form of these automated coding assistants. 930 00:44:18,640 --> 00:44:22,640 And so on to really be able to do the work of multiple people 931 00:44:22,640 --> 00:44:24,400 potentially. Yeah. 932 00:44:24,400 --> 00:44:27,760 So definitely, those skills are still kind of like the most 933 00:44:27,760 --> 00:44:30,360 important things, right? Especially if you are engineers 934 00:44:30,440 --> 00:44:33,000 listening to this and worrying about your jobs getting 935 00:44:33,000 --> 00:44:35,880 disappeared simply because getting replaced by a IS. 936 00:44:36,320 --> 00:44:39,680 So how about leaders themselves? So I know some leaders maybe 937 00:44:39,680 --> 00:44:42,240 have decades of experience. Maybe they are up and coming 938 00:44:42,240 --> 00:44:44,200 leaders as well, like engineering managers and all 939 00:44:44,200 --> 00:44:46,920 that. Do you see a certain skill set 940 00:44:46,920 --> 00:44:49,960 or attitude that must change in this EI era as well? 941 00:44:50,680 --> 00:44:54,560 Yeah, I do think that every time there's a technology disruption, 942 00:44:54,960 --> 00:44:59,040 it can be a little bit scary simply because perhaps the 943 00:44:59,040 --> 00:45:02,120 processes and tools and behaviours that have served you 944 00:45:02,120 --> 00:45:04,960 really well in your career to date, it might be sort of under 945 00:45:04,960 --> 00:45:07,000 threat or, you know, about to be transformed. 946 00:45:07,400 --> 00:45:10,840 And that can be exciting. And it can also be kind of scary 947 00:45:10,840 --> 00:45:12,680 and threatening like, Oh my God, what do I do? 948 00:45:13,000 --> 00:45:16,520 I can't depend on the same set of skills and habits and 949 00:45:16,720 --> 00:45:19,600 behaviors that have got me to where I am in my career. 950 00:45:20,120 --> 00:45:22,400 And I think that the people who can sort of take that 951 00:45:22,400 --> 00:45:25,640 uncertainty on board and re step into the unknown are the ones 952 00:45:25,640 --> 00:45:27,440 who can thrive. And I think that applies at a 953 00:45:27,440 --> 00:45:29,920 company level as well. Every single company out there 954 00:45:29,920 --> 00:45:31,960 at this point is wondering like what do we do about AI? 955 00:45:32,640 --> 00:45:35,440 And I think it's the ones that are bold and are willing to 956 00:45:35,440 --> 00:45:38,440 disrupt themselves. Like I, I just spoke to the 957 00:45:38,440 --> 00:45:40,280 company earlier on this week and said we're essentially 958 00:45:40,280 --> 00:45:43,320 disrupting ourselves. Like we started this company as 959 00:45:43,320 --> 00:45:46,040 a traditional SaaS provider. Now we're becoming an AI company 960 00:45:46,040 --> 00:45:48,120 and that is a process of disruption. 961 00:45:48,600 --> 00:45:52,840 And I would look at Apple in the years of Steve Jobs, if you want 962 00:45:52,840 --> 00:45:55,360 to look at an example of a company that's has the courage 963 00:45:55,600 --> 00:45:58,840 and the sort of the fortitude to disrupt itself repeatedly like 964 00:45:58,840 --> 00:46:02,120 Apple did that you remember, like the iPod that used to be 965 00:46:02,120 --> 00:46:04,920 the best thing ever, like Apple disrupted their own product with 966 00:46:04,920 --> 00:46:07,280 the iPhone. And I think that that is the 967 00:46:07,280 --> 00:46:11,160 kind of thinking that is needed here when this new technology 968 00:46:11,160 --> 00:46:13,800 wave comes along at the company level as well as the management 969 00:46:13,800 --> 00:46:16,920 individual leader level is giving your teams. 970 00:46:17,120 --> 00:46:19,680 First of all, have possessing that courage to step into the 971 00:46:19,680 --> 00:46:22,520 unknown yourself, but also inspiring your teams to do so 972 00:46:22,720 --> 00:46:25,160 and saying hey, this is you know, if we don't do this, 973 00:46:25,160 --> 00:46:27,440 someone else will and we will get disrupted. 974 00:46:27,960 --> 00:46:30,960 And you can look at the history of this, like if you look at 975 00:46:30,960 --> 00:46:33,280 things like the Innovator's Dilemma, you know, Clayton 976 00:46:33,280 --> 00:46:35,880 Christensen's book that talks about this is like a natural 977 00:46:35,880 --> 00:46:37,560 law, like companies get disrupted because not all of 978 00:46:37,560 --> 00:46:39,560 them have the fortitude to actually jump on new 979 00:46:39,560 --> 00:46:41,160 technologies and disrupt themselves. 