1 00:00:05,720 --> 00:00:10,465 In this episode of Translating Proteomics, host Parag Malik discusses 2 00:00:10,505 --> 00:00:14,521 the role of AI and machine learning in biotech with special 3 00:00:14,593 --> 00:00:19,161 guests Matt McElwain from Madrona Venture Group and Vijay Pandey 4 00:00:19,193 --> 00:00:24,045 from Andreessen Horowitz. Their fascinating conversation covers 5 00:00:24,425 --> 00:00:28,633 advances that have enabled biotech to make use of AI in machine 6 00:00:28,689 --> 00:00:32,867 learning, how researchers are applying AI and machine learning 7 00:00:32,931 --> 00:00:38,411 in biotech, and the future of AI and machine learning in biotech. 8 00:00:38,603 --> 00:00:42,135 And now, here are Parag, Vijay and Matt. 9 00:00:42,955 --> 00:00:46,395 Hey everyone. Welcome to another episode of Translating Proteomics. 10 00:00:46,515 --> 00:00:49,931 We are really fortunate today to have two celebrities with us. 11 00:00:50,043 --> 00:00:55,163 We have Matt McElwain who is a is managing director at Madrona Venture 12 00:00:55,219 --> 00:00:58,459 Group, where he focuses on helping companies that apply machine 13 00:00:58,507 --> 00:01:02,505 learning and cloud computing to solve problems and create vibrant, 14 00:01:02,585 --> 00:01:06,177 successful businesses. Matt has frequently been recognized as 15 00:01:06,201 --> 00:01:09,777 one of the country's most successful and influential VCs and 16 00:01:09,841 --> 00:01:13,033 recently been immersed in the development of companies exploring 17 00:01:13,089 --> 00:01:17,689 the intersection of life and data science. We are also fortunate 18 00:01:17,737 --> 00:01:22,257 to have Vijay Pandey with us. Vijay was the Henry Dreyfus professor 19 00:01:22,281 --> 00:01:25,281 of Chemistry, Structural Biology and Computer Science at Stanford 20 00:01:25,353 --> 00:01:30,009 University. He's known for his pioneering work on folding at home, 21 00:01:30,137 --> 00:01:33,509 a distributed computing project that simulates protein dynam. 22 00:01:33,697 --> 00:01:37,205 His research focuses on understanding how proteins and other 23 00:01:37,245 --> 00:01:41,093 biomolecules move and interact really at the intersection of life 24 00:01:41,149 --> 00:01:45,469 sciences and computing. And since 2014, Vijay has been a general 25 00:01:45,517 --> 00:01:50,333 partner at Andreessen Horowitz where He leads the a16z bio plus 26 00:01:50,389 --> 00:01:54,705 health fund, which invests in the life sciences and healthcare. 27 00:01:55,245 --> 00:01:59,405 Both Vijay and Matt have worked with many founders using artificial 28 00:01:59,445 --> 00:02:01,997 intelligence and machine learning to tackle problems in the 29 00:02:02,021 --> 00:02:06,001 life sciences. And I'm very excited to chat with them about the 30 00:02:06,033 --> 00:02:10,169 ways that biotech and really the world are changing as a result 31 00:02:10,217 --> 00:02:16,657 of AI. I also mentioned that both that Matt and Vijay are both 32 00:02:16,721 --> 00:02:20,073 members of Nautilus Board of directors and I'm so delighted to 33 00:02:20,089 --> 00:02:22,565 have them here today. Welcome both. 34 00:02:23,345 --> 00:02:24,729 Great to be here. Thank you. 35 00:02:24,897 --> 00:02:25,925 Thanks very much. 36 00:02:26,905 --> 00:02:30,901 Well, I'd love to start off with hearing from both of you just 37 00:02:30,933 --> 00:02:36,253 what your perceptions are, your definitions are of AI and machine 38 00:02:36,309 --> 00:02:39,605 learning. Because when I go out in the world and I ask that question, 39 00:02:39,765 --> 00:02:43,389 I get such an incredibly wide range of answers. So I'm going to 40 00:02:43,397 --> 00:02:46,853 start with you, Matt. Matt, what is artificial intelligence and 41 00:02:46,869 --> 00:02:47,745 machine learning? 42 00:02:48,045 --> 00:02:50,285 Oh, well, it's interesting. I was just at a conference the last 43 00:02:50,325 --> 00:02:54,125 two days and I think so many people are trying to define this, 44 00:02:54,205 --> 00:02:59,183 but for me it's where a machine can help you with either 45 00:02:59,239 --> 00:03:03,495 the predictions or the recommendations or the categorizations 46 00:03:03,655 --> 00:03:06,919 and ultimately the reasoning. And it's in the word reasoning that 47 00:03:06,927 --> 00:03:11,719 you get people to get into big debates about can a machine actually 48 00:03:11,807 --> 00:03:15,175 reason? When do we go from predictions of different forms to 49 00:03:15,215 --> 00:03:19,167 reasoning? And so it's a spectrum of different things that 50 00:03:19,351 --> 00:03:23,111 machines, more specifically the algorithms that are leveraging 51 00:03:23,143 --> 00:03:27,257 the computer and then data that they are, is used to both train 52 00:03:27,321 --> 00:03:31,565 and then for inference is the whole category of artificial intelligence. 53 00:03:32,545 --> 00:03:35,857 Yeah, Vijay, what color would you add to that? 54 00:03:35,961 --> 00:03:38,857 Yeah, I think it's. I'd say something very similar. I think when 55 00:03:38,881 --> 00:03:41,609 you think about machine learning, you think about methods, 56 00:03:41,697 --> 00:03:45,873 classical methods like SVMs or random forest, you know, which, you 57 00:03:45,889 --> 00:03:49,329 know, even linear regression to some degree is a very, very primitive 58 00:03:49,377 --> 00:03:53,603 form of machine learning. It's making a kind of a small model of 59 00:03:53,779 --> 00:03:58,131 something that's predictive behavior. AI is generally associated 60 00:03:58,163 --> 00:04:00,691 with big models, these foundation models, and that gets 61 00:04:00,723 --> 00:04:04,347 the intelligence and reasoning that Matt talked about. And I think 62 00:04:04,371 --> 00:04:09,155 the big shift is that machine learning in our space was many small 63 00:04:09,195 --> 00:04:13,787 models that maybe a company would use. And what is at hand is 64 00:04:13,811 --> 00:04:17,587 the potential to have maybe a few very large models that with things 65 00:04:17,611 --> 00:04:20,891 like transfer learning and other capabilities, can do more than 66 00:04:20,923 --> 00:04:24,055 what, more than the sum of a few models could do. 67 00:04:24,795 --> 00:04:26,923 That's really interesting. We're going to come back to that 68 00:04:27,019 --> 00:04:31,427 in a little bit about how AI and ML have evolved over the last 69 00:04:31,571 --> 00:04:35,819 year. 5 years, 10 years. But maybe let's go all the way back to 70 00:04:35,827 --> 00:04:40,427 the beginning. And I'd love to understand how did each of you get 71 00:04:40,491 --> 00:04:43,891 interested in AI and ML? I'll start with you first for this one, 72 00:04:43,923 --> 00:04:44,771 Vijay. 73 00:04:44,963 --> 00:04:48,227 Yeah, so actually probably when I was in college is when I started. 74 00:04:48,291 --> 00:04:51,187 I think that's when neural nets were starting to get hot. So 75 00:04:51,211 --> 00:04:56,401 this was 1988 to 1992 to date me. And so, but those were like single 76 00:04:56,433 --> 00:04:58,961 layer neural nets. People were starting to play around with that. 77 00:04:58,993 --> 00:05:02,825 A lot of fun stuff done at the Santa Fe Institute. And that was 78 00:05:02,945 --> 00:05:05,945 hot. But it kind of came and went because there are only certain 79 00:05:05,985 --> 00:05:09,017 things that neural nets could do. And I remember, and it seems 80 00:05:09,041 --> 00:05:12,145 almost quaint now that there are these philosophical discussions 81 00:05:12,185 --> 00:05:15,897 of could you can't handle xor, you know, it can't handle something 82 00:05:15,961 --> 00:05:18,409 very nonlinear then like it's never going to work, it's never going 83 00:05:18,417 --> 00:05:22,439 to be useful for anything. And that led to various winters. And 84 00:05:22,447 --> 00:05:25,519 I think a lot of us put that stuff on the shelf for a while. And 85 00:05:25,527 --> 00:05:27,751 then it came back in terms of machine learning and machine learning 86 00:05:27,783 --> 00:05:29,887 was Definitely very interesting. And it was probably 87 00:05:29,951 --> 00:05:35,127 around, I don't know, maybe 2012 or 2013 when deep learning started 88 00:05:35,191 --> 00:05:39,215 appearing. And, you know, it's people like Andrew and Daphne at 89 00:05:39,255 --> 00:05:42,927 Stanford, but also, you know, many others. And we got into that 90 00:05:43,031 --> 00:05:45,567 naturally as well. That's when it started getting interesting that 91 00:05:45,591 --> 00:05:48,887 you could learn features. And for me probably was Andrew's paper 92 00:05:49,031 --> 00:05:52,401 with a Google team and Jeff Dean about learning from YouTube 93 00:05:52,513 --> 00:05:56,017 and that you'd have various neurons that would pop up and there'd 94 00:05:56,041 --> 00:05:58,521 be a neuron that's a human face. And of course, since it's trained 95 00:05:58,553 --> 00:06:02,313 on YouTube, there's a neuron that's a cat and so on. Since so 96 00:06:02,329 --> 00:06:06,049 many cats on YouTube. And that idea that we can go from machine 97 00:06:06,097 --> 00:06:09,977 learning where the human gives the features to deep learning, where 98 00:06:10,001 --> 00:06:13,593 the computer learns the features, that is taking the human 99 00:06:13,649 --> 00:06:15,817 more and more out of the loop. And that's just been the progression 100 00:06:15,841 --> 00:06:18,805 as we've gone along where basically is really learning. 101 00:06:19,125 --> 00:06:22,341 Yeah, yeah, maybe I'll pick up on that. Because, you know, having 102 00:06:22,373 --> 00:06:26,797 been inventor for almost 25 years now, you know, it really was 103 00:06:26,901 --> 00:06:31,685 that period of time in 2012, 2013, you know, you had Imagenet, 104 00:06:31,805 --> 00:06:36,237 you had, you know, you know, increasing digitization of data around 105 00:06:36,301 --> 00:06:39,893 which you could train models on. You had the Netflix competition 106 00:06:40,029 --> 00:06:43,605 that had gone on and people were stepping back and realizing 107 00:06:43,645 --> 00:06:47,471 it first in the research community that this was a time when 108 00:06:47,503 --> 00:06:51,391 there might be another whole opportunity to advance the practical 109 00:06:51,463 --> 00:06:54,711 applications of machine learning and what came to be known 110 00:06:54,743 --> 00:06:57,535 as, you know, kind of deep learning. Right around, right around 111 00:06:57,575 --> 00:07:01,359 that time, I worked with a professor named Carlos Guestrin at 112 00:07:01,367 --> 00:07:04,103 the University of Washington. He's actually now at Stanford. He 113 00:07:04,119 --> 00:07:06,767 was the Amazon professor of machine learning at the time. And 114 00:07:06,791 --> 00:07:10,615 he and his grad students had an open source project. And we saw 115 00:07:10,655 --> 00:07:13,455 the traction that open source project, which was called Graph Lab, 116 00:07:13,495 --> 00:07:17,311 was getting and decided to see if we could build a commercial success. 