Colossal Biosciences Uses Quantum AI to Target 95% Automated Bioengineering

Colossal Biosciences利用量子AI力争实现95%生物工程自动化 首席执行官Ben Lamm预计，通过量子计算、AI文献解析和湿件机器人技术，基因编辑将从90%的科学发现转向95%的自动化工程。 在本次炉边谈话中，首席执行官Ben Lamm详细介绍了Colossal Biosciences如何利用全栈AI系统、比较基因组学和多模态模型来重构猛犸象、渡渡鸟和恐狼等灭绝物种的基因组。他解释了AI文献解析如何规避严重的动物福利风险，例如筛选基因变体以避免白恐狼出现失明和失聪。Lamm重点介绍了Colossal在建立全球生物库、执行多基因编辑以及深入开展公众与原住民咨询方面的工作，例如针对新宣布的恐鸟（Moa）项目与毛利人展开合作。展望未来，他设想量子计算、生物反馈回路和实验室机器人技术将推动研究从科学发现转向自动化工程，为更广泛的生态系统恢复和全球生物多样性保护提供支持。 Colossal Biosciences 在 2,000 英亩的生态保护区内成功将幼年雌狼 Khaleesi 融入两只成年雄狼的族群中，展示了其恐狼复活项目在动物福利与行为整合上的实际进展。 (0:04) 为重建猛犸象和渡渡鸟等严重降解的古代基因组，Colossal 结合多模态 AI、3D 颅骨扫描与近百种近缘鸟类泛基因组比对，精确定位控制表型特征的关键功能基因。 (2:42) 在恐狼基因工程中，AI 文献解析技术提前预警白毛基因变异会导致现代犬科动物失明和失聪，促使研究人员在活体编辑前更换为安全的替代基因变异。 (5:02) Colossal 致力于将量子计算、湿实验模拟与自动化机器人相融合，推动生物学研究从以往以发现为主的模式，转型为工程化占比高达 95% 的高确定性研发。 (13:56) 面对 2050 年前高达 50% 生物多样性可能灭绝的危机，Colossal 联合多国政府推出全球“生物保险库”计划，采用 ATAC-seq 和全基因组测序预测濒危种群的近亲繁殖与遗传瓶颈。 (17:14)

Ben Lamm of Colossal Biosciences outlines how the company combines artificial intelligence, comparative genomics, and cellular engineering to reconstruct extinct species while transferring those technologies to endangered wildlife conservation. Rather than operating strictly as a software firm, Colossal is building an end-to-end biological engineering pipeline supported by predictive computational models, animal welfare safeguards, and Indigenous partnerships. AI-Driven Reconstruction and Welfare Screening in Genomes (2:42) Colossal deploys artificial intelligence to reconstruct ancient DNA degraded over millennia. For its woolly mammoth program, the team analyzes more than 120 genomes ranging from 3,000 to 1.5 million years old, executing comparative genomics against the Asian elephant, which shares 99.6% of its DNA sequence. For the dodo, where researchers possess only a single genome, Colossal built an LLM-based multimodal framework integrating computer vision, deep learning, and natural language processing. By sequencing approximately 100 related bird species and analyzing 3D skull scans and MRIs, the team created a bird pan-genome to isolate the genes governing cranial-facial morphology. AI literature parsing also operates as an animal welfare safeguard. When analyzing two ancient dire wolf genomes aged 73,000 and 12,000 years, genomic analysis verified that dire wolves had white coats. However, AI models parsing millions of scientific papers flagged that the primary gene variant causing white coats induces deafness and blindness in modern canids. This allowed researchers to substitute a safer alternative gene variant before conducting physical engineering, protecting the health of animals housed on their 2,000-acre ecological preserve. Transitioning Biology from Discovery to Scalable Engineering (11:52) Colossal is shifting biological research from observational discovery into automated engineering. Lamm notes that while biology historically operated at 90% discovery and 10% engineering, modern computational tools have brought that ratio closer to 50/50, with a future target of 95% engineering and 5% discovery. The company's platform can execute more than 75 simultaneous genetic modifications in cell lines, utilizing a machine-learning recommendation engine that evaluates past experimental outcomes to advise scientists on whether to insert payloads, knock out sequences, or swap single base pairs. To scale this model, Colossal is pursuing the convergence of AI, quantum computing, and laboratory robotics. Lamm states that quantum computing provides the concurrent processing power required to simulate complex biological systems and wetware experiments in silicon, calculating probabilistic success rates before entering the physical laboratory. Automated robotic systems will then perform the physical lab edits, freeing human researchers to focus on higher-level data analysis rather than manual trial-and-error. Cultural Governance, Ethics, and Global Stakeholder Engagement (6:25) Because synthetic biology introduces profound ethical concerns, Lamm reports that more than 50% of Colossal's executive time is dedicated to public transparency, government outreach, and consultations. Delivering massive genetic payloads into living cells creates serious moral hazards if applied to human germline editing, necessitating strict distinctions between therapeutic disease elimination and elective trait enhancement. To guide responsible application, Colossal engages federal agencies, international bodies, and operates a dedicated $50 million conservation foundation. Cultural consultation is similarly mandatory prior to lab work. Before initiating work on the moa—a 12-foot, wingless flightless bird native to New Zealand that was hunted to extinction roughly 600 years ago—Colossal partnered with the Ngāti Wai Research Centre and Māori representatives. Because the species is revered as sacred taonga (cultural treasure), securing Indigenous consent and collaboration was required before commencing genomic research, setting a precedent for future reintroductions alongside local governments and private landowners. Deploying Genomic Bio-Vaults to Prevent Mass Extinction (17:14) Lamm acknowledges that re-engineering extinct species is vastly more expensive and complex than preserving existing biodiversity, yet traditional conservation alone cannot keep pace with forecasts indicating that up to 50% of global biodiversity could be lost by 2050. In response, Colossal and its foundation are establishing global bio-vaults in collaboration with international governments. Unlike conventional bio-banks that merely freeze tissue, these bio-vaults execute full genome sequencing and ATAC-seq (Assay for Transposase-Accessible Chromatin) to generate comprehensive population diversity maps. These genomic maps allow conservationists to monitor the biological health of small, endangered populations—such as herds of 200 rhinos—by identifying inbreeding risks and anticipating genetic bottlenecks before they manifest. Colossal aims to apply the tools refined on flagship de-extinction candidates like the mammoth, dire wolf, and dodo to active conservation initiatives, including field programs protecting the northern white rhino and the vaquita. While Colossal has demonstrated proof-of-concept steps—such as integrating the young wolf Khaleesi into an established pack and engineering 'woolly mice'—its ten-year vision relies on moving de-extinction from computational simulations to stable wild populations. The long-term viability of the initiative will depend on whether automated genomic platforms can be safely translated into living ecosystems and scaled fast enough to backstop accelerating global biodiversity loss.

