Jackson Lab Professor Uses OpenAI Codex to Model Immune Cell Systems

杰克逊实验室教授利用 OpenAI Codex 模拟免疫细胞系统 人工智能推理模型的指数级提升将使医学界能够模拟人类生物系统，在十年内治愈所有重大疾病，并在 15 年内逆转人类衰老。 Derya Unutmaz 博士详细介绍了其将 Codex、GPT-5.5 及推理系统等 AI 模型整合到免疫学和生物医学研究中的历程。他展示了简化流式细胞术分析、模拟 T 细胞受体通路以及自动化 CRISPR 靶点筛选的定制应用。展望未来，Unutmaz 倡导开发数字孪生和自主 AI 智能体集群，以彻底革新药物研发、高度个性化癌症疗法以及细胞工程。 生物系统的巨大复杂性——数万亿个组成部分以及数十亿个同时发生的反应——超出了人类的认知能力，使得 AI 驱动的模拟成为科学探索的必要途径。 (0:10) 侧重推理能力的 AI 模型（如 o1-preview、o3 和 Pro 模型）通过实现跨领域科学推理和可靠的假设生成，标志着一个重大转折点。 (2:15) 借助 Codex，研究人员无需传统的软件工程团队，即可快速构建定制的领域专属应用和交互式工具。 (5:44) Unutmaz 展示了通过 Codex 构建的定制工具，可用于流式细胞术分析、免疫细胞 T 细胞受体信号传导模拟以及动态细胞环境建模。 (9:17) 生物工程的终极目标是可编程生物学，即将人体细胞视同软件代码，可以对其进行设计、编译并应用于临床治疗部署。 (18:25)

Dr. Derya Unutmaz argues that biological complexity exceeds the limits of unassisted human cognition, requiring artificial intelligence models, digital twins, and autonomous agent systems to model living systems and accelerate therapeutic discovery. The Evolution from Pattern Matching to Reasoning (0:10) Dr. Derya Unutmaz dates his conviction that biology demands computational tooling to the early 1990s, when he recognized that trillions of cellular components and billions of simultaneous reactions overwhelm unaided human analysis. While deep learning enabled parallel processing of large biological datasets and milestones like AlphaFold demonstrated computational modeling of structures, early large language models such as GPT-4 remained largely restricted to literature synthesis and routine text generation rather than reliable experimental prediction. A substantive shift occurred with the release of reasoning-focused systems. When Unutmaz evaluated OpenAI's o1-preview in September 2024 by prompting it to translate an analogy between a survival video game and tumor-immune dynamics into a wet-lab protocol, the system designed a nuanced experimental setup. Subsequent frontier systems, including GPT-5.5 Pro, demonstrated an ability to forecast unpublished laboratory results with full accuracy, indicating to Unutmaz that models are beginning to replicate the biological intuition typically acquired through decades of experimental experience. Codex-Driven Software for Wet-Lab Analysis (9:17) Rather than relying on generic software development pipelines that require months of engineering, Unutmaz utilizes AI coding tools such as Codex to program tailored scientific software in minutes. Demonstrating an internal tool for flow cytometry—a method where individual fluorescently tagged cells pass through lasers to measure up to 20 markers—Unutmaz built a web interface that parses multi-parametric data into 2D dot and contour plots, profiles CD4+ helper and CD8+ cytotoxic T cells, and automatically calculates optimal antibody panels for targeted subsets like TH17 or memory cells. He has also applied generative coding to cellular simulation and genetic engineering. Using Codex, Unutmaz built a browser-based T-cell receptor simulator that models real-time molecular phosphorylation, transcription factor activation, and downstream responses across variable antigen affinities and PD-1 checkpoint inhibition. For genetic engineering, he developed a native macOS application in Swift that queries public genomic databases for genes such as CD4, isolates and ranks 20-to-22 nucleotide CRISPR target windows across 2,000-base-pair sequences, and compiles batch target libraries ready for direct submission to commercial oligonucleotide synthesis vendors. Digital Twins and Simulated Clinical Trials (21:13) Unutmaz defines a biological digital twin as a comprehensive computational replica of an individual patient, integrating host genetics, environmental exposures, circulating metabolites, hormones, and trillions of gut microbes. In his assessment, contemporary medicine relies on broad-population treatments—such as statins that benefit only a fraction of recipients—because standard diagnostics cannot compute personalized biological variance. By simulating these networks, clinicians could model individual patient responses to targeted immunotherapies, balancing anti-tumor efficacy against dangerous autoimmune tissue destruction. Unutmaz forecasts that high-fidelity digital twins could compress clinical trial evaluation periods from 5 to 10 years down to 5 to 10 days, predicting that exponential compute scaling will make all diseases treatable within a decade and enable human age reversal within 15 years. While acknowledging skepticism from peers who observe historical progress moving linearly, he argues that critics underestimate exponential advances in computing and reasoning models. Automating Discovery Through Science 2.0 (29:17) Looking forward, Unutmaz outlines a transition to 'Science 2.0,' where the traditional cycle of manual experimental design and analysis is replaced by networked AI agents coupled with automated robotic laboratories. In this framework, coordinator and specialist agents formulate hypotheses, screen out low-probability experimental paths in simulation, guide robotic physical execution, and instantly interpret incoming data streams, shifting the principal investigator's role toward high-level strategy. Addressing concerns regarding workforce displacement in research laboratories, Unutmaz cites Jevons paradox, contending that dropping the computational and experimental cost of science to near zero will expand overall demand for bioengineering. Because humanity currently understands only an estimated 10% of total biological systems, he argues that agent-driven discovery will enable engineers to design synthetic cell therapies and novel biological tissues on personal workstations rather than downsizing research output. Unutmaz's forward projections rely on continuous exponential leaps in AI reasoning, software generation, and computational capacity to simulate the trillions of interacting components inside the human body—infrastructure that current computing power cannot yet support.