117 00:07:17,383 --> 00:07:20,503 That company was commercially successful and then went on to be 118 00:07:20,519 --> 00:07:24,303 acquired by Apple. And what's neat is at the time, Apple had five 119 00:07:24,359 --> 00:07:28,015 machine learning services inside an iPhone. This is like 2016 120 00:07:28,175 --> 00:07:31,191 today. There are hundreds of them, all kinds of ways that you 121 00:07:31,223 --> 00:07:32,515 use them within your phones. 122 00:07:34,175 --> 00:07:36,943 That's really interesting. And Matt, one of the things that you 123 00:07:36,999 --> 00:07:42,795 and Vijay both commented on was data and availability of data. 124 00:07:42,915 --> 00:07:47,947 So things like YouTube, suddenly you had millions, billions 125 00:07:48,011 --> 00:07:53,547 of videos you had throughout the Internet, billions of pictures 126 00:07:53,611 --> 00:07:58,395 of cats and dogs that really gave you structured data to learn 127 00:07:58,435 --> 00:08:02,891 upon. And so that was often well annotated. It was controlled 128 00:08:02,923 --> 00:08:06,535 it was organized, it was in a, it was in a commonplace imagenet. 129 00:08:07,245 --> 00:08:12,325 How much do you think our boon in machine learning today has been 130 00:08:12,365 --> 00:08:17,465 driven by those efforts to organize, annotate, coalesce data? 131 00:08:17,885 --> 00:08:21,173 How much do you think it's been driven by advances in compute 132 00:08:21,309 --> 00:08:25,141 that have made it possible to operate on large data? And how much 133 00:08:25,173 --> 00:08:29,069 do you think it's been driven by advances in algorithms that have 134 00:08:29,157 --> 00:08:31,185 changed the way that we interact with that data? 135 00:08:33,015 --> 00:08:36,395 Well, the very unsatisfying answer is yes. 136 00:08:37,855 --> 00:08:39,555 That is very unsatisfying. 137 00:08:39,975 --> 00:08:46,767 But all kidding aside, for me, one of the things that, and I know 138 00:08:46,791 --> 00:08:50,375 we'll get into it, what got me interested in applying some of these 139 00:08:50,415 --> 00:08:54,615 machine learning capabilities as compute advance, as algorithmic 140 00:08:54,655 --> 00:08:59,151 approaches advance, was the fact that the data was becoming so 141 00:08:59,183 --> 00:09:03,831 much higher resolution and lower cost and much more ubiquitous 142 00:09:03,903 --> 00:09:08,367 in the life sciences. And so when you know, you go from it costs 143 00:09:08,391 --> 00:09:14,007 $1 billion to sequence, you know, a genome to $100 and, you know, 144 00:09:14,111 --> 00:09:17,327 fast approaching far less than that, you realize that you've got 145 00:09:17,391 --> 00:09:22,399 data abundance. And with data abundance, you need algorithmic approaches 146 00:09:22,447 --> 00:09:25,807 to get to the needles in the haystack, the insights and the data 147 00:09:25,871 --> 00:09:29,471 abundance. And these were the kind of macro insights that we were 148 00:09:29,503 --> 00:09:34,421 seeing in that 2012, 2013 timeframe. And so I do think it is 149 00:09:34,453 --> 00:09:37,397 the combination of data and then some data sets that you just 150 00:09:37,421 --> 00:09:39,845 couldn't, you know, you didn't have the processing power even back 151 00:09:39,885 --> 00:09:44,037 then. You know, now that, you know, many generations of, you know, 152 00:09:44,181 --> 00:09:47,677 significantly Nvidia chips, but even beyond just the chips themselves. 153 00:09:47,741 --> 00:09:50,501 So think about the brilliance of them buying Mellanox and being 154 00:09:50,533 --> 00:09:53,877 able to have, you know, these, these compute systems that you could 155 00:09:53,901 --> 00:09:58,491 run in parallel at incredible speeds and enable, you know, different 156 00:09:58,523 --> 00:10:01,571 kinds of processing. So unfortunately, I think the answer 157 00:10:01,603 --> 00:10:04,051 is yes, but, but Vijay will undoubtedly have some better insights 158 00:10:04,083 --> 00:10:04,611 than that. 159 00:10:04,723 --> 00:10:07,107 Well, so, you know, all this was happening while my kids were 160 00:10:07,131 --> 00:10:10,347 growing up. So I was watching human learning, you know, at the 161 00:10:10,371 --> 00:10:14,059 same time from, from little kids onward. And like I remember, 162 00:10:14,187 --> 00:10:17,875 you know, we had cats in the house and when they first saw a dog, 163 00:10:17,915 --> 00:10:20,739 they called it a cat. And that's pretty good semi supervised 164 00:10:20,787 --> 00:10:24,355 learning, right? They clustered the animals, the quadrupeds, 165 00:10:24,395 --> 00:10:27,887 and so that's pretty good. And one label in there for cat. And so 166 00:10:27,971 --> 00:10:30,831 reasonable generalizations of cat turns out not to quite to be 167 00:10:30,863 --> 00:10:34,039 true. But like, I think the big question, I think on a lot of 168 00:10:34,047 --> 00:10:37,511 our lines was are we really going to have supervised labels for 169 00:10:37,543 --> 00:10:41,543 enough stuff that we could really learn? And do we really learn 170 00:10:41,599 --> 00:10:45,719 that way. And the self supervised where you can mask out 171 00:10:45,767 --> 00:10:51,231 words or mask out atoms or mask out nucleic acids, that starts 172 00:10:51,263 --> 00:10:53,671 getting really interesting because now you can take all of this 173 00:10:53,703 --> 00:10:57,701 unlabeled data, which we have tons of, whether that be Internet 174 00:10:57,733 --> 00:11:01,885 or DNA sequences, and actually learn on it, predict next word or 175 00:11:01,925 --> 00:11:05,325 next sequence and so on. And that starts get, that's I think really 176 00:11:05,365 --> 00:11:09,301 what was a huge revolution and obviously you need a lot of data 177 00:11:09,333 --> 00:11:12,317 for that. But there's different types of data. Having all 178 00:11:12,341 --> 00:11:15,781 of the supervised learning would be hard. Being able to do it 179 00:11:15,813 --> 00:11:19,005 self supervised I think was a key moment, at least for me. Going 180 00:11:19,045 --> 00:11:21,693 from there's no way this is going to work to like, oh I can, 181 00:11:21,749 --> 00:11:23,585 I can start to see how this is going to work. 182 00:11:24,955 --> 00:11:27,635 That's really interesting. So when you talk about the types of 183 00:11:27,675 --> 00:11:33,227 data and there is a bit of a continuum from fully unstructured 184 00:11:33,291 --> 00:11:40,115 data and all the way to highly structured, annotated, labeled, connected 185 00:11:40,155 --> 00:11:43,619 to an ontology. So you have these sort of two ends of the spectrum. 186 00:11:43,787 --> 00:11:48,571 And what it felt like was there was a time where it was believed 187 00:11:48,603 --> 00:11:52,083 that in order to do effective learning you really needed to be 188 00:11:52,099 --> 00:11:54,723 on that far end of the spectrum. You needed, you needed 189 00:11:54,779 --> 00:11:58,603 pristine data, you needed it to be incredibly well annotated, 190 00:11:58,779 --> 00:12:05,227 structured and again grounded to an ontology. And then it feels 191 00:12:05,251 --> 00:12:09,075 like the world has pushed to the other side. Now that we're able 192 00:12:09,115 --> 00:12:12,163 to use unstructured data, we're able to do unstructured to 193 00:12:12,179 --> 00:12:16,353 structured data conversion, we're able to self supervised to 194 00:12:16,369 --> 00:12:21,281 some extent. And so how have you seen that playing out in the 195 00:12:21,313 --> 00:12:21,885 world? 196 00:12:22,785 --> 00:12:25,929 Yeah, I think some of that might be related to the bitter lesson 197 00:12:25,977 --> 00:12:28,377 too for like if you have enough data and compute, you just 198 00:12:28,401 --> 00:12:31,593 plow through things and that there's just not that much structured 199 00:12:31,649 --> 00:12:34,761 data. And even the variation in structure makes it unstructured 200 00:12:34,873 --> 00:12:37,177 in a sense if everything was structured identically, that'd be 201 00:12:37,201 --> 00:12:41,001 one thing. And so in the end the English language or pick whatever 202 00:12:41,033 --> 00:12:44,033 your favorite language is, but let's say pick English is structured 203 00:12:44,089 --> 00:12:48,729 enough that I, I think once the algorithms got enough to handle 204 00:12:48,777 --> 00:12:52,129 that then you can and transfer that learning onto all the other 205 00:12:52,177 --> 00:12:55,921 domains. So I, you know, it's kind of like there's this history 206 00:12:55,953 --> 00:12:58,617 though. Like you think about computer vision where you like old 207 00:12:58,641 --> 00:13:01,505 school computer vision was very smart and then you just replace 208 00:13:01,545 --> 00:13:05,017 that with CNNs and then visual transformers and so on that I think 209 00:13:05,041 --> 00:13:07,761 that's the progression. But that's I think what we really wanted 210 00:13:07,793 --> 00:13:11,267 when we Thought of AI. I, I mentioned the kid thing. I, I watched 211 00:13:11,291 --> 00:13:14,347 my kids like learn English. I didn't have some structured thing. 212 00:13:14,411 --> 00:13:17,595 I mean, there was school, but it wasn't like, like an ontology 213 00:13:17,675 --> 00:13:20,187 or something like that. And they were just picking things up 214 00:13:20,211 --> 00:13:22,491 and that was a fantasy. And I think the big question in my mind 215 00:13:22,523 --> 00:13:26,195 was, how are we going to teach it to learn from sort of real life? 216 00:13:26,235 --> 00:13:29,255 Because there's plenty of real life and not a lot of structure. 