The Information's TITV
00:10:34 · Anne Wojcicki, Stephanie Palazzolo
Like, I want everyone who's studying lung cancer to move it forward. And you see, like, when every single person has access to resources, we can make a huge difference. And ultimately, the goal for us is like, I want to have a significant impact on prevention and treatments for diseases. And I think the way that we can do that best is as a nonprofit. And [00:10:57] S0: talk to me about, like, how much has it helped having some of these, you know, large language models, frontier AI models? How much have those helped specifically in analyzing this genetic data versus, you know, older forms of AI that you could have used in the past?
00:10:57 · Stephanie Palazzolo, Anne Wojcicki
[00:10:57] S0: talk to me about, like, how much has it helped having some of these, you know, large language models, frontier AI models? How much have those helped specifically in analyzing this genetic data versus, you know, older forms of AI that you could have used in the past? [00:11:12] S5: What's interesting about AI is, you know, those are language models. And so specifically, what we have is we have a DNA language model. And so AI, particularly in those earliest versions, like, did not really understand DNA. And so part of what we've been working onand we had a paper that came out last year at NeurIPS about, like, how can you actually use AI and these models to actually look at all of the data and really be able to improve risk prediction? So there's two things that we see as a huge opportunity by working with partners or by training models ourselves. And that is going to be one, can you just really interpret the genome? And so you hear about all kinds of families that are wondering, like, I have a child who's sick and no one can figure out why. Or people... Like my sister will say, like, did I have a genetic risk factor for lung cancer? Like, how do we actually read your whole genome and how can we understand it? I think that leveraging these large language models by training data on a large number of data sets, like, I think we're going to be able to really be able to understand why some people have disease and how do you predict what you're really at risk for.
00:11:12 · Anne Wojcicki, Stephanie Palazzolo
[00:11:12] S5: What's interesting about AI is, you know, those are language models. And so specifically, what we have is we have a DNA language model. And so AI, particularly in those earliest versions, like, did not really understand DNA. And so part of what we've been working onand we had a paper that came out last year at NeurIPS about, like, how can you actually use AI and these models to actually look at all of the data and really be able to improve risk prediction? So there's two things that we see as a huge opportunity by working with partners or by training models ourselves. And that is going to be one, can you just really interpret the genome? And so you hear about all kinds of families that are wondering, like, I have a child who's sick and no one can figure out why. Or people... Like my sister will say, like, did I have a genetic risk factor for lung cancer? Like, how do we actually read your whole genome and how can we understand it? I think that leveraging these large language models by training data on a large number of data sets, like, I think we're going to be able to really be able to understand why some people have disease and how do you predict what you're really at risk for. [00:12:22] S0: And talk to me about the process of training this sort of DNA language model. I imagine a big part of it is using a different, you know, different training data sets than the ones that are probably used for...
00:12:22 · Stephanie Palazzolo, Anne Wojcicki
[00:12:22] S0: And talk to me about the process of training this sort of DNA language model. I imagine a big part of it is using a different, you know, different training data sets than the ones that are probably used for... [00:12:33] S5: Yeah. We're not...... Flawed, but I mean, we're not... We're actively... So we did a bunch of work previously. And again, pre Chapter 11, we are doing a bunch of work, and that's where we published this paper on NeurIPS. We are actively looking now at what is that right path forward for us. So we're not doing anything today, but we're really focusing on, like, there's a huge amount of opportunities. We... We're working in small bits internally, but we think a lot about, like, who's the right partner for us? Where do we wanna be putting our efforts? Like, what are the first priorities? So one thing that we're doing a lot of work on now is, you know, actually getting whole genome data. So 23andMe works on SNP data, meaning, like, we look at a 100,000,000 variants in humans. As you start to get long read data and as you start to get the whole genome, you're gonna be able to make all kinds of new discoveries. So it's about get... Putting that right data set together. We'll identify the right partners, and then we'll start to move forward. [00:13:30] S0: Mhmm. Who... Like, what are some examples of, you know, partners that you would consider? Like, are these other model developers? Like, obviously, I know, like, Anthropic, for instance, has gotten very into AI for bio. Who might be some of the partners that you would think of
00:13:30 · Stephanie Palazzolo, Brittany Greenfield, Anne Wojcicki
[00:13:30] S0: Mhmm. Who... Like, what are some examples of, you know, partners that you would consider? Like, are these other model developers? Like, obviously, I know, like, Anthropic, for instance, has gotten very into AI for bio. Who might be some of the partners that you would think of [00:13:46] S4: to work [00:13:47] S0: with on this sort of thing? [00:13:48] S5: Interestingly, there hasn't... Like, when we went to NeurIPS, we were the only genetics paper there. And so that was last year. So it's been really interesting to see, you know, Anthropic, OpenAI, Google, like, everyone is coming. Like, everyone looks at biology and health care now, and it's super interesting to people. So there's a lot of really interesting players out there, you know, a lot of expertise in little pockets of some of these companies. I don't know. We'll have to see sort of like who makes the most sense. And I think for us as a nonprofit, what's always top priority for us is that mission. Like, are we doing... How can we fulfill the mission that our customers have entrusted us with? Like, I wanna really understand risk prediction. I wanna be able to give that information back to individuals. I wanna be able to predict disease because then you can actually do something about it and try to prevent.
00:13:48 · Anne Wojcicki, Stephanie Palazzolo
[00:13:48] S5: Interestingly, there hasn't... Like, when we went to NeurIPS, we were the only genetics paper there. And so that was last year. So it's been really interesting to see, you know, Anthropic, OpenAI, Google, like, everyone is coming. Like, everyone looks at biology and health care now, and it's super interesting to people. So there's a lot of really interesting players out there, you know, a lot of expertise in little pockets of some of these companies. I don't know. We'll have to see sort of like who makes the most sense. And I think for us as a nonprofit, what's always top priority for us is that mission. Like, are we doing... How can we fulfill the mission that our customers have entrusted us with? Like, I wanna really understand risk prediction. I wanna be able to give that information back to individuals. I wanna be able to predict disease because then you can actually do something about it and try to prevent. [00:14:37] S0: And I guess talking a bit about the business and the business model, obviously, you're a nonprofit now. I think a big concern in the past has been that perhaps the business model as even as a for profit wasn't necessarily sustainable. And I know you've talked a bit about cutting back on certain areas and really like kind of focusing on these key, on these kind of key areas for 23andMe now. But why is like a nonprofit the right way to go? And are are you concerned about some of these sustainability questions coming up now