杰克逊实验室（The Jackson Laboratory）免疫学家 Derya Unutmaz 教授指出，生物系统包含数万亿个相互作用的部件，远超人类认知极限，而具备推理能力的 AI 模型与代码生成工具正在重塑生物学研究，推动实验科学向预测性与可编程医学转型。 生物复杂性与前沿推理模型的转折 (0:10) Unutmaz 的学术生涯始于 1990 年代初从医学院毕业，当时他意识到生物体内存在数万亿个相互作用的组件与数十亿次生化反应，仅凭人类大脑无法解析大量疾病为何难以治愈。这一认知促使他早期尝试编写代码建模生物功能，并一路关注深度学习与 AlphaFold、ChatGPT 等里程碑的发展。他认为生命科学必须依赖具备高并行处理能力的 AI 模拟来进行复杂数据分析。 在模型测试历程中，早期的 GPT-4 虽然大幅缩短了推荐信撰写和文献梳理的时间，但无法用于预测实验结果。转折点发生在 2024 年 9 月测试 o1-preview 时，Unutmaz 设计了一个跨学科提示词，要求模型将海岛大逃杀生存游戏的机制类比为免疫细胞对抗肿瘤，并据此设计科学实验。该模型输出的深度与洞察力让他意识到推理模型已经能够用于假设生成与实验设计；随后在 GPT-5.5 Pro 上，模型对其实验室未发表的复杂实验得出了 100% 准确的预测，展现出资深科学家需要 30 年实验经验才能积累的生物学直觉。 利用 Codex 构建专用科研软件与仿真工具 (5:44) Unutmaz 将自身形容为“Codex 成瘾者”，由于并非传统软件工程师，他过去即便编写简单的贪吃蛇游戏也需要数月，而现在则通过 Codex 在数分钟内生成专用科研工具。在流式细胞术（Flow Cytometry）工作流中，细胞需经过荧光抗体标记并通过激光阵列以分析各亚群；针对多达 20 种标记的复杂数据，他使用 Codex 独立开发出一款能够进行二维门控、生成等高线图、分析 CD4+ 与 CD8+ T 细胞并自动推荐最佳抗体面板的 Web 应用。 除了数据分析，他还利用 AI 构建生物仿真模型。Unutmaz 演示了在浏览器端运行的 T 细胞受体（TCR）模拟器，研究人员可调节配体结合亲和力、抗原剂量与共受体信号，实时观察下游磷酸化模式、转录因子激活以及肿瘤清除或自身免疫损伤等结果，并能加入 PD-1 抑制剂或小分子药物以观察信号通路的实时阻断。在一项测试中，他将 AI 生成的《自然》（Nature）期刊模拟免疫图谱封面直接输入 Codex，由模型自动解析图像实体并生成了包含可点击细胞亚型、抗 PD-1 机制开关以及生物学年龄滑块的动态交互模拟器。 CRISPR 基因设计与可编程生物学愿景 (18:49) 在基因组编辑领域，Unutmaz 回顾了自己从 25 至 30 年前开发早期基因编辑方法到如今使用 CRISPR 的演进。以 T 细胞表面的 CD4 基因为例，其序列长度可达 2000 个核苷酸，但 CRISPR 工具需要精准锁定其中 20 至 22 个核苷酸的目标区域。为此，他使用 Swift 编写了一款原生 macOS 应用程序（且正在开发 iPad 版本），输入基因名称即可自动从公共数据库获取 DNA 原始序列，筛选并排序目标区域，随后直接复制至商业合成供应商的订购单中，该软件还支持输入基因列表进行批量文库设计。 这一系列自主工具开发指向了生物医学工程的终极目标——将人体细胞视作可编写的高级代码。Unutmaz 认为，未来的生物学家将像软件工程师编写程序一样，在计算机上设计、模拟并部署治疗性细胞行为，建立一套整合型的“生物学 Codex”。 数字孪生驱动的超个性化医疗与临床试验变革 (21:13) Unutmaz 将“数字孪生”（Digital Twin）定义为对整个人体生物系统的全量模拟模型，涵盖遗传基因、数万亿肠道微生物、循环代谢物、激素和环境暴露。由于人体复杂性，传统医学长期采用“千人一方”的模式，导致类似他汀类药物仅对少数服药者有效。而在癌症等具有极高遗传异质性的疾病中，部分智能靶向药仅对 1% 携带特定突变的肺癌患者有效，澳大利亚已有计算机科学家利用 ChatGPT 和 Grok 针对宠物犬肿瘤突变设计个性化 mRNA 疫苗。 通过数字孪生进行虚拟测试，研究人员能够预测不同治疗方案的疗效与副作用，防止免疫疗法出现细胞衰竭或因过度激活引发组织损伤。基于 AI 模拟能力，Unutmaz 预测新药研发与临床试验周期有望从 5 至 10 年压缩至 5 至 10 天；他同时预测，指数级增长的技术进展有望在未来 10 年内使所有主要疾病变得可治疗，并在 15 年内实现人类年龄逆转。尽管目前全球算力尚不足以模拟人体数万亿组件，但他预计超级智能与算力基础设施将在未来 5 至 10 年内弥补这一缺口。 科学 2.0：自主 Agent 协作与研发范式迁移 (29:17) 在被称为“科学 2.0”（或 3.0）的未来图景中，科学家花费数周规划实验、数月手动分析数据的传统循环将被终结。取而代之的是专用 AI Agent 网络：假说生成 Agent 负责构思并在数千种虚拟条件下进行模拟筛选，剔除低概率路径；随后高置信度方案被送入自动化机器人实验室进行批量物理实验；专门的数据分析 Agent 则实时处理返回数据并反馈给协调 Agent，形成持续自主的科研闭环，而首席研究员（PI）的角色则转变为高层战略指挥。 针对科研岗位是否会被替代的担忧，Unutmaz 引用杰文斯悖论（Jevons Paradox）指出，科学研究成本的暴跌与效率的剧增将成倍放大对生物工程的需求。由于人类目前仅理解约 10% 的生物系统，AI 工具将赋能数千名生物工程师在电脑前设计全新细胞类型、人造组织和靶向疗法，使原本耗时数年的药物发现任务缩减至数小时。 Unutmaz 强调，生命科学实验有 95% 至 98% 最终以失败告终，这种严苛训练培养出高度的抗挫力与持续迭代的主观能动性。随着 AI 试验的资金与时间成本趋近于零，科研与各行业的核心竞争力在于好奇心与行动力。不过，其对于“十年内治愈所有疾病”及“数天内完成临床模拟”的激进预测，仍高度依赖于未来 5 至 10 年内超智能算力能否真正攻克整个人体多尺度生物系统的模拟难题。

## Background 本视频于2026年8月3日由OpenAI官方YouTube频道发布。视频记录了杰克森实验室（The Jackson Laboratory）教授、免疫学专家兼医学博士Derya Unutmaz与OpenAI开发者体验主管Romain Huet的深度对谈，展示如何使用Codex自主开发生物科研工具，并展望AI助力科学家仿真实验的未来愿景。 ## TL;DR 在访谈中，免疫学家Derya Unutmaz展示了他如何借助Codex完全从零搭建起专业级单细胞流式细胞分析套件、T细胞受体信号通路仿真器，以及原生Swift编写的CRISPR文库设计工具。他深入剖析了前沿模型（从o1到GPT-5.5）展现出的深邃科学推理能力与顶级专家级的科研直觉，并阐述了利用AI构建生物“数字孪生”的终极构想，该设想有望将长达5至10年的临床试验压缩至数天之内。面对科研与技术范式的剧变，他强调试错成本的断崖式下降正在赋予每个人前所未有的创造力，而在AI重塑人类文明的黄金时代中，能动性与好奇心是唯二关键的品质。 ## Atmosphere & style 整场对话洋溢着极富感染力的极客激情与科学乐观主义，既有硬核前沿生物医学的严谨探讨，又穿插着关于算力催更和跨界实验的默契笑谈。当Derya回忆起前沿推理模型完美命中未公开实验结果、近乎展现出沉淀三十年的科研直觉时，其真情流露的震撼与动容令观众深切共鸣。整场节目节奏流畅而热烈，不仅是一次顶尖构建者开发工作流的精彩展示，更是一段见证AI如何赋能人类认知边界、引发生命科学深刻范式转移的精神洗礼。 ## Bulletin - [22:19](https://www.youtube.com/watch?v=4sexN3yE8xg&t=1339s) 医学免疫学家 Derya Unutmaz 提出借助多组学数据与免疫系统图谱构建人体“数字孪生”（Digital Twin），预言未来 5 到 10 年内超级算力将把长达数年的临床试验压缩至 5 到 10 天，实现高保真虚拟临床验证。 - [26:46](https://www.youtube.com/watch?v=4sexN3yE8xg&t=1606s) 前沿模型展现出实验推演的质变飞跃，Unutmaz 透露 GPT-5.5 具备了媲美在实验室摸爬滚打 30 年资深科学家的“隐性直觉”，以 100% 的准确率成功预测了复杂的未公开生物学实验结果。 - [09:17](https://www.youtube.com/watch?v=4sexN3yE8xg&t=557s) 得益于 Codex 的自然语言编程支持，非软件工程背景的科研人员已能自主构建出处理超 10 万单细胞事件的原生流式细胞分析软件，以及可自动化批量设计导向 RNA 文库的原生 Swift CRISPR 设计工具。 - [29:17](https://www.youtube.com/watch?v=4sexN3yE8xg&t=1757s) 生命科学正迈向由多智能体协作与机器人湿实验闭环构成的“科学 2.0/3.0”时代，杰文斯悖论表明实验门槛与成本的骤降将推动生物工程与合成生物学需求呈爆发式增长。 ## Key takeaways - [00:10](https://www.youtube.com/watch?v=4sexN3yE8xg&t=10s) 人体系统包含数万亿构件与海量非线性生化反应，复杂性远超人力极限，促使免疫学家在 90 年代初便意识到生命科学的终极突破必须依赖 AI 构建因果模型。 - [02:15](https://www.youtube.com/watch?v=4sexN3yE8xg&t=135s) 在早期测试 o1-preview 时，通过将《堡垒之夜》等大逃杀游戏机制隐喻为免疫系统清除肿瘤的逻辑，模型展现出的深层科学推理能力让科研人员确信 AI 已跨越知识整合走向机理假设推演。 - [05:44](https://www.youtube.com/watch?v=4sexN3yE8xg&t=344s) 以往生物学专用软件开发周期漫长且高度依赖商业分析套件，而借助 Codex，科学家可直接通过自然语言交互手搓出满足个性化实验流式细胞分析需求的全功能工具。 - [12:28](https://www.youtube.com/watch?v=4sexN3yE8xg&t=748s) 现代工程如飞机制造在实装前均需经过严密的气动仿真，而生物学目前缺乏先验模拟工具；AI 的远大愿景是从单个虚拟细胞逐步跃迁至整个人体的系统级数字孪生。 - [13:26](https://www.youtube.com/watch?v=4sexN3yE8xg&t=806s) 在浏览器端完全由 Codex 构建的 T 细胞受体（TCR）动态仿真器，能模拟调节配体亲和力与剂量，实时推演关键信号通路的激活抑制与转录因子磷酸化修饰谱图。 - [16:01](https://www.youtube.com/watch?v=4sexN3yE8xg&t=961s) 静态视觉图像可借助 AI 迅速转化为功能性沙盒应用，Codex 仅凭一张《自然》封面概念图即可自动构建出支持模拟 PD-1 靶向阻断与年龄衰老变量的动态细胞互作系统。 - [18:25](https://www.youtube.com/watch?v=4sexN3yE8xg&t=1105s) 基于 Swift 打造的原生 macOS/iPadOS 工具能够快速调取 DNA 序列数据库并计算脱靶评分，批量导出 CRISPR 导向 RNA 文库并直连商业合成机构完成湿实验闭环。 - [21:13](https://www.youtube.com/watch?v=4sexN3yE8xg&t=1273s) 传统临床受限于复杂性往往千人一面，数字孪生旨在整合个体的全基因组、蛋白质、代谢物及微生态，推动医疗范式从“治疗孤立疾病”转向“治疗患病的生命整体”。 - [22:19](https://www.youtube.com/watch?v=4sexN3yE8xg&t=1339s) 数字孪生将颠覆传统临床试验，通过在虚拟模型中以 5 到 10 天的时间替代漫长的人体测试，为患者量身定制实现 100% 疗效与 0% 副作用的专属疗法。 - [24:39](https://www.youtube.com/watch?v=4sexN3yE8xg&t=1479s) 肿瘤治疗展现出极高突变异质性，AI 虽已被用于为个体（甚至宠物犬）定制 RNA 疫苗，但如何精确调控免疫系统杀伤力并规避致命的自免疫风暴仍是关键壁垒。 - [26:46](https://www.youtube.com/watch?v=4sexN3yE8xg&t=1606s) 许多科学家因短期幻觉便过早放弃 AI 工具，却忽视了技术指数级跃迁；GPT-5.5 在未公开复杂实验中的精确预测，证明前沿模型已领悟到书本之外的高维科研经验与实验直觉。 - [29:17](https://www.youtube.com/watch?v=4sexN3yE8xg&t=1757s) 在智能体科研闭环中，多 Agent 集群协同生成假设并预测成功率，机器人自动化执行湿实验并即时反馈，人类科学家的核心角色将演进为设定探索目标的领航者。 - [29:17](https://www.youtube.com/watch?v=4sexN3yE8xg&t=1757s) 根据“杰文斯悖论”，当 AI 大幅降低生命科学实验成本时，科研需求不会枯竭，反而会激发成千上万名工程师在计算机前模拟构建全新的细胞类型与人工组织。 - [32:40](https://www.youtube.com/watch?v=4sexN3yE8xg&t=1960s) 生物学实验常年伴随着 95% 到 98% 的高失败率，面对 AI 时代的高速重塑，科研与跨界从业者最核心的品质在于抗挫折的“能动性”（Agency）与“好奇心”（Curiosity）。 - [34:06](https://www.youtube.com/watch?v=4sexN3yE8xg&t=2046s) AI 将软件乃至物理制造的试错成本降至极低，阻碍创新的不再是昂贵的迭代代价，而是人们是否敢于不断提出“如果……会怎样”并动手付诸实验的勇气。 - [34:53](https://www.youtube.com/watch?v=4sexN3yE8xg&t=2093s) 坚决反驳针对人工智能的负面悲观情绪，AI 不会毁灭人类而是文明的超级放大器，身处一线的 AI 构建者正驱动人类历史上最伟大的黄金时代转型。 ## Detailed walkthrough ### [00:10](https://www.youtube.com/watch?v=4sexN3yE8xg&t=10s) 序幕：当医学免疫学家跨界 AI 构建者 Romain Huet 热情开场，欢迎杰克森实验室（The Jackson Laboratory）教授、医学博士（M.D.）Derya Unutmaz 来到《Builders Unscripted》。Romain 指出 Derya 是一位非常独特的构建者，与平时接触的纯软件开发者截然不同——他拥有深厚的医学背景，横跨生物科学、生物工程，并带着癌症与免疫学等硬核领域的专业积淀大步迈入 AI 领域。 面对 Romain 关于“何时初次意识到生物学与科学必须依赖 AI”的提问，Derya 回忆起自己刚从医学院毕业的时刻。当时他目睹了生物系统令人望而生畏的复杂度：人体内包含数万亿个不同部件和数十亿级同时发生的生化反应，传统医疗面对海量疾病常常束手无策。正是在 20 世纪 90 年代初，由于深感这种复杂性凭借人力无法穷尽，他开始对人工智能产生浓厚兴趣，甚至尝试在 90 年代写代码探索 AI，畅想未来有朝一日能用 AI 构建生物模型。 此后，从深度学习革命席卷而来、展现出大规模并行处理海量信息的能力，到 AlphaFold 攻克蛋白质结构预测，再到 ChatGPT 引爆通用大模型，生物学对 AI 的需求逐步化为现实，而这一切的认知起点，正是他刚走出医学院的那一刻。 --- ### [02:15](https://www.youtube.com/watch?v=4sexN3yE8xg&t=135s) 从 o1-preview 到科学推理的“顿悟时刻” 作为 OpenAI 社区的早期深度参与者，Derya 始终活跃在 X（原 Twitter）上，长期坚定倡导 AI 将如何重塑人类文明。2024 年 9 月，OpenAI 邀请他在早期阶段上手测试首个具备高级推理能力的模型——o1-preview。 Derya 分享了一个令人动容的测试细节。由于平时酷爱电子游戏，他热衷于将游戏世界的机制与生物学研究进行跨学科映射。他以《堡垒之夜》等“大逃杀”（Battle Royale）孤岛生存对抗类游戏为隐喻，向 o1-preview 抛出了一个跨学科构想：“如果将大逃杀游戏机制映射到免疫系统，你会如何设计一套免疫细胞抗击癌症的推演方案？”此前包括 GPT-4 在内的模型大多只能给出浅层回答，但 o1-preview 却给出了极其深邃、充满洞见的解答，其严密的推理深度甚至让 Derya 当场热泪盈眶（"almost made me emotional"），团队后续甚至针对该方案展开了真实的实验。 虽然 GPT-4 时代大语言模型在文献检索、知识整合以及将原本耗时一小时的推荐信撰写压缩至五分钟等方面已表现优异，但科学家仍无法全信其推演实验结果的能力。而 o1-preview 的推理突破，让 Derya 真正确信：AI 赋能科学探索已成不可逆转的必然趋势。 - **测试时间**：2024 年 9 月受邀参与 o1-preview 早期测试。 - **跨界 Prompt 灵感**：将 Battle Royale（大逃杀）孤岛生存对抗模式类比为人体免疫细胞识别与围剿肿瘤的过程。 - **演进路线**：从早期的文献检索与日常文书处理，跃迁到 o1-preview、o1-pro 及 o3 的科学实验假设与机理推理。 --- ### [05:44](https://www.youtube.com/watch?v=4sexN3yE8xg&t=344s) “Codex 重度成瘾者”与科研工具自主开发 谈到当下的工作流演变，Derya 戏称自己已是彻底的“Codex 认证成瘾者”（"certified Codex addict"）。他的晨间日常固定为：起床第一件事倒一杯咖啡，紧接着查看 Codex 为他构建的工具与代码。过去几个月里，为了验证脑海中迸发出的无数模拟实验、应用与小游戏点子，甚至等待 Codex 通宵运行的结果，他几乎处于夜不能寐的兴奋状态。 过去生物学家即使掌握编程，搭建一个复杂的专用分析软件也需要耗费数周乃至数月。而在现代实验室中，高度依赖昂贵且复杂的专业分析软件。以免疫学不可或缺的流式细胞术（Flow Cytometry）为例，科研人员使用荧光标记分子与特异性抗体结合不同细胞表面受体，再将数十万个细胞逐个通过高能激光束激发，产生庞大而复杂的多通道荧光数据。为了统计分析特定抗击癌症的细胞、自身免疫病细胞比例及各类表型，过去几十年科学家只能依赖成熟的第三方商业分析套件。Derya 曾萌生“亲自手搓一套流式分析软件”的疯狂念头，在屡次失败后，他最终完全借助 Codex 构建出了一款全功能的流式细胞分析软件。 ```mermaid flowchart LR accTitle: Flow Cytometry Analysis with Codex accDescr: Pipeline showing cells labeled with fluorescent antibodies, laser interrogation, and analysis via Codex built app. subgraph LAB ["实验室检测流程"] A["荧光抗体标记细胞<br/>(如 CD4 / CD8 等受体)"] B["高能激光流式分析仪<br/>(逐个激发荧光信号)"] A --> B end subgraph DATA ["数据产出"] C["数十万级单细胞<br/>多通道离散数据点"] B --> C end subgraph APP ["Codex 构建的原生软件"] D["交互式门控截取<br/>(Gating)"] E["等高线图与散点图<br/>实时极速渲染"] F["细胞亚群百分比统计<br/>(如杀手 T 细胞比例)"] C --> D D --> E D --> F end ``` --- ### [09:17](https://www.youtube.com/watch?v=4sexN3yE8xg&t=557s) 演示实操：百分之百由 Codex 编写的单细胞分析套件 Derya 在现场打开笔记本电脑，展示了他完全使用 Codex 打造的流式细胞数据交互分析应用。软件界面实时载入了一个大型数据样本，其中屏幕上的每一个像素点都代表一个真实的人体细胞： - **单细胞多维分析**：该样本涉及多达 20 种不同的特异性荧光标记抗体，能够精确对应不同细胞表面的受体结合状态。 - **细胞亚群分类**：他特别演示了免疫学中的核心 T 细胞亚群——携带 CD4 分子的辅助性 T 细胞，以及携带 CD8 分子、能够直接识别并杀伤肿瘤的细胞毒性 T 细胞（Killer T Cells）。 - **极速交互与渲染**：他在图表上自由划定门控区域（Gating），软件瞬时输出选定群体的精准百分比分布与统计学分析，并能秒级自由切换等高线图（Contour plots）与多色散点图。尽管包含超过 100,000 个单细胞事件，前端处理与数据渲染依旧流畅无比。 - **智能辅助功能**：Derya 甚至扩展了一个智能推荐面板，用户只需在下拉框中选择目标细胞类型（例如幼稚 T 细胞 / Naive T Cell），软件便会自动提示并高亮区分度最高的表面 Marker 组合。 面对 Romain 的赞叹，Derya 确认该应用代码 100% 由 Codex 自动生成。在开发早期虽然遇到不少图表渲染与功能逻辑问题，但他通过自然语言持续给出反馈（例如“我看不太清这张图表的某些部分，写一个 Prompt 修复渲染”），AI 便反复迭代优化，最终完成了专业商用级软件的核心功能。 ### [11:31](https://www.youtube.com/watch?v=4sexN3yE8xg&t=691s) 从流式细胞分析面板到生物学编程工具 德里亚·乌努特马兹（Derya Unutmaz）向罗曼·休特（Romain Huet）展示了他利用代码构建的细胞标记物筛选工具。在日常免疫学实验中，科学家需要从错综复杂的细胞亚群中筛选出特定细胞，例如中枢记忆细胞（central memory cell）、初始 T 细胞（naive T cell）或 TH17 辅助性 T 细胞。通过该应用选择特定标志物后，研究人员便能直接回到流式细胞术（flow analysis）中开展精准分析。罗曼感叹，即使具备工程背景，普通人要从零搭建出这样精巧的抽象系统通常也需要数周甚至数月的时间。 德里亚坦然回应自己并不是传统意义上的软件工程师，而是一名生物医学工程师：“如果让我自己闭门写几个月代码，顶多写出一个贪吃蛇游戏；凭空做出这类应用在过去根本就是天方夜谭。”得益于 Codex，他如今能够快速拼装出一个个趁手的小工具来服务日常科研。面对罗曼关于是否尝试过更多模型辅助日常工作的询问，德里亚直言，AI 对生物学最深远的颠覆性影响，在于对高阶复杂生命系统的全真模拟。 - **应用场景**：流式细胞术多色荧光标记物组合（Panels）的快速设计与细胞亚型筛选。 - **开发者背景**：生物医学工程专业背景，无专业前端与软件工程开发经验，完全依赖自然语言与 Codex 协作开发。 --- ### [12:28](https://www.youtube.com/watch?v=4sexN3yE8xg&t=748s) 终极构想：从虚拟细胞到“数字孪生” 德里亚打了一个极具冲击力的比方：人类在制造飞机时，绝不会把零件胡乱拼凑在一起就祈祷它能飞上天，而是会在制造前开展极尽详实的空气动力学计算机仿真；然而在生物学领域，科学家至今仍缺乏这种先验模拟能力，因为生物体的组成部件和变量实在过于庞大复杂。他的长期科研愿景，正是借助 AI 构建“虚拟细胞”（Virtual Cells），完整复刻一个活体免疫细胞的所有生化行为，进而演进到组织模拟，最终实现涵盖整个人体系统的“数字孪生”（Digital Twin）。 在构想宏图的同时，德里亚也不忘向坐在一旁的 OpenAI 开发者体验主管打趣施压：“要做到这一点，我们需要庞大得多的算力，所以你们最好加大马力推进算力研发。”罗曼心领神会地笑答“意见已收悉（Point taken）”。随后德里亚展示了一个初步实践的 Demo：一个针对 T 细胞表面核心传感器——T 细胞受体（TCR）的仿真程序。 ```mermaid flowchart LR accTitle: 从微观受体到数字孪生的跨尺度生物模拟路径 accDescr: 展示了从单个TCR受体模拟开始，逐步扩展至虚拟免疫细胞、人体器官组织，最终达成个性化数字孪生体的演进层次。 subgraph Micro["微观分子与通路"] A["T细胞受体 (TCR) 仿真"] --> B["复杂信号转导网络\n(磷酸化/转录因子)"] end subgraph Meso["细胞与组织尺度"] B --> C["虚拟免疫细胞\n(Virtual Cells)"] C --> D["器官与组织网络\n(Tissues)"] end subgraph Macro["宏观全人体尺度"] D --> E["全生命系统数字孪生\n(Digital Twin)"] end ``` --- ### [13:26](https://www.youtube.com/watch?v=4sexN3yE8xg&t=806s) 破解生死开关：单受体信号通路的动态仿真器 德里亚将界面切换至他构建的 T 细胞受体仿真工具。TCR 是坐落于免疫细胞表面的核心分子哨兵，也是人体健康的终极裁决者——受体与配体结合的亲和力大小以及接收到的协同信号，往往直接左右机体的生死：“信号强度的微小偏差，可能诱发自身免疫病，可能决定能否彻底清除恶性肿瘤，可能成功消灭被病毒感染的细胞，也可能因为免疫风暴失控而引发致命损伤。”在这一微小受体的内部下方，潜藏着极为庞大且交织的分子信号通路。 在演示中，用户可以像调音台一样精确调整配体的性质、结合质量与作用剂量，然后点击运行仿真。该系统不仅能精确预判哪些级联下游分子会被激活、哪些会被抑制，甚至能够动态呈现关键转录因子的磷酸化修饰谱图。当德里亚引入一个抑制性分子并改变输入信号时，应用会即时推演并提示原通路已完全阻断，转而触发了截然不同的生物学事件链。 罗曼对此深感震撼，指出绝大多数开发者使用 Codex 或 GPT-5.5 主要是搭建常规网站、数据看板或移动端外壳，但德里亚是在浏览器原生环境中直接构建了一套严密的、具有前向推演能力的生物学因果引擎。德里亚笑着透露这就是自己每天早晨的动力来源：“这也是为什么我每天喝完咖啡就早早起床，心想：‘今天我要往仿真器里再塞点新东西，看看会发生什么。’” - **输入变量调控**：配体结合亲和力（Affinity）、配体暴露剂量（Dose）、多受体互作与协同刺激信号。 - **输出模拟特征**：分子级联激活状态、抑制性分子干预反馈、转录因子特异性磷酸化模式（Phosphorylation patterns）。 --- ### [16:01](https://www.youtube.com/watch?v=4sexN3yE8xg&t=961s) 从 Nature 封面幻想图到可交互的活体细胞沙盒 德里亚分享了另一个纯粹出于科研好奇心诞生的创意项目。