217 00:13:29,835 --> 00:13:32,251 Yeah. Just to pick up one other piece here, it's like to your, 218 00:13:32,283 --> 00:13:36,381 Back to your earlier question, you've got unstructured data that's 219 00:13:36,413 --> 00:13:42,317 in multimode too. You know, image data, video data, audio data, 220 00:13:42,381 --> 00:13:46,573 textual data, et cetera. And a lot of people ask, well, why, why 221 00:13:46,629 --> 00:13:50,149 now? What's so different now? Well, part of it is this algorithmic 222 00:13:50,197 --> 00:13:54,245 approaches to have generative capabilities on that multimodal, 223 00:13:54,285 --> 00:13:58,637 multi structured data. And then on top of that you can do that 224 00:13:58,781 --> 00:14:03,239 through a multiplicity of user interfaces. And so I can have this 225 00:14:03,287 --> 00:14:07,727 generative output and I can summon it with my voice, let alone 226 00:14:07,791 --> 00:14:11,887 a textual prompt or even an image prompt. And so we've really 227 00:14:12,031 --> 00:14:17,015 gotten all to that level and we're not even at the earliest days, 228 00:14:17,095 --> 00:14:21,495 in my opinion of the agentic capabilities on top of that multimodal 229 00:14:21,535 --> 00:14:24,235 set of interfaces. I think that's what's coming next. 230 00:14:25,055 --> 00:14:28,635 So I'd like to push on that a little bit just to understand because 231 00:14:29,065 --> 00:14:33,177 many people say that the transition for the Internet was, 232 00:14:33,321 --> 00:14:36,425 you know, there were web pages, they were in a structured 233 00:14:36,465 --> 00:14:41,241 format, people looked at them in some way, and then they maybe 234 00:14:41,273 --> 00:14:45,497 got to the Links browser, which was a textual interface of 235 00:14:45,521 --> 00:14:51,017 them, and that the transition point came with mosaic 0.1 beta or 236 00:14:51,041 --> 00:14:55,093 whatever, what the very first one was, and that that was the. It 237 00:14:55,109 --> 00:14:59,173 was really an interface. It wasn't as much about the infrastructure, 238 00:14:59,229 --> 00:15:03,205 it was really about the interface that unlocked and transitioned. 239 00:15:03,245 --> 00:15:08,501 And one, an argument has been made that really with the Chat and 240 00:15:08,533 --> 00:15:12,469 Chat, ChatGPT and others, that it was suddenly giving an interface 241 00:15:12,517 --> 00:15:18,525 to the sophisticated AIs. That was the transformer that our poorly 242 00:15:18,565 --> 00:15:22,661 chosen term, that was the thing that drove the transition in 243 00:15:22,693 --> 00:15:26,261 adoption. It wasn't even that, yes, the models had been getting 244 00:15:26,293 --> 00:15:29,281 better and better and getting better, but it was our ability to 245 00:15:29,313 --> 00:15:33,841 interact with them that that really led the transition. How would 246 00:15:33,873 --> 00:15:35,245 you both react to that? 247 00:15:36,465 --> 00:15:39,169 Well, I think that that interface moment was important. I 248 00:15:39,177 --> 00:15:41,857 was actually talking to somebody who was at OpenAI right 249 00:15:41,881 --> 00:15:44,353 around that launch. In fact, he told me he was the first paying 250 00:15:44,409 --> 00:15:47,265 customer when they actually realized that they were onto something 251 00:15:47,305 --> 00:15:51,081 with ChatGPT, which I think completely surprised them, the amount 252 00:15:51,113 --> 00:15:57,515 of adoption that ChatGPT got. But the interface mattered. The interface 253 00:15:57,555 --> 00:16:02,135 also mattered with Mosaic, as you say. But Netscape didn't survive. 254 00:16:02,635 --> 00:16:04,955 And there's a whole bunch of reasons for that. And I think the 255 00:16:04,995 --> 00:16:09,335 interesting question is what are going to be the surviving companies 256 00:16:09,755 --> 00:16:14,467 in this world of generative AI and agentic capabilities? Our view 257 00:16:14,491 --> 00:16:17,811 is that it's actually not going to be at the model layer, but 258 00:16:17,843 --> 00:16:21,283 it is going to be at the application and agentic layer that 259 00:16:21,299 --> 00:16:24,163 the disproportionate percentage of the value is going 260 00:16:24,179 --> 00:16:28,907 to be captured. And while the interfaces are interesting, it's 261 00:16:28,931 --> 00:16:32,523 about actually solving a problem. Now, what's cool about ChatGPT 262 00:16:32,619 --> 00:16:35,043 or perplexity or some of the things that Google's now trying to 263 00:16:35,059 --> 00:16:38,123 embed into their services is they actually solve problems. You 264 00:16:38,139 --> 00:16:41,859 can put a prompt in a text prompt or, you know, increasingly 265 00:16:41,907 --> 00:16:45,323 a voice prompt in a responsive way, and you can get amazingly good 266 00:16:45,379 --> 00:16:49,621 and generally accurate results. And so that's a nice step, 267 00:16:49,653 --> 00:16:51,749 but I think it's a step on the journey that we're going to look 268 00:16:51,797 --> 00:16:54,765 back and we're going to realize that it was some things beyond 269 00:16:54,805 --> 00:16:57,213 that and maybe even some players that aren't even imagined 270 00:16:57,269 --> 00:16:58,665 yet that are the big winners. 271 00:16:59,085 --> 00:17:01,909 Well, I think the thing about the interface was that for people 272 00:17:01,957 --> 00:17:05,277 who are outside of AI that weren't paying attention to what 273 00:17:05,301 --> 00:17:07,877 was happening, I think they were just shocked what was going 274 00:17:07,901 --> 00:17:10,741 on, anything. And even for those of us that had played with 275 00:17:10,773 --> 00:17:15,705 this, GPT 3.5 was a big step over Bird and even a big step over 276 00:17:15,745 --> 00:17:19,369 GPT 2 and probably 3. So like there was some inflection point in 277 00:17:19,377 --> 00:17:23,025 the back end. It's really interesting to try to figure out 278 00:17:23,065 --> 00:17:26,913 what is the right analogy for these large language models. You 279 00:17:26,929 --> 00:17:29,393 know, I think some people talk about them as being analogous to 280 00:17:29,409 --> 00:17:32,193 microprocessors and maybe there'll be many because, you know, 281 00:17:32,209 --> 00:17:36,673 there's intel and Zilog and so on at Motorola. But like another 282 00:17:36,729 --> 00:17:38,897 version would be is. Let's go back in time and think about the 283 00:17:38,921 --> 00:17:43,233 various revolutions we had. You know, mobile, there's two companies 284 00:17:43,289 --> 00:17:46,641 that dominate mobile, like Apple and Google. There's cloud, 285 00:17:46,713 --> 00:17:49,769 there's three companies that dominate cloud. Is this going to 286 00:17:49,777 --> 00:17:52,297 be something like that? Because it takes like mobile and 287 00:17:52,321 --> 00:17:55,665 cloud, it takes a huge amount of capital to build up these infrastructures. 288 00:17:55,745 --> 00:17:58,665 There's only so many companies on the planet that can do that. Maybe 289 00:17:58,745 --> 00:18:03,225 a startup will make it into that sort of high rarefied air. But 290 00:18:03,265 --> 00:18:05,697 like it will because it requires so much capital. It's going 291 00:18:05,721 --> 00:18:11,869 to be the usual suspects of Amazon or Google or Microsoft and 292 00:18:11,917 --> 00:18:15,453 we'll see. I think to Matt's point, the applications is what's 293 00:18:15,509 --> 00:18:18,045 interesting. It's like, you know, there's the iPhone, but then 294 00:18:18,125 --> 00:18:21,677 all the companies, the Airbnbs and the Lyft and Ubers and all that 295 00:18:21,701 --> 00:18:24,981 stuff, that's where a lot of money is made. And we're not even 296 00:18:25,093 --> 00:18:28,701 starting that really yet. I think that's just starting. Another 297 00:18:28,733 --> 00:18:31,565 way I think about it is like if we think AI is going to replace 298 00:18:31,645 --> 00:18:36,163 jobs, chatgpt for 20 bucks a month is like crazy cheap and it 299 00:18:36,179 --> 00:18:38,811 doesn't make sense economically. And I think that reflects 300 00:18:38,843 --> 00:18:42,335 the gap between where we are now and what we'd like to do. 301 00:18:43,035 --> 00:18:46,843 Yeah, that makes sense. I'd love to transition now to chat a 302 00:18:46,859 --> 00:18:52,055 little bit about biology and hear your perspectives about how. 303 00:18:52,755 --> 00:19:01,465 Let's start where we are today and how has AIML leading up to today 304 00:19:02,125 --> 00:19:07,225 changed our understanding of proteins and biology in general? 305 00:19:08,965 --> 00:19:12,613 Yeah, I think, you know, this has obviously been a evolution and 306 00:19:12,629 --> 00:19:14,853 a revolution in different ways. I think the big thing about 307 00:19:14,909 --> 00:19:18,125 biology in the last few decades is we finally have tons of 308 00:19:18,165 --> 00:19:21,549 data. You know, starting with genomics in the 90s and obviously 309 00:19:21,637 --> 00:19:26,387 proteomics today and all of the other omics in between. That 310 00:19:26,451 --> 00:19:30,043 large, large scale data speaks exactly to what we were talking about 311 00:19:30,099 --> 00:19:34,115 for AI, that you need to have this data and this stuff, you know, 312 00:19:34,195 --> 00:19:37,923 it's not on the Internet, it has to be measured and so on. So 313 00:19:37,979 --> 00:19:42,203 that's a huge thing, that it's interesting that all of this robotics 314 00:19:42,259 --> 00:19:45,027 and all this new technology and CRISPR that can be applied to 315 00:19:45,051 --> 00:19:49,635 this are coming right smack dab with the advances in AI and like 316 00:19:49,715 --> 00:19:53,595 AI needs the data and the data needs the AI because you can't understand 317 00:19:53,675 --> 00:19:56,367 just all the raw data. So I think it's kind of beautiful that 318 00:19:56,391 --> 00:19:59,023 they're coming together and you could imagine that they're kind 319 00:19:59,039 --> 00:20:01,303 of pushing each other because those who are interested in AI are 320 00:20:01,319 --> 00:20:04,007 going to push for more robotics and those that are doing 321 00:20:04,031 --> 00:20:06,223 all the robotics data will push for how we're going to understand 322 00:20:06,279 --> 00:20:09,215 it. I think that's the beautiful moment we're in today. 323 00:20:09,255 --> 00:20:12,103 It will be interesting to see if one gets more of a lead than the 324 00:20:12,119 --> 00:20:14,463 other, but I suspect they'll stay pretty close. 