Zuckerberg Philanthropy Consolidates Science Units to Accelerate AI Disease Cures

扎克伯格慈善机构整合科学部门以加速AI疾病疗法研发 Chan Zuckerberg Initiative（CZI）正将其分散的中心统一在人工智能领导之下以构建虚拟细胞模型，并认为大规模开源计算将缩短在世纪末治愈所有疾病的时间线。 在过去十年中，CZI已从传统的科研资助转向开发开源的、规模达1亿美元的基础设施项目，例如与更广泛科学界合作构建的CELLxGENE图谱。通过将前沿生物学与前沿人工智能相结合，CZI正在开发分层虚拟细胞模型，使研究人员能够在投入昂贵的湿实验之前，在计算机中模拟复杂的生物系统并检验高风险假说。为了最大化这种飞轮效应，CZI正在人工智能领导下整合其业务，转变现代实验室空间以优先发展大规模计算集群而非扩大物理占地面积，并致力于让全球研究人员都能平等获取这些资源。 Chan Zuckerberg Initiative（CZI）成立时确立了一项世纪跨度的使命，旨在帮助科学界在21世纪末之前治愈、预防或管理所有疾病。 (0:00) 像NIH拨款这样的传统政府资助在长期、昂贵的工具开发方面留下了巨大空白，CZI旨在通过资助周期为10至15年的基础设施项目来填补这一空白。 (3:42) CZI的核心战略依托于“工具优先”假说，认为重大科学突破从根本上受限于观测手段，且需要由新工具来开启。 (3:08) CZ Biohub通过将前沿生物学与前沿人工智能相结合来弥合文化鸿沟，从而生成专为训练AI模型从零设计的大规模、高度特异性的数据集。 (7:03) CZ Biohub Network在战略上分布于三个地点——旧金山、芝加哥和纽约——各自侧重于生物学、成像和组织工程的不同交叉领域。 (7:03)

Mark Zuckerberg and Priscilla Chan argue that curing, preventing, or managing all human diseases can be dramatically accelerated by redirecting philanthropy toward long-term tool building, open-source biological data, and domain-specific artificial intelligence. By integrating computational models directly with experimental biology, the Chan Zuckerberg Initiative aims to replace empirical trial-and-error medicine with predictive, virtual cellular simulations. Addressing the Infrastructure Gap in Basic Science (3:08) Traditional biomedical research largely relies on short-term government grants, such as NIH awards to individual laboratories, leaving a structural deficit for long-horizon, high-cost engineering. Zuckerberg argues that major breakthroughs in biology historically depend on novel observation instruments, comparing current biological research without advanced tools to writing code without the ability to step through and debug it. Drawing on Chan's clinical experience treating pediatric patients with uncharacterized conditions and genetic variants of unknown significance, CZI directed its philanthropy toward 10-to-15-year infrastructure projects requiring $100 million to $1 billion that are too large for standard academic grants and too slow for venture capital. Open-Source Data Standardization and Single-Cell Mapping (11:35) Recognizing that biology still lacked a foundational reference system comparable to chemistry's periodic table, CZI funded single-cell transcriptomics and introduced CELLxGENE, an open-source tool developed to solve data annotation bottlenecks. By offering an accessible annotation platform, the initiative established standardized data and metadata formats across the scientific community. While CZI directly funded roughly 25% of the resulting single-cell atlas, external researchers contributed the remaining 75% due to the utility of the common format. Standardized single-cell datasets enable researchers to compare healthy and diseased cells to identify exact mechanisms of pathology. For instance, an early-stage biotechnology company utilized CELLxGENE to analyze millions of cells to isolate aberrant gene expression in fibroblasts associated with idiopathic pulmonary fibrosis, demonstrating how open-source biological maps can streamline drug target discovery and help forecast off-target side effects. Virtual Cells and Predictive Biological Modeling (18:34) To reduce the high financial and temporal costs of physical laboratory work, CZI is constructing hierarchical virtual cell models to serve as digital model organisms. Much like fruit flies have functioned as biological benchmarks, these in silico models allow scientists to simulate hypotheses—spanning molecular protein structures up to complex systems like the immune system—before executing physical wet-lab experiments. This computational derisking enables investigators to pursue higher-risk scientific inquiries without jeopardizing grant funding. Rather than depending on a single generalized artificial intelligence, the virtual cell framework integrates specialized biological models. These include Variant Former, an AI trained on paired CRISPR edits to forecast downstream cellular modifications, diffusion models that synthesize rare cellular states, and spatial imaging models. CZI is also directing efforts toward biological reasoning models designed to assess cause-and-effect dynamics rather than mere statistical correlations. Unifying Compute Clusters and Interdisciplinary Biohubs (28:09) CZI is consolidating its biohubs, software operations, and artificial intelligence research under a unified philanthropy led by AI researcher Alex Rives, reflecting their view that computation must tightly couple with experimental design. Borrowing organizational structures from physics, the Biohub model physically co-locates biologists directly beside software engineers to establish immediate feedback loops between wet-lab instruments, such as cryo-EM, and computational model generation. Under this model, computational capacity functions as the primary lab resource. While standard academic research groups typically access only tens of GPUs, the Biohub established a 1,000-GPU cluster with plans to expand toward 10,000 GPUs. External academic scientists are invited to pitch research questions requiring this computational scale, expanding access to high-performance infrastructure across institutions including UCSF, Stanford, and UC Berkeley. Zuckerberg and Chan forecast that coupling unified biohubs with domain-specific AI will allow the scientific community to achieve their disease prevention and management goals significantly sooner than their original 2100 target; however, the practical translation of these tools still depends on closing the gap between computational simulations, experimental verification, and commercial pharmaceutical drug delivery.