在看到图像生成模型 Image 2.0 发布后，他被其处理极端复杂科学画面的能力所吸引。当时顶尖学术期刊《自然》（Nature）刚发表了一篇关于“人类免疫图谱”（Immune Atlas）的里程碑论文，德里亚便随手给模型输入提示词，让它基于该图谱设计一张虚拟的《自然》封面图。生成的结果令人赞叹，画面中精细勾勒出的大量免疫细胞亚型，正是他过去 30 年科研生涯中反复打交道的对象。 惊喜之余，德里亚将这张静态图片直接喂给了 Codex，只下达了一条宏观指令：“我希望你理解这张图片，并把它转变成一个可交互的仿真网页应用。”在没有提供复杂代码规范或微观逻辑指导的情况下，Codex 自动生成了一套流畅生动的动态交互界面。在演示中，屏幕上的不同免疫细胞正在发生实时接触与迁移，点击特定细胞即可调出记忆细胞（Memory cell）或效应细胞（Effector cell）的详细生化属性。 应用甚至允许用户动态模拟临床关键靶点与年龄变量：例如阻断著名的癌症免疫检查点靶点 PD-1 后肿瘤浸润细胞的杀伤变化，或是滑动年龄轴观察衰老对免疫细胞机能的系统性削弱。罗曼指出，这种将静态视觉直接跃迁为功能性沙盒的模式，展现出了极其强悍的科学教育与深度探索潜力。 --- ### [18:25](https://www.youtube.com/watch?v=4sexN3yE8xg&t=1105s) 编程生命：用 Swift 与 Codex 打造原生 CRISPR 设计器 德里亚进一步提出“将细胞视为可编程代码软件”的核心哲学，并期盼未来能够出现专门面向生物编程的专用“Codex”。早在近 30 年前，他就曾参与过早期基因编辑技术的探索；而如今有了精准的 CRISPR 基因剪刀工具，科学家理论上能够靶向修改任何基因、修复致病突变或改变表达量。然而实际工程面临严峻的计算瓶颈：面对一条长达数千碱基对（Nucleotides）的目标基因，要在哪个具体区段施加切割才能确保极致特异性、规避脱靶风险，需要大量计算与算法筛选。 市面现存工具无法满足其个性化工作流，于是德里亚借助 Codex 开发了一款运行于 macOS 的原生 Swift 应用程序。在演示中，输入 T 细胞核心标志物 CD4 基因，应用立刻自动抓取 NCBI 数据库中的完整 DNA 序列，快速演算并圈出数十个长度在 20 到 22 个核苷酸的潜在切割靶区，并按照有效性与脱靶评分进行智能降序排列。 更具实用价值的是其高通量文库设计（Design Library）功能。科研人员输入一系列基因名称列表，点击按钮即可批量生成多套对应的 CRISPR 导向 RNA（gRNA）序列组合。德里亚只需一键复制导出的候选序列，提交给线上引物合成公司生产试剂，数天后即可直接在湿实验（Wet Lab）中验证编辑效果。该应用完全使用 Swift 原生代码编写，且德里亚正着手将其迁移适配至 iPadOS，以方便在实验室超净台前移动操作。 ```mermaid flowchart TD accTitle: 基于原生 macOS 应用的 CRISPR 文库快速设计与合成流程 accDescr: 展示了从输入目标基因名称开始，系统如何自动调取数据库、算法智能打分筛选，再到直接对接商业化引物合成机构完成湿实验闭环。 A["输入目标基因\n(例如 CD4 基因)"] --> B["自动拉取 DNA 序列\n(在线生物数据库)"] B --> C["计算识别所有 20-22nt 靶点\n(特异性与脱靶潜能评分)"] C --> D["智能排序并生成\nCRISPR 候选列表"] D --> E["一键生成批量基因文库\n(Design Library)"] E --> F["导出序列直连合成供应商\n(送交湿实验功能验证)"] ``` --- ### [21:13](https://www.youtube.com/watch?v=4sexN3yE8xg&t=1273s) 为什么要构建数字孪生：从治疗疾病到治疗患者 罗曼将话题重新引回“数字孪生”（Digital Twin）的科学前景，询问这一概念在免疫学与肿瘤生物学交叉领域究竟何时能够真正落地。德里亚深入剖析了为何传统医学无法仅凭外在表征解决深层疾病：人类生物系统绝不仅是血压、体温等体外读数，其底层是由错综复杂的免疫网络、成千上万种代谢物、肠道内数以万亿计的庞大共生菌群以及内分泌激素交织而成的超复杂巨系统。 德里亚强调，每一个人的基因组背景与所处后天环境的动态交织，决定了他何时会生病、会患上何种病症，以及对特定的靶向药或免疫疗法是否产生响应。当今医疗范式受限于生物系统的非线性复杂性，往往只能进行千人一面的滞后干预。而引入 AI 与数字孪生体的根本目的，正是为了在疾病真正发生之前做出精确预测，并最终颠覆临床思维——不再去孤立地“治疗疾病本身”，而是真正针对每一个独立的个体去“治疗患病的生命整体（Treat the patient, not the disease）”。 ### [22:19](https://www.youtube.com/watch?v=4sexN3yE8xg&t=1339s) 数字孪生与终极个性化医疗：临床试验从十年压缩至数天 长期以来，现代医学普遍采用一种粗放的给药方式：将同一种药物用于成百上千患有同种疾病的患者。例如，数百万人都在服用他汀类降脂药（Statins），但实际上该药仅对其中一小部分人真正产生预期效果。Derya Unutmaz 提出了一个颠覆性的愿景：AI 能否彻底仿真人类生物学系统？通过整合个体的全基因组、代谢物、蛋白质以及庞大复杂的免疫系统，为每个人构建一个独一无二的生物学数字孪生（Digital Twin）。 一旦能够建立这种高保真模型，医生和科学家便可以在数字空间中反复探问“如果……”：如果我们改变某种生活健康参数会怎样？如果我们给予某种特定药物会怎样？最终，AI 将根据个人的生物学特征直接为患者量身定制专属疗法。医学将走向近乎 100% 疗效且 0% 副作用的“完全个性化”阶段。更震撼的是，临床试验周期将从漫长的 5 至 10 年骤降为短短的 5 到 10 天，AI 将在虚拟环境中代替人类受试者完成绝大部分临床试验验证。 面对外界认为“十几年内逆转衰老、人类活到几百岁”如同天方夜谭、纯属科幻小说的质疑，Unutmaz 反驳指出，大众习惯了传统抗癌研究耗费 50 年才取得微小进展的线性思维，却忽略了 AI 正在以指数级速度狂飙。尽管全球现有的全部算力加起来也无法模拟由数万亿生物构件组成的真实生命体，但随着算力的超常规爆发以及预计在未来 5 到 10 年内通用超级智能的到来，借助 AI 仿真数字孪生并在其上进行医学试验将彻底颠覆整个医疗范式。 ```mermaid flowchart TD accTitle: AI 驱动的生物数字孪生与个性化治疗流程图 accDescr: 展示从患者多组学数据输入、构建虚拟数字孪生，到模拟实验替代传统人体临床试验的闭环流程。 subgraph Input ["患者生物学特征采集"] A["多组学数据 (基因组/代谢物/蛋白质)"] B["个体免疫系统图谱"] end subgraph Twin ["AI 驱动的数字孪生构建"] C["超级算力虚拟仿真模型"] D["模拟干预测试与假设验证"] end subgraph Outcome ["极速临床与精准诊疗"] E["虚拟临床试验 (从5-10年缩短至5-10天)"] F["100% 疗效、0% 副作用的定制药物"] end A --> C B --> C C --> D D --> E E --> F ``` --- ### [24:39](https://www.youtube.com/watch?v=4sexN3yE8xg&t=1479s) 肿瘤学靶向药与 AI 设计的个性化 RNA 疫苗 肿瘤治疗是当下临床中最贴近个性化医疗的领域。即便是同一种癌症类型，不同患者体内携带的基因突变也千差万别。以肺癌为例，肿瘤专科医生首要的工作便是对患者发生突变的基因进行测序，因为突变位点的差异直接决定了用药方案。例如，某种靶向“聪明药”（Smart drug）可能只对 1% 的特定突变肺癌患者有效，而对剩余 99% 的患者完全无效。未来 AI 的目标，是能够针对每个人体内独特的所有突变位点即时生成专有药物。 Unutmaz 举了一个引人注目的实例：澳大利亚的一位计算机科学家利用 ChatGPT 和 Grok，成功为自己患癌的爱犬定制设计了一款 RNA 疫苗，这正是极度个性化医疗的雏形，因为该疫苗完全是针对那只狗身上特有的癌症突变定制的，目前类似机制的研究正在广泛开展中。 然而，决定癌症疗效的不仅仅是基因突变，个体的免疫反应同样关键。虽然以激活机体自身免疫杀伤肿瘤为代表的免疫疗法（Immunotherapy）引发了一场治疗革命，但它并不能让所有患者受益：为什么有些人的免疫细胞能精准识别并清除癌细胞，而另一些人的免疫细胞却会陷入“耗竭”状态？与此同时，过度激活免疫系统极易引发致命的全身毒副作用。如果 AI 能够彻底解析患者免疫系统的反应边界与调控机制，癌症治疗将迎来全面突破。 - 传统肿瘤靶向药往往高度特异：仅对 1% 携带特定基因突变的患者极具疗效，对其他 99% 毫无效果。 - 计算机科学家已借助大语言模型（如 ChatGPT 和 Grok）成功设计出针对宠物犬癌症突变的定制 RNA 疫苗。 - 免疫系统虽是抗击癌症的终极武器，但具有极高危险性，一旦过载激活可能引发严重的自体损伤乃至死亡。 --- ### [26:46](https://www.youtube.com/watch?v=4sexN3yE8xg&t=1606s) 从质疑到震撼：GPT-5.5 展现的“30 年实验室直觉” 作为一名受过严格医学训练的医学博士（MD），Unutmaz 对 AI 和 Codex 的激进拥抱常让身边的同行觉得他“完全疯了”。从 GPT-3.5 问世起他就不断力推 AI 工具，但大多数科学家非常迟疑。人类的大脑天然难以理解这种指数级跃迁，许多人在一年半前试用过某个版本（如 GPT-4），遇到幻觉或回答平庸后就轻易下结论将其弃用，却忽视了 AI 世界里一年半已相当于几个世纪的演进，更无法体会模型在持续迭代中所产生的质变。 谈及模型能力的跨越，Unutmaz 回忆起近期使用 GPT-5.5 Pro 模型的经历，甚至激动到几乎落泪，感叹“这怎么可能”。此前版本在综合知识、提取模式和整合信息方面已经足够出色，但 GPT-5.5 带来的震撼在于，它似乎跨越了某种阈值——展现出了唯有在实验室摸爬滚打 30 年才可能沉淀出的“科研直觉”。 在长期的科研生涯中，存在大量无法简单从书本或文献中查找到的隐性认知与直觉经验。过去 Unutmaz 经常和学生打赌某个复杂实验的走向，而他凭借深厚直觉的预测往往能百分之百言中。在最近一次极为复杂的未公开实验中，GPT-5.5 竟然以 100% 的准确率完美预测了实验结果。这种拥有顶级专家直觉般推理能力的表现，标志着前沿模型在专业科研场景中实现了从单纯知识检索到高维逻辑领悟的飞跃。 --- ### [29:17](https://www.youtube.com/watch?v=4sexN3yE8xg&t=1757s) 科学 2.0/3.0：AI Agent 与自动化实验室重塑科研范式 放眼未来几年的科研日常，Unutmaz 断言生命科学将迎来“科学 2.0 乃至 3.0”的范式颠覆。过去科学家花数周策划实验、耗费数月分析数据的漫长周期已一去不复返。取而代之的是由多智能体协同驱动的科研新体系： ```mermaid flowchart LR accTitle: 科学 2.0 智能体循环工作流 accDescr: 展示 AI 科学家群体生成假设、仿真筛选、机器人湿实验与极速数据闭环的流转过程。 subgraph Ideation ["假设与筛选"] A["多智能体集群<br/>发散生成海量假设"] --> B["筛选最优假设<br/>并仿真预测成功率"] end subgraph Execution ["执行与自动化"] B --> C["自动化机器人实验室<br/>高通量执行湿实验"] end subgraph Feedback ["反馈与迭代"] C --> D["分析 Agent 极速解构数据"] D --> E["主控 Agent (Master) 调整理论"] E --> A end ``` 在这一全新架构中，科学家的核心角色转变为给智能体下达宏观目标（例如“去攻破肺癌机制”），充当指明探索方向的领航者。面对“AI 替代实验后科研人员是否会失业”的担忧，Unutmaz 引用了“杰文斯悖论”（Jevons paradox）：当研究效率极大提升、门槛大幅降低时，科研的需求不仅不会缩减，反而会呈爆发式增长。人类目前对生物学的理解仅仅掌握了不到 10%，在更强大的 AI 工具加持下，研究者将迅速从基础发现迈向更广阔的生物工程（Bioengineering），未来将有成千上万名工程师坐在计算机前模拟构建全新的细胞类型与人工组织。 --- ### [32:40](https://www.youtube.com/watch?v=4sexN3yE8xg&t=1960s) 面对失败的韧性与 AI 时代唯二重要的品质：能动性与好奇心 不仅是生物学，材料科学、物理学以及早已被加速的化学制药领域（过去需要数年的研发现在几个小时即可完成），乃至下游的临床治疗全栈链路都将被 AI 全面重塑。对于非生命科学领域的普通大众和跨界从业者，Unutmaz 也分享了自己多年科研历程中淬炼出的宝贵心法。 做生物学实验最大的特点就是与失败为伴，实验室里 95% 至 98% 的尝试最终都会以失败告终。正是这种职业特性赋予了他强大的抗挫韧性、好奇心以及尝试新鲜事物的强大能动性（Agency）。他在使用 Codex 和 AI 编码时，展现在公众面前的光鲜成果背后同样堆积着海量未跑通的代码与并不好用的应用原型，但他从未停下脚步。 