325 00:20:14,639 --> 00:20:17,343 Yeah, I agree. I mean, you know, the wet lab is always going 326 00:20:17,359 --> 00:20:21,319 to be the wet lab, but the degrees of Automation and the scale 327 00:20:21,367 --> 00:20:25,447 at which you can, you can, you know, test out small molecules or 328 00:20:25,471 --> 00:20:28,383 you can test out protein to protein interactions, you know, in 329 00:20:28,399 --> 00:20:31,431 your wet lab and then take that data and feed it into the models 330 00:20:31,463 --> 00:20:35,223 that you're training and do that in an iterative way. I think 331 00:20:35,239 --> 00:20:38,503 that the automation part's a little bit underappreciated. But 332 00:20:38,519 --> 00:20:41,775 the other thing is that I like this notion that Vijay's bringing 333 00:20:41,815 --> 00:20:45,279 up of kind of how one thing pushes the other. And I'll tell a 334 00:20:45,287 --> 00:20:49,591 little story, you know, of course, you know, now lots of people 335 00:20:49,623 --> 00:20:53,483 are aware of David Baker in the Institute of Protein Design and 336 00:20:53,499 --> 00:20:56,771 I remember six years ago being in his. It's about six years ago 337 00:20:56,883 --> 00:21:00,259 that after winning the protein folding prediction competition for 338 00:21:00,307 --> 00:21:03,771 years, you know, he and his team lost to Demis and the team at 339 00:21:03,803 --> 00:21:08,275 open at DeepMind, you know, Google. And the great thing about 340 00:21:08,315 --> 00:21:11,227 David Baker was he's like, well, it's time for my teams to really 341 00:21:11,251 --> 00:21:15,139 learn deep learning. And now the push and pull back and forth 342 00:21:15,187 --> 00:21:19,565 from Rosettafold and Alphafold and the various generations of those 343 00:21:19,715 --> 00:21:24,169 has made the work better and the capabilities better, moving beyond 344 00:21:24,217 --> 00:21:28,129 even protein predictions. And I think it's why they deservedly 345 00:21:28,177 --> 00:21:31,777 won the Nobel Prize. Interestingly that they won it in 346 00:21:31,841 --> 00:21:35,153 chemistry. Right. There's no such Nobel Prize. 347 00:21:35,209 --> 00:21:37,369 Chemistry is natural for that though, right? I mean, what else? 348 00:21:37,457 --> 00:21:40,977 I think it's the right choice, but I think that the degree to which 349 00:21:41,081 --> 00:21:46,801 their chemistry research and insights and capabilities were dependent 350 00:21:46,833 --> 00:21:52,397 upon machine and deep learning and then actually, you know, it was 351 00:21:52,421 --> 00:21:56,545 what was remarkable to me and. But appropriately so. Yes. 352 00:21:57,525 --> 00:22:01,261 So, so Vijay, your, your own background and mine is as well was 353 00:22:01,333 --> 00:22:05,557 very much in physics based models of folk. 354 00:22:05,701 --> 00:22:06,165 Yeah. 355 00:22:06,245 --> 00:22:14,215 And so the, I'm curious, your perception of this jump from physics 356 00:22:14,295 --> 00:22:21,303 based potentials to statistical potentials to now AI 357 00:22:21,399 --> 00:22:25,951 based weights and models. 358 00:22:26,063 --> 00:22:29,935 Yeah. So it's kind of fun because also like we'll talk about 359 00:22:29,975 --> 00:22:32,335 going full circle because there's physics in all of this stuff 360 00:22:32,415 --> 00:22:37,519 and I think the rationale for doing physics of proteins is that 361 00:22:37,607 --> 00:22:39,945 the structure should be a starting point, that these things 362 00:22:39,985 --> 00:22:42,729 have conformational change. It could be small side chains or it 363 00:22:42,737 --> 00:22:45,865 could be large allosteric changes. And so that's going to be 364 00:22:45,905 --> 00:22:49,081 key. And actually I think nothing has brought that up more 365 00:22:49,113 --> 00:22:52,353 to the forefront than having all these structures predicted. And 366 00:22:52,369 --> 00:22:54,929 these structures are predicting the experiments that are 367 00:22:55,017 --> 00:22:58,793 available, which is X ray and cryo m and so on. But actually the 368 00:22:58,809 --> 00:23:02,145 other parts are some ways the really hard parts. Like how do you 369 00:23:02,185 --> 00:23:04,385 understand whether How a protein will change when the drug 370 00:23:04,425 --> 00:23:07,061 goes there and you have allosteric conformational change. 371 00:23:07,193 --> 00:23:10,533 So actually we're starting to see actually physics come back in, 372 00:23:10,589 --> 00:23:13,581 in new types of foundation models for dynamics. And actually 373 00:23:13,613 --> 00:23:16,573 it's very much in line with what we see for in other areas of 374 00:23:16,589 --> 00:23:21,381 an AI. Like self driving cars can be trained on simulations. Which 375 00:23:21,453 --> 00:23:24,101 sounds almost like you've got like computers playing Grand Theft 376 00:23:24,133 --> 00:23:26,629 Auto or something like that. But like they learn from that and 377 00:23:26,637 --> 00:23:30,045 then transfer learn onto real cars. You can imagine doing that 378 00:23:30,165 --> 00:23:32,925 or, and people are very much thinking about doing that on the 379 00:23:33,005 --> 00:23:37,917 protein side. But the fun, fun part is that you know, the Hamiltonian 380 00:23:37,941 --> 00:23:40,877 I've dealt with most at the spin glass Hamiltonian, it's used 381 00:23:40,901 --> 00:23:43,829 for the thermodynamics of proteins, it's used for many things. 382 00:23:43,877 --> 00:23:46,973 It's also exactly the same Hamiltonian from neural networks. 383 00:23:47,109 --> 00:23:51,925 You know, there's physics to understand, deep learning and AI. 384 00:23:52,045 --> 00:23:55,261 And I think the irony is that now we're seeing this new area where 385 00:23:55,293 --> 00:23:59,205 computer scientists are dealing with a physical object. An 386 00:23:59,245 --> 00:24:03,757 LLM and a transformer is a simulation of data. It's a stochastic 387 00:24:03,781 --> 00:24:07,021 simulation. And this is why it's somewhat alien to computer scientists 388 00:24:07,053 --> 00:24:10,021 and why they say oh like we can't understand it. But actually 389 00:24:10,093 --> 00:24:13,693 it's basically a physical process and can be understood very 390 00:24:13,709 --> 00:24:16,061 much through statistical mechanics. And many people are doing 391 00:24:16,093 --> 00:24:19,237 that. And so again physics has come back and so therefore I don't 392 00:24:19,261 --> 00:24:22,765 think it's a shock that a lot of the leaders in AI actually have 393 00:24:22,805 --> 00:24:24,625 come with theoretical physics backgrounds. 394 00:24:25,925 --> 00:24:29,171 Yeah, that's really interesting. And I think this, this 395 00:24:29,323 --> 00:24:35,451 harmonious cycle as you described, between grounded physically 396 00:24:35,523 --> 00:24:40,187 motivated entities, models that are built in a wide variety 397 00:24:40,251 --> 00:24:45,339 of ways and experiment and a loop between them. I think there 398 00:24:45,347 --> 00:24:48,227 was a, there was a time where it was believed very much again that 399 00:24:48,251 --> 00:24:52,955 you had AI off in one world and you had science off in another 400 00:24:53,035 --> 00:24:59,675 world. But it feels like there is, there is, we're moving to a place 401 00:24:59,715 --> 00:25:05,355 it's still hard, where we're taking advantage of knowledge to 402 00:25:05,435 --> 00:25:09,115 train better models and to train models with lower amounts of 403 00:25:09,155 --> 00:25:14,243 data. And accordingly those models and the understanding of them 404 00:25:14,379 --> 00:25:17,995 are helping us feel out where the boundaries of knowledge are and 405 00:25:18,035 --> 00:25:21,427 maybe coming back again and informing our understanding of the 406 00:25:21,451 --> 00:25:22,415 natural world. 407 00:25:23,635 --> 00:25:26,531 I think that's right. And I think the magic that we have now, 408 00:25:26,563 --> 00:25:30,267 that we didn't have before is with large models you can have transfer 409 00:25:30,331 --> 00:25:33,763 learning where possibly you could do zero shot learning, where 410 00:25:33,779 --> 00:25:37,827 it just knows the answer for something new or few Shot learning, 411 00:25:37,851 --> 00:25:40,603 where you put a, do a few experiments and you can understand 412 00:25:40,659 --> 00:25:43,107 things. I think the big problem for machine learning, let's 413 00:25:43,131 --> 00:25:47,307 say in drug design, like 20 years ago when my team was starting 414 00:25:47,331 --> 00:25:51,143 doing that at Stanford, you know, we would need maybe like in 415 00:25:51,159 --> 00:25:54,527 the beginning 50 actives or something like that to train. And 416 00:25:54,551 --> 00:25:58,063 in any program, if you have 50 actives, you don't need machine learning. 417 00:25:58,119 --> 00:26:01,255 You're basically done. And so getting to the point where you could 418 00:26:01,295 --> 00:26:04,775 do few shot or zero shot learning, which you could even do 419 00:26:04,815 --> 00:26:07,527 before deep learning, but you know, getting to that point was really 420 00:26:07,551 --> 00:26:10,327 the key point. And that's what's available now. And that's 421 00:26:10,351 --> 00:26:12,135 why I think it's really interesting is that there's so much 422 00:26:12,175 --> 00:26:15,871 knowledge in the model that only a small bit of additional experiments 423 00:26:15,903 --> 00:26:16,995 is often needed. 424 00:26:17,385 --> 00:26:20,001 I think there's two other things that are quickly coming down 425 00:26:20,033 --> 00:26:24,393 the path, which is I can take some of these frontier larger models 426 00:26:24,569 --> 00:26:28,465 and then I can take my own data set, my domain specific data 427 00:26:28,505 --> 00:26:32,073 set could be a proteomics data set, for example, and I can train 428 00:26:32,129 --> 00:26:36,297 a smaller model and I can leverage that in more efficient ways 429 00:26:36,401 --> 00:26:39,161 for the domain specific problems that I'm trying to solve. 430 00:26:39,313 --> 00:26:43,201 And then the other area that is we're seeing happen quickly is, 431 00:26:43,273 --> 00:26:49,081 and you certainly, if folks have played with ChatGPT's 4.0 version 432 00:26:49,113 --> 00:26:52,905 of their model, is this concept of chain of thought reasoning. 