980 00:46:42,480 --> 00:46:44,360 Yeah. So I think thanks for pointing 981 00:46:44,360 --> 00:46:46,520 that out, right? So definitely it's something 982 00:46:46,520 --> 00:46:48,960 that for some of us leaders, right, have to think about 983 00:46:48,960 --> 00:46:51,800 disrupting yourself, disrupting your team and disrupting your 984 00:46:51,800 --> 00:46:54,200 company, right? So I think those things can be 985 00:46:54,200 --> 00:46:57,960 healthy, especially if it can also revolutionize the way 986 00:46:57,960 --> 00:47:00,160 things are done. So how about the juniors here, 987 00:47:00,160 --> 00:47:02,280 juniors who probably are thinking whether software 988 00:47:02,280 --> 00:47:05,040 development is going to be their career or someone who is 989 00:47:05,040 --> 00:47:08,360 studying software engineering. But I think that they now have 990 00:47:08,360 --> 00:47:10,360 to switch. Do you still think there's a 991 00:47:10,360 --> 00:47:12,760 room for them to actually enter the industry? 992 00:47:12,920 --> 00:47:16,360 Because I heard so many times that people say simply because 993 00:47:16,360 --> 00:47:19,800 you can be more productive now, maybe we need less junior 994 00:47:19,800 --> 00:47:21,760 engineers. So what is your view here? 995 00:47:21,760 --> 00:47:26,840 I think that there's never a shortage for people who are able 996 00:47:26,840 --> 00:47:30,600 to create technology, and AI makes that easier. 997 00:47:31,040 --> 00:47:35,440 And so, yes, you could argue the field of programming is becoming 998 00:47:35,440 --> 00:47:37,680 more commoditized because of all this automation. 999 00:47:38,520 --> 00:47:42,520 But like I said before, AI can't necessarily take the place of 1000 00:47:42,520 --> 00:47:45,600 having good product sense or having the combination of 1001 00:47:45,880 --> 00:47:48,400 programming skills and design skills and so on. 1002 00:47:49,080 --> 00:47:51,480 There's different ways of framing this, but I would say 1003 00:47:52,160 --> 00:47:54,440 thinking of it as being a force multiplier, like you're able to 1004 00:47:54,440 --> 00:47:57,400 do more, whether it's as an entrepreneur or whether it's as 1005 00:47:57,760 --> 00:47:59,640 an engineer who's part of a larger organization. 1006 00:48:00,120 --> 00:48:03,240 If you're able to wield these tools well, that'll serve you 1007 00:48:03,240 --> 00:48:05,680 very well. And I think we're still also in 1008 00:48:05,840 --> 00:48:07,880 the early stage of this. So if you are able to, for 1009 00:48:07,880 --> 00:48:12,400 instance, start a site project, a personal project, and put that 1010 00:48:12,480 --> 00:48:15,800 in GitHub or put that on your resume and show that you have 1011 00:48:15,800 --> 00:48:18,160 the ability to do this. Even if it's some personal 1012 00:48:18,160 --> 00:48:21,160 project you just happen to be really excited about, definitely 1013 00:48:21,160 --> 00:48:23,600 do that. Because that shows that you can 1014 00:48:23,600 --> 00:48:26,040 do it and that shows that you can adapt and that you have some 1015 00:48:26,040 --> 00:48:30,680 level of mastery or facility in this area and that will serve 1016 00:48:30,680 --> 00:48:33,240 you well as you look for jobs. Because the employers that we're 1017 00:48:33,240 --> 00:48:36,480 looking for people who can do this now and they're still kind 1018 00:48:36,480 --> 00:48:39,040 of hard to come by. Maybe not for too many more 1019 00:48:39,320 --> 00:48:42,400 months or years, but there's an opportunity to jump in there if 1020 00:48:42,400 --> 00:48:45,400 you really are proactive and potentially stand out from. 1021 00:48:45,400 --> 00:48:46,520 Other candidates because of that. 1022 00:48:47,760 --> 00:48:49,840 Yeah. So I think definitely don't get 1023 00:48:49,840 --> 00:48:51,960 this heartened especially if you're really passionate about 1024 00:48:51,960 --> 00:48:54,800 software engineering, right, creating application products 1025 00:48:54,800 --> 00:48:57,280 and even like enterprise software still kind of like the 1026 00:48:57,280 --> 00:48:59,480 go to things that you are passionate about. 1027 00:48:59,680 --> 00:49:02,000 So I think that don't get this heartened and maybe a try out 1028 00:49:02,000 --> 00:49:03,640 stuff just like Patrick mentioned. 