Mark Zuckerberg 与 Priscilla Chan 阐述了 Chan Zuckerberg Initiative (CZI) 的战略核心：通过长期慈善资金开发开源标准化工具与大规模计算基础设施，将前沿人工智能与基础生物学深度融合，以期在本世纪末甚至更早实现治愈、预防或管理所有人类疾病的目标。 填补科研资助空白：以工具驱动基础科学 (3:08) CZI 将战略建立在“重大科学突破多由新观测工具驱动”的历史规律之上，类似显微镜之于微生物学、望远镜之于天文学。Mark Zuckerberg 指出，当前生物学研究缺乏调试代码般的直观观测手段。Priscilla Chan 结合其在加利福尼亚大学旧金山分校（UCSF）儿科临床的观察表示，现代医学常受限于“试错式治疗”，医生面对未知意义变异（VUS）等罕见或复杂病症时往往缺乏基础机制解释。然而，传统美国国立卫生研究院（NIH）式的短期分散科研基金无法支持大规模工程开发，风险投资亦难以承受漫长回报周期。CZI 选定这一慈善生态位，通过为耗资 1 亿至 10 亿美元、跨越 10 至 15 年的底层工具研发提供全额资助，构建推进基础科学的“希望管道”，并将成果全部开源交付全球科学界。 数据网络效应：绘制生物学元素图谱 (15:55) 针对生物学界长期缺乏“元素周期表”式通用基准图谱的现状，CZI 最初开发了单细胞数据标注工具 CELLxGENE 以解决标注效率瓶颈。该工具的普适性意外引发了强劲的网络效应：全球研究者在采用同一工具的过程中，自发统一了数据格式与元数据标准。在过去十年积累的数百万单细胞数据库中，CZI 仅直接资助了约 25% 的数据，其余 75% 均由国际科研社区自愿上传贡献，其数据标准化程度与质量控制远超以往的公共数据库（如 GEO）。该开源平台大幅降低了跨学科交互门槛，例如支持免疫学家探索其理论在神经退行性疾病中的关联；同时它已被生物科技初创企业用于药物靶点发掘，例如在特发性肺纤维化（IPF）研发中精准分离并分析特定成纤维细胞的基因表达。 虚拟细胞与模拟：从突变预测到因果推理 (18:34) CZI 正在构建多层级的“虚拟细胞”（Virtual Cell）模型，试图从蛋白质分子、细胞结构逐步向上模拟至复杂的虚拟免疫系统。该模型被定位为现代生物学的数字“新果蝇”，使研究人员能够在耗时昂贵且高风险的实体湿实验之前，先在计算机内（in silico）以极低成本进行模拟运算与假设筛选，从而大幅降低探索风险，支持科学家挑战更高难度的科研课题。技术路径上，CZI 避开单一的通用大模型，转向组合训练专业模型：包括利用成对 CRISPR 编辑数据预测细胞转归的“Variant Former”、模拟罕见细胞形态的扩散模型，以及结合成像研究所数据的空间表征模型。其终极前沿是打造生物推理模型，使系统超越单纯的统计关联，具备推演生物机制因果逻辑的能力。 机构重塑与算力基建：打造闭环协作网络 (28:09) 为消弭人工智能与生物学之间的行业认知隔阂，CZI 宣布将其软件团队、区域 Biohub 与 AI 实验室整合成统一运营的慈善实体，并任命前 Meta AI 蛋白质研究专家、Evolutionary Scale 联合创始人 Alex Rives 牵头领导。这一架构实现了“AI 模型预测与湿实验数据采集”之间的紧密闭环：当模型出现盲区时，湿实验团队可利用冷冻电镜等高端工具定向生成补全数据。在组织运作上，Biohub 借鉴高能物理领域的跨学科协作范式，通过物理工位并置促成生物学家与工程师的高频交流，打破了 UCSF、斯坦福大学与加州大学伯克利分校之间的体制壁垒。此外，Biohub 将算力视作新型核心实验空间，正在将其 GPU 集群规模从 1,000 块扩展至 10,000 块级别，并向外部学术机构开放计算资源申请。 Zuckerberg 与 Chan 认为，AI 技术的突破使攻克全人类疾病的 century-scale 目标有望大幅提前。然而整套设想的高效运转依赖于完整的转化链条：慈善基金加速基础科学并开源工具，生物技术公司据此研发新型疗法，制药企业进行规模化生产，公共卫生系统负责全球分发。虚拟细胞预测与湿实验闭环能否经受住真实世界复杂疾病临床转化的严苛检验，仍是该长周期战略面临的核心考验。

[0:00] Priscilla Chan: Most scientists couldn't look at us with a straight face, [3:08] Mark Zuckerberg: Without those kind of tools, it's kind of like you're coding without being able to step through the code and debug things. [6:29] Mark Zuckerberg: frontier biology paired with frontier AI. [11:35] Priscilla Chan: most diseases should be thought of as rare diseases [15:55] null: come for the annotation, stay for the virtual cell model [18:34] null: new fruit fly [24:06] null: reasoning model over biology. [28:09] null: significantly sooner. [34:53] null: it doesn't matter what the org chart is... you guys need to sit next to each other and until you get this thing to work. [38:40] null: They don't want space. Yeah. They just want GPUs, agents. [40:27] null: completely envious of people working for for-profit companies

[0:00] this is a a a space that I mean that [0:02] there's just going to be a huge amount [0:03] of leverage with AI. It still seems like [0:05] there could be a lot more effort in this [0:07] space around building tools and it's [0:09] kind of this crazy thing that we're you [0:11] know here in you know 2025 and there's [0:15] not the kind [music] of periodic table [0:16] of elements equivalent for biology. We [0:18] think that this is like probably one of [0:20] the most important [music] sets of tools [0:22] that you need to build. When we first [0:24] set out that the goal to cure and [0:26] prevent disease by the end of the [0:27] century, people like honestly most [0:29] scientists couldn't look at us with a [0:30] straight face. [laughter] [0:32] >> And that's crazy. [0:34] >> Yes. And it was true because if you just [0:37] decided to spend the money funding the [0:39] next best [music] grant for every single [0:41] lab in the country, like you there was [0:44] no pathway to that being true. The [0:45] biology folks, I think, looked [music] [0:47] at it as if it were crazy ambitious. And [0:49] then the AI folks are like, well, that's [0:52] kind of boring. That's just [0:54] automatically going to happen. I know. [0:56] It's like, okay, there's something in [0:57] between there that needs to be bridged. [1:02] >> Mark Priscilla, welcome to the Asz [1:04] podcast. [1:05] >> Thanks for having us. [1:06] >> Yeah, great to be here. Excited. [1:07] >> All right. Excited to have you. You're [1:09] doing exciting stuff. [1:10] >> Yeah. Well, to that end, almost a decade [1:12] ago, you guys started the Chan [1:13] Zuckerberg initiative with the mission [1:15] and intent to cure, prevent, manage all [1:17] disease by the end of this century. [1:19] There's a lot of missions that you guys [1:21] could have poured your time and [1:22] resources into. Why don't we talk about [1:24] take us behind the conversations of why [1:26] you guys picked this one? Maybe [1:27] Priscilla, why don't we start with with [1:28] you and you hear your side of the story. [1:31] >> It always surprises people when I talk [1:33] about how we work in basic science [1:34] research. Um, I trained as a [1:36] pediatrician and people always think [1:37] like, oh, it must be about medicine. And [1:40] for me, it w, you know, I went into [1:43] medicine because I wanted to improve [1:45] people's lives. I wanted to make a [1:47] difference. I wanted to be able to help [1:48] others. And I think training as a [1:51] pediatrician at UCSF, I met a lot of [1:54] patients and frankly like little kids [1:57] and families for which like we just had [2:00] no idea what the problem was. and they [2:03] might have like a specific gene that [2:06] they could name if they were lucky. Um, [2:08] or they could be grouped into a bunch of [2:10] other diseases and there'd be a general [2:12] sort of PDF they'd print out like this [2:14] is what we know. And then it was my job [2:16] as an intern or resident to try to [2:18] translate like like a few lines of [2:22] information to how we were supposed to [2:24] take care of the patient. And for me, [2:26] that's when I really like realized the [2:28] power of basic science and how we need [2:31] to work on basic science to advance the [2:35] forefront of what's possible and without [2:37] that there's sort of I think of it as [2:40] the pipeline of hope. [2:41] >> Yeah. And why did you think um [2:45] you could cure all disease? Because [2:46] that's like a very like aggressive goal. [2:49] >> Um do you want to do you want to answer [2:51] that one? [2:51] >> Yeah. Well, well, I mean we're not going [2:53] to cure all diseases to be clear. I mean [2:55] the the strategy is to help scientists [2:58] and the scientific community cure all [3:00] diseases. So the strategy is really one [3:03] of accelerating the pace of basic [3:05] science and the theory that we had was [3:09] if you look at the history of science [3:10] most major breakthroughs are basically [3:13] preceded by the invention of a new tool [3:16] to observe phenomena in a new way. [3:18] Right? So think about things like the [3:19] microscope, right? Being able to observe [3:21] bacteria or you know other fields, the [3:24] telescope or [3:25] >> um you know, but it's [3:27] >> just to use an engineering example, you [3:29] know, it's without those kind of tools, [3:32] it's kind of like you're coding without [3:33] being able to step through the code and [3:35] debug things, right? So it's um that's [3:38] like the old days [3:39] >> when you [laughter] [3:41] so so our our whole approach on this is [3:44] basically let's help build tools that [3:47] will accelerate the pace of the whole [3:49] field and I think that that there's a [3:52] niche that I think fits that because if [3:55] you look at how funding works in science [3:57] you know the vast majority of funding [3:58] comes from the government and NIH [4:00] grants. it's parcled out into these [4:02] relatively small grants that allow [4:04] individual investigators to investigate [4:07] usually pretty near-term things. Um, and [4:10] the development of these kind of new [4:12] types of tools, whether it's imaging or [4:14] building now a lot of AI things like [4:16] virtual cell models, um, are longer [4:19] term, often times more expensive to [4:22] develop. So think about like on the [4:24] order of a hundred you know maybe you [4:27] know hundred million to a billion [4:29] dollars over a um over a 10 to 15 year [4:32] period and then you you try to unlock [4:34] those tools and give them to the [4:36] scientific community to accelerate the [4:37] pace. So