Unutmaz 给所有人的终极建议是：**在 AI 时代，能动性（Agency）和好奇心（Curiosity）是真正唯二重要的品质。** 永远不要畏惧失败，不要对技术心生恐惧，而要抓住一切机会向 AI 追问“如果换一种方式做会怎样？”，在持续实验与动手构建中拥抱变革。 ### [34:06](https://www.youtube.com/watch?v=4sexN3yE8xg&t=2046s) 试错成本的骤降与拥抱实验的勇气 在谈到技术的深层影响时，Derya Unutmaz 指出，过去人们往往不敢轻易提出“如果……会怎样？”（what if）这样的设想，因为探索未知的财务和时间成本实在太高昂了。他以企业搭建网站为例：以往一家公司可能需要耗费数千美元去外包建站，哪怕交付出来的成品并不完全符合心意，大家通常也会妥协，心想“行吧，凑合能用就行，我可不敢再去乱改了”。然而如今在 AI 的赋能下，思维模式发生了彻底颠覆——你随时可以发问“如果我把这个地方改成那样会如何？”短短几分钟之内，一个全新改版的网站便能呈现在眼前；甚至对于实体产品，你也可以在瞬间完成全新设计并直接交由 3D 打印机制造出来。 Derya 坚信，这种低门槛快速迭代的能力可以应用到现实世界的方方面面。现在的关键在于人们是否拥有敢于动手实验的勇气。既然开展一项新实验、验证一个新想法的成本已经降到了前所未有的超低水平，那么“为什么不去试一试呢？”这种极其简单而切实的日常改变，完全可以无限延展并深刻改变生活与工作中的一切。 - **建站范式的转变**：过去耗费数千美元且因改动代价过高而只能“凑合使用”的静态网站，现在几分钟内即可通过 AI 重新构想并即时迭代。 - **物理与数字融合**：不仅限于代码和网页，创作者还可以在几分钟内完成产品设计，并无缝衔接 3D 打印等物理制造流程。 - **创新的核心前提**：阻碍创新的不再是昂贵的试错代价，而是人们是否具备拥抱低成本实验的魄力。 ```mermaid flowchart LR accTitle: 传统模式与 AI 时代的创新成本对比 accDescr: 对比传统高成本低尝试与 AI 赋能下低成本高频实验的差异 subgraph Past["传统模式"] A["高昂资金与时间成本"] --> B["不敢轻易修改"] B --> C["妥协于'凑合能用'"] end subgraph Present["AI 赋能时代"] D["提出'What If'设想"] --> E["几分钟内完成生成与迭代"] E --> F["极低成本的大胆实验"] end ``` ### [34:53](https://www.youtube.com/watch?v=4sexN3yE8xg&t=2093s) AI 带来的黄金时代与人类最伟大的转型 对于当下社会上对人工智能的恐慌与悲观论调，Derya 表达了极其鲜明且坚定的反对立场。他坦言自己看到了太多负面情绪，人们整天忧心忡忡地谈论“AI 要干这个坏事了”、“AI 又要夺走那个了”。但他所坚信的恰恰完全相反——AI 不会毁灭人类，而是将极大地增强人类自身的能力，带领整个人类文明跨入一个前所未有的“黄金时代”。 在 Derya 看来，身处前沿的 AI 研究人员“全都是英雄”，因为他们当前所从事的工作正在促成“人类历史上最伟大的转型”。Romain Huet 对此感同身受，并赞叹这是结束本期对谈最美妙、最充满正能量的方式。Romain 表示，非常期待看到 Derya 接下来如何进一步推动 Codex 以及顶尖前沿模型的能力边界，将生物免疫学世界与先进算力深度融合，早日实现细胞“数字孪生”（digital twin）的远大构想。双方相约在数月之后再次录制新一期节目以跟进科研进展，访谈在对 Derya 加州之行的温馨祝愿与轻松愉快的笑声中圆满落幕。 - **乐观主义的未来愿景**：Derya 坚决反驳外界的负面悲观情绪，坚信 AI 是增强人类潜能的关键杠杆，必将开启属于人类的“黄金时代”。 - **致敬科研先驱**：Derya 直言 AI 研究员是推动“人类最伟大转型”的真正英雄。 - **后续研究展望**：双方约定在几个月后重聚回访，继续跟踪利用 Codex 与前沿大模型推进免疫学模拟、构建数字孪生细胞系统的突破性进展。 ## Notable quotes > [02:15](https://www.youtube.com/watch?v=4sexN3yE8xg&t=135s) 如果将大逃杀游戏机制映射到免疫系统，你会如何设计一套免疫细胞抗击癌症的推演方案？ — Derya Unutmaz > [05:44](https://www.youtube.com/watch?v=4sexN3yE8xg&t=344s) certified Codex addict（彻底的 Codex 认证成瘾者） — Derya Unutmaz > [11:31](https://www.youtube.com/watch?v=4sexN3yE8xg&t=691s) 如果让我自己闭门写几个月代码，顶多写出一个贪吃蛇游戏；凭空做出这类应用在过去根本就是天方夜谭。 — Derya Unutmaz > [12:28](https://www.youtube.com/watch?v=4sexN3yE8xg&t=748s) 要做到这一点，我们需要庞大得多的算力，所以你们最好加大马力推进算力研发。 — Derya Unutmaz > [13:26](https://www.youtube.com/watch?v=4sexN3yE8xg&t=806s) 信号强度的微小偏差，可能诱发自身免疫病，可能决定能否彻底清除恶性肿瘤，可能成功消灭被病毒感染的细胞，也可能因为免疫风暴失控而引发致命损伤。 — Derya Unutmaz > [13:26](https://www.youtube.com/watch?v=4sexN3yE8xg&t=806s) 这也是为什么我每天喝完咖啡就早早起床，心想：‘今天我要往仿真器里再塞点新东西，看看会发生什么。’ — Derya Unutmaz > [16:01](https://www.youtube.com/watch?v=4sexN3yE8xg&t=961s) 我希望你理解这张图片，并把它转变成一个可交互的仿真网页应用。 — Derya Unutmaz > [21:13](https://www.youtube.com/watch?v=4sexN3yE8xg&t=1273s) 不再去孤立地“治疗疾病本身”，而是真正针对每一个独立的个体去“治疗患病的生命整体（Treat the patient, not the disease）”。 — Derya Unutmaz > [26:46](https://www.youtube.com/watch?v=4sexN3yE8xg&t=1606s) 这怎么可能 — Derya Unutmaz > [32:40](https://www.youtube.com/watch?v=4sexN3yE8xg&t=1960s) 在 AI 时代，能动性（Agency）和好奇心（Curiosity）是真正唯二重要的品质。 — Derya Unutmaz > [34:06](https://www.youtube.com/watch?v=4sexN3yE8xg&t=2046s) 为什么不去试一试呢？ — Derya Unutmaz > [34:53](https://www.youtube.com/watch?v=4sexN3yE8xg&t=2093s) 全都是英雄……人类历史上最伟大的转型 — Derya Unutmaz ## Mentioned Derya Unutmaz, Romain Huet, OpenAI, 杰克森实验室 (The Jackson Laboratory), Codex, AlphaFold, ChatGPT, Grok, o1-preview, o1-pro, o3, GPT-4, GPT-5.5 Pro, Image 2.0, 《自然》 (Nature), 《堡垒之夜》 (Fortnite), CRISPR, T 细胞受体 (TCR), NCBI 数据库, Swift, macOS, iPadOS, 杰文斯悖论 (Jevons paradox), 流式细胞术 (Flow Cytometry) ## Video description Derya Unutmaz, M.D., an immunologist and professor at The Jackson Laboratory, chats with Romain Huet about building tools for biology with Codex and a future where AI helps scientists simulate experiments. 00:00 Builders Unscripted Episode 5 09:17 Cell analysis with Codex 13:26 Simulating immune cells 29:48 The future of AI-powered science [Source](https://www.youtube.com/watch?v=4sexN3yE8xg)

[2:15] Derya Unutmaz: 如果将大逃杀游戏机制映射到免疫系统，你会如何设计一套免疫细胞抗击癌症的推演方案？ [5:44] Derya Unutmaz: certified Codex addict（彻底的 Codex 认证成瘾者） [11:31] Derya Unutmaz: 如果让我自己闭门写几个月代码，顶多写出一个贪吃蛇游戏；凭空做出这类应用在过去根本就是天方夜谭。 [12:28] Derya Unutmaz: 要做到这一点，我们需要庞大得多的算力，所以你们最好加大马力推进算力研发。 [13:26] Derya Unutmaz: 信号强度的微小偏差，可能诱发自身免疫病，可能决定能否彻底清除恶性肿瘤，可能成功消灭被病毒感染的细胞，也可能因为免疫风暴失控而引发致命损伤。 [13:26] Derya Unutmaz: 这也是为什么我每天喝完咖啡就早早起床，心想：‘今天我要往仿真器里再塞点新东西，看看会发生什么。’ [16:01] Derya Unutmaz: 我希望你理解这张图片，并把它转变成一个可交互的仿真网页应用。 [21:13] Derya Unutmaz: 不再去孤立地“治疗疾病本身”，而是真正针对每一个独立的个体去“治疗患病的生命整体（Treat the patient, not the disease）”。 [26:46] Derya Unutmaz: 这怎么可能 [32:40] Derya Unutmaz: 在 AI 时代，能动性（Agency）和好奇心（Curiosity）是真正唯二重要的品质。 [34:06] Derya Unutmaz: 为什么不去试一试呢？ [34:53] Derya Unutmaz: 全都是英雄……人类历史上最伟大的转型

[2:15] Derya Unutmaz: o1-preview just gave me this response that almost made me emotional. I was like, 'Oh my God.'... That was the moment where I realized when it starts to reason, then you get things that are going to be extremely relevant for science. [0:10] Derya Unutmaz: I can call myself a Codex addict. I'm definitely a certified one. [9:17] Derya Unutmaz: To make these apps would be kind of like a dream... I said, 'Why not make one for myself?' It was a crazy idea because it's very complex. In fact, I failed a lot... But since 5.5, now I have a completely working version of this. [16:06] Derya Unutmaz: I want you to understand this image and create an interactive simulator website. [24:39] Derya Unutmaz: There was a case from Australia, this computer scientist using ChatGPT and Grok, created an RNA vaccine for the dog... that's as personalized as you can get because that RNA vaccine was created just for those cancer's mutations. [32:40] Derya Unutmaz: In the age of AI, really agency and curiosity are the only things that matter. Don't be scared. Constantly try, constantly experiment with AI... because the cost of experimentation is so low. Why shouldn't you do that?