433 00:26:53,065 --> 00:26:55,329 You could think of it conceptually like a decision tree, 434 00:26:55,377 --> 00:27:00,925 right, Where I can actually go out to different kind of tree paths 435 00:27:01,825 --> 00:27:05,265 within the models and maybe even a compilation of models and 436 00:27:05,305 --> 00:27:08,553 come back with different answers and have something that we 437 00:27:08,569 --> 00:27:11,663 might think of as a router that ranks those answers and actually 438 00:27:11,719 --> 00:27:16,335 gives me back the one that's the best fit answer. And you know, 439 00:27:16,375 --> 00:27:18,799 that's I think, a little bit of a colloquial way to explain it. 440 00:27:18,887 --> 00:27:21,903 But this idea of chain of thought reasoning, which is by the 441 00:27:21,919 --> 00:27:24,615 way, way more computationally intensive. So again, back to your 442 00:27:24,655 --> 00:27:27,327 questions. It is a yes, because without that computational 443 00:27:27,391 --> 00:27:31,039 power, these kinds of models wouldn't actually function today. 444 00:27:31,167 --> 00:27:35,695 So that combination of sort of model distillation and domain specific 445 00:27:35,735 --> 00:27:38,527 models plus chain of thought reasoning, again, I think are leading 446 00:27:38,551 --> 00:27:40,247 us to what's next to be possible. 447 00:27:40,391 --> 00:27:43,583 Well, I think Matt's point brings up one thing in my head, which 448 00:27:43,599 --> 00:27:45,863 is that I think the thing in computer science, which is many of 449 00:27:45,879 --> 00:27:48,767 our favorites, is the composability, how you can take parts 450 00:27:48,791 --> 00:27:52,871 and build on top. So like you could have autoencoders and decoders, 451 00:27:52,903 --> 00:27:55,167 you put a bunch of them together, you get a transformer. 452 00:27:55,271 --> 00:27:57,599 Now that you have the transformer, you can come up with 453 00:27:57,647 --> 00:28:00,247 answers. But now you can do something on top of that with chain 454 00:28:00,271 --> 00:28:02,983 of thought and you can just see the layers and layers and layers 455 00:28:03,039 --> 00:28:06,043 of building, which is like anything else that people have done 456 00:28:06,059 --> 00:28:09,491 in computer science. Further, further layers of abstraction going 457 00:28:09,563 --> 00:28:13,003 higher and higher up. That's the thing that's just amazing to 458 00:28:13,019 --> 00:28:16,971 imagine, like where how much more you could get done in like even 459 00:28:17,043 --> 00:28:20,619 2 years, 5 years, 10 years. You can just imagine this going on 460 00:28:20,667 --> 00:28:22,215 for actually quite a while. 461 00:28:22,795 --> 00:28:25,707 So I'd like to come back to one point that we've been talking 462 00:28:25,731 --> 00:28:30,187 about a little bit. But the volume of data available for the 463 00:28:30,211 --> 00:28:34,153 protein structure predictions was large. But it's not Internet 464 00:28:34,209 --> 00:28:40,017 large. It's, you know, tens of thousands of observations maybe on 465 00:28:40,041 --> 00:28:46,657 that, on that scale. But we had encodings of that data that in 466 00:28:46,681 --> 00:28:50,617 terms of pairwise distances and things like that, that were very 467 00:28:50,681 --> 00:28:56,905 natural to convert into a machine learning framework for other 468 00:28:56,985 --> 00:29:02,357 domains of biology. Where do you think the, you know, A, do you 469 00:29:02,381 --> 00:29:08,453 think we need to necessarily have what volume of data and B, what 470 00:29:08,509 --> 00:29:11,425 encodings have we have we figured out? 471 00:29:12,565 --> 00:29:16,381 You know, representation is always tricky, but I think as we 472 00:29:16,413 --> 00:29:19,813 talked about, we can go less and less structured. I think you 473 00:29:19,829 --> 00:29:23,805 can imagine almost like whatever human language equivalent 474 00:29:23,925 --> 00:29:27,157 is used to disc discuss it, that could be used almost like the 475 00:29:27,181 --> 00:29:30,501 LLM at the minimum could be user interface between the way we 476 00:29:30,533 --> 00:29:33,637 think about it and the way it gets encoded, such that maybe we 477 00:29:33,661 --> 00:29:37,365 don't have to be so clever about how we featurize these things 478 00:29:37,405 --> 00:29:42,085 that that can get taken care of. You know, I think in terms of 479 00:29:42,125 --> 00:29:45,469 the all the variety of data, that's where it gets really interesting, 480 00:29:45,517 --> 00:29:48,605 right, because you can have massive, massive multimodal data 481 00:29:48,645 --> 00:29:51,661 sets. And the question is, is there interesting transfer learning? 482 00:29:51,693 --> 00:29:55,007 And there's fun stuff you see with like inpainting of cells based 483 00:29:55,031 --> 00:29:59,711 on genomic data and just combining imaging and genomics and 484 00:29:59,863 --> 00:30:03,767 proteomics and all this stuff, you start to have all the readouts 485 00:30:03,791 --> 00:30:07,895 of a cell. If you can start to predict the future of a cell from 486 00:30:07,935 --> 00:30:11,595 its sort of current state. Now it's basically a cell simulator. 487 00:30:14,775 --> 00:30:18,463 Yeah, I mean, I think that the other thing that you're seeing happen 488 00:30:18,599 --> 00:30:21,407 is all throughout the supply chain of the data that I'm dealing 489 00:30:21,431 --> 00:30:26,707 with and the multimodal data types, how do I create kind of annotation, 490 00:30:26,771 --> 00:30:32,495 how do I enable encoding in ways that is a lot more efficient 491 00:30:32,915 --> 00:30:36,803 and a lot more automated than it's been with companies like Scale 492 00:30:36,859 --> 00:30:39,907 AI and others that have had to do it in a far more manual way. So 493 00:30:40,011 --> 00:30:43,295 again, I think that it's hard to look at just this point in time. 494 00:30:44,675 --> 00:30:50,087 Another data point here is just 18 months ago we were at like, 495 00:30:50,151 --> 00:30:55,415 you know, $36 per million tokens processed and then three months 496 00:30:55,455 --> 00:31:00,351 ago we were at $4 per millions token process and we're rapidly coming 497 00:31:00,383 --> 00:31:04,119 down to $1 per million tokens process and in some cases even better 498 00:31:04,167 --> 00:31:08,023 than that. And so you just keep having these, you know, you 499 00:31:08,039 --> 00:31:11,127 know, kind of exponentially declining, in this case cost curves 500 00:31:11,231 --> 00:31:14,685 that support our ability to do everything from, you know, kind of 501 00:31:14,775 --> 00:31:19,305 data prep and data annotation to, you know, the compute and then 502 00:31:19,345 --> 00:31:21,921 ultimately unlocking the value from inference perspective. 503 00:31:22,033 --> 00:31:24,521 Well, and the fun thing is like that's beyond just the Moore's 504 00:31:24,553 --> 00:31:30,481 law of Moore transistors. That's specific hardware for inference. 505 00:31:30,593 --> 00:31:34,065 That's new tricks of learning smaller models from larger models. 506 00:31:34,225 --> 00:31:38,845 And once we're doing inference on like a billion person world scale, 507 00:31:39,145 --> 00:31:42,451 all this stuff will be so critical that I think we have even 508 00:31:42,523 --> 00:31:46,531 possibly just started the beginning of what could be optimized. 509 00:31:46,643 --> 00:31:49,427 And that's going to be fun because all that optimization for 510 00:31:49,451 --> 00:31:52,855 other things can be put directly into what we do in biology. 511 00:31:54,435 --> 00:31:59,379 Interesting question. So biology data is different. It's in 512 00:31:59,507 --> 00:32:03,883 scale, it's different in resolution, it's different in timescale 513 00:32:03,939 --> 00:32:10,027 from things like Wikipedia and English language data. From both 514 00:32:10,051 --> 00:32:13,795 of your perspectives, what is it about biology that maybe makes 515 00:32:13,835 --> 00:32:21,003 it different than other domains, financial data, text, textual 516 00:32:21,059 --> 00:32:29,051 data, and makes it either both harder or easier for AI? 517 00:32:29,163 --> 00:32:31,015 I think it's definitively harder. 518 00:32:32,755 --> 00:32:34,415 Why do you think it's harder, Matt? 519 00:32:35,075 --> 00:32:39,779 What I think is harder is that, you know, if the ultimate goal, 520 00:32:39,867 --> 00:32:44,979 right, of trying to deeply understand biological data is to 521 00:32:45,027 --> 00:32:51,131 help us live longer, healthier lives, then the fact that there's 522 00:32:51,163 --> 00:32:56,051 a fair bit of non determinism, even though there's a lot of determinism, 523 00:32:56,123 --> 00:32:59,507 there's a fair bit of non determinism and it's sometimes it's 524 00:32:59,531 --> 00:33:02,643 idiosyncratic. Why does something mutate or not mutate? You 525 00:33:02,659 --> 00:33:07,541 know, why, why does, you know, this cancer, you know, express these 526 00:33:07,573 --> 00:33:10,733 proteins and not those proteins? And how does, and how does 527 00:33:10,749 --> 00:33:16,185 it evolve over time? And I think we don't sufficiently understand 528 00:33:16,485 --> 00:33:20,285 the unpredictability of biology and I think that makes it 529 00:33:20,365 --> 00:33:23,773 harder to work with the data sets. But that's. 530 00:33:23,909 --> 00:33:27,613 Yeah, I think that's right. I think another way to phrase the same 531 00:33:27,709 --> 00:33:31,877 concept is like that a Lot of the current generative tasks are 532 00:33:31,901 --> 00:33:35,687 very robust to mistakes. Like, if the picture's a little off, even 533 00:33:35,711 --> 00:33:38,695 if I end up having four fingers instead of five, like, we're 534 00:33:38,735 --> 00:33:42,551 kind of okay with that, you know, and life sciences and healthcare. 535 00:33:42,583 --> 00:33:45,415 If I end up with four fingers, I'm gonna be really pissed. You know, 536 00:33:45,455 --> 00:33:48,591 like, it's something where. Or you put the methyl group in the wrong 537 00:33:48,623 --> 00:33:51,479 place and no longer inactive. You know, there's these activity 538 00:33:51,527 --> 00:33:55,415 cliffs. And so that's something where now we're in a space 539 00:33:55,455 --> 00:33:59,551 where small things can have big changes. And those activity cliffs 540 00:33:59,583 --> 00:34:02,817 have always haunted computational drug designers because, 541 00:34:02,881 --> 00:34:07,089 like, there's. It's not a very simple surface, the latent space. 