1029 00:49:04,000 --> 00:49:07,960 So we have talked a lot about you know, AI in clinical trials, 1030 00:49:08,080 --> 00:49:10,960 AI in general and also software engineering and engineering 1031 00:49:10,960 --> 00:49:12,880 teams. Is there anything else that you 1032 00:49:12,880 --> 00:49:15,840 think you want to share with us today that we haven't talked 1033 00:49:15,840 --> 00:49:17,760 about? Maybe just coming back to the 1034 00:49:17,760 --> 00:49:22,840 theme of career development, like some of the most successful 1035 00:49:23,320 --> 00:49:26,080 and interesting careers I've seen of all the people I've 1036 00:49:26,080 --> 00:49:29,000 worked with have been people who really focus on just trying 1037 00:49:29,000 --> 00:49:30,160 things out. There. 1038 00:49:30,160 --> 00:49:32,480 Was this one maybe an inspirationalist degree? 1039 00:49:32,760 --> 00:49:35,240 Like there was this one really junior product manager I worked 1040 00:49:35,240 --> 00:49:40,400 with when I first joined Google and I encouraged them to sort of 1041 00:49:40,400 --> 00:49:43,200 become a user of our application, which at that time 1042 00:49:43,200 --> 00:49:46,000 was this kind of e-commerce product. 1043 00:49:46,240 --> 00:49:48,720 And so I said, hey, you should probably sign up as a merchant, 1044 00:49:48,720 --> 00:49:51,240 start your own little sort of fake business and kind of try 1045 00:49:51,240 --> 00:49:53,760 out the product. And he ended up really getting 1046 00:49:53,760 --> 00:49:55,920 into e-commerce. Like he built this company that 1047 00:49:55,920 --> 00:49:59,440 ended up becoming successful and he ended up sort of becoming AVP 1048 00:49:59,440 --> 00:50:02,720 at Shopify. And I just kind of, I was giving 1049 00:50:02,720 --> 00:50:05,400 my talk earlier on this year at the company on career. 1050 00:50:05,400 --> 00:50:08,200 Development and I actually spoke to him and just said, hey, just 1051 00:50:08,200 --> 00:50:10,400 how would you? Sum up like your success because 1052 00:50:10,400 --> 00:50:13,040 it's like such an incredible. Career path that he followed. 1053 00:50:13,560 --> 00:50:15,640 As an entrepreneur and now senior executive and. 1054 00:50:15,680 --> 00:50:18,000 Like all these things. And he just said, you just got 1055 00:50:18,000 --> 00:50:20,720 to try things out. And he reflected back to me, 1056 00:50:20,720 --> 00:50:23,840 like AI just enables so many more people to do this now. 1057 00:50:24,280 --> 00:50:26,640 Like to do the kind of things that he did that at the time 1058 00:50:26,640 --> 00:50:29,560 required software engineering expertise, machine learning 1059 00:50:29,560 --> 00:50:32,600 expertise. Like now you can potentially 1060 00:50:32,600 --> 00:50:35,000 vibe code some of the stuff that he did really easily. 1061 00:50:35,360 --> 00:50:38,200 And so the opportunity for people who have ideas that they 1062 00:50:38,200 --> 00:50:42,560 observe problems in the world, the gap between observing that 1063 00:50:42,560 --> 00:50:46,880 problem and trying to solve it is now much, much smaller than 1064 00:50:46,880 --> 00:50:48,960 it was. And so whether you're trying to 1065 00:50:48,960 --> 00:50:51,040 impress some big company that you want to get a job at or 1066 00:50:51,040 --> 00:50:52,760 whether you actually want to start a new company. 1067 00:50:53,160 --> 00:50:54,480 It's just much easier to do that. 1068 00:50:54,760 --> 00:50:56,720 So I would. Sort of, you know, learn from 1069 00:50:56,720 --> 00:51:00,000 his experience and maybe be inspired by his story because, 1070 00:51:00,520 --> 00:51:04,120 you know, he was really, really understoring the fact that it's 1071 00:51:04,120 --> 00:51:06,720 just much easier to try things out these days. 1072 00:51:06,720 --> 00:51:09,320 And trying things out is often what leads to great inventions 1073 00:51:09,320 --> 00:51:12,120 and innovations in the world. Thanks for sharing such 1074 00:51:12,120 --> 00:51:14,640 inspirational story. I think definitely worth to 1075 00:51:14,640 --> 00:51:17,200 think about, right? So especially you having cross 1076 00:51:17,200 --> 00:51:19,480 functional skills back when you mentioned about cross functional 1077 00:51:19,480 --> 00:51:21,040 thing, right? If you have cross functional 1078 00:51:21,040 --> 00:51:23,360 skills, definitely the opportunities and the 1079 00:51:23,360 --> 00:51:26,160 possibilities are there for us to give it a try, right? 