[0:10] Derya, thank you so much for being here. [0:13] Thank you. I'm very excited. [0:14] Likewise. [0:15] Very psyched to have you, because you're obviously a very unique kind of builder. [0:20] Quite unlike the builders we usually talk to. [0:23] You have a medical background, but you also work in, [0:27] you know, bioscience, bioengineering, and you're pushing into AI [0:32] in a way that, like most builders don't, coming at it with this real depth [0:38] that you have in so many topics, like from biology to cancer to immunology. [0:43] So very excited to dive in. [0:45] Thank you. [0:45] Thank you very much. [0:46] Maybe the first question, like if you if we rewind, like, [0:50] when did you first realize that biology and science were going to need AI? [0:55] Yeah, I think that was, right after I graduated from medical school. [0:59] When I realized the complexity of the biological system. [1:03] In fact, after I graduated, I went into biomedical research [1:07] because I really wanted to understand the biology and I so that, [1:12] you know, at the time, well, we couldn't treat a lot of diseases. [1:16] And because of the complexity and then more I went into the research [1:20] aspect of it, the more I realized, oh my God this is incredible. [1:24] How are we going to solve this? [1:26] There are trillions of different parts and biological systems and made billions [1:32] of reactions that are happening and it just seemed so overwhelming. [1:36] And that's when I got interested in AI, artificial intelligence, early 90s. [1:42] Then I realized, maybe someday, you know, [1:45] we could actually build models using AI. [1:48] And I got very interested in, trying to code with AI during the 90s and [1:52] and of course, you know, fast forward with deep learning revolution. [1:56] Then I was so excited because for the first time, [2:00] I could see that deep learning is able to [2:03] process this massive amount of information in a sort of a parallel fashion. [2:08] And of course, AlphaFold and then ChatGPT. [2:11] Yeah. [2:11] but the first moment was when I finished my medical school. [2:15] Amazing. [2:15] I And since that moment of launching ChatGPT, [2:17] it feels like you very much being in our community, right? [2:20] We've been testing models. [2:21] I remember you doing some work with o1-preview. [2:23] Yes. [2:24] When we had the very first, like, reasoning models. [2:26] Like what was the first, like, reaction when you started to have your hands on that? [2:30] I still remember this was September 2024. [2:33] Actually, OpenAI approached me, I guess because I was, [2:37] very active on X and that I would Twitter in the old days, [2:42] you know, talking about the AI, [2:44] you know, how this was going to change humanity. [2:48] You know, at the time there were a lot of skeptical people [2:50] and I think there still are. [2:51] But, you know, I truly believed in it. [2:53] And I went all-in in AI and I guess, OpenAI was interested for me to try. [3:00] I guess that the first reasoning model and I still remember it, [3:05] when I gave it a very complex [3:10] question on immunology in my topic. [3:14] But the question, actually, I even remember the prompt, [3:17] and I'm very interested in games, video games. [3:21] So I like to sort of cross [3:23] virtualize games to, to science. [3:26] And one of the games that I would play is, you know, this [3:30] sort of like a survival game when you're in a, in an island [3:34] and, you know, you're fighting. [3:35] I won't give the names of the companies, but you understand the type of games. [3:40] And so it's kind of a battle royale games. [3:43] And, you know, in a way, the immune system is kind of [3:46] like a battle royale for against tumors. [3:50] And I said, you know, think of, [3:53] how would, [3:55] if you just imagine battle royale games and immune system, how would you develop, a scenario [4:02] where the immune cells would fight against cancer? [4:05] It's a completely different field. [4:07] And in fact, we actually did experiments related to that later on. [4:11] And o1-preview just gave me this response that almost made me emotional. [4:16] I was like, oh my God. [4:18] Before that, you know, GPT 5.40 you know, it wouldn't [4:22] it wouldn't really give you that kind of a deep, insightful, response. [4:27] And, yeah, [4:29] I still remember that day that was, you know, special. [4:31] Incredible. [4:32] Was that the moment you would say what you thought [4:35] Okay, like, AI is now inevitable for science? [4:38] Absolutely. [4:39] I mean, even before that, you know, especially after 4 [4:43] GPT 4 came out, you know, it was extremely useful. [4:47] I would tell my colleagues, because there's so much information, [4:52] especially in biology, you can't really keep up with it. [4:54] And, you know, for searching literature [4:57] and really synthesizing, the knowledge, [5:01] of course, you know, doing mundane things like writing your letters. [5:04] You know, I read a lot of letters of recommendation [5:07] and things that we used to take an hour would take five minutes. [5:10] But, it wasn't at the point where you could fully trust it [5:15] or ask questions like, so what would be the outcome of an experiment for example? [5:20] And I think o1-preview was that moment where I realized [5:25] when it starts to reason, [5:27] then you get, things that's going to be extremely relevant for science. [5:31] So of course, after that we had the pro version and o3, which got better. [5:35] And now we have unbelievable models alone and that, you know, [5:39] mentioned about my current experience and which is getting better and better. [5:44] Yeah. I mean, like, maybe speaking of the current moment, like, fast forward to today. [5:47] I read a couple months ago your tweets, about your new routine were like, [5:51] you're going to start your days now with first the morning coffee. [5:54] But second kind of Codex has to be doing something for you. [5:59] Yes. [6:00] So I can call myself a Codex addict. [6:03] I'm definitely a certified one. [6:07] Yeah. So the first thing in the morning when I wake up because [6:11] what for Codex enabled is, you know, there's all these ideas [6:16] that, you know, whether you want to make a simulation, [6:20] whether you want to make a certain app or you want to make a game or whatever it is. [6:26] You know, in the past, you had to know how to code. [6:29] Even if you knew how to code, it would take you weeks or months to do it. [6:33] But now, I mean, I get a lot of ideas and the moment I wake up, get my coffee, say, okay, I gotta try this idea. [6:43] You know, how is it going to work out. [6:45] And sometimes, actually, Codex is working overnight, [6:48] so I want to see what the what the results are. [6:50] And I haven't slept that much in the last couple of months because of that. [6:56] Fascinating. [6:57] So, yeah. [6:58] I love Codex. [7:00] And I'm curious, like, for someone like you that has so much, [7:03] like, real depth in like, immunology, oncology, cancer, T-cells, like, [7:09] how do you approach, like, using Codex to bring these worlds together? [7:14] And in a way that's also useful to your day to day? [7:16] Yeah. [7:17] Yeah. I can, even show specific examples. [7:20] I've been, trying to build, you know, some apps [7:24] that are not very complex, but extremely useful for day to day bases. [7:30] I have a bunch of examples of that. [7:32] But then also we, we are very much dependent on software for analysis. [7:38] As I said, the biology is very complex. [7:40] Whether there's genetics, immunology, we, for example, [7:44] do a lot of analysis called flow cytometry. [7:47] Basically, this is like our window to the [7:50] world of cells, mainly immune cells. [7:52] But you can analyze any cells. [7:54] So we have these machines, there's a big laser. [7:57] We label the cells with particular markers, [8:00] fluorescent markers, because there's like hundreds of different types of cells. [8:04] So how do you know, what type of cells there are in your blood in your tissues. [8:08] So we label them, we pass them through this laser. [8:10] And the laser, you know, analyzes thousands of cells [8:14] and then creates, some sort of a data, [8:17] that will tell me this is an immune cell that can fight against cancer. [8:21] This one can cause auto immunity or whatever the function is. [8:25] But we have to take that data and put it into specialized software. [8:29] Take all that thousands or sometimes hundreds of thousands of data points, [8:34] individual cells and the software will make that into graphs [8:39] and then so you can say, okay, give me [8:41] the percentage of these cells there and how did they link to that and so on. [8:45] So this is very sophisticated software, [8:47] which we've been using for the last couple of decades. [8:50] And you know, one day I said, why not make one for myself? [8:56] It was a crazy idea because it's very complex. [8:58] In fact, I failed a lot. [9:01] I tried to make some things work. [9:04] But since 5.5, now I have a completely working, version of this. [9:10] That's amazing. [9:11] I would love to see it if you have [9:12] Yeah. [9:13] on your laptop to show me. [9:14] I can show you some of these examples. [9:17] So, start with this. [9:20] And I have different versions of this [9:22] but I'll show you one of these applications. [9:25] So basically, I already uploaded a file and you can see that these are [9:31] actually each of these dots is a single cell. [9:34] And these are the fluorescent molecule colors, if you like. [9:38] So whatever the antibody, it has that first molecule, [9:42] it will label that particular cell type. [9:45] And actually I can choose here. [9:47] Okay. [9:48] Well, change me because you can see these are [9:50] all of these are different antibodies that are labeled different. [9:53] It's very complex because there's 20 different molecules [9:57] and each of them binding to a receptor. [9:59] And their combination will define, will give you [10:02] a particular subset of cells that will we defined by those markers alone. [10:07] For example, my favorite cells are these T cells [10:12] that have, what we call the CD4 molecule. [10:15] And some of them have what's called the CD8 molecule. [10:18] The ones that have CD8 or are killer cells. [10:21] They will go and kill the lymphocytes. [10:23] And so I can, put gates here and I can say, okay, well, you know, what's the, [10:28] what's the percentage of cells that are CD8 positive or CD4 positive? [10:33] So it will create all kinds of statistic analysis. [10:36] But, you know, this is actually, really, complex software because I can change [10:41] the contour, plots, the graphics, it's a different kind of representation. [10:45] Remember, there's like 100,000 events here and it's doing so fast, [10:49] which is incredible because I didn't think that [10:52] it would be optimized this way. [10:54] And you built all of this with Codex? [10:56] 100% with Codex. Yeah. [10:58] Incredible. [10:58] I mean, it took me a while because, you know, some things wouldn't work but [11:02] especially with 5.5, I would say, okay, well, I can’t see the graphs [11:05] very well and, you know, fix that and write me a prompt for that. [11:09] And it would just go and work and work and you know. [11:11] So I made this little app, [11:13] which basically, you can actually choose the cell type you want. [11:17] Okay. So I say I want a naive T cell, and it will show me [11:20] all the markers that I can potentially choose. [11:24] And in fact, it will even show me, for example, you know, the ones that are more relevant for that cell type [11:29] versus the other one. [11:31] this becomes extremely useful for developing these, what we call panels. [11:36] I'm looking for the central memory naive cell and a T cell and TH17 cells. [11:41] Okay. [11:41] If I choose these markers then I can go back and do my flow analysis. [11:46] That's amazing. [11:47] I mean, like you’re obviously an engineer by training but, like, [11:50] you would have taken you, I'm sure, like weeks of work, if not months of work [11:53] to even do, build this abstract. [11:56] So I'm not a software engineer. [11:57] I'm a biomedical engineer. [11:58] Sure. [11:58] So I can't really [12:00] write, I could probably write a snake game if I worked months. [12:07] To make these apps would be kind of like a dream. [12:11] So, I mean, I make these, like, little kind of apps that [12:16] help