542 00:34:07,257 --> 00:34:09,473 But what's interesting is, I think with enough data, you start 543 00:34:09,489 --> 00:34:12,089 to understand the latent space, especially if you have a 3D 544 00:34:12,137 --> 00:34:14,873 representation of the protein, because a lot of that actually is 545 00:34:14,889 --> 00:34:17,953 much more obvious, whether something fits or not or how it interacts 546 00:34:17,969 --> 00:34:22,745 or so on. So I think the complexity is so much greater than, 547 00:34:22,865 --> 00:34:27,025 say, just an LLM reproducing text and just predicting the next 548 00:34:27,065 --> 00:34:27,765 token. 549 00:34:28,775 --> 00:34:32,159 Yeah, I mean, I think that the two other things I would add to this 550 00:34:32,207 --> 00:34:35,839 is, you know, we talk a lot about models, in this case, physical 551 00:34:35,887 --> 00:34:39,151 models in the biology world. You know, a mouse model as an example. 552 00:34:39,223 --> 00:34:42,079 Right. And there's the, you know, sort of the euphemism of, we 553 00:34:42,127 --> 00:34:44,887 cured a lot of mice of certain kinds of cancer, but that doesn't 554 00:34:44,911 --> 00:34:49,151 then translate into the humans as a treatment strategy. I think 555 00:34:49,183 --> 00:34:51,911 another way to look at the data is to say, well, gosh, as we 556 00:34:51,943 --> 00:34:55,631 get to deeper resolutions of the data, what more can we learn, 557 00:34:55,663 --> 00:34:59,399 to Vijay's point, about its complexity and, you know, proteins 558 00:34:59,447 --> 00:35:02,623 is a good example, again, because it's not just about the proteins. 559 00:35:02,679 --> 00:35:06,767 It's about the proteo forms. And so the more we can unlock the 560 00:35:06,791 --> 00:35:11,007 power of data and compute and algorithmic approaches, we can start 561 00:35:11,031 --> 00:35:14,887 to appreciate the role that proteoforms play in different kinds 562 00:35:14,911 --> 00:35:15,995 of disease states. 563 00:35:16,455 --> 00:35:19,527 Yeah. I mean, just to weigh in with my own perspective, I agree 564 00:35:19,551 --> 00:35:23,581 with you, Matt. I think biology is harder. And I think the 565 00:35:23,613 --> 00:35:28,677 point that you raised Vijay about the bar for good enough, that's 566 00:35:28,701 --> 00:35:33,245 one of the major reasons why biology is harder. That if we're 567 00:35:33,285 --> 00:35:39,837 making a diagnosis that really matters, if we're trying to decide, 568 00:35:39,901 --> 00:35:44,225 hey, you should take this drug or that drug, and we guess wrong. 569 00:35:44,565 --> 00:35:51,047 And we guess wrong confidently, which is many new emerging 570 00:35:51,111 --> 00:35:55,463 tools, they fail incredibly confidently. And what we really want 571 00:35:55,479 --> 00:35:57,527 them to do is we want them to throw up their hands and say, sorry, 572 00:35:57,591 --> 00:36:02,775 don't know, not giving you an answer. And that's challenge. And 573 00:36:02,775 --> 00:36:07,007 then the third reason I think it's harder. I have four. The third 574 00:36:07,191 --> 00:36:13,503 is the scales of data that we're trying to integrate from microsecond 575 00:36:13,559 --> 00:36:17,639 timescales, operating on multiple phosphorylations on a single 576 00:36:17,687 --> 00:36:22,689 protein molecule up to a metabolite which is present for again 577 00:36:22,737 --> 00:36:26,041 a microsecond or moving from one place to another all the way 578 00:36:26,073 --> 00:36:29,737 up to the fact that I did something to myself when I was 10 579 00:36:29,921 --> 00:36:36,121 and I'm paying for it when I'm 65. So you have this convolution 580 00:36:36,193 --> 00:36:39,873 of things at very, very short time scales and very, very long time 581 00:36:39,929 --> 00:36:44,273 scales that we need to consider and we just simply don't 582 00:36:44,289 --> 00:36:46,681 have data connecting all of those timescales. 583 00:36:46,873 --> 00:36:50,233 And I think I very much agree with you on the life science side, 584 00:36:50,369 --> 00:36:54,633 there's one bit of optimism on the healthcare side, which is that 585 00:36:54,809 --> 00:36:58,145 for better or worse, the job of a doctor is really, really hard 586 00:36:58,265 --> 00:37:01,721 to take in all of this data and make a diagnosis. And as there's 587 00:37:01,753 --> 00:37:05,889 more data to ascertain, it becomes more of a data science problem 588 00:37:06,017 --> 00:37:09,625 than a like, I'm going to use my gut feeling and you know, it's 589 00:37:09,665 --> 00:37:13,977 kind of like using TurboTax versus doing all my taxes on with 590 00:37:14,001 --> 00:37:17,083 a pencil and paper. I'm going to make mistakes. The computer is 591 00:37:17,099 --> 00:37:20,475 going to be really helpful. And that bar for medicines, I think 592 00:37:20,515 --> 00:37:23,211 really intriguing for things like diagnosis and so on, because 593 00:37:23,243 --> 00:37:26,579 that stuff is going to get really hard. And actually, you know, 594 00:37:26,667 --> 00:37:30,467 as you know, we can also put other layers on top of LLMs and have 595 00:37:30,491 --> 00:37:33,659 mixtures of experts such that I think we can deal with hallucinations 596 00:37:33,747 --> 00:37:37,491 in a variety of ways even with low latency. And I think we're seeing 597 00:37:37,523 --> 00:37:40,851 that now roll out. It's maybe a little less elegant because we 598 00:37:40,883 --> 00:37:44,155 have to have classifiers on top of everything else. But maybe 599 00:37:44,195 --> 00:37:46,835 it's also not unlike the thinking fast and slow, the type 600 00:37:46,835 --> 00:37:50,447 of one and type two systems. And so if you think about it, our 601 00:37:50,471 --> 00:37:53,079 brains evolved and it might not be the most elegant solution 602 00:37:53,127 --> 00:37:57,275 either, but it works. And maybe we can take that approach here. 603 00:37:58,015 --> 00:38:03,015 Yeah, and I think we're also seeing a journey where the data science 604 00:38:03,095 --> 00:38:06,439 capabilities on a set of patients that we can use and analyze 605 00:38:06,487 --> 00:38:09,959 can also, for at least some set of patients, you know, using 606 00:38:10,047 --> 00:38:13,257 more precise or precision medicine approaches, actually said, 607 00:38:13,281 --> 00:38:15,881 well, this actually should be the first line treatment versus the 608 00:38:15,913 --> 00:38:18,297 second or the third line treatment and you're starting to 609 00:38:18,321 --> 00:38:21,249 see that. For example, I was just briefed the other day on some 610 00:38:21,297 --> 00:38:25,657 progress in immunotherapies, blood cancer, you know, related immunotherapies, 611 00:38:25,721 --> 00:38:29,089 where the immunotherapies are moving up to second and even first 612 00:38:29,137 --> 00:38:30,805 line, you know, strategies. 613 00:38:31,225 --> 00:38:34,529 Well, and imagine a world where we could measure enough to 614 00:38:34,577 --> 00:38:39,819 have a real readout. So that might be your genomic baseline, obviously, 615 00:38:39,867 --> 00:38:42,371 maybe proteomics and other things and metabolomics that you 616 00:38:42,403 --> 00:38:46,531 measure and then you have that over time. And AI is very good at. 617 00:38:46,563 --> 00:38:49,675 You can mask out from data sets so you can predict the future. 618 00:38:49,835 --> 00:38:52,915 And so let's say we could say like Vijay, if your behavior, your 619 00:38:52,955 --> 00:38:56,971 diet and exercise doesn't change, this is your future and lay 620 00:38:57,003 --> 00:38:59,595 that out for you. And basically over time, you watch that 621 00:38:59,635 --> 00:39:02,835 future come to pass. I think it could be one of the most dramatic 622 00:39:02,875 --> 00:39:06,993 things to get people to change their behavior because we have the 623 00:39:07,009 --> 00:39:10,137 ability to look into the crystal ball of the future. And I 624 00:39:10,161 --> 00:39:12,729 think that could be great or it could be quite scary, but. And 625 00:39:12,777 --> 00:39:15,977 it could be very actionable if we do those things early. We just 626 00:39:16,001 --> 00:39:18,929 need the data. Actually, the AI part, ironically, for some of 627 00:39:18,937 --> 00:39:21,645 this stuff may be straightforward. The data isn't there. 628 00:39:22,705 --> 00:39:25,425 That's interesting. So I think what I just heard you say is that 629 00:39:25,505 --> 00:39:28,745 people, instead of investing in AI companies, they should actually 630 00:39:28,825 --> 00:39:34,217 invest in healthcare fitness clubs because the AI is going to 631 00:39:34,241 --> 00:39:36,821 tell people to go to the gym more and people are actually, actually 632 00:39:36,853 --> 00:39:37,733 going to listen. 633 00:39:37,869 --> 00:39:41,345 Why not both? Like you can always do both. Yeah, okay. Yeah. 634 00:39:41,805 --> 00:39:47,565 So I'm curious, just in both of your seats as investors, how are 635 00:39:47,605 --> 00:39:52,973 you seeing founders and new companies in biotech incorporating 636 00:39:53,029 --> 00:39:54,785 AI into their companies? 637 00:39:56,525 --> 00:40:00,533 Yeah, I can jump in. Like, I think there's a spectrum. Right. 638 00:40:00,589 --> 00:40:04,917 So I think right now we see machine learning literally in everything. 639 00:40:05,021 --> 00:40:09,069 So there's like, that's like table stakes. It might be simple 640 00:40:09,117 --> 00:40:12,901 machine learning, but I don't think there's, there's not a biotech 641 00:40:12,933 --> 00:40:15,349 that doesn't have some bioinformatics or some machine learning 642 00:40:15,397 --> 00:40:18,917 in there. But then I think then there's people that are sort 643 00:40:18,941 --> 00:40:21,757 of using more advanced technologies and then there are the 644 00:40:21,781 --> 00:40:24,717 people on the bleeding edge and those are just different types 645 00:40:24,741 --> 00:40:27,941 of companies. And so for the bioinformatics that's become such 646 00:40:27,973 --> 00:40:31,465 table stakes for just doing interesting things, that that's there. 