that's that's kind of the the [4:39] theory [4:39] >> right and and there it seems like [4:41] there's also something that that's is [4:44] you don't really get credit for the [4:45] tools in a lot of ways. I mean, we've [4:47] been noted, well, we have companies that [4:48] use your tools and they're very happy [4:50] about it. But, um, you know, I didn't [4:53] even know that that was the case. And [4:55] so, [4:55] >> that's why it's philanthropy. [4:56] >> Yeah. Well, it is, but most people do [4:59] philanthropy to get credit, too. I mean, [5:01] you know, like that's a you know, that [5:02] that's kind of a part of it. So, [5:06] how did you I guess did you think about [5:08] that or were you just like, no, like [5:10] this is going to work and if it works, [5:12] that's all we need. We're super focused [5:15] on like actually making every scientist [5:17] better and and beyond science like [5:20] startups, startup founders because I the [5:23] point is we can't do this alone. And [5:26] when we first set out that the goal to [5:28] cure and prevent disease by the end of [5:30] the century, people like honestly most [5:32] scientists couldn't look at us with a [5:34] straight face [laughter] [5:35] >> and crazy. [5:37] >> Yes. And it was true because if you just [5:40] decided to spend the money funding the [5:42] next best grant for every single lab in [5:45] the country like you there's no pathway [5:48] to that being true. But if you forced [5:51] people to really think about this and [5:53] like okay what is the most credible [5:55] pathway to doing this and what are the [5:57] barriers to that credible pathway then [6:00] we sort of got somewhere right? They [6:01] were like, well, like there's no shared [6:03] tools or like we don't have we're not [6:06] working on big projects and building the [6:08] right data sets. And we're like, okay, [6:10] well then we can start doing something [6:12] about that. Um, and so that's where the [6:14] idea of building shared tools cuz no one [6:16] right now in the science. [6:17] >> Well, that's so interesting. So [6:19] basically, you're like, we're going to [6:20] cure all disease and they're like, [6:22] >> yeah, can't be done. Why can't it be [6:23] done? Well, because we don't have the [6:25] tools. Okay, that's pretty that's a [6:28] pretty cool sequence. [6:29] >> Yeah. Yeah. I mean, there's also this [6:30] funny thing where the the biology folks, [6:32] I think, looked at it as if it were [6:34] crazy ambitious. And then the AI folks [6:36] are like, well, that's kind of boring. [6:39] That's just automatically going to [6:40] happen. I know it's like, okay, there's [6:42] something in between there that needs to [6:44] be bridged. And if you can like kind of [6:46] use the the kind of modern AI tools in [6:49] order to build the types of tools that [6:51] biologists need. So that's a big part of [6:53] how we think about our work is um [6:55] >> AI has got to be the most overestimated [6:57] and underestimated technology ever like [6:59] simultaneously. So weird. I mean, yeah, [7:01] we'll probably like the internet early [7:02] on, but but we kind of think about [7:04] ourselves and the work that we're doing [7:06] at the Biohub as frontier biology paired [7:09] with frontier AI, right? So, there's [7:11] there are labs that do frontier AI that [7:14] uh basically, you know, are building the [7:16] most advanced models. Um and then there [7:20] are lots of biological research [7:22] organizations that that effectively do [7:25] very leading edge [7:26] >> research to build um you know to either [7:30] discover new data sets or or or looking [7:32] to certain challenges. [7:34] >> But so far there hasn't been anyone [7:36] who's tried to do both of those at once. [7:38] And when you look at I mean even [7:40] something like AlphaFold which is [7:41] amazing right it's it was built off of [7:43] this data set that was a public data set [7:45] that had been produced decades ago right [7:48] and um [7:50] >> what what I think you have the [7:52] opportunity to do if you do both of [7:53] those together is produce specific data [7:56] sets for the purpose of training AI [7:58] models to build virtual cells that can [8:00] do specific things [8:01] >> right [8:02] >> so I think that that's like a a pretty [8:04] interesting zone to be in [8:06] >> and of all the things that that we've uh [8:10] that we've worked on. You know, actually [8:11] when when we started CZI, we we kind of [8:13] actually focused on a number of areas [8:14] and what we found is just that the [8:16] science research has had by far the [8:18] biggest return. So, we've just doubled [8:19] down on it over and over and over until [8:21] now we're at the point that, you know, [8:23] we're 10 years in and Biohub is really [8:25] the like main focus of of our of our [8:28] philanthropy at this point. [8:30] >> Um, but yeah, I mean, that's kind of [8:32] that's basically the focus. Maybe you're [8:34] not giving yourselves enough credit [8:35] because you're sort of saying, "Well, [8:36] there's bite-size science. We didn't [8:38] want to do that. There's century scale [8:40] science and that seemed like a long time [8:42] horizon, but achievable, ambitious." But [8:44] you've actually identified, you know, [8:46] which I think is really fantastic, grand [8:48] scientific challenges [8:50] >> that are right in between. They're 10 to [8:52] 15 year horizons, at least per kind of [8:55] the way you communicate about them and [8:56] the way you energize [8:57] >> the scientific community about them. 10 [8:59] to 15 is kind of an interesting time [9:00] horizon sort of like similar to the time [9:03] horizon of a venturebacked company [9:05] similar to the time horizon on which a [9:07] team can work together for that period [9:08] of time I think it's how did you get to [9:10] that number and then how are you [9:12] thinking about the challenges that you [9:14] take on in each 10 to 15 year wave [9:17] because that's concrete achievable [9:21] you know you build a lot of credibility [9:22] around it the way that you've announced [9:24] those challenges [snorts] [9:26] >> well I'm curious how you guys think [9:27] about it but for us when we looked at [9:29] the grand challenges for on the 10 to 15 [9:32] year time horizon it needs to be like [9:36] when you look at it you're like I see a [9:38] path [9:39] >> right [9:39] >> not everything needs to be solved for us [9:42] to take it on in fact if everything's [9:43] solved then that feels like that should [9:45] just go [9:46] >> ambitious enough [9:46] >> yeah like you like we we have we have [9:49] some risk appetite right so we want [9:51] things where we're like there's a [9:53] credible pathway someone who is at the [9:56] helm who can do this And there's enough [9:59] ambiguity where we feel like we could [10:01] take on that risk and if we do it like [10:04] the the returns could be higher than [10:06] even expected and the way we modeled [10:09] that from you know in the biohubs is we [10:13] we have three biohubs. We have one in [10:15] San Francisco, one in Chicago, one in [10:17] New York. The one in New York works on [10:19] cell engineering. You know can we [10:21] engineer cells to go in and detect [10:23] signals, read it out or to take certain [10:25] actions. In Chicago, we're building [10:27] tissues and looking at uh tissue cell [10:31] communications within tissues. And then [10:33] in San Francisco, we're looking at deep [10:35] im imaging and uh transcrytoics. [10:38] And that work the locations are not by [10:42] accident. We also look at the partner [10:44] universities because we have folks who [10:46] come to the biohubs to do this work [10:47] collaborative, interdisciplinary [10:50] um and sort of unconstrained by the [10:53] traditional lab. But we also build off [10:56] of the labs at these academic institutes [11:00] that support the work. And so uh that's [11:02] how we sort of choose the grand [11:04] challenge and um and the locations. And [11:08] then the sort of layering and the uh [11:12] large language models and AI coming into [11:14] the picture has been so interesting [11:17] because we were already building tools [11:19] to measure interesting data