Colossal Biosciences首席执行官Ben Lamm表示，物种去灭绝技术正依托人工智能、量子计算与多重基因编辑，从传统的生物学“发现”范式向概率可预测的“工程学”范式转变；该技术体系的核心不仅在于重建物种表型，更在于将其底层能力外溢至全球生物多样性防御体系中。 多模态AI重构古基因组与健康筛查 (2:42) 在去灭绝研发过程中，古代DNA样本历经数千年至百万年通常已严重降解断裂。Lamm介绍，Colossal的猛犸象项目通过分析超过120个年龄在3000年至150万年之间的古基因组，借助AI比对算法将其与共享99.6% DNA序列的亚洲象进行比较基因组学分析，从而精准分离出决定猛犸象特征的功能基因，避免盲目编辑。针对仅留存单个可用基因组的渡渡鸟，团队研发了整合计算机视觉、深度学习和自然语言处理的多模态AI系统，测序约100种近缘鸟类构建泛基因组，并结合头部3D扫描与核磁共振成像（MRI）建模，以锁定调控鸟喙和颅骨结构的基因密码。 AI文献挖掘同样被用于规避动物福利风险。通过对两具分别距今7.3万年和1.2万年的古恐狼基因组测序，团队证实恐狼毛色确为白色。然而，AI模型在系统检索数百万页既往科研文献后提示，导致该白色表型的特定基因突变会在现代犬科动物中引发耳聋与失明。基于这一预警，科研团队及时筛选并替换为安全的替代基因变异体，在确保动物机体健康的前提下达成了预期表型目标。 从科学发现向计算驱动工程闭环转型 (11:52) Lamm指出，生命科学研发范式正在发生根本性重塑：生物学研究过去通常由90%的科学发现与10%的工程化构成，现代AI的介入已将二者比例推进至50/50左右，未来则有望演进为95%的工程化与仅5%的发现，核心发现流程将被计算系统自动消化。目前Colossal的编辑平台已能在活细胞系内同时执行超过75次基因修改，包括大片段DNA插入、特定序列敲除以及单碱基置换。 为持续提升编辑效率，Colossal建立了机器学习推荐反馈机制，将真实的实验结果持续输入算法引擎，反向指导科学家选取达成目标表型的最佳工具组合。面向未来，Lamm规划了AI、量子计算与湿实验机器人的三位一体架构：利用量子计算的高并发算力在硅基环境中高精度模拟复杂生物反应，对数千种基因编辑组合及其潜在病理缺陷进行概率推演，随后指导自动化实验机器人仅针对高概率成功方案（如80%成功率）执行物理合成，从而大幅压缩试错周期。 伦理边界、原住民治理与动物福利实践 (6:25) 伴随基因递送载量与多重编辑能力的提升，合成生物学带来了不可忽视的伦理边界问题。Lamm指出，当技术外溢至人类种系编辑或在疾病治疗与功能增强之间产生分歧时，将直接冲击对人类身份的定义。为此，公司高管超过50%的时间用于公共教育、政策沟通以及与原住民群体的权益协商。例如在新宣布的恐鸟（Moa，一种约600年前灭绝的12英尺高新西兰特有无翼鸟）项目中，Colossal在启动基因组工作前，先行与Ngāti Wai研究中心和毛利人代表达成协作机制，以尊重原住民将该物种视为神圣文化宝藏（taonga）的历史权益。 在活体繁育与野外承载方面，团队强调动物福利是衡量项目的首要指标。在公司拥有的2000英亩生态保护区内，雌性恐狼幼崽Khaleesi成功融入由两只成年雄狼构成的既有群落。Lamm说明，虽然物种社会化整合属于动物行为学课题而非算法决策，但在高风险的群落引入中保障生命安全与健康，是去灭绝工程从实验室概念转化为实体生态运作的前提。 生物保险库与全球物种保护技术外溢 (17:14) 针对全球在2050年前可能损失高达50%生物多样性的严峻预测，Lamm坦言，逆向复活已灭绝物种的技术复杂度和财务成本远高于保护现存物种，但传统保护手段的推进速度已难以抵消当前的物种消亡节奏。为此，Colossal设立了5000万美元的专项保护基金，并联合多国政府推出全球“生物保险库”（bio-vaults）计划。 该计划突破了传统生物样本库仅冷冻组织标本的局限，通过引入ATAC-seq（染色质转座酶可及性测序）与高精度全基因组测序技术，绘制极危物种的种群遗传多样性图谱。该技术能够对体量极小（如仅存200头犀牛）的孤立种群进行健康监测，预警近亲繁殖并预测基因瓶颈。目前，去灭绝研发中积累的高通量基因工具已逐步外溢至北方白犀牛和小头鼠海豚等物种的野外抢救工作中；Lamm设定的十年期目标是在与地方政府、土地所有者及原住民合作将旗舰物种安全重引入生境的同时，使先进基因测序与编辑技术成为全球生态保护的基础设施。 本场对话展现了Colossal Biosciences将物种去灭绝包装为跨学科系统工程的技术野心，其以AI文献筛查与量子模拟为核心的研发闭环，正试图重塑现代遗传学工作流程。然而，从硅基模拟成功到复杂生态系统的物理野化之间仍存在巨大鸿沟，75次以上多重编辑的潜在脱靶累积效应、以及在野生环境中建立长期自维持种群的可行性，仍是该技术路径有待长期实证检验的核心疑问。

## Background 本视频由 Ai4 频道于 2025 年 8 月 24 日发布在 YouTube 平台。内容为 Colossal Biosciences 联合创始人兼首席执行官 Ben Lamm 与 NBC 新闻副主编 Jason Abbruzzese 展开的炉边对谈。视频简介指出，本次对谈深入探讨了 Colossal 如何利用 AI 创新推进合成生物学、实现突破性发现，解析了物种灭绝复活（de-extinction）的运作机制及其对 AI 时代的思考。 ## TL;DR 对谈全面梳理了 Colossal Biosciences 在物种灭绝复活与合成生物学领域的核心进展。Ben Lamm 分享了恐狼幼崽合群、长毛小鼠培育及渡渡鸟多模态基因组研究的最新成果，阐明了从古 DNA 拼接、细胞工程到动物抚育的全系统工程架构。他指出 AI 与未来量子计算正在推动生命科学发生范式跃迁，使研发从漫长的经验探索转变为高达 95% 的可预测工程化设计，并能在湿实验前有效排查基因缺陷。此外，Lamm 还阐述了与新西兰毛利原住民的深度协作、未来十年的物种重引入蓝图，以及应对全球生物多样性危机的“生物金库”（Bio Vaults）计划。 ## Atmosphere & style 对谈节奏紧凑、氛围活跃且充满前沿科技的探索感，主持人幽默的切入与 Ben Lamm 坦诚、高密度的专业分享形成了绝佳互动。现场既有将科幻概念落地为现实工程的振奋与自豪（如轻松谈及恐狼毛色与长毛小鼠），又始终保持着对动物福利、生殖系编辑伦理与原住民文化主权的清醒敬畏。整场交流让听众深切感受到生物技术工业化进程的澎湃动力，以及面对全球物种消亡危机时兼具理想主义与工程务实精神的紧迫感。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=sQ0ZliED2gA&t=4s) 物种灭绝复活并非单纯依赖资本或单点技术，而是涵盖古DNA样本获取、计算生物学与AI、细胞系基因编辑以及专业动物抚育的完整全系统工程体系。 - [02:35](https://www.youtube.com/watch?v=sQ0ZliED2gA&t=155s) AI 在破碎古DNA基因组重建与比较基因组学中起决定性作用，既能完成猛犸象百余个基因组的特异性比对，也能通过多模态架构攻克渡渡鸟单基因组数据匮乏难题。 - [04:49](https://www.youtube.com/watch?v=sQ0ZliED2gA&t=289s) 多模态 AI 具备海量科学文献交叉挖掘能力，在恐狼毛色研究中成功预警并规避了伴发失明与耳聋的缺陷突变，从算法源头保障动物福利。 - [06:25](https://www.youtube.com/watch?v=sQ0ZliED2gA&t=385s) 高通量基因编辑工具已能实现单次75个以上位点的精准递送与修改，随着技术演进，生殖系基因编辑与增强技术的伦理边界亟待社会理性共识与规范。 - [10:37](https://www.youtube.com/watch?v=sQ0ZliED2gA&t=637s) 灭绝复活技术落地深刻牵涉原住民文化主权与国际合作，例如新西兰恐鸟项目需在毛利部落文化宝藏（Taonga）治理框架下推进。 - [11:52](https://www.youtube.com/watch?v=sQ0ZliED2gA&t=712s) 基于历史实验数据构建的闭环 AI 推荐系统，能够根据目标表型自动匹配最佳基因编辑路径与工具组合，大幅降低实验试错成本。 - [13:56](https://www.youtube.com/watch?v=sQ0ZliED2gA&t=836s) AI、量子计算高并发算力与实验室自动化的结合，将推动生命科学从漫长试错的“科学探索”范式彻底转变为虚拟模拟主导的“工程化设计”范式。 - [15:46](https://www.youtube.com/watch?v=sQ0ZliED2gA&t=946s) 公司高度聚焦于猛犸象、渡渡鸟与袋狼等旗舰物种，年内重点展示鸟类基因组学技术演进，暂不盲目扩张新增物种线。 - [16:28](https://www.youtube.com/watch?v=sQ0ZliED2gA&t=988s) 未来十年的核心愿景是将复活物种放归自然栖息地并持续量化其生态修复效应，同时将基因与细胞工程工具普惠化应用于全球濒危物种救援。 - [17:10](https://www.youtube.com/watch?v=sQ0ZliED2gA&t=1030s) “生物金库”计划超越传统的被动组织冷冻，通过 ATAC 测序与全基因组分析主动评估种群健康并预警遗传瓶颈，打造抵御2050年生物多样性危机的科技防线。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=sQ0ZliED2gA&t=4s) 恐狼幼崽近况与灭绝复活的全系统工程 对谈开场时，主持人 Jason Abbruzzese 打趣道这可能是他主持过的“最离奇但又最酷的 AI 炉边对话”。随后，他询问了备受公众瞩目的“恐狼”（Direwolf）复活项目的最新进展。Ben Lamm 分享了一个令人振奋的“幼崽更新”（pupdate）：就在前一天，团队成功将雌性幼狼 Khaleesi 融入到了两只年龄稍长的雄狼 Romulus 和 Remus 的狼群中。三只恐狼目前生活在一片占地 2,000 英亩的生态保护区内，相处非常融洽。Lamm 坦言，将一只雌狼引入年长雄狼的领地一度让人十分紧张，但动物福利是团队一切工作的核心，这次顺利合群标志着巨大的成功。 谈及如何将这种十到二十年前还只存在于科幻小说中的概念变成现实，Lamm 笑称“这全都是 George Church 的‘错’”，是这位知名遗传学家最先构想利用基因工程、AI 与计算生物学来复活物种，并将其反哺于现代医疗和自然保护。Lamm 强调，很多人以为 Colossal 只是融了很多钱的纯 AI 公司，但实际上他们必须从零构建一整套完整的工程系统体系。 ```mermaid flowchart TD accTitle: Colossal Biosciences 全系统工程架构 subgraph S1["数据获取与计算"] A["古DNA核心<br/>(Ancient DNA Core)"] --> B["计算生物学与AI核心<br/>(AI & Comp Bio Core)"] end subgraph S2["基因编辑与细胞工程"] B --> C["近缘物种比对与靶点设计"] C --> D["细胞系基因工程改造<br/>(Genetic Engineering)"] end subgraph S3["繁育与生态保护"] D --> E["动物抚育团队<br/>(Animal Husbandry)"] E --> F["生态保护区野化与监测"] end ``` - **恐狼幼崽近况**：Khaleesi 与雄狼 Romulus、Remus 成功在 2,000 英亩的生态保护区组建狼群。 - **系统工程核心环节**：涵盖古 DNA 样本获取核心、AI 与计算生物学核心、细胞系基因编辑平台以及专业的动物抚育团队。 --- ### [02:35](https://www.youtube.com/watch?v=sQ0ZliED2gA&t=155s) AI 在古 DNA 重建与渡渡鸟多模态研究中的核心角色 Lamm 明确指出，没有人工智能，Colossal 的工作根本无法开展。古 DNA 样本在出土时往往极其破碎，AI 首先用于古基因组的拼接与重建。例如在猛犸象项目中，团队掌握了超过 120 个基因组样本，时间跨度从 3,000 年前一直延伸到 150 万年前。借助 AI 进行比较基因组学分析，团队能够在亚洲象与猛犸象高达 99.6% 相似度的基因组中，精准定位“究竟是哪些基因造就了猛犸象”，避免盲目乱改。 然而并非所有项目都有上百个基因组可用。在渡渡鸟（Dodo）项目中，科学家手里只有“唯一一份”基因组序列。为了攻克这一瓶颈，团队开创性地构建了融合计算机视觉、自然语言处理与大语言模型（LLM）的多模态 AI 架构： ```mermaid flowchart LR accTitle: 渡渡鸟项目多模态AI基因识别流程 subgraph Inputs["多源异构数据采集"] G["唯一渡渡鸟基因组 +<br/>约100种相关鸟类测序"] S["全鸟类头骨 MRI 与 3D 扫描"] L["海量鸟喙形态学文献库"] end subgraph AIPlatform["多模态大模型框架"] P["泛基因组构建 (Pan-genome)"] M["CV + NLP + 深度学习系统"] end subgraph Output["编辑指导"] T["锁定影响颅面与鸟喙形态的候选基因"] end G --> P S --> M L --> M P --> M M --> T ``` - **猛犸象基因组对比**：拥有 120 余个年代跨度从 3,000 年到 150 万年的基因组，通过 AI 筛选 0.4% 的关键特异性差异。 - **渡渡鸟多模态攻关**：在单基因组限制下，通过测序约 100 种相关鸟类、进行全头骨 3D/MRI 扫描并结合海量文献，反推颅面发育基因。 --- ### [04:49](https://www.youtube.com/watch?v=sQ0ZliED2gA&t=289s) 数据洪流与动物福利：AI 如何识别潜在基因缺陷 在回答“为什么人类研究员无法靠肉眼完成这些工作”时，Lamm 指出最根本的阻碍在于数据体量与多维交叉关联的复杂性。他以恐狼的毛色研究为例：团队分析了分别距今 73,000 年和 12,000 年的两组恐狼基因组，数据明确证明恐狼实际上是白色的——这印证了《权力的游戏》中的设定，却引来社交媒体上网民的争论，因为有人在 2015 年画过红色的恐狼插画，Lamm 幽默地回应：“可恐狼在 2015 年又没活着。” 更关键的是，原始计算数据指示了一条能让毛色变白的基因编辑位点。但当多模态 AI 系统全面检索并解析海量犬类和现生狼的医学文献时，AI 发出了红色预警：该突变变异体在现代犬科动物中极易伴发失明和先天性耳聋。如果仅靠传统数据分析，团队极有可能直接实施编辑并导致动物健康灾难。依托 AI 预警，团队最终找到了另一条既能呈现白色外观又能保障听力与视力健康的基因突变路径。 - **恐狼毛色事实**：通过 7.3 万年前与 1.2 万年前的古基因组分析，证实恐狼确为白色。 - **避免基因缺陷**：AI 文献挖掘揭示了某一白毛基因与失明、耳聋的强相关性，促使团队更换为安全的编辑方案，保障了动物福利。 --- ### [06:25](https://www.youtube.com/watch?v=sQ0ZliED2gA&t=385s) 基因编辑伦理、技术溢出与公众透明沟通 随着基因工具的飞速发展，Colossal 目前已经能够实现单次同时进行 75 个以上的精准基因编辑，并输送大分子载荷。这种级数的技术飞跃引发了公众对伦理边界的担忧。Lamm 认为，社会正逐渐走出对 AI“又哭又闹、悲伤恐惧”的情绪阶段，步入理性的接受与落地期。但在合成生物学领域，当这类技术与 AI 以及未来的量子计算交汇时，人类必须直面深刻的道德难题——例如是否允许对人类胚胎进行可遗传的生殖系基因编辑（Germline Editing）。消除严重遗传病与进行智力增强之间的界限究竟该划在哪里，将从根本上重新定义“何为人类”。 Lamm 也谈到了公众沟通的痛点。他曾在社交媒体上开玩笑说，自己每天都被问及一模一样的 10 个问题，完全可以训练一个大模型来替自己回答；结果真有网友做出了一个“猛犸象聊天机器人”。事实上，Colossal 不仅将研发的所有技术向生态保护领域开放，还成立了规模达 5000 万美元的保护基金。为了让外界理解这些跨越国界的技术，团队将一半以上的时间花在了科普、与政府立法者沟通以及与原住民社区的合作上。 - **高通量编辑能力**：依托模型指导，单次实验可完成 75 个以上的基因位点编辑及大片段递送。 - **生态保护基金**：专门设立 5,000 万美元基金，将灭绝复活技术直接赋能濒危物种保护。 --- ### [10:37](https://www.youtube.com/watch?v=sQ0ZliED2gA&t=637s) 新西兰恐鸟（Moa）项目与原住民文化认同 在谈及最新公布的恐鸟（Moa）复活项目时，Lamm 纠正了公众的常见误解：恐鸟在遗传学上虽然不能直接算作恐龙，但它是人类历史上曾并存生活过的“最接近恐龙形态”的生物。这是一种生活在新西兰、身高可达 12 英尺（约 3.65 米）的巨型无翼走禽，在大约 600 年前被毛利人猎杀至灭绝，曾对当地森林冠层与植被生态调控起着关键作用。 由于恐鸟属于毛利文化中的珍贵宝藏（Taonga）和神圣物种，Colossal 在正式启动该项目前，专程前往新西兰与 Ngāi Tahu 研究中心及毛利部落展开深入协作，全面阐明技术边界并获得原住民的许可与支持。Lamm 强调，灭绝复活绝非单纯的实验室科学，它深刻牵涉到跨国界协作与原住民文化主权。 - **恐鸟生物特征**：高约 12 英尺、完全没有翅膀的新西兰本土巨鸟，约 600 年前灭绝。 - **文化合作机制**：与毛利人及 Ngāi Tahu 研究中心紧密合作，遵循原住民文化准则推进古 DNA 复原。 --- ### [11:52](https://www.youtube.com/watch?v=sQ0ZliED2gA&t=712s) 前沿探索：AI 推荐系统、量子计算与湿实验自动化模拟 在展望 AI 应用的前沿时，Lamm 透露团队正在测试一套用于指导基因编辑的闭环反馈推荐系统。在细胞系构建过程中，插入大片段、小片段敲除或改变单个碱基都有不同的编辑模式与效率差异。团队利用前期积累的海量实验数据训练模型，当科学家想要达到特定的生物学表型时，推荐系统能直接给出最优的编辑工具和路径组合建议。 展望更长远的未来，Lamm 对 AI、量子计算与机器人技术的交汇寄予厚望。尽管行业里总调侃“量子计算永远还有两年”，但他坚信三者的结合将彻底颠覆生物学研究范式：系统能够在虚拟环境中运行高保真的“湿实验模拟”（Simulated Wetware），筛选出成功率超过 80% 的编辑方案，再指挥实验室机器人自动完成物理操作，科学家最终只需对产物进行验证分析，从而成百倍地加速生命科学研发。 - **智能编辑推荐系统**：基于历史细胞系编辑数据，自动为特定目标匹配最佳基因编辑工具。 - **终极科研范式**：量子计算 + AI 模拟湿实验，结合实验室机器人实现高度自动化的高通量基因工程。 ### [13:56](https://www.youtube.com/watch?v=sQ0ZliED2gA&t=836s) 量子计算与 AI 结合：从“科学探索”迈向“工程化设计” 主持人 Jason Abbruzzese 顺着 Ben Lamm 多次提及的量子计算话题切入，询问量子与 AI 的结合为何如此令人兴奋。Lamm 解释道，虽然量子在通信、国防或加密领域有广泛且吸引眼球的应用，但对于 Colossal Biosciences 而言，其最大价值在于提供超大规模的“并发算力”（concurrent horsepower），能够在湿实验（wetware）环境之前进行真正的大规模实验室级模拟。 通过将过往验证有效的实验数据持续回流至 AI 系统，团队能够在量子模拟环境中并发测试数以亿计的基因编辑组合，并结合海量文献快速排查潜在缺陷。例如，针对恐狼（dire wolf）基因组分析时，系统可以预测将某些古代基因导入现代狼体内是否会导致耳聋或失明等遗传缺陷，最终直接输出一份指导报告，明确告知实验人员“去尝试这五种最优组合即可”，从而免去了盲目试错的过程。Lamm 指出，生命科学过去是“90% 依靠探索发现，10% 依靠工程实现”，目前已演进至约五五开，而未来借助 AI 和量子计算，将彻底转变成“95% 依靠工程设计，仅 5% 依靠探索发现”，因为绝大部分的发现阶段将在数字系统内部完成。 ```mermaid flowchart TD accTitle: AI 与量子驱动的基因工程研发流程 accDescr: 展示从历史数据与文献输入到量子并发模拟，再到精准工程实验的完整流程。 subgraph Input["数据与文献输入"] D1["湿实验验证数据回流"] D2["全量基因与科学文献"] end subgraph Simulation["AI 与量子并发模拟"] S1["高并发基因编辑概率计算"] S2["致病与表型缺陷预测\n(如排查恐狼致盲/致聋基因)"] end subgraph Execution["实验室精准工程落地"] E1["生成精准实验报告\n(推荐 Top 5 实验路径)"] E2["湿实验高效验证与工程化合成"] end Input --> Simulation Simulation --> Execution ``` - **算力需求**：生物基因组合空间极其庞大，需要量子计算提供海量的高并发算力支持多维度实验模拟。 - **范式转移**：从过去依赖漫长野外/湿实验试错的“科学探索”（Scientific Discovery），转向由算法先导预测、精准实施的“科学工程”（Scientific Engineering）。 --- ### [15:46](https://www.youtube.com/watch?v=sQ0ZliED2gA&t=946s) 研发管线进展与旗舰物种路线图 谈到公司接下来的管线与物种规划时，Lamm 回顾了近期引发外界广泛关注的成果，包括成功培育出的长毛小鼠（Woolly Mice）——他还幽默地提到两只小鼠“奇奇与蒂蒂（Chip and Dale）目前生活得非常好”——以及恐狼相关的基因组研究突破。 面对关于近期是否会有全新物种发布的疑问，Lamm 明确表示，今年不会再宣布新的物种项目，团队正高度聚焦于既定的几大旗舰物种（猛犸象、袋狼、渡渡鸟）。不过，鸟类基因组学（avian genomics）研究团队在技术上取得了多项重要突破，预计在年底前会向公众展示相关阶段性进展。虽然这并不意味着渡渡鸟会在今年诞生，但底层基因工具的演进依然令人振奋。 - **近期成果**：长毛小鼠“Chip and Dale”状态良好，恐狼基因组研究获积极反响。 - **年度规划**：本年度不再公开新增物种，重点展示鸟类基因组学技术突破，全力推进现有核心旗舰项目。 --- ### [16:28](https://www.youtube.com/watch?v=sQ0ZliED2gA&t=988s) 十年愿景：物种重引入、生态评估与底层技术普惠 在描绘 Colossal 未来十年的发展蓝图时，Lamm 表示，如果一切按计划推进，届时公司将已成功复活所有核心旗舰物种。更重要的是，他们将与原住民群体、私有土地所有者以及各国政府深度合作，把这些物种真正放归到原有生态栖息地中，并持续监测与评估它们对当地生态系统带来的积极改变。 与此同时，Lamm 希望复活物种过程中衍生出的各类底层生物技术，能够实现从保育试点向全球常态化应用的跨越。这些技术不仅用于拯救极度濒危的物种（如北部白犀牛、加湾鼠海豚 Vaquita），还应全面开放并赋能整个生物保护与医疗生态，成为全人类可共同利用的通用工具。 - **十年目标矩阵**： - **物种复活与野化**：实现旗舰物种复活，并与政府及原住民社区合作完成生态重引入。 - **生态效应量化**：持续评估复活物种对碳循环、植被与生态恢复的实际正向影响。 - **技术普惠化**：将保育繁育、基因编辑与细胞工程技术规模化应用于全球濒危物种拯救。 --- ### [17:10](https://www.youtube.com/watch?v=sQ0ZliED2gA&t=1030s) 终结灭绝与“生物金库”：现代版诺亚方舟的守护机制 针对“这是否意味着灭绝时代的终结”这一提问，Lamm 直言，从经济和技术成本来看，复活并重建一个灭绝物种的代价，远远高于直接保护现存物种。然而现实非常严峻：到 2050 年，全球可能损失高达 50% 的生物多样性，传统保护手段的推进速度已经赶不上物种消亡的速度。 为此，Colossal 及其基金会正在全球范围推进“生物金库”（Bio Vaults）计划。这并非传统意义上“仅把生物组织塞进冰箱冷冻”的被动生物银行，而是与多国政府合作的主动防护工程：在采集样本后，立即进行全套 ATAC-seq（染色质可及性测序）和全基因组测序，构建群体遗传多样性图谱。即便某个区域尚存 200 头犀牛，系统也能深入分析其近亲繁殖程度、评估健康状况，并精准预测未来的“遗传瓶颈”。Lamm 坦言，尽管作为一家去灭绝公司的 CEO 说出“拯救现有物种远比复活灭绝物种更划算”听起来有些奇特，但改变公众与消费者行为极为缓慢，人类迫切需要建立起这样一座高科技的“诺亚方舟”作为兜底屏障。 - **生物多样性危机**：预计在 2050 年前，地球可能失去 50% 的生物多样性，现行保护手段滞后于消亡速度。 - **生物金库（Bio Vaults）工作模式**： - **超越传统冷冻**：全面结合 ATAC 测序与高深度全基因组测序。 - **遗传图谱构建**：实时评估种群健康状况与近亲繁殖风险。 - **遗传瓶颈预警**：提前预测种群基因衰退，为精准繁育与基因拯救提供决策依据。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=sQ0ZliED2gA&t=4s) 最离奇但又最酷的 AI 炉边对话 — Jason Abbruzzese > [00:04](https://www.youtube.com/watch?v=sQ0ZliED2gA&t=4s) 这全都是 George Church 的‘错’ — Ben Lamm > [02:35](https://www.youtube.com/watch?v=sQ0ZliED2gA&t=155s) 究竟是哪些基因造就了猛犸象 — Ben Lamm > [04:49](https://www.youtube.com/watch?v=sQ0ZliED2gA&t=289s) 可恐狼在 2015 年又没活着。 — Ben Lamm > [06:25](https://www.youtube.com/watch?v=sQ0ZliED2gA&t=385s) 又哭又闹、悲伤恐惧 — Ben Lamm > [11:52](https://www.youtube.com/watch?v=sQ0ZliED2gA&t=712s) 量子计算永远还有两年 — Ben Lamm > [13:56](https://www.youtube.com/watch?v=sQ0ZliED2gA&t=836s) 去尝试这五种最优组合即可 — Ben Lamm > [13:56](https://www.youtube.com/watch?v=sQ0ZliED2gA&t=836s) 90% 依靠探索发现，10% 依靠工程实现 — Ben Lamm > [13:56](https://www.youtube.com/watch?v=sQ0ZliED2gA&t=836s) 95% 依靠工程设计，仅 5% 依靠探索发现 — Ben Lamm > [15:46](https://www.youtube.com/watch?v=sQ0ZliED2gA&t=946s) 奇奇与蒂蒂（Chip and Dale）目前生活得非常好 — Ben Lamm > [17:10](https://www.youtube.com/watch?v=sQ0ZliED2gA&t=1030s) 拯救现有物种远比复活灭绝物种更划算 — Ben Lamm ## Mentioned Ben Lamm, Jason Abbruzzese, Colossal Biosciences, NBC News, Ai4, George Church, 恐狼 (Dire Wolf), 猛犸象 (Woolly Mammoth), 渡渡鸟 (Dodo), 恐鸟 (Moa), 袋狼 (Thylacine), 长毛小鼠 (Chip and Dale), 北部白犀牛 (Northern White Rhino), 加湾鼠海豚 (Vaquita), 《权力的游戏》 (Game of Thrones), Ngāi Tahu 研究中心, 毛利人 (Māori) ## Video description Ben Lamm, Co-Founder/CEO, Colossal Biosciences Jason Abbruzzese, Assistant Managing Editor, NBC News Abstract This session will explore how Colossal makes use of AI innovation to advance synthetic biology and power breakthrough discoveries. Learn how the de-extinction process works, and get an inside look into how Co-Founder and CEO Ben Lamm thinks about the AI era. [Source](https://www.youtube.com/watch?v=sQ0ZliED2gA)

[0:04] Jason Abbruzzese: 最离奇但又最酷的 AI 炉边对话 [0:04] Ben Lamm: 这全都是 George Church 的‘错’ [2:35] Ben Lamm: 究竟是哪些基因造就了猛犸象 [4:49] Ben Lamm: 可恐狼在 2015 年又没活着。 [6:25] Ben Lamm: 又哭又闹、悲伤恐惧 [11:52] Ben Lamm: 量子计算永远还有两年 [13:56] Ben Lamm: 去尝试这五种最优组合即可 [13:56] Ben Lamm: 90% 依靠探索发现，10% 依靠工程实现 [13:56] Ben Lamm: 95% 依靠工程设计，仅 5% 依靠探索发现 [15:46] Ben Lamm: 奇奇与蒂蒂（Chip and Dale）目前生活得非常好 [17:10] Ben Lamm: 拯救现有物种远比复活灭绝物种更划算

[6:25] Ben Lamm: temper tantrum phase [6:25] Ben Lamm: acceptance phase of grief [8:52] Ben Lamm: taonga [11:52] Ben Lamm: wetware [13:56] Ben Lamm: concurrent horsepower [0:04] Ben Lamm: pupdate [8:52] Ben Lamm: Woolly Mammoth chatbot [15:46] Ben Lamm: woolly mice [17:14] Ben Lamm: bio-vaults [17:14] Ben Lamm: Noah's Ark

[0:04] Uh Ben, thank you for being here for [0:06] what uh I'm hoping will be maybe the [0:08] weirdest uh uh fireside [0:10] >> applications of AI using de-extinction. [0:13] >> Yeah, I mean deextinction a word. I'm [0:15] not even sure when that first came onto [0:17] the scene, but and we just saw a good [0:20] teaser trailer, [0:21] >> which we had all the rights for the [0:22] >> all the I'll check with our legal [0:24] department. [0:24] >> I don't want to get like yeah some ex [0:27] nerds on it. as an NBC Universal [0:29] employee. I that's that's good to hear. [0:31] Um well, when people hear deextinction, [0:33] I think obviously science fiction is the [0:35] first thing that comes to mind, but [0:36] we're now starting to see it in the real [0:38] world. The first thing I need an update [0:40] on is how are the direwolves doing? [0:42] >> Good. We just gave our first uh pupdate [0:45] yesterday. Uh and so we uh we we showed [0:47] uh we I think we launched it with USA [0:49] Today and a few others. I think NBC we [0:51] sent it to as well, which I think [0:52] they're covering, which is exciting. uh [0:53] and uh we just integrated Khesi uh our [0:56] girl into the pack with Romulus and [0:58] Remis and it went well and so one of the [1:00] things that you know this isn't [1:01] obviously AI related but animal welfare [1:04] is like key to everything that we do and [1:06] so the integration was a little bit [1:07] nerve-wracking right it's like we're [1:08] we're integrating a female wolf into a [1:11] pack of two males that are older and it [1:14] went great and they're all interacting [1:16] great they're on a 2,000 acre ecological [1:18] preserve and uh now they're completely [1:20] integrated as a pack which is awesome [1:22] And I I I think they're a great real [1:24] world example of of what you guys are [1:26] doing, which is, you know, a kind of [1:29] science that even as of 10 to 20 years [1:32] ago was again relegated to to science [1:34] fiction. How did you go from [1:38] whatever the first idea was for doing [1:40] this to having actual real living [1:43] >> direwolves [1:44] on a on a preserve? [1:47] >> We uh well, this is really all George [1:48] Church's fault. So it it was his idea in [1:51] terms of like how we could use genetics [1:52] and and AI and computation biology to [1:54] bring back species as well as use those [1:57] technologies for you know healthcare as [1:59] well as conservation. And what's [2:00] interesting is uh we had to build the [2:03] entire system. So sometimes people are [2:04] like well you guys raised a lot of money [2:06] uh are you guys just an AI company? How [2:07] are you guys looking at this? But [2:09] fundamentally we had to build the entire [2:10] system. So we've got an AI core, we've [2:13] got a computational biology core, we've [2:14] got an ancient DNA core that goes out [2:17] and procure samples, and then we have to [2:19] use all these technologies to take those [2:20] samples, compare them to the closest [2:22] living relatives, and then engineer them [2:24] into cell lines using genetic [2:26] engineering. Uh, and then we have an [2:28] animal husbandry team. And so there's [2:30] all these different components that we [2:31] had to go build for this systems model [2:33] for deextinction and conservation. [2:35] >> And, and you touched on this briefly, [2:36] but let's get like down into the [2:38] nitty-gritty. How is AI [2:40] being used in this process? [2:42] >> We couldn't do like to your point uh we [2:45] couldn't do our job without AI. So we're [2:47] using AI in the uh ancient DNA [2:50] reconstruction of the genome. So when we [2:51] get genomes are extremely broken. Uh [2:54] like so for the mammoth project we have [2:56] over 120 