647 00:40:32,005 --> 00:40:34,685 You know, naturally, with my background, I get very excited about 648 00:40:34,725 --> 00:40:37,493 sort of the pushing the envelope and that the pushing the 649 00:40:37,509 --> 00:40:41,053 envelope is in all spectrums, all stages of designing a drug, whether 650 00:40:41,109 --> 00:40:45,701 we're finding new Targets hitting those targets in novel ways. 651 00:40:45,813 --> 00:40:48,837 And then even now we're starting to see AI for clinical trials 652 00:40:48,861 --> 00:40:50,853 and so on. And that's, I think, been the holy grail for a 653 00:40:50,869 --> 00:40:55,237 while. The whole part of this is a data exercise. And if you have 654 00:40:55,261 --> 00:40:59,453 a latent space model that can describe human disease biology, you 655 00:40:59,469 --> 00:41:02,353 could imagine using that all the way through. Even from target 656 00:41:02,409 --> 00:41:05,025 identification, through trials, and even to personalized 657 00:41:05,065 --> 00:41:08,121 medicine and precision medicine on the other side. I mean, 658 00:41:08,193 --> 00:41:11,505 basically, if you have a human biology simulator, it could be, it 659 00:41:11,545 --> 00:41:13,925 would be really the revolutionary thing that's missing. 660 00:41:14,545 --> 00:41:18,369 Yeah, the one thing I'll add to that is, you know, we, given our 661 00:41:18,417 --> 00:41:22,089 background, have this preference for platform companies. 662 00:41:22,257 --> 00:41:25,777 And by platform here I'm meaning a platform that has data, 663 00:41:25,921 --> 00:41:30,835 a mix of wet lab and dry lab capabilities and data modeling all, 664 00:41:30,915 --> 00:41:33,427 you know, all built into it so that you can have some kind of an 665 00:41:33,451 --> 00:41:36,059 intelligent application. It could be again, protein to protein 666 00:41:36,107 --> 00:41:40,203 interactions, it could be small molecule discovery, etc. That 667 00:41:40,219 --> 00:41:43,491 is a little bit out of favor right now in the life sciences world, 668 00:41:43,683 --> 00:41:46,867 but we'll take the long term view on that because we believe that 669 00:41:46,931 --> 00:41:50,035 that's going to be an essential ingredient of the companies 670 00:41:50,075 --> 00:41:53,931 that are going to win long term. What's harder is to get your 671 00:41:53,963 --> 00:41:56,523 head around how those companies are going to capture value 672 00:41:56,579 --> 00:42:00,439 versus create value. They're going to create value at a minimum 673 00:42:00,487 --> 00:42:03,255 by partnering with a number of other life sciences companies and 674 00:42:03,295 --> 00:42:06,435 pharma companies and leveraging their distinct capabilities. 675 00:42:06,735 --> 00:42:09,599 And they may also have, and we encourage them to have, you know, 676 00:42:09,647 --> 00:42:12,135 some exploration of their own internal programs that they could 677 00:42:12,175 --> 00:42:15,327 partner out at some point in time. We like that because it gives 678 00:42:15,351 --> 00:42:19,063 some diversification to the, to the investment, but we also think 679 00:42:19,079 --> 00:42:22,967 that it more fully leverages the platform. And you know, at the 680 00:42:22,991 --> 00:42:25,503 moment that's a bit out of favor, but I think give it a little 681 00:42:25,519 --> 00:42:29,383 bit of time here, give, give us some success stories. I think 682 00:42:29,399 --> 00:42:32,207 you're going to see that that comes back into favor and I'd. 683 00:42:32,231 --> 00:42:34,863 Double down on that. I agree with Matt's characterization that 684 00:42:34,879 --> 00:42:38,687 it's out of favor. And I largely view this as a macroeconomic, 685 00:42:38,791 --> 00:42:42,655 a reflection of going from a low interest rate phenomena to a 686 00:42:42,695 --> 00:42:46,951 high interest rate behavior that we have now. But you know, there's 687 00:42:47,063 --> 00:42:50,511 boom bust cycles in all investment cycles and some of the 688 00:42:50,543 --> 00:42:54,817 best companies are typically built in the bus times. You know, 689 00:42:54,841 --> 00:42:58,441 all the big names you think about are typically started then. 690 00:42:58,473 --> 00:43:02,473 And so the platform companies that can survive now and build now, 691 00:43:02,609 --> 00:43:05,017 those will be the real killers. On the other side that will 692 00:43:05,041 --> 00:43:08,041 do things that when that comes, when the pendulum swings back, 693 00:43:08,073 --> 00:43:12,273 as it will, as it has many times, those companies will be really 694 00:43:12,329 --> 00:43:15,985 positioned to do well. And so it's just a natural part of the business 695 00:43:16,025 --> 00:43:19,885 cycle. And I can see why people will want assets short term. 696 00:43:19,925 --> 00:43:23,373 But ideally the best platforms generate a bunch of assets too. 697 00:43:23,509 --> 00:43:28,661 Yep. Well, that's a great place to chat about that. We've talked 698 00:43:28,693 --> 00:43:31,133 so much about all the different types of AI and where it 699 00:43:31,149 --> 00:43:35,825 can be used. I'm curious from each of you, what is your favorite 700 00:43:36,845 --> 00:43:41,749 and or what you believe is the most impactful application of AI 701 00:43:41,797 --> 00:43:44,025 ML that you have seen? 702 00:43:46,095 --> 00:43:49,895 Well, I think there's a, there's so many different things 703 00:43:49,935 --> 00:43:53,343 to choose from there. I think this on a personal level, things 704 00:43:53,359 --> 00:43:56,663 like Perplexity and chatgpt are, they're just magic. I mean I 705 00:43:56,679 --> 00:43:59,543 was using it this morning on some, some things that it just gave 706 00:43:59,559 --> 00:44:02,871 me better answers for, you know, and, and so you're, you're 707 00:44:02,903 --> 00:44:05,855 really just loving that component. I think that the other 708 00:44:05,895 --> 00:44:09,079 piece that, that we're seeing is, you know, that there are companies 709 00:44:09,127 --> 00:44:14,139 that are doing what seemingly are narrow tasks and solving problems 710 00:44:14,187 --> 00:44:18,051 that are focused and narrow and doing that really well and through 711 00:44:18,083 --> 00:44:22,491 the interaction with customers, getting invited into adjacencies, 712 00:44:22,643 --> 00:44:26,811 other areas that they can help solve even better. Whether that's 713 00:44:26,843 --> 00:44:29,491 a company that makes, you know, kind of all the different types 714 00:44:29,523 --> 00:44:33,347 of meeting communications I get, we're on a Zoom or we're on 715 00:44:33,371 --> 00:44:37,155 a Slack thread or on how do I take all of that, summarize it, synthesize 716 00:44:37,195 --> 00:44:40,731 it and make it in context, available in a more productive way. 717 00:44:40,803 --> 00:44:43,979 You know, an instant meeting recap, for example, or a company 718 00:44:44,027 --> 00:44:47,659 that's just advancing the state of, you know, marketing content, 719 00:44:47,787 --> 00:44:50,883 you know, you know, for big enterprises. And there's just. I 720 00:44:50,899 --> 00:44:54,211 could go down a very long list here. I'll give you one more example. 721 00:44:54,323 --> 00:44:59,171 Runway ML, super cool company. Basically you can sort of speak a 722 00:44:59,203 --> 00:45:03,035 video into existence. And now there's coming along some, some new 723 00:45:03,075 --> 00:45:06,387 players and chat GPT or OpenAI has got their own version that they're 724 00:45:06,411 --> 00:45:10,121 trying to get out as well. But just to think about the capabilities 725 00:45:10,273 --> 00:45:13,385 in that case, leveraging diffusion models to be able to speak 726 00:45:13,425 --> 00:45:14,845 a video into existence. 727 00:45:15,545 --> 00:45:18,697 So Matt gave so many good examples. I'll just mention, like 728 00:45:18,721 --> 00:45:21,657 the two that come to mind is that so as you having dinner with 729 00:45:21,681 --> 00:45:26,033 a friend who's does a lot of programming and you know, I started 730 00:45:26,089 --> 00:45:28,257 programming again just to play with the new tools and that's been 731 00:45:28,281 --> 00:45:33,417 fun. But between coding and writing, those two things, like many 732 00:45:33,441 --> 00:45:36,521 of us actually have coded with a partner. That's often called like 733 00:45:36,553 --> 00:45:40,537 extreme programming or you'd write with a writing partner. Now, 734 00:45:40,641 --> 00:45:43,665 like, these LLMs are like the best coding partner or writing partner, 735 00:45:43,705 --> 00:45:46,409 and they're available all the time, just on my laptop, anytime. 736 00:45:46,537 --> 00:45:48,849 And it's just kind of mind blowing how much you can get done, 737 00:45:48,897 --> 00:45:52,705 how quickly. But the thing that I think will be really exciting 738 00:45:52,745 --> 00:45:55,905 is kind of what I was alluding to before, the ability to use health 739 00:45:55,945 --> 00:45:59,385 records to predict someone's health. I am very curious about that, 740 00:45:59,425 --> 00:46:02,209 because that's all such a crapshoot, you know, to understand 741 00:46:02,297 --> 00:46:05,177 things and like, to even talk about, like, okay, what if I did 742 00:46:05,201 --> 00:46:08,113 this, what if I did that? And then finally, I think there's a lot 743 00:46:08,129 --> 00:46:12,097 of sort of common wisdom in medicine that I bet won't hold up 744 00:46:12,121 --> 00:46:16,137 to the data. And there's starting to be like little sort of 745 00:46:16,161 --> 00:46:19,545 cracks that are emerging. Like, is your cholesterol level really 746 00:46:19,585 --> 00:46:23,377 relevant for heart disease? Actually more people have heart attacks 747 00:46:23,401 --> 00:46:26,769 with low cholesterol than high cholesterol and so on. So, like, 748 00:46:26,857 --> 00:46:29,905 all of these things that have been common wisdom in medicine that, 749 00:46:29,945 --> 00:46:32,521 you know, go back to Framingham from like, what, 30 years 750 00:46:32,553 --> 00:46:36,501 ago. Like, now we finally have lots of data, let's actually see 751 00:46:36,533 --> 00:46:40,653 what's there. And these, I bet a lot of these rules of thumb may 752 00:46:40,709 --> 00:46:43,813 actually be wrong and will turn things upside down. And we've 753 00:46:43,829 --> 00:46:47,621 seen this all the time. That data can change things. I just really 754 00:46:47,653 --> 00:46:50,829 want to see what the data suggests, not sort of what one paper 755 00:46:50,877 --> 00:46:53,445 said 30 years ago. And everyone has just kind of now assumed 756 00:46:53,485 --> 00:46:54,225 is right. 757 00:46:55,445 --> 00:46:58,989 Yeah. And so I'd like to carry that theme just a little bit further. 