building the [11:22] data sets but we didn't really know what [11:24] to do with them yet. Um and large [11:27] language models coming onto the scene [11:29] we're like wow we can make sense of all [11:31] of this now. I'm curious what you view [11:35] success as in the therapeutic realm. So, [11:38] you know, we think a lot about [11:40] understanding biology and sometimes we [11:42] bet on startups that want to unlock [11:43] completely new biological areas, [11:45] diseases where we don't know what's [11:46] going wrong. And then there's another [11:48] group of folks who kind of say, hey, [11:51] okay, now that we understand what's [11:52] going wrong, let's fix it. Um, let's [11:54] come in with a drug. Let's come in with [11:55] a new type of chemistry, a new type of [11:57] antibbody. How do you what do you think [11:59] success for the CZ Biohub looks like 10, [12:04] 20, 50 years from now in terms of the [12:08] new medicines that you've enabled? [12:09] >> We want there to be like an explosion of [12:11] a community who are building these um [12:14] just the new wave of what it means to be [12:17] deploying precision medicine. like we [12:20] like I think for rare diseases and [12:21] common diseases alike, you're really [12:24] talking about individual biology that we [12:26] sort of lump together. Um and uh they [12:31] and we often don't know how it happens, [12:33] right? We know that you have this [12:34] mutation or the worst nightmare is you [12:36] have a variant of unknown significance. [12:38] What does that even mean? [12:40] >> The horrible us. [12:41] >> Yes. Horrible. And you're like you tell [12:43] someone you kind of know something but [12:45] we don't know what it means. But if you [12:47] look at the way we've been able to look [12:49] at variants and look at single cell [12:51] transcrytoics, we're starting to be able [12:53] to say, okay, this variant actually [12:55] impacts this set of downstream cells and [12:58] then we start looking at the proteins [12:59] that get expressed and how it looks [13:02] similar or different to what a healthy [13:04] cell would look like. Then you can start [13:06] targeting. Okay, like let's look at that [13:08] as a target. And you both know the [13:11] specificity of the target you want to [13:13] build based on the ability the ability [13:15] to connect mutation to protein [13:18] expression as well as to be able to [13:20] predict off target effects. What are the [13:23] side effects? Because you also know [13:25] where else that drug will be able to [13:28] interact with the body. And and so those [13:31] are rare like and and but I really think [13:34] most diseases should be thought of as [13:37] rare diseases because each one of our [13:39] biology is different and right now we [13:42] just get lumped right we get lumped [13:44] based on age demographics ancestry if [13:47] we're lucky uh to have that level of [13:49] understanding but truly each one of our [13:52] biology is different and say like if you [13:53] look at hypertension or depression like [13:56] we kind of just go by trial and error [13:58] and saying like let's just try that drug [14:00] and see what happens. happens. But what [14:01] should really happen is being able to [14:04] precisely and accurately and quickly [14:06] treat people by looking at individuals [14:08] biology. We want to enable the basic [14:10] science and we would be thrilled if [14:12] people picked up the models that we [14:14] build to be able to build the [14:16] diagnostics, the therapeutics that need [14:18] to come. [14:20] >> You've built amazing data sets. I have [14:22] to say like I mean you may not hear the [14:24] feedback from the startup community and [14:26] the pharma community and the R&D [14:28] community but it's there because you've [14:30] committed to open source and so people [14:33] may not be they may not all be writing [14:35] papers but they are using those tools. [14:38] Um there's a startup in our portfolio [14:40] working on idiopathic pulmonary [14:42] fibrosis. The name tells you how vexing [14:44] the disease is. It's idiopathic. We [14:46] don't know why [laughter] it happens. [14:47] The IPF is named that way. And so, you [14:51] know, he was telling me that he used [14:53] your cell by gene atlases to look at [14:55] millions of single cells in patients [14:58] with disease, without disease, try to [15:00] pinpoint the fibroblasts, double click [15:02] on the fibroblasts and their gene [15:04] expression. It's try to, you know, use [15:06] that to inform, hey, where could I go [15:08] after a new drug target in this disease [15:10] that's fundamentally a strange clump of [15:12] idiopath, you know, idiopathic um [15:15] origin. So um I think there's a huge [15:18] there's a huge group of innovators who [15:20] are who love the tools, the [15:21] visualizations, the query systems and [15:24] really the software approach that you [15:25] built to making that data incredibly [15:27] accessible. So [15:28] >> cell by gene is like almost an accident [15:30] though. [15:32] >> Tell us more. [15:33] >> So do you want to share a little bit [15:34] about cellene or do you want me to [15:36] start? Well, I mean, I don't know which [15:37] part you want to get into, but I mean, [15:38] but the cell atlas work overall, I mean, [15:40] it's kind of this crazy thing that [15:42] we're, you know, here in, you know, 2025 [15:45] and there's not the kind of periodic [15:47] table of elements equivalent for [15:49] biology, right? So, that was sort of a [15:51] lot of the inspiration of it was all [15:53] right, how do we both through work that [15:55] we're going to do in the Biohub and [15:56] through other grants um be able to pull [15:58] together and standardize a format where [16:00] you can have all this data. And when we [16:02] were starting off, we didn't even [16:05] necessarily have in mind that we were [16:07] going to use that to build virtual cell [16:09] models. I think that's sort of just come [16:11] into focus as the AI work has advanced, [16:14] but that's a very exciting thing. We [16:16] should definitely spend a bunch of time [16:17] on the virtual cell models, but I'm not [16:19] sure what you wanted to get into on the [16:20] cell atlas. [16:21] >> Well, the single cell work is was one of [16:24] our first RFAS 10 years ago we started [16:26] and we were like, okay, we think this is [16:28] possible. We actually funded the [16:29] methodology for it to to standardize how [16:33] it was going to be done. So that was 10 [16:34] years ago. And we then were we seated a [16:38] few labs to start building out that data [16:41] set. But we were like there are like [16:43] millions or billions of different cell [16:45] types and different permutations. Like [16:47] how are we going to do this? And um [16:50] especially with like a burgeoning [16:52] technique. And so we ended up um seating [16:55] a few groups and they started doing work [16:57] and then they told us they had a [16:59] problem. There was a uh there was a [17:02] bottleneck in their workflow because [17:04] they couldn't annotate the data fast [17:06] enough. Um and so we built cell by gene [17:09] was an annotation tool. That's the [17:11] original source of this. So we built the [17:14] annotation tool to make it easy for [17:16] people to who are doing single cell [17:18] science to be able to annotate the data. [17:21] And then we put we put the data that we [17:22] collected publicly so people could [17:24] share. But because everyone started [17:25] using the same annotation tool, everyone [17:28] was standardized then on the same data [17:30] formats [17:31] >> and then there started being a a [17:33] community around the tool and they [17:35] wanted to share back and build the [17:37] atlas. So now after 10 years there are [17:40] millions of cells that have been built [17:42] into this uh shared resource for the [17:44] entire scientific community. We only [17:46] funded about 75% of it. Sorry that's [17:49] wrong. We've only funded 25% of it. 75% [17:53] came from the broader community saying [17:55] this is useful and there's an easy way [17:58] for us to standardize and build the same [18:01] metadata. [18:02] >> That's right. [clears