genomes ranging from 3,000 [2:58] years old to 1.5 million years old. We [3:01] then use AI to actually do all the [3:03] comparative genomics uh which is pretty [3:05] interesting. We have to try to [3:06] understand what genes made a mammoth a [3:07] mammoth. Because for us, if you look at [3:09] an animal and you look at an Asian [3:11] elephant, which is 99.6% the same as as [3:14] a woolly mammoth, we want to identify [3:16] the genes that really made a mammoth a [3:18] mammoth, right? And so we don't want to [3:19] go in and just edit everything. We want [3:21] to be really thoughtful about it. And so [3:22] we're using AI in that process, too. [3:24] We're even using AI. Uh this is kind of [3:26] weird. Not every single project we have [3:28] unlimited genomes for. Like the Dodo [3:30] project, we only have one genome. [3:32] >> The Dodo project being [3:33] >> the Dodo project. We're bringing back [3:34] the [3:34] >> bring bring back the job. Yeah. And so [3:35] we only have one genome and so we don't [3:37] have like 120 plus genome. So, so it's [3:39] hard from a comparative genomics [3:40] perspective and uh so this is [3:43] interesting where we're looking at kind [3:44] of like converging AI and using kind of [3:46] this mo multimodal approach where we're [3:48] taking vision uh artificial intelligence [3:51] and even natural language processing and [3:53] pulling into kind of this kind of uh uh [3:55] system that we built using kind of this [3:57] LM framework from research where we're [3:59] insert we're pulling in all the data [4:01] from all this bird beak morphology uh [4:04] and then we're also doing all the [4:05] sequencing of about a hundred different [4:07] species of birds and that that are [4:09] related and some closely and some not [4:11] closely related to the dodo. And then we [4:13] actually did MRIs and we did full 3D [4:15] scans of every single uh bird in in [4:18] their skull that that in that kind of [4:20] group so that we could build a pan [4:22] genome so then we could use our AI [4:24] system to understand okay how do we [4:26] isolate genes that we think would change [4:28] the cranial facial morphology of the [4:30] bird. So, so we have to so we're [4:33] building these kind of like uh systems [4:34] that kind of uh cross vision, cross deep [4:37] learning and cross uh even uh LLMs where [4:40] we're pulling in data from all the [4:41] scientific research to really understand [4:44] you know where we where we can even uh [4:46] start our editing for the for that [4:47] because we only had one GM. Well, I [4:49] think you kind of just touched on where [4:50] I wanted to go next, which was, you [4:51] know, why couldn't humans be doing that [4:54] kind of thing to ask it very simply, [4:56] what is challenging about what you're [4:59] doing that AI is good at? [5:02] >> Yeah. Well, I mean, it's it's just a ton [5:03] of data, right? And so, um, uh, and it's [5:07] hard for us to parse it. So, so, for [5:09] example, when we did the, uh, [5:10] comparative genomics for the direwolf, [5:13] we had two direwolf genomes. One was [5:15] 73,000 years old, one was 12,000 years [5:17] old. Uh, we saw that, you know, [5:18] direwolves were white. We found out that [5:20] they were white. We looked at it. We [5:21] looked the day they're white. Game of [5:23] Thrones got it correct. Uh, lots of [5:25] people on social media yelled at us [5:27] because they saw a picture that someone [5:28] painted of them as red and they're like, [5:29] "No, we saw a picture that someone [5:31] painted and they were red uh in 2015, [5:33] right?" And I was like, "Well, they [5:34] weren't alive in 2015." And so, but what [5:36] we found, and this is where AI was [5:38] really helpful to us, what we found is [5:40] is is in the data, it's like, "Go make [5:42] this genome, this edit in the genome to [5:44] make them white, right? And here's the [5:46] gene specifically that that made them [5:48] white. But then when we pulled in all [5:49] this literature, we actually found in [5:52] dogs and in other wolves, some of those [5:54] same gene variants actually caused [5:56] blindness and deafness. And that's not [5:58] something that that the computational [6:00] analysis showed us in the data. And so [6:02] pulling in all this extra data and [6:04] having AI parse it, it kind of gave us [6:06] this red flag of saying, hey, if you go [6:08] make this change, there could be an [6:09] animal welfare issue to it. And so we [6:11] then use that data said okay we're going [6:12] to have the same outcome but we're going [6:14] to use this gene variant over here [6:15] because it's safe and and that's [6:17] something that you know the data of us [6:19] just looking at the raw data without [6:21] leveraging AI to kind of look at all the [6:23] ingestable data uh would we would have [6:25] known. [6:25] >> Yeah. Uh now you you kind of just [6:27] touched on where humans start entering [6:29] the process and the decisions they have [6:31] to make. Um, what kind of work does your [6:33] company do and do you do thinking about [6:36] any of the moral or ethical or societal [6:38] ramifications of these technologies? [6:40] Obviously, I think people hear them, [6:42] including myself, and they start going, [6:45] >> "That's that's that's scary stuff. What [6:47] else could it be used for beyond doing [6:50] things like [6:51] >> it is it is crazy." And for us as it [6:54] relates to synthetic biology, we're [6:56] working with the federal government. [6:58] We're working with indigenous people [6:59] groups. We're working with a lot of [7:00] different collaborative groups around [7:01] the world to actually think about how we [7:03] build frameworks for leveraging [7:04] synthetic biology. Being that AI is core [7:07] to our business, but we're not just [7:08] purely an AI company. Um, I wouldn't say [7:10] that that we're out there saying this is [7:12] like ethical frameworks for AI, and [7:13] there's a lot of people that are doing [7:15] that. Thank god we're past the temper [7:17] tantrum phase of AI where everyone's [7:18] just sad and scared and weird. Um, and [7:20] now we're in the point where people are [7:21] like, okay, now we it's here. We're in [7:23] the acceptance phase of grief, right? [7:25] And so like we're like actually working [7:26] through it and imple implementing it [7:28] which is awesome. Uh but but for us we [7:30] have to think about kind of like what [7:31] are the what are the results of taking [7:33] the insights from data from the AI [7:36] systems in the data that we get to apply [7:38] to whether it's animal welfare in the [7:39] case of like the direwolves or in cases [7:41] of human genome engineering. And so if [7:43] you look at Colossal, we're making 75 [7:45] plus edits at a time um using kind of [7:48] our system and and and our models and [7:50] that actually help us design these [7:52] guides of how we are delivering large [7:54] payloads into cells and other labs are [7:56] doing too. So that's quite a big step [7:58] function, right? Right. And so so when [8:00] we think about that and we think about [8:01] then the application of those [8:02] technologies to humans uh and you know [8:05] there's now companies wanting to do [8:06] editing in embryos for humans you know [8:08] there are ramifications to the [8:10] advancement of these technologies and AI [8:12] and then eventually quantum are just [8:14] going to massively accelerate that. And [8:16] so on one hand you can say we want to [8:17] cure disease states on the other hand is [8:19] like okay how much germline editing do [8:21] we want to do in the gene pool that can [8:23] be passed on the next generation. And so [8:25] those are like really hard not just [8:28] regulatory issues but moral hazard [8:30] issues. It's like where do we draw the [8:32] line? If we can eradicate a disease in [8:34] in a child do we do it or do we say you [8:36] know if we can you know increase [8:38] intelligence do we do it? So these are [8:40] the things that I think that really [8:42] synthetic biology genome engineering and [8:44] uh AI and eventually quantum like that [8:47] kind of convergence is going to really [8:49] change uh and make us really think about [8:51] what it means to even be human. And when [8:53] you try to communicate to to to people [8:54] outside of the the company or even [8:56] outside of AI. Yeah. Um you had a funny [8:58] post on Twitter recently about how you [9:01] >> I get a lot of feedback from my post. So [9:02] I I make less posts these days. [9:04] >> Yeah. But it was a good post because it [9:06] was saying like you get the same 10 [9:08] questions over and over. You could train [9:09] an LL. [9:10] >> Oh my l [9:10] >> to answer them. I mean [9:11] >> so you know what's interesting I said [9:12] that and then someone made a woolly [9:14] mammoth chatpot. I didn't make that. [9:16] Like someone made a woolly mammoth [9:17] chatpot. They're like okay. And then [9:18] they they made it. Yeah, I I did make a [9:20] joke saying that like we get asked to [9:22] because it's interesting, right? It's [9:23] like we we'll talk about like [9:24] conservation. Like all the technologies [9:26] we make are available for conservation. [9:28] We're deploying technologies into [9:29] conservation. We even raised a [9:31] foundation with $50 million for for [9:33] conservation. And then people will be [9:34] like, "You guys should do stuff for [9:36] conservation." It's like, "Well, there's [9:37] a button on our website says [9:38] conservation. We have a foundation. We [9:40] we mention it every time we open our [9:41] mouths." So I was like, "Maybe I'll just [9:43] make an LLM and like like we'll just [9:45] take down all of our stuff and you can [9:46] just talk to a bot." Well, and I was [9:48] going to say, how crucial is it to be [9:49] communicating with the public right now [9:51] about like how this technology works, [9:53] the safeguards being put around them, [9:55] the thought that is being put into it? [9:56] >> I think it's really important. And so, [9:58] and so one of the things that when I [10:00] started the company with George, I [10:01] thought we would spend a lot of time in [10:02] the lab, maybe some like, you know, [10:04] dungeons and like subbins like the [10:07] Smithsonian, but I and maybe some field [10:09] work, but I never thought that we would [10:10] spend, you know, half or more than half [10:12] our time in the field, right? And so, we [10:15] spend a lot of time educating. We spend [10:16] a lot of time with uh uh lawmakers. We [10:18] spend a lot of time with governments [10:19] because the work that we're doing [10:21] doesn't, you know, it transcends [10:23] borders. It it transcends geopolitical [10:26] uh frameworks that that currently are [10:28] existing, right? You know, like mammoths [10:30] don't see borders. You know, do saw a [10:32] little bit of border because they were [10:33] an island, right? But some of these [10:34] species don't. And so we have to think [10:36] about that. And then we're also spending [10:37] a lot of time uh with even indigenous [10:39] people groups, right? And so we just [10:41] announced our MOA project. And you know, [10:43] before we ever would work on the MOA [10:45] project, we went and worked with the [10:46] Nahu Research Center and the the Maui [10:49] people because that's part of their [10:50] Tonga and their sacred species. [10:52] >> I was going to say, what is the MOA [10:53] project? [10:53] >> So, so the MOA is awesome. Uh