758 00:46:59,117 --> 00:47:02,765 And I'll mention one of the places where I think there's a tremendous 759 00:47:02,845 --> 00:47:09,229 opportunity that we haven't seen yet in AI and bio is in experimental 760 00:47:09,277 --> 00:47:15,949 design. You mentioned sort of partner writing for writing, writing 761 00:47:16,037 --> 00:47:19,997 some prose or writing code. But we haven't yet gotten to the 762 00:47:20,021 --> 00:47:23,581 place where we're partner writing experimental designs and 763 00:47:23,653 --> 00:47:27,501 we're sitting with the AI and saying, hey, well, I have this question, 764 00:47:27,653 --> 00:47:31,547 I have this hypothesis, I'd like to test it. You know, I. How 765 00:47:31,571 --> 00:47:34,003 should I design my experiment? Should I design it this way, should 766 00:47:34,019 --> 00:47:37,323 I design it that way? What's the best way? And then, you know, 767 00:47:37,379 --> 00:47:41,055 sending it off to the cloud lab to actually have it done and 768 00:47:41,435 --> 00:47:42,171 come back. 769 00:47:42,283 --> 00:47:44,827 I think you're closer than either of us on this. Like, when 770 00:47:44,851 --> 00:47:46,015 is that going to happen? 771 00:47:47,155 --> 00:47:51,843 I think soon. I think we have, we have the cloud lab infrastructure 772 00:47:51,939 --> 00:47:57,641 aligning that we can send design documents around to them and 773 00:47:57,673 --> 00:48:04,805 say, hey, execute this thing. I think we have some concept of statistical 774 00:48:05,825 --> 00:48:10,433 design and variance that has been built in. I think what we haven't 775 00:48:10,489 --> 00:48:13,817 taught is, hey, I want to test this variable that is impacted by 776 00:48:13,841 --> 00:48:18,033 these 12 things. And so the concept of what experiment provides 777 00:48:18,089 --> 00:48:22,473 what, that's the piece that's currently missing. We just have data 778 00:48:22,529 --> 00:48:26,707 that came from an experiment, but we don't have. Why would I reason. 779 00:48:26,851 --> 00:48:30,955 Why would I choose to do a Western versus a Coomassie stain 780 00:48:31,035 --> 00:48:35,051 versus a proteomics experiment versus a proteopharm experiment? 781 00:48:35,123 --> 00:48:39,243 So that reasoning layer is what's currently missing. I don't 782 00:48:39,259 --> 00:48:42,051 think it's an intractable problem, but I think it's. That's 783 00:48:42,083 --> 00:48:46,043 the. That's currently, in my view, the missing link between that, 784 00:48:46,219 --> 00:48:49,499 that, that chat bot that's helping me design my next experiment. 785 00:48:49,547 --> 00:48:52,175 And what would that get you, would get you autonomous labs? 786 00:48:52,475 --> 00:48:53,215 Yes. 787 00:48:53,995 --> 00:48:56,619 Well, I think it would also get you potentially efficiency. I 788 00:48:56,627 --> 00:49:01,971 mean, to take it, maybe push it just one place further. Why couldn't 789 00:49:02,003 --> 00:49:06,283 I computationally run all those experiments on some system, 790 00:49:06,379 --> 00:49:09,683 some model that's been trained to do, you know, any of those types 791 00:49:09,699 --> 00:49:13,363 of experiments, like just even gene synthesis, as an example. And 792 00:49:13,419 --> 00:49:16,323 I say, okay, that's a part of a broader experiment, but I want 793 00:49:16,339 --> 00:49:19,749 to see what the gene synthesis is. I can do that computationally, 794 00:49:19,917 --> 00:49:23,557 probably as accurately as a lot of CROs, you know, today. And 795 00:49:23,581 --> 00:49:25,677 then I take all these different components together to 796 00:49:25,701 --> 00:49:29,749 then actually figure out what experiments I want and need to do 797 00:49:29,837 --> 00:49:32,837 at some point, you know, in some validated wet lab. 798 00:49:32,981 --> 00:49:36,021 Well, so carrying that one step further, Matt, I think even 799 00:49:36,053 --> 00:49:40,397 if the experiments are not very accurate, what you could do 800 00:49:40,461 --> 00:49:43,927 is with the sensitivity analysis around that model is you 801 00:49:43,951 --> 00:49:47,743 could see, hey, here's the place where I have the greatest uncertainty, 802 00:49:47,919 --> 00:49:52,431 where if I poke this a little or a lot, you know, huge things happen. 803 00:49:52,583 --> 00:49:57,303 And so here are the six experiments I should do to remove 804 00:49:57,359 --> 00:50:00,903 that uncertainty. And so I think that that's a piece of that, 805 00:50:01,039 --> 00:50:01,855 that efficiency. 806 00:50:01,935 --> 00:50:04,655 Well, especially that's a mindset where you're now doing experiments 807 00:50:04,695 --> 00:50:07,415 in an active learning framework rather than like a human 808 00:50:07,495 --> 00:50:11,313 hypothesis framework. And that would be a huge paradigm shift. 809 00:50:11,489 --> 00:50:16,905 Yeah, yeah. All right, so last, last question. Following up 810 00:50:16,945 --> 00:50:23,441 on this, let's look out 10 years. So what are you most excited 811 00:50:23,513 --> 00:50:27,441 about coming into the world? We've just talked about this one 812 00:50:27,553 --> 00:50:33,169 pivot in how we do experiments 10 years from now. What are you most 813 00:50:33,217 --> 00:50:38,005 excited about in the intersection between AI and bio health. 814 00:50:39,255 --> 00:50:42,351 I can go in first. So like, so I think the thing that I'm particularly 815 00:50:42,383 --> 00:50:46,095 excited about is that as what you can do with AI you see it very 816 00:50:46,135 --> 00:50:49,487 generally, but it's going to really impact bio. You see it with 817 00:50:49,511 --> 00:50:53,047 co pilots right now where like one person with AI can do the work 818 00:50:53,071 --> 00:50:56,527 of like three or five people. What that means though is that you 819 00:50:56,551 --> 00:51:00,191 can now have smaller groups do the work of what much larger groups 820 00:51:00,223 --> 00:51:03,631 do you know, you think about like a team of 200 versus a team 821 00:51:03,663 --> 00:51:07,637 of 20. Is a 200 person team really 10 times more effective? You 822 00:51:07,661 --> 00:51:10,557 know, people generally sort of have a rule of thumb, a square to 823 00:51:10,581 --> 00:51:14,197 end. So maybe like three times more effective. You can have 20 people 824 00:51:14,261 --> 00:51:18,749 with AI, maybe they're more effective than a 200 person team. 825 00:51:18,877 --> 00:51:21,557 And that's going to really change the social dynamics of how 826 00:51:21,581 --> 00:51:25,805 we work and how we do things and how we build startups. Like these 827 00:51:25,845 --> 00:51:29,213 startups will be leaner actually. People might get more equity 828 00:51:29,269 --> 00:51:31,173 because there's not much sharing. Maybe they don't need as 829 00:51:31,189 --> 00:51:35,451 much capital, maybe that means they can go quicker. Like I think 830 00:51:35,483 --> 00:51:38,731 just how we do our work is going to be revolutionized and it's 831 00:51:38,763 --> 00:51:42,579 not replacing people as much as it's really empowering them. Much 832 00:51:42,627 --> 00:51:47,435 like you know, if I had to dig a ditch by hand versus a backhoe, 833 00:51:47,595 --> 00:51:49,763 I guess you're replacing people and really just making the 834 00:51:49,779 --> 00:51:53,547 job a lot easier with the backhoe. And I think what hopefully 835 00:51:53,611 --> 00:51:58,531 especially do is take all the stuff that we hate doing and the 836 00:51:58,563 --> 00:52:02,021 grunge work and we always see this today with LLMs and really make 837 00:52:02,053 --> 00:52:05,397 that so seamless that that's the minimum but the fantasy which 838 00:52:05,421 --> 00:52:08,893 I think is very possible. We have small teams of great people 839 00:52:08,989 --> 00:52:11,533 empowered to do things and we're just doing the exciting parts 840 00:52:11,549 --> 00:52:14,013 of it. I don't think that's that far off. I think you talked 841 00:52:14,029 --> 00:52:16,585 about 10 years easily in 10, maybe sooner. 842 00:52:17,445 --> 00:52:21,445 Yeah. Let me take this from another direction, which is I'm excited 843 00:52:21,485 --> 00:52:27,887 about the idea of end user side agents. Now we can all relate 844 00:52:27,911 --> 00:52:31,271 to the concept of like a shopping agent or a travel agent, 845 00:52:31,383 --> 00:52:35,367 but I'm going to have my own health agent. I'm going to have the 846 00:52:35,431 --> 00:52:38,471 data about all the different, you know, tests that have been run 847 00:52:38,503 --> 00:52:42,519 on me and you know, other kinds of, you know, kind of, you 848 00:52:42,527 --> 00:52:46,703 know, health data that's ingested into this, you know, personal 849 00:52:46,799 --> 00:52:51,183 agent, this personal model that's going to enable me to interact 850 00:52:51,239 --> 00:52:55,143 with the provider side, the supplier side the doctors of different 851 00:52:55,199 --> 00:52:59,595 types and advise me on what I should be doing that nudge to go 852 00:52:59,715 --> 00:53:05,283 get the workout in and what I should be how I should be making 853 00:53:05,339 --> 00:53:08,547 choices about my personal health and that seems a little bit 854 00:53:08,571 --> 00:53:11,731 amazing to all of us because how hard it is to get our own personal 855 00:53:11,803 --> 00:53:16,467 health data today. But this is not a technological limitation and 856 00:53:16,491 --> 00:53:19,859 I think the demand side of the end users in all these areas is going 857 00:53:19,867 --> 00:53:22,873 to not only pull for them to have the control of the agent but 858 00:53:22,889 --> 00:53:25,165 it's actually going to change the business models too. 859 00:53:26,785 --> 00:53:31,377 That's great. I think this has been a really fun discussion. Thank 860 00:53:31,401 --> 00:53:35,865 you all both for joining me today and I know we could probably 861 00:53:35,905 --> 00:53:39,433 talk for another six hours but I'm really appreciative. Thanks so 862 00:53:39,449 --> 00:53:40,005 much. 863 00:53:40,505 --> 00:53:41,297 Thank you. 864 00:53:41,441 --> 00:53:42,965 Thanks very much. It was a lot of fun. 865 00:53:45,785 --> 00:53:49,567 We hope you enjoyed the Translating Proteomics podcast brought 866 00:53:49,591 --> 00:53:53,551 to you by Nautilus Biotechnology. To contact us or for 867 00:53:53,583 --> 00:53:57,967 further information please email TranslatingProteomics Nautilus 868 00:53:58,071 --> 00:53:58,215 bio.