throat] [18:03] >> It's like an interesting what you'd call [18:05] a network effect, right? [18:06] >> Yeah. I was going to say it sounds like [18:08] the internet. Yeah. [18:10] >> Come for the annotation, stay for the [18:12] stay for the virtual cell model. [18:13] >> Well, it was very important when we were [18:14] getting started with the work to have [18:16] everyone who was doing it have a [18:18] consistent format. So that way it could [18:20] be used and portable. And then once that [18:22] kind of took off as as the way that it [18:24] would get done, then other people just [18:25] found it valuable. [18:26] >> Yeah. And even relative to prior data [18:28] bases like GIO and and whatnot, they're [18:30] just simply not as standardized or QC. [18:32] >> Yeah. Control. [18:33] >> Yeah. [18:34] >> Let's get into virtual cells. One of the [18:36] the great challenges that the grandchild [18:37] you would focus on. Um maybe talk about [18:39] what is the promise or the hope and [18:40] maybe some of the challenges or where [18:42] we're at with it. [18:42] >> Yeah. I mean, we think that one of this [18:44] is going to be one of the most important [18:46] tools at this point is basically [18:48] building up the kind of hierarchy from [18:50] proteins to um to [18:54] just different structures within the [18:56] cell to whole to like whole like a [18:58] virtual immune system or different [19:00] levels of hierarchy. And we think that [19:03] this is going to end up being like a [19:04] very important set of tools for people [19:06] to effectively generate hypotheses for [19:09] for different science work. um you know [19:12] even before you get to the point where [19:13] you're really running full experiments [19:15] in it you can come up with some um [19:17] estimate of how that might run um it [19:19] will be useful for some of the precision [19:21] medicine type um examples that Priscilla [19:25] was talking about a few minutes ago but [19:27] we think that this is like probably one [19:30] of the most important sets of tools that [19:32] you need to build um and it's not a [19:35] single thing right so there's different [19:37] angles to to come at this from the cell [19:39] atlas data is helpful for understanding [19:42] things on a cellular level. Um, one of [19:45] the the kind of most important things [19:47] that we're doing right now, the the um [19:49] there's this this great company, [19:51] Evolutionary Scale, who actually had a [19:53] bunch of researchers who'd formerly [19:54] worked at Meta on protein folding [19:56] models, um, is joining a Biohub and and [20:00] Alex Reeves, the the, uh, leader of it, [20:02] is actually going to be the the kind of [20:03] head of the whole science program, which [20:06] is actually kind of interesting. [20:07] >> Yeah. when you think about it where it's [20:08] like you have AI and biology coming [20:10] together and really it's like an AI [20:12] person who understands biology is [20:13] running it rather than a biologist who [20:15] has some understanding of AI. I think [20:16] just kind of speaks a little bit to [20:18] where we think the the relative um [20:21] weight of these things is. But [snorts] [20:23] I mean we basically view, you know, like [20:25] Priscilla was saying with the different [20:27] biohubs. Then New York doing cellular [20:29] engineering will basically make it so [20:32] that you can have cells that can record [20:34] different things that are going on [20:35] around the body and and share that data [20:37] and then you can build that into models. [20:39] The Chicago Biohub being able to record [20:42] inflammation um and and basically study [20:44] that in order to kind of help understand [20:47] um like that. That's a that's a [20:49] different data set. We have the imaging [20:50] institute which is we just trained our [20:52] our first set of models around that [20:54] which are the first like spatial models [20:57] around understanding like the way that [21:00] that kind of cells look in different [21:02] states and eventually just like you have [21:05] this analogy on the um kind of the [21:09] industry side around language models [21:10] where you have different capabilities [21:12] and then over time you train them into [21:13] models and it gets more and more [21:14] general. [21:15] >> That's kind of the idea here. So we'll [21:17] we'll we'll build the biohubs around [21:19] grand biological challenges. The biohubs [21:22] will build tools that will generate [21:23] novel data sets. We will build models [21:26] based on those and then eventually [21:28] combine the models into an increasingly [21:30] general view of a virtual cell that will [21:32] be useful um both for scientists and [21:35] hopefully startups and companies that [21:37] are working on finding drugs which is [21:39] not our part of the whole thing but but [21:41] I think is obviously a really important [21:42] part of what needs to happen. [21:44] >> Yeah. And you know, you guys think about [21:46] risk all the time in terms of when you [21:48] make investments like I think the [21:52] promise of being able to do virtual [21:54] biology using a virtual cell model is [21:56] you can actually take on riskier ideas. [21:59] right now like grant funding can be hard [22:01] to come by and the wet lab work is [22:03] expensive and slow and it's not just you [22:06] know money it's also time and so you [22:10] have to choose something that you think [22:11] is going to have some likelihood of [22:13] success to keep your lab career going [22:16] and so it naturally lends people to take [22:20] on like some risk but not a lot of risk [22:22] because they need to make sure that they [22:24] are hitting like a certain percentage of [22:27] the time to make tenure or publish or [22:30] whatever they need to do. But if you had [22:32] a virtual cell model where you could [22:33] simulate really highquality biology, you [22:36] could actually then start testing and [22:38] tinkering on the computational side and [22:41] like ask riskier questions, things that [22:43] would have been expensive and t costly [22:45] in terms of time and resources to do in [22:47] the lab and actually see if there is [22:50] promise doing the experiments in [22:52] silicone before you make the time and [22:54] money investment in the wet lab. [22:55] >> Do you think of it kind of like a model [22:57] organism? [22:58] >> Yeah. like it's the new fruit of fly. [23:00] >> Yeah. [laughter] [23:01] I was going to ask given the complexity [23:03] of a cell, like how close um like how [23:07] accurate do you think you'll get the [23:09] model too? I mean just assuming I mean [23:11] maybe you get it to like a perfectly [23:12] accurate representation of a cell, but [23:14] like how accurate to be useful with the [23:17] virtual cell have to be? [23:18] >> I think it will obviously iterate and [23:20] get better and better because right now [23:22] we we like right now we're still just [23:23] talking about uh transcrytoics. are [23:26] expanding into different ways of looking [23:28] at the cell, but you get more and more [23:32] accuracy and but I don't think it needs [23:34] to be 100% accurate to be useful because [23:37] you just want to be able to derisk the [23:39] idea on the front end a little bit. Um, [23:42] and the more and more you derisk it, the [23:44] the more efficient it gets obviously, [23:46] but it will be useful if you even get [23:48] directional signal. And yes, I do. We do [23:51] think about it like as a a model [23:53] organism, but in a way that's like has [23:57] fidelity to the human body, like you [23:59] know, like I don't want to [24:00] >> All models are wrong. Some are useful. [24:02] >> Yeah. [24:03] >> This is hopefully has has utility on [24:05] certain access. [24:06] >> Exactly. And just like the language [24:08] models, you build in specific [24:10] capabilities. So it's not so for [24:12] example, you know, one of the models [24:13] that uh we're we're publishing is is [24:17] variant former, right? basically you [24:19] know makes it so that um it's trained on [24:22] a bunch of effectively pairs of you you [24:24] have a cell you apply crisper to it in a [24:26] place you see what comes out at the [24:28] other