it is uh [10:57] you know, I'd argue it's the closest [10:58] thing. It's it's not genetic. I'll be [11:00] careful about saving. It is not [11:02] genetically the closest thing to a [11:03] dinosaur, but it is probably the closest [11:05] thing that humans live next to a [11:06] dinosaur. It's a 12- foot flightless [11:09] bird that didn't have wings. uh that was [11:11] native to to New Zealand. And uh it's [11:13] awesome. It was it was really helpful to [11:15] the uh vegetational overgrowth of of [11:17] canopies. Uh the Mauy people actually [11:19] hunted it to extinction um you know a [11:21] couple about 600 years ago and uh we [11:24] actually have the DNA to bring it back. [11:26] But even in that process, not only [11:27] educating uh lawmakers and and others in [11:31] the field and journalists and the public [11:32] at large, we have to go work with the [11:34] indigenous people groups to make sure [11:36] they understand the application of the [11:38] technologies, where the limits are and [11:39] also get their permission and buy in. So [11:41] we have to spend a lot of time, I'd say [11:44] half of our time is spending on [11:45] education and transparency, which I [11:48] think is really uh important even if I [11:50] occasionally make jokes on Twitter. [11:52] >> Fair enough. Uh the applications of AI I [11:54] think is is what is most interesting [11:57] about what you guys are doing, right? Um [11:59] you know a lot of people here I think [12:01] know that AI is more than just a chat [12:03] GPT or a bot. [12:05] >> Um but talk to me about some of the [12:07] applications you guys are starting to [12:09] pursue on the bleeding edge. What are [12:11] you excited about starting to really [12:13] focus on? So one of the things that that [12:15] we are just now testing uh and it's [12:17] working so far uh but it's still the [12:19] early days is how can we uh create [12:22] feedback loops from so we make a lot of [12:24] cell lines right and we make a lot of [12:26] different edits and we're using [12:27] different combinations edits and [12:28] different edits uh whether you're [12:30] inserting a big piece of DNA or just [12:32] knocking on a small piece or just [12:33] changing one little rung on that you [12:35] know twisted ladder of DNA uh we we like [12:38] different ones have different editing [12:40] modalities and different editing [12:41] modalities have different efficiencies [12:42] and so we're combining all that in in a [12:44] pretty interesting way. Um, but if we [12:46] just do that randomly and and one cell [12:48] line works and one cell line doesn't, [12:49] you know, how do we know if you know [12:51] what we're doing is the right way? And [12:53] so so feeding all this data into a [12:55] feedback loop because we're now creating [12:57] enough data that we can actually train a [12:59] system. So we're now actually have a [13:00] recommendation system that's actually [13:02] saying, okay, for this type of edit, if [13:04] you want to produce this type of result, [13:05] you need to use these tools, which is [13:07] really interesting. and where I think [13:09] it's going. Uh, and this is where, you [13:11] know, I'm hopeful, you know, we hear [13:13] quantum is two years away every two [13:14] years, but I'm very excited about [13:16] quantum and AI because I I think that [13:18] eventually we can do simulated wetwware [13:20] or or lab work uh using AI and [13:23] eventually robotics. And so if we can [13:26] empower AI and quantum and AI into [13:28] eventual robotics, I'm not talking [13:30] about, you know, the Tesla robots like [13:31] the laundry stuff. School, but [13:33] ultimately I think it'd be really [13:34] interesting to say, you know, how can we [13:36] run simulated uh experiments all, you [13:39] know, in a system and then say, okay, [13:41] here's like the 80% probability if you [13:44] make these edits, then have robots do [13:46] that in in there so that then the [13:47] scientists can just do an analysis of [13:49] that final product, right? And so that's [13:51] where it's going and that's where we're [13:53] spending a lot of time at kind of the [13:54] intersection of of AI and robotics and [13:56] then simulation and and hopefully [13:58] getting ready for quantum. [13:59] >> All right. Well, if if you brought up [14:01] quantum a few times, then we're going to [14:02] have to go there. Uh walk me through why [14:05] quantum and AI uh is is something to get [14:08] excited about. What about quantum? [14:10] >> It's in you know the theory, right? And [14:13] people are doing this. There's there's [14:14] great companies out there doing this, [14:15] but the application of being able to run [14:18] so many simultaneous experiments in a [14:20] true simulation, right? You hear about [14:22] like the fact that like we're in [14:23] simulation, some people believe that. [14:25] Um, some people don't. Uh, but but [14:27] ultimately being able to simulate a lab [14:29] environment in a wetware environment. [14:31] And there's many examples of using [14:32] quantum for communication and defense [14:34] and others that are much more [14:35] interesting probably than than what we [14:36] want to use it for. But the ability for [14:38] us to, you know, look at like what is [14:40] the probab pro probabilistic uh success [14:42] rate of these edits using the data that [14:45] we already know is working that we've [14:46] been feeding back into our AI system and [14:48] then being able to simulate all these [14:50] experiments at once so that we can [14:52] simulate all these different gene [14:53] combinations pulling in the literature [14:55] right so that we can look for uh issues [14:57] like you know the genes that could cause [14:58] deafness or