on a daily basis. [12:18] That's pretty incredible. [12:19] And on T cells, like, tell me more. [12:20] Have you, like, tried to play with more models to kind of help in your day to day work using Codex [12:26] or even things like ImageGen? [12:28] Yeah. So one of the things that I'm very passionate about and I think [12:33] when I say AI is going to make tremendous impact on especially biology, [12:37] is to be able to simulate biological systems because they're so complex. [12:44] For example, if you're building an airplane, right, you don't just say, [12:48] okay, if I put these parts together, I hope that I'm going to have an airplane. [12:52] You do a simulation, right? [12:54] So how is it going to be a or the aerodynamic and all that stuff? [12:57] For, for any kind of complex things. [12:59] But for biology, we don't have that because there's so many parts. [13:04] And my goal is that at some point we'll be able to build what's called virtual cells. [13:09] We will be able to mimic a complete immune cell, [13:13] by AI and eventually tissues and eventually [13:17] what I call the digital twin, like a complete human being. [13:21] We're going to need a lot more compute. [13:23] So you guys better invest developing that. [13:25] Point taken. [13:26] But, you know, just to start, I began with an app like this. [13:30] This is basically one receptor. [13:33] It's called the T cell receptor that's on the top [13:36] of these immune cells, the T cells that I just showed you. [13:39] And it's the main receptor, but is one receptor is extremely complex. [13:43] Depending on the affinity of the molecules it sees, [13:47] depending on the other signals it sees, [13:51] that’s the decider. That will decide sometimes whether you're going to live or die. [13:56] Because the strength of the signal could mean that you get autoimmune disease. [14:00] It could mean that you can clear the tumors. [14:03] It could mean that you kill the virally infected cells [14:06] or it could mean it will cause too much damage and kill you. [14:10] So, the decision point is there and there is very complex events [14:15] that happen underneath that receptors, [14:18] huge signaling pathways. [14:21] And so I built this just to be able to simulate, [14:25] for example, if I have only the T cell receptor. [14:28] And then, you know, if the ligand is a certain quality [14:31] and the dose is this much, I can actually control all of that here. [14:36] And then I can say, okay, run me a simulation based on that. [14:39] It will run me a simulation. [14:41] Wow. [14:41] It will tell me what molecules are going to get activated, [14:45] which ones will not get activated and it will even show me [14:48] the phosphorylation patterns of what's called the transcription factors. [14:53] And I can say, okay, well, what if I just put pressure. [14:57] There’s an inhibitory molecule and then I choose a different signal, what would happen? [15:03] And then it will show me. [15:05] Okay, well, okay. This pathway stopped now. [15:07] So you're going to get a different type of events. [15:10] So basically you can just expand this. [15:12] You can say, okay, what if I just put a small molecule [15:15] that will inhibit this molecule, what will be the output? [15:19] Right. [15:19] And then what if I just add more receptors to here? [15:22] And how are they going to interact with each other? [15:25] I love that because it's not even just like visualizing data sets or like searching in data sets. [15:30] This is like a complete like application that you build to, like identify the inputs [15:35] and outputs in the browser of every like particular like cell and scenario. [15:39] Incredible. [15:40] Yeah. [15:41] I mean, we see so many builders who are obviously like using tools [15:44] like Codex and GPT 5.5 to build like websites, web applications, mobile apps. [15:49] But this is like so intricate and so advanced. [15:53] For the biology. [15:54] Yeah. [15:54] I'm that's why I wake up early after coffee. [15:57] I say, okay. [15:58] So I'm going to add this to, to this simulator, see what happens. [16:01] And so and then the other thing, which was kind of a fun thing to do, [16:06] because after Image 2.0 came out, [16:10] I just fell in love with it, you know, I can make extremely complex images. [16:15] For example, this image, I think I shared it on X as well. [16:19] This was about an immune atlas that was published [16:24] in this journal called Nature it’s top journal in the world. [16:29] And so I basically asked image to said, you know, create me [16:33] a cover page of Nature just for fun based on this atlas. [16:38] And it made this unbelievable image where you can see all these immune cells [16:43] are the ones that I've actually worked for the past 30 years. [16:46] So like, oh my God, you know, this is so detailed and so beautiful. [16:50] And then I said, what if I just take that image and put it into Codex [16:54] and tell it to sort of make it live in an interactive website. [17:01] And that's what I did. [17:02] And then Codex just came up with these beautiful animations [17:06] and how these cells are interacting with each other. [17:09] I can click on them. [17:10] It shows me what you know. [17:12] This is a memory cell. [17:14] This is an effector cell. [17:15] It gives me all kinds of information about them. [17:17] You know, if I block it, for example, this molecule, PD-1, is very important in cancer therapy. [17:23] If you block this molecule, you can cure a lot of cancers called checkpoint inhibitor. [17:27] And you know, this was a totally a fun thing to do. [17:30] But then I realized, wow, I can even change, for example, [17:33] based on your age, how your immune cells are going to behave. [17:37] This is like a mini simulator here. [17:39] It all started with just one image. [17:42] Just this one image. And GPT managed to generate it. [17:44] Exactly. And I didn't say, okay, do this or do that. [17:47] I said, I want you to understand this image [17:51] and create an interactive simulator website. [17:54] That's all I did. [17:56] And Codex just built this for me. [17:58] Amazing. [17:58] Yeah. [17:59] It's also such a great tool for education as well. Right. [18:01] Because if you want to learn into, like, dig into any topic, [18:05] learn about that topic. [18:06] Yeah. [18:07] You can generate image but also just generate like simulators. [18:10] Exactly. [18:11] For that topic, it is pretty incredible. [18:12] Sounds like you're building so many apps like thank you fro sharing. [18:14] This is like a window into your work and how you think and how you use Codex [18:20] in a way that, like, I have not used Codex to build these apps so far. [18:23] So this is fascinating for me. [18:25] Let me just show one more, one more app, which I think is very useful as a biomedical engineer. [18:32] What we want to do is we want to manipulate the cells. [18:36] In a way the cells are kind of like [18:39] programed code software. [18:41] And I'm hoping that one day we'll have a Codex for biology where we can [18:45] we can totally program the cells. [18:47] In fact, we started doing that, [18:49] with these technologies to genetically edit cells. [18:53] I actually, developed some of that 25 years ago. [18:57] 30 years ago. [18:58] And now we have this thing called CRISPR. [19:00] I'm sure a lot of people have heard of it. [19:02] Basically, CRISPR is you can target any gene, you can change [19:07] fixed mutations, you can delete the gene, you can overexpress it. [19:10] It's like a genome engineering. [19:12] But the problem is that again, this is very complex. [19:15] So you have a gene that's let's say 2000 words nucleotides. [19:21] Where are you going to target? [19:23] Right. [19:24] So you have to have some sort of a computation [19:27] to say that that's going to be specific. [19:30] It's going to do this effect or not. [19:33] And there are some tools. [19:34] Like then I said, well, I want my own, you know, so that I can [19:37] so I built this application where you can actually choose any gene you want. [19:42] So let's say this is a CD4 gene, which is I mentioned about that [19:46] on the surface of T cells. [19:48] It immediately pulls the sequence from the database. [19:52] So thats the DNA of the CD4? [19:53] Yeah, exactly. [19:54] And then it gives me all the potential targets. [19:58] So these are 20, 22 nucleotide, regions. [20:01] But it's a very big, big gene. [20:04] And then it ranks them. Says, okay, well, you should probably choose this one because it's good. [20:08] And I can just say, okay, add that there, add that there [20:11] and then I can copy it and then give it to an online company. [20:16] And they will synthesize that and send it to me and I will engineer and do the experiment. [20:21] I can visualize that. [20:23] It shows me all the regions where, these are being targeted. [20:26] so there are some features here that are not present in other applications. [20:31] That's the beauty. [20:32] For example, I said, okay, can you create a library for me? [20:36] So what if I have, you know, multiple genes [20:39] and I want many different CRISPR targets. [20:42] So I built this, version where I can just write the genes name here [20:46] and all I do is just click on design library. [20:49] It will make, different CRISPR molecules for me. [20:51] That's amazing. And this is a native app that we're looking at? [20:54] On your Mac, right? [20:54] This is a native Mac OS app that's built with Swift and actually [20:58] I'm going to make an iPad version too, so that's [21:01] That’s really cool. [21:02] Well, thank you so much for sharing that. [21:04] This is incredible to see the behind the scenes of your work. [21:06] Yeah. [21:07] Now you understand why after coffee I [21:09] I totally understand the routine now. [21:11] It makes so much more sense to me. [21:13] One thing that you mentioned earlier that I'd love to double click on [21:16] is this idea of like a digital twin, right? [21:18] And as you have like your vantage point [21:21] in like immunology, cancer, biology. [21:25] When do you think this idea of a digital twin becomes viable? [21:29] First of all, why digital twin and why we need AI? [21:33] The reason we need the digital twin [21:35] and when I say digital twin, I mean, a person's complete biological system. [21:40] So our biology is not just what you measure outside. [21:45] As I showed you, there's the immune system, there's our metabolites. [21:48] There's our trillions of bacteria in our gut and hormones and so on. [21:53] So it's very, very complex system. [21:55] And of course there's your genetics and your genetics and your environment decide [22:00] pretty much everything whether you're going to get sick or when are you going to get sick. [22:04] Whether you're going to respond to a given treatment. [22:06] Can we predict before the disease happens? [22:10] And it has to be very personalized. [22:12] In fact, you have to treat the patient not the disease [22:17] but because the biology is so complex. [22:19] We have been giving the same drug with the same disease to millions of people. [22:24] For example, statins are used by millions of people but it only affects a small portion of them. [22:31] So can we come to a point where AI can completely simulate the biology [22:39] and your biology with your genome, with your metabolites, [22:43] with your proteins, your immune system. [22:47] And if you can get that point, then we can start to ask this question. [22:51] What if I make this change in my health? [22:54] Or what if I give them this drug? [22:56] Or maybe, I will basically order a treatment [23:01] just for you based on your biology, what the AI is going to tell us. [23:06] And so we're getting to a point of complete personalization [23:11] where the drugs will have nearly 100% effect with 0% side effect. [23:17] And that means that the clinical trials that we do are going to be accelerated [23:22] from 5 to 10 years to maybe 5, 10 days. [23:27] Literally, AI is going to do the clinical trial for you. [23:31] And that's the incredible acceleration because when I say [23:34] in the next decade or so, we're going to treat all of the diseases. [23:39] Eventually, within 15 years, when reverse aging and people [23:43] will be able to live hundreds of years, people say, well, that sounds crazy. [23:47] That's science fiction. [23:49] You know, it took us 50 years, for cancer just to be able to treat a little