1080 00:51:26,520 --> 00:51:29,560 And vibe coding, yeah, it could be one way to actually build 1081 00:51:29,560 --> 00:51:32,480 POCS and kinda like test the market early before you actually 1082 00:51:32,480 --> 00:51:35,160 build the actual product. So Patrick, thank you so much 1083 00:51:35,160 --> 00:51:37,600 for sharing today. So I learnt a lot, especially 1084 00:51:37,600 --> 00:51:39,840 from those inspirational stories that you mentioned. 1085 00:51:40,200 --> 00:51:43,160 But before I let you go, I have one last question that I always 1086 00:51:43,160 --> 00:51:45,520 ask my guests. I call this the three technical 1087 00:51:45,520 --> 00:51:47,880 leadership wisdom. You can think of it just like an 1088 00:51:47,880 --> 00:51:49,520 advice that you want to give to the listeners. 1089 00:51:49,520 --> 00:51:51,920 What are the top three things that you want to share today? 1090 00:51:52,640 --> 00:51:54,800 Definitely, I'll return to the curiosity. 1091 00:51:54,880 --> 00:52:00,000 Like that's super important. I would say definitely following 1092 00:52:00,000 --> 00:52:03,400 your excitement and all things like just don't settle. 1093 00:52:03,400 --> 00:52:06,200 For doing work that you don't find and personally inspiring. 1094 00:52:06,720 --> 00:52:08,600 That you don't feel like, oh, I'm excited to get to work in 1095 00:52:08,600 --> 00:52:10,320 the morning. Like if you don't feel that way, 1096 00:52:10,320 --> 00:52:12,880 then think of other things that you might get more excited 1097 00:52:12,880 --> 00:52:15,360 about. And it sort of sounds obvious, 1098 00:52:15,360 --> 00:52:19,320 but I think that people get conditioned after a while to not 1099 00:52:19,320 --> 00:52:22,120 doing that and just settling. And I think that it's important 1100 00:52:22,120 --> 00:52:25,040 to really be excited about your work for a number of different 1101 00:52:25,040 --> 00:52:27,920 reasons. And I think the third thing 1102 00:52:27,920 --> 00:52:32,480 would be definitely, I always advise people to find a mentor. 1103 00:52:33,040 --> 00:52:36,160 And that also sounds kind of obvious, but I think the 1104 00:52:36,160 --> 00:52:38,240 majority of people don't actually actively seek out a 1105 00:52:38,240 --> 00:52:40,520 mentor. And it can be super, super 1106 00:52:40,520 --> 00:52:43,080 rewarding. Like I keep coming back to some 1107 00:52:43,080 --> 00:52:45,560 of the people in my own. Career who have really helped 1108 00:52:45,560 --> 00:52:47,480 me, even if it's just a sentence. 1109 00:52:47,480 --> 00:52:51,560 Or a catch phrase or a certain just way of doing one thing. 1110 00:52:51,800 --> 00:52:55,640 It can change your life over time really by just helping to 1111 00:52:55,640 --> 00:53:00,080 reframe problems or deal with, you know, communication issues 1112 00:53:00,080 --> 00:53:02,760 or whatever the case may be over time in your career. 1113 00:53:02,760 --> 00:53:04,360 That can have a really compounding effect. 1114 00:53:04,760 --> 00:53:08,880 So seek out at least one mentor, ideally multiple over the course 1115 00:53:08,880 --> 00:53:09,880 of your career. And it's really. 1116 00:53:09,880 --> 00:53:12,200 Going to help you. Yeah, thanks for pointing that 1117 00:53:12,200 --> 00:53:13,840 out. So you mentioned like even a 1118 00:53:13,840 --> 00:53:16,600 simple sentence phrase and all that could actually change 1119 00:53:16,600 --> 00:53:18,920 someone's lives, right, And can change your life as well. 1120 00:53:19,200 --> 00:53:22,000 So I think for people who have been around in the industry as 1121 00:53:22,000 --> 00:53:23,400 well. So be a mentor, I think I would 1122 00:53:23,400 --> 00:53:26,200 say as well, because you never know how much impact that you 1123 00:53:26,200 --> 00:53:29,480 can make to someone's lives. So, Patrick, if people love this 1124 00:53:29,480 --> 00:53:32,520 conversation, they wanna reach out to you or maybe learn more 1125 00:53:32,520 --> 00:53:33,840 from you. Is there a place where they can 1126 00:53:33,840 --> 00:53:36,360 find you online? Yeah, The best way is LinkedIn. 1127 00:53:36,480 --> 00:53:38,360 You can find me on LinkedIn pretty easily. 1128 00:53:38,960 --> 00:53:39,840 Right. OK. 1129 00:53:39,840 --> 00:53:41,720 So thank you so much for sharing today, Patrick. 1130 00:53:41,720 --> 00:53:44,160 I really learned a lot. And again, thank you so much. 1131 00:53:44,160 --> 00:53:46,040 It was a real pleasure to be here with Henry. 1132 00:53:46,040 --> 00:53:46,400 Thank you.