side so it's it basically is able [24:30] to make that kind of a prediction like [24:31] okay if you have this edit that you're [24:33] doing to to a cell what is likely going [24:35] to happen um another one of the models [24:38] is it's this diffusion model basically [24:40] you can describe a type of cell that you [24:42] would like it to simulate and it will [24:43] just produce a kind of synthetic model [24:46] of of of the cell Um, again, I mean, [24:50] it's kind of interesting because to [24:52] Priscilla's point before about how [24:54] everyone is different and and like and [24:56] different cells have have kind of um, [24:59] you know, you want to be able to [25:00] simulate these kind of rare [25:02] configurations. Um, having at least a [25:05] synthetic version of what that could [25:06] look like is interesting and then you [25:08] can test against that. The cryo model I [25:10] think is interesting because it's [25:11] spatial. So it kind of gives you a sense [25:12] of there are all these different models [25:14] that you can have that allow you to um [25:18] basically look at different kinds of [25:19] things and then you just train them in [25:20] to be increasingly general over time. [25:23] >> Wow. Very interesting. And is the is the [25:26] modeling technology basically LLMs or [25:29] like like is there is there a reasoning [25:31] model? Is it like a just [25:32] >> Oh, that's actually Yeah. I know that's [25:33] a fascinating one too and because one of [25:35] the new models um I think this one is [25:38] very early but it's um it's it's [25:40] basically the first reasoning model over [25:42] biology. So the the idea is that um [25:45] yeah, you you you effectively have these [25:47] models that that kind of simulate world [25:48] models in different ways and then you [25:50] want it to be able to not just um be [25:55] able to spit out correlations, right, in [25:57] terms of like what it's found, but [25:59] actually be able to kind of reason [26:00] through how things would would evolve [26:03] and why things would happen. Um I think [26:05] that one's quite early but it's uh but [26:08] it is interesting conceptually as what I [26:11] think is clearly going to be an [26:13] important direction [26:14] >> um in terms of how these models evolve. [26:16] >> Yeah. No it because that's what I was [26:20] thinking you know that if it doesn't [26:21] work the next question you have is why? [26:24] >> Yeah. [26:24] >> You know like [26:25] >> but I think what you find in reasoning [26:27] the the analogy [26:28] >> you're married to your hypothesis. [26:29] [laughter] [26:30] Well, yeah. Sure. Sure. Yeah. I mean, [26:31] the the uh [26:33] >> Yeah. I thought I thought you're saying [26:35] if if the reasoning model doesn't work, [26:36] why? I mean, I think the [26:39] kind of way in No, it's I mean, the [26:41] language model analogy for that would be [26:43] you need better kind of world models or [26:45] or better pre-trained models in order to [26:47] get the reasoning to be good. But, but [26:48] it's yeah, you just you build more [26:51] >> you build more capabilities into it. And [26:52] I think that there's probably an order, [26:54] too. So the work that Alex and the [26:57] evolutionary scale folks worked on is a [26:59] lot of it is protein um which is [27:02] interesting because that's at a kind of [27:06] smaller resolution obviously than the [27:07] cellular data the cell atlas but [27:10] >> part of the hypothesis is that you can [27:12] look at all these different cells and [27:13] you can kind of simulate how they might [27:15] behave but you're going to have a [27:17] somewhat shallow understanding unless [27:19] you actually have this hierarchical [27:21] understanding of what um how the sub [27:24] components of the cells are going to [27:26] interact. So [27:27] >> our view is that you basically want to [27:29] build up a state-of-the-art protein [27:32] model and then have that be a part of [27:34] the state-of-the-art cellular model and [27:36] then once you have that you build things [27:38] like the virtual immune system which [27:40] allows you to simulate um much more [27:42] complicated systems. But it's sort of [27:43] this like hierarchical approach to [27:45] building up these these uh virtual [27:47] models. That makes a lot of sense [27:49] because also as you get into [27:51] personalization, you've got like common [27:54] proteins combining into a unique cell. [27:57] So that [27:59] makes it like from a systems standpoint [28:02] that makes it like much more manageable. [28:04] That that makes a lot of sense. [28:05] Interesting. [28:06] >> Yeah. [28:07] >> Yeah. Know it's it's it's very [28:08] fascinating stuff. [28:09] >> Yeah. [28:09] >> So you guys are announcing some big news [28:11] this week. Do you want to give us a [28:12] sneak preview? Well, I the big news is [28:16] uh thinking about how we are going to be [28:19] coming together as one team. Um and you [28:22] know in the past we have done we've run [28:25] biohubs and we've done built software [28:28] we've done some AI research but all of [28:30] it has been really thinking about has [28:32] been a little bit decentralized but now [28:35] under Alex's leadership we are going to [28:37] come together as the biohub a uh an [28:41] operating philanthropy where we are [28:43] doing the science um in service of a [28:46] singular goal together and how do we [28:49] actually advance the state of biology [28:52] and research um at the intersection of [28:54] AI and biology. [28:56] >> Amazing. Alex is amazing. So, [28:58] >> yeah. No, he's great. And then and then [28:59] the other thing is that the piece that [29:00] that I mentioned earlier, which is just [29:02] Yeah. I mean, CCI has focused on a [29:04] number of different things. We've really [29:05] just found over time that we we feel [29:07] like we've been able to make the biggest [29:08] difference in science. So, we've just [29:10] kept on doubling down on it and we're [29:12] going to continue doing work in [29:13] education. We're going to continue [29:15] supporting local communities and and in [29:17] those different pieces. But going [29:19] forward, the biohub is really going to [29:21] be the main thrust of our philanthropy [29:23] and we're very excited about that [29:25] because I think that this is there there [29:26] has been you know when we started [29:28] >> the mission to see if we could help the [29:30] scientific community cure and prevent [29:32] diseases by the end of the century. I do [29:34] think with the advances in AI that [29:36] should be possible to do significantly [29:38] sooner and that is a very worthy and [29:41] important and very exciting goal that we [29:44] think we kind of have a unique place in [29:45] the ecosystem that we can help empower [29:47] others to make fast progress on that. So [29:50] there there's obviously like plenty of [29:53] advantages to decentralization from a [29:55] management communication overhead and so [29:57] forth and so like what are you trying to [30:00] add by adding this kind of new [30:03] layer/unification [30:05] on top like what what are the outputs [30:07] and then I guess what are the [30:08] complexities to that because that's um [30:10] I'm sorry to ask a CEO question. [30:12] >> No no I I mean [laughter] I'm like super [30:16] you want to go for it then I can jump [30:18] in. [30:18] >> Yeah. So there are obviously amazing [30:20] groups doing frontier AI and a lot of [30:22] groups doing uh great frontier biology [30:26] and where we think we can do uniquely is [30:29] actually tie these two together and we [30:32] are we've funded data sets we've built [30:34] data sets we're like building the [30:36] instrumentation now to be able to look [30:39] at the cell whether it's you know for at [30:41] the tissue cell communication our cryoEM [30:44] where we can look at the cell at nearly [30:47] atomic level. So we have the ability to [30:51] not only build the data sets but [30:54] actually shape and form them the way we [30:56] want based on what we see as necessary [30:59] to complement the existing body of [31:01] knowledge. And so we have amazing teams [31:04] doing that work and we're building these [31:06] AI models. And so what the reason to do [31:09] it together is then we can actually [31:12] complete the flywheel like you know the [31:15] model is looking like it has some gaps [31:17] and blind spots in this area. Okay, who [31:19] do we talk to? How do we build um the [31:22] next data set? And you know we're seeing [31:24] this in the lab like the metadata is [31:27] going to be so rich that we can feed [31:29] back into the way that we do this [31:30] modeling. Yeah. [31:38] I think it's going to be incredibly [31:39] powerful. And it's it's more than it's [31:42] more than just like, you know, writing [31:43] down a spec and saying like please [31:44] deliver this. Like these people need to [31:47] be sort of working shouldertoshoulder [31:49] and shaping uh each other's work for [31:52] this to actually um be the more and more [31:56] accurate model of how the human cell [31:57] works. [31:58] >> Well, yeah. It's so interesting because [31:59] that is exact like that's has been the [32:01] biggest surprise in the industry for us [32:03] in AI world like forget biology for one [32:06] second is that the domain [32:10] specific models have been like super [32:13] interesting like the original thesis [32:15] were like there's just some AIs are get [32:17] so smart they're going to be smarter [32:19] than everybody at everything but um [32:21] >> like on video models like every video [32:24] model is best at something but not [32:26] everything And so knowing what problem [32:29] you're solving actually turns out to be [32:31] sort of ironically very important in AI [32:34] um because you can actually get to a way [32:38] better result. Yes. [32:39] >> If you put the two together like yeah [32:41] we're we're seeing that over and over [32:43] over again uh in a way that that is [32:47] >> I would say very counterintuitive to the [32:49] whole narrative kind of going into it. [32:51] >> And in biology it used to be the or at [32:53] least you know one assumption was well [32:54] the data sets aren't on the internet. So [32:56] part of the reason you need a domain [32:58] specific model is that the data sets are [33:00] not public. you guys are kind of bucking [33:02] that trend too by creating a lot of [33:04] open- source access to the data and then [33:06] even then it sounds like you're betting [33:09] you know on the trend that we're seeing [33:10] in other industries but still there will [33:12] be nuance in how you annotate that data [33:14] curate that data [33:14] >> well and how you talk to a scientist [33:16] right like so because you have to not [33:18] only know the the data and the model and [33:20] so forth but like the conversation is [33:22] what we keep finding out ends up being [33:24] very very important right [33:25] >> so rich and so important how you [33:27] actually [33:27] >> a scientist isn't going to talk to it [33:28] like you know I talked to chat at GPT or [33:30] whatever. So, this is the fly you can [33:32] talk to. [33:33] >> Yeah. Yeah. Yeah. That that's really [33:35] that's super exciting. [33:36] >> And the user interface is actually [33:38] really important. Um you talked about uh [33:41] you guys have a founder who's using Cell [33:43] by Gene. That user interface was [33:45] intentionally designed to not need to [33:47] have a computational or really a very [33:50] deep biological background to be able to [33:52] use because you want people coming from [33:53] different fields to look at the problem. [33:56] It's like look here, help us solve [33:58] problems here. And so building that user [34:01] interface in a way where it's not a very [34:03] high barrier to entry to be able to poke [34:05] around and learn something and bring [34:07] knowledge back to your work, that's [34:09] intentional. And we're really hoping [34:11] when we build these virtual models um [34:14] that we get to a place where we can [34:17] allow a lower and lower barrier entry [34:20] for people to say like you know like I [34:22] have some knowledge about this maybe I [34:24] can contribute. Um a very pertinent [34:26] example is turns out I think immunology [34:29] has a ton to do with neuro degeneration [34:31] right but [34:32] >> seems like immunology is behind all this [34:34] so might be part of your century vision. [34:36] >> Uh so you need to be able to allow the [34:40] immunologists to come in and [34:41] [clears throat] understand neuro [34:42] degeneration and understand how their [34:44] world fits in. And so the more you lower [34:46] the barrier to entry allows people to [34:48] actually think in a sort of truly [34:51] collaborative and interdisciplinary way. [34:53] So will the Biohub grow as a team? Like [34:57] will you employ more people at the [34:58] Biohub proper or are you moving towards [35:01] more of a network model with more sites, [35:04] more labs, more communitydriven data [35:07] sets? Like which which is the thrust? Or [35:09] maybe it's both. [35:10] >> Probably a little of both. And we've [35:11] added new biohubs over time. Um and then [35:16] we're also building up more of this like [35:18] central AI team. [35:20] >> Cool. So um but I don't I think that [35:23] these organizational questions of how do [35:25] you set this up are fascinating and a [35:27] lot of our approach is sort of informed [35:30] by [snorts] what the rest of the field [35:33] is doing because I you kind of think [35:35] about science as it's this portfolio [35:37] right society has a portfolio of stuff [35:39] that it's trying to do and as in terms [35:41] of philanthropy you want to [35:44] >> be the most additive that you can be by [35:46] trying to figure out what else is [35:47] underrepresented. So science by default [35:50] is very decentralized, right? It's like [35:52] kind of the the way that granting has [35:54] worked, the way that I think scientists [35:55] by default want to work. [35:57] >> Um [35:58] >> so I think a lot of what we've found is [36:00] that figuring out ways to encourage [36:03] collaboration in um ways that otherwise [36:08] seem very simple but weren't happening [36:10] before can unlock a lot of value. So the [36:12] very first Biohub what we did there were [36:15] two kind of interesting things. One was [36:17] it was this collaboration between UCSF, [36:19] Stanford and Berkeley and there are all [36:21] these really smart people at all these [36:23] different places who previously I guess [36:25] in theory they could have figured out a [36:27] way to work together but there was not [36:28] really a formal construct for them to do [36:30] that and this just allowed a lot more [36:32] collaboration. Mhm. [36:33] >> The other one is cross- discipline. [36:36] Basically having biologists sit next to [36:39] engineers and this view that like these [36:41] two disciplines are things that need to [36:43] um and I I don't know. I mean I'm sure [36:46] you know you've seen this in a lot of in [36:47] a lot of the companies but like [36:49] >> it's there's so many interesting [36:51] >> in the companies they always like set [36:52] them apart. [36:53] >> Well, it's interesting. No, it's [36:54] interesting how many organizational [36:55] questions or problems you can fix just [36:58] by having two teams sit together, right? [37:00] It's like it doesn't matter what the or [37:01] chart is or like whatever. It's like you [37:03] guys need to sit next to each other and [37:05] until you get this thing to work and [37:08] >> um [37:09] >> that's something I really believe in. [37:10] So, [37:10] >> and you have 10 you have 10 to 15 years. [37:12] >> Well, no, it's all like communication is [37:14] such an underrated problem in general. [37:16] Yeah. Uh [clears throat] in in all kinds [37:19] of in building anything or solving [37:20] anything. So, [37:22] >> that's a that's pretty neat. [37:24] >> Yeah. Yeah. And it's it's just really [37:25] kind of simple stuff, but but I think [37:27] it's um [37:28] >> it's sort of novel as a model. And one [37:30] of the things that's so we've now copied [37:32] this [37:33] >> from the first Biohub to the Biohub [37:35] network and expanded it to other models, [37:37] but it's also just been neat to see um [37:41] other folks who are working in the field [37:43] also