blindness from the [14:59] direwolves in in and if you implement [15:01] them in today's wolves those are the [15:03] things that I think are super [15:04] interesting because then you can [15:06] essentially get that data and insight in [15:08] a report saying okay go try these five [15:10] experiments right versus us just being [15:12] in the field of discovery and I really [15:14] think that that empowers us to move from [15:16] the world of scientific discovery to [15:19] scientific engineering like right now [15:20] you know it used to be like 90% [15:22] discovery 10% engineering I'd say it's [15:24] more like 50/50 or 6040 now but I think [15:27] in the future it's going to be like 95% [15:29] engineering you know because of these [15:31] tools and technologies and 5% discovery [15:33] because the discovery is going to happen [15:34] in the system [15:35] >> that's well and and for lack of a better [15:37] term, you just need more horsepower. [15:38] >> Yeah, we just need more horsepower and [15:40] we need concurrent horsepower [15:41] >> and that's what Quantum can offer. [15:43] >> Yeah. For us, it can offer a lot more [15:45] other things like security to other [15:46] people. Yeah. [15:46] >> Sure. For you guys. Um, in terms of, you [15:49] know, what else you got coming down the [15:51] pipe? Is there anything else? Any other [15:52] animals that we can expect in the the [15:54] near future? [15:55] >> Yeah, we surprised the world with the [15:56] woolly mice. We surprised and Chip and [15:59] Dale doing quite well. Uh, and then um [16:01] we surprised the world with with the [16:02] direwolves and you know, it was [16:04] interesting feedback. It was great. And [16:06] then uh uh no, we're very focused on our [16:08] flagship species. There's nothing else [16:09] to announce yet. Um but uh I think that [16:13] we'll show progress before the end of [16:14] the year. Not a dodo, but we'll show [16:16] progress on our aven genomics group uh [16:18] before the end of the year, which we're [16:19] pretty excited about. We've had a couple [16:20] cool breakthroughs there. Um and then uh [16:22] but we're not going to be announcing any [16:23] more species uh this year. [16:25] >> Yeah, fair enough. And [16:26] >> well, only only two this year. [16:28] >> Yeah. And in in 10 years, where are you [16:30] guys hoping to be? Like what does this [16:32] company look like down the line? [16:33] >> Yeah. I I I think in 10 years if if [16:35] everything goes right, we've brought [16:37] back uh all of our flagship species. [16:39] We've successfully worked with [16:40] indigenous people groups, private land [16:42] owners, and governments to put them back [16:43] into their environments. We're testing [16:45] the applications and the imple in the [16:47] and the implications of of of how [16:49] they're affecting the environment [16:50] hopefully in a positive way. And then I [16:52] hope that the technologies that we've [16:53] developed have have expanded beyond just [16:56] conservation where they're not just [16:58] being field tested like with the [17:00] northern white rhino and vikita type [17:02] projects but actually being implemented [17:04] across the board for everyone. So that [17:05] that that would be where I hope we are [17:08] in 10 years and I think that we're [17:09] currently on pace to do it [17:10] >> and and would you say that [17:14] that means the end of extinction? Like [17:16] is that is that is that even in the [17:17] realm? Yeah, I mean I look it is more [17:19] expensive and this is it's a great [17:21] question. It is way more expensive to [17:23] bring back something uh and re-engineer [17:26] and rebuild extinct species for today [17:28] than just preserve today, right? Like I [17:30] I I would hope that we would change, you [17:32] know, people's views. But we are in in a [17:34] world where 50% of biodiversity could be [17:36] lost between now and 2050. That's not [17:39] that far away. And so another issue, [17:41] another uh implication that that we're [17:43] we're doing right now is the [17:44] implementation of uh bio vaults. So [17:47] we're working with several governments [17:48] around the world to actually uh uh [17:50] protect their species uh take their [17:52] samples and and not just put them in a [17:54] fridge like some people are doing in bio [17:56] banking but actually go do full analysis [17:58] on and do uh all the ATAC on them do all [18:00] the different uh genome sequencing build [18:02] a population diversity map so that we [18:04] say okay yeah there may be 200 rhinos [18:06] over here but are they healthy what are [18:08] they inbreeding where where are there [18:10] going to be forecasted genetic [18:12] bottlenecks so so that's a big [18:13] initiative that that we have and our [18:15] foundation has uh to do that so that [18:17] it's almost like a Noah's Arc type [18:19] mission. Um because I I do think, you [18:21] know, it's going to take a long time to [18:22] change it's going to be much harder to [18:24] change consumer behavior uh than to [18:27] advance the technologies. And so um you [18:29] know I still think it's cheaper and [18:31] better to save what we've got than what [18:32] we do. And that's weird coming from the [18:34] guy that runs a deextinction company, [18:35] but I think it's much better to do that [18:36] than what we do. But but I think we need [18:38] a safeguard and and we just don't have [18:40] it because we know that while modern [18:42] conservation works, it's not working at [18:44] the speed at the speed of which were, [18:46] you know, eradicating [18:47] >> Absolutely. Ben, thank you so much. This [18:48] was wonderful. I think we just exactly [18:50] nailed time. [18:50] >> Oh, awesome. Thanks. [18:52] [Applause]