bit. [23:55] What they don't calculate is that AI is [23:59] advancing exponentially. [24:01] And this is also assuming that the compute is going to greatly increase [24:04] because even the current available compute [24:08] if you put all of the compute in the world, it’s not enough to simulate biological systems. [24:13] There are trillions of components. [24:15] So if we get to that point, which I believe we'll be able to do in the next [24:19] 5, 10 years and we'll have the superintelligence then. [24:23] Then basically we simulate this digital twin with AI and then we [24:28] we do an experiment, not on humans [24:31] but with AI for your biology. [24:35] And that will change medicine. [24:38] That will change everything. [24:39] So if we take the example, for instance, [24:40] of like a cancer patient, say, and if we had the digital twin, [24:44] what could a doctor do differently versus what they do today? [24:48] We did just basically try different hypotheses [24:50] and experiments on that digital twin and see how that reacts. [24:53] And that's how we treat the stable group and the treatment group almost? [24:57] Exactly. [24:57] Actually, cancer and oncology is as close as we [25:01] we are today for personalization because [25:05] even the same cancer type have many different mutations. [25:09] And so you probably if you were a lung cancer patient, the first thing [25:14] your doctor, oncologist, would do is to sequence the genes that are mutated [25:18] because depending on which genes mutated, you have a different drug for that. [25:22] For example, 1% of lung cancers can use a particular drug and [25:27] that drug is very effective in that 1%, but it's not effective at the 99%. [25:31] So these are smart drugs that companies are developing. [25:35] So we can come to a point where we can literally create a drug [25:41] for all the mutations that you have. [25:44] In fact, there was a case from Australia, this computer scientists [25:48] using ChatGPT and Grok, you know, [25:51] created an RNA vaccine for the dog. [25:54] Right. Yes, yes. [25:55] so that's as personalized as you can get because it was that RNA [25:59] vaccine was created just for those cancers’ mutations. [26:04] And there's a lot of trials going on with that. [26:07] So I think it will change completely because, as I said, based on [26:12] not only your genes and mutations but your immune system. [26:16] For example, the immune system is extremely effective in killing cancer cells. [26:21] And this has been a revolution, what we call immunotherapy. [26:24] But it doesn't work in everybody. [26:26] Why is that? [26:27] Why is some people's immune cells able [26:29] to recognize the cancer and kill, but others don't? [26:33] Some of them become exhausted and whatnot. [26:36] So if you could figure that out and of course [26:39] there are also side effects because immune system is very dangerous. [26:42] If you activate it too much, you can actually cause a lot of harm. [26:46] I'm curious, like you seem so AI pilled and so Codex pilled, while also being, like, an MD. [26:52] Like, what are the people around you thinking about AI? [26:55] Like, do you try to get them also, like, to adopt these tools as fast as you are? [26:59] Yeah, I tried very hard and I think they think that I'm totally crazy. [27:04] Although I think nowadays they're starting to see [27:08] the potential. [27:10] And I've been saying this for, since GPT 3.5 came out. [27:16] I think people are very hesitant [27:20] and I can understand that because this is something so new [27:25] that human mind is not able to comprehend this exponential advance. [27:30] For example, people use GPT 4.0 a year and a half ago, which is like ages in AI world. [27:38] And say, well, you know, it was hallucinating too much. [27:41] It didn't really answer to me very well. [27:44] In fact, even, the difference between GPT 5.4 and 5.5 is day and night. [27:50] If you constantly experiment and then assume that it's going to get better and better. [27:57] I trust AI models, [27:59] even for things that I've been working on for 30 years [28:03] because recently, GPT 5.5 Pro model gave me this report [28:08] that I almost cried, I said, how is this possible? [28:11] I remember seeing your post online about that experience. [28:13] Exactly. [28:14] Because that latest model crossed the threshold. [28:17] Of course, GPT 5 and Pro and 5.4, they were fantastic [28:23] understanding the knowledge and the patterns and putting them together. [28:29] But what 5.5 did was [28:31] almost as if it had the same experience that I had working in the lab [28:37] for 30 years because there are certain things that are intuition. [28:41] You just know it's not out there in the literature. [28:44] You feel that, [28:45] for example, I would bet with my students, I say, do this experiment. [28:49] I'll bet you it's going to work this way. [28:51] And they would bet sometimes and they would lose 100% of the time. [28:55] Because, you know, I just know that it's going to work. [28:59] You’ve had so many reps that it’s just your intuition. [29:01] Somehow you have that intuition. [29:03] And that's what I felt with 5.5 that it [29:06] it predicted the outcome of an experiment, very complex experiment [29:11] we had done, at 100% level, which was incredible. [29:17] And I'm curious, like, if we continue on this rate of progress [29:20] that you are experiencing on your side, [29:23] what do you imagine, like your day to day looking like in a few years from now? [29:28] Like, what will be like unmistakably different for people like yourself and researchers around you in this world of biology? [29:35] Well, again, you know, some of the things I say [29:38] may sound very radical but fortunately for me [29:40] but maybe, unfortunately, those who don't believe in this [29:44] is that there's going to be a complete fundamental shift. [29:48] In fact, I call this science 2.0 or 3.0. [29:51] The way we do science is going to completely change. [29:55] The days where we would, for example, [29:58] come up with an idea, plan for experiments for weeks [30:02] and then analyze them for months, those days are over. [30:07] We have to and [30:09] students or other scientists have to realize that we are in a very accelerated timescale, [30:16] that the way we do science is going to be a bunch of AI agents [30:20] that will come up with the hypothesis and they're already able to come up with hypotheses because, again, [30:25] the number of things you can come up with are almost endless. [30:29] You know, how do you know that hypothesis better than that one? [30:31] Then, you know, you get the best ideas. [30:35] Then, it will simulate the experiments for you. [30:38] Because, again, I can do a thousand experiments [30:41] but which ones going to work? I don't know that. [30:43] If it could drive me to a point, says, okay, this type experiment [30:48] likely to work more because of this in this, then I can just focus on that. [30:52] The chances of that experiment working is going to increase tremendously. [30:56] And the data I get, I have almost instant response. [31:00] I'll give it to other AI agents and they will immediately [31:03] analyze and feed it back to the to the master. [31:07] And that master is going to make a new hypothesis and plan for a new experiment. [31:10] So I guess my role is going to be just telling the agents, go figure this out. [31:17] This is the direction you should go. [31:19] I want to cure lung cancer and work on that. [31:23] Go explore. [31:24] Exactly. [31:25] And then, you know, somebody still has to do the experiment. [31:27] But then I think the labs are going to get automated. [31:30] This is already happening. [31:32] And so there's going to be a bunch of robots doing lots of, lots of, experiments. [31:36] When people say, well, you know, but will I still have a job or whatnot? [31:41] You know, there's something called Jevons paradox. [31:45] If we are able to do so much, [31:47] then we will be able to do so much more, especially in biology. [31:51] As I said, you know, [31:52] we can get to a point of bioengineering, but we only know, like 10% of biology. [31:57] Imagine how much more, how fast we can learn. [32:00] And then once we have that power, like I'm building apps now, [32:04] then I'll be able to build new cell types, new tissues. [32:09] So there will be maybe thousands of bioengineers, [32:12] you know, in front of their computers, simulating and building. [32:16] So, but this is going to fundamentally change not just biology, [32:20] I think every field physics and material sciences [32:23] and chemistry, already for drug discovery. [32:27] You can do it in a matter of hours what took years to do. [32:31] And doing clinical trials for doctors treating patients, [32:36] all of that stack is going to change and accelerate. [32:40] And speaking of these other fields, actually, for people [32:43] who are not in science like you are, but maybe like in different fields, [32:47] what would you advise them about your journey? [32:50] Like you've been at this for a long, long time and like, you know, thinking [32:54] about AI and having this curiosity to build with the AI every step of the way, [32:59] what would you like to advise them [33:01] in terms of how do you have to rethink their own field and their own work? [33:06] Yeah. So I have one advantage. [33:08] My my job, my work, my life [33:11] depends on constantly doing experiments. [33:14] And this is very painful because especially in biology, [33:17] 95 to 98% of the experiments you do fail. [33:21] So I'm very used to failure. [33:23] That's why it's called experiment. [33:24] You say, okay, well try this, try that. [33:27] And you develop a certain resilience and agency and curiosity just to try. [33:32] And that's how I was so excited about Codex. [33:36] And I have to, of course, I showed you the things that worked [33:39] but there were a lot of failures before that. [33:41] Many apps that I did didn't work very well. [33:44] Didn't come out very well. [33:45] But you don't, you shouldn't give up. [33:47] And my advice to people would be in the age of AI, [33:51] really agency and curiosity are the only things that matter. [33:55] You don't be scared. [33:57] Constantly try, constantly experiment with AI. [34:00] You know, say, what if I can do it this way, [34:03] you know, because you can now ask that what if question. [34:06] Before it was just too expensive to ask that. [34:10] Let’s say you have a a website for your company. [34:13] You know, you probably spent thousands of dollars [34:16] to have it made and then it wasn't perfect. [34:18] But you say, okay, well, you know, it's good enough. [34:20] I don't want to touch it. [34:22] Now you can say, what if I can just change this and make it this way? [34:26] Just a matter of minutes you have a new website or a new product [34:30] that you can design and then 3D printed, for example. [34:34] I think everything this can be applied [34:37] but then you have to have the courage [34:39] to experiment because the cost of experimentation is so low. [34:43] Why shouldn't you do that? [34:44] Yeah. [34:45] These are very simple daily things but I think you can really extend that [34:50] to pretty much anything in your life. [34:53] You just have to embrace it and look at AI is something incredibly positive. [34:59] And I see a lot of negativity. [35:01] AI is going to do this. [35:03] AI is going to do that. [35:04] I believe just the opposite. [35:05] It's going to really enhance us to a golden age. [35:11] And I'm really super excited about it. [35:14] And I think, you know, AI researchers are for me are all heroes, by the way. [35:19] Because this is going to be the greatest transformation of humanity. [35:24] And I'm super excited for the future. [35:27] Thank you so, so much, Derya. [35:29] It feels like it's amazing, you know, way to close the conversation. [35:33] Very positive. [35:35] And we can't wait to see what are you going to do next. [35:38] And how are you going to push Codex and these frontier models to, like, [35:42] bring all of these worlds together and work on this, like, digital twin idea. [35:47] Absolutely. [35:48] Happy to come back for another episode. [35:52] We'd love to have you back in a few months to see how much progress you've made. [35:55] Until then, enjoy your time in California. [35:57] Thank you. [35:58] Thank you so much. [35:58] Thank you Deyra. [35:59] This was a lot of fun.