adopt similar models because it's a [37:45] pretty intuitive thing. [37:46] >> But you know, at some point you'll reach [37:48] the point where you know, actually it's [37:50] really good to have decentralized work [37:51] too, right? So it shouldn't be that like [37:53] we're not saying that this is like the [37:54] way that all science should work. We're [37:55] just saying that there's a space for [37:57] this. It can unlock a lot of value [37:58] because it for whatever reason hasn't [38:01] been the default. [38:02] >> Yeah. And we still rely on like [38:04] >> Yeah. There's famous like stories in the [38:06] MIT lab about that. That's how they [38:08] invented lasers and so forth is they put [38:10] a bunch of people from different [38:12] departments in the same [38:13] >> the lab. Yeah. Well, actually physics is [38:16] where we got a lot of the inspiration. [38:17] like physics has just historically been [38:20] like labs have just rallied around big [38:23] projects and big shared resources. Um [38:26] and we will you know we are relatively [38:29] centralized but we still depend on a lot [38:31] of labs who are doing sort of exact [38:33] frontier work or complimentary work to [38:36] come together to support this. There's [38:37] that. But one more thought on your [38:40] expansion question is like and maybe [38:41] this is like the uh modern AI lab. We [38:44] are not expanding like a lot of square [38:46] footage per se, but we're expanding our [38:49] compute. Um [38:50] >> yeah, [38:51] >> the research they don't want employees [38:52] working for them. They don't want space. [38:54] Yeah. They just want GPUs, [38:56] >> agents. So it's just like in a sense [38:58] that's new lab space. Um it's much more [39:01] expensive than wet lab space. [39:03] >> And you guys have always been creative [39:05] on that. Even in the last few years, [39:06] you've created ways to share access to [39:09] compute. You've enabled academic labs to [39:12] you know um I forgot the name of your [39:15] program kind of like [39:17] >> scientists and residents or something [39:18] like that [39:20] rental kind of hoteling. [39:22] >> The core of it is clusters. um you know [39:25] if you look at individual labs they'll [39:26] have like [39:28] >> like a large lab would have tens of GPUs [39:32] >> um and we were the first to really build [39:34] a large scale compute cluster um a [39:37] thousand now we're we have plans to move [39:39] to the 10,000 range and that one [39:44] requires a different type of project [39:46] obviously you're are able to ask [39:48] different types of questions [39:50] >> um and uh it's a resource that we use [39:53] but also we've invited scientists to [39:56] apply and say like what question do you [39:58] have that uh could use this amount of [40:01] resource and be able to uh stem uh sort [40:04] of seed collaborations that way [40:06] >> and so if a scientist is out there [40:07] listening like who's not employed by the [40:10] biohub or working at the biohub but [40:12] wants to collaborate with the biohub [40:14] >> that you're going to create interesting [40:17] >> interesting doors [40:19] >> to utilize the resources that's awesome [40:21] >> yeah I mean the GPUs are somewhat zero [40:22] sum Right. So that the data isn't. So [40:26] [laughter] yeah. [40:26] >> Yeah. Fair enough. [40:27] >> Yeah. So you're about to celebrate 10 [40:30] years um doing this. As as you look out [40:33] in the years to come, what else can you [40:36] tell us about either things that you're [40:37] thinking about for the future or maybe [40:39] even principles or a northstar that's [40:41] going to guide how you guys grow and [40:42] evolve going forward? [40:44] You know, it's been really interesting [40:47] in the past 10 years because I actually [40:50] spent the first few years completely [40:52] envious of people working for for-profit [40:54] companies because there's so much [40:56] clarity. Like the market will tell you [40:59] whether or not it's private or public [41:01] will tell you if you're doing a good [41:02] job. [41:03] >> If they think you're doing a good job, [41:04] >> if they think you're [laughter] [41:05] they're not always right. [41:07] >> They're not always different. But I was [41:10] still envious cuz that was I was like I [41:12] craved that feedback like am I doing a [41:14] good job? And you know 10 years in you [41:18] the reason why we're doubling down on [41:19] biology is like not only did we achieve [41:22] what we said we were going to do and [41:24] when we set out to set out on these [41:27] projects it actually delivered more than [41:29] we thought we were going to. And I was [41:31] like okay that's a signal I can latch on [41:33] to and like that's a signal I we can [41:35] really continue doubling down and doing [41:37] more of that. And so I think it's uh [41:40] continuing to tolerate the early [41:42] ambiguity when you're like, "Okay, I'm [41:44] gonna do more of this." Um and uh and [41:48] being patient, but uh uh being willing [41:52] to have a long time horizon, but be [41:54] impatient at the same time. because it's [41:56] all those iterations along the way that [41:58] have sort of allowed us to get to this [42:00] place where you know to get lucky ready [42:04] having built data data sets to take [42:06] advantage of AI and large language [42:08] models that's because of all the work [42:09] that we have been doing and so being [42:12] able to continue moving forward in this [42:14] ambiguity and sometimes lack of signal [42:17] on a big goal like I think we've sort of [42:19] set the DNA for that. [42:21] >> Amazing. [42:21] >> Oh, no pun intended. [laughter] [42:23] >> Yeah. But we get to see how many people [42:24] use the tools and the feedback. Yeah. [42:27] Yeah. [42:27] >> Yeah. You have customers which is pretty [42:29] cool. [42:29] >> Yeah. [42:30] >> For philanthropy. Like that's awesome. [42:32] >> Yeah. No, it's it's one of the fun [42:33] things about building tools is like you [42:36] kind of get to see [42:37] >> Yeah. [42:38] >> How valuable do people find the tools? [42:40] Do people use the tools in order to [42:42] publish important work? [42:43] >> Right. Right. Right. Right. Yeah. And [42:45] well, I mean our feedback is they're [42:47] awesome. [42:48] >> Feedback [42:49] and and completely unique by the way. So [42:52] like [42:53] >> the the other thing is like what would [42:55] you use if you didn't have this? It's [42:57] like there's nothing. [42:58] >> No. Yeah. It's a real it's a real kind [43:00] of void. I mean there's this whole [43:02] pipeline that that needs to exist from [43:04] accelerating basic science to funding a [43:06] lot of people to use it to then you can [43:08] get into the biotechs that basically can [43:11] start to work on on on basically coming [43:14] up with novel therapies and then you get [43:15] the pharma companies that do them at [43:17] scale. And then there's a space for [43:19] philanthropy on the other side of public [43:20] health of basically taking the the [43:23] therapies and and kind of bring them out [43:25] to everyone in the world. But this is a [43:27] a space that and that there's just going [43:29] to be a huge amount of leverage with AI [43:32] and it is um yeah it's it still seems [43:35] like there could be a lot more effort in [43:38] the space around building tools and just [43:40] accelerate the whole thing a lot better. [43:41] >> Yeah. And I do think it is the place [43:44] where you are completely unique. Right. [43:47] The other things there are other people [43:48] who can do that but there's nobody doing [43:51] what [43:51] >> that's got good good founder market. [43:53] >> Yes. Founder market fit. [laughter] I [43:54] mean if we didn't exist would it be a [43:57] problem? Yes. Like those questions uh [44:00] really land you know as a VC [44:02] >> like one of us as an engineer the other [44:04] one scientist doctor. [44:06] >> Yeah very happy this direction. [44:08] >> Yeah [44:08] >> we thank you very much not only for our [44:11] companies but for us as humans um for [44:14] working on this work. It's amazing work. [44:17] Thank you. [44:17] >> Thank you guys. [44:18] >> Thank you so much. [44:30] [music]