Chai Discovery Boosts AI Antibody Hit Rates to 16% in Scaling Bet

押注Scaling定律，Chai Discovery将AI抗体命中率提升至16% Chai Discovery认为，将纯粹的Scaling定律与简化的模型架构应用于生物学，可将药物设计周期从9个月缩短至9天。 Chai Discovery联合创始人Josh Meier和Matt McPartlon认为，药物设计同样遵循“苦涩的教训”（the bitter lesson）：扩展数据、模型和算力带来的收益，远胜于人工构建复杂的生物学特性。通过从不可预测的“药物发现”转向意图驱动的“药物设计”，Chai-2将从头（de novo）抗体设计的成功率从0.1%提升至16%。Chai并不推进自研药物管线，而是作为一家平台基础设施公司，与Eli Lilly和Novartis等大型制药企业展开合作。通过模型架构的极简设计、合成结构生成以及实体实验室数据反馈，他们的愿景是打造一套分子CAD套件，将药物研发迭代周期从9个月压缩至9天。 Chai Discovery 将生物学视为可工程化的问题，坚持“苦涩的教训（bitter lesson）”，通过剔除复杂的手工子模块、极致简化模型架构，依靠扩展数据、算力和参数规模来解锁底层生物规律。 (0:00) 从头抗体设计实现跨越式突破，Chai-2 模型将抗原结合命中率从行业此前约 0.1% 提升至 15%–16%，让研究人员能从密集的反馈信号中实现快速性能爬升与优化。 (15:00) Chai 的终极愿景是构建分子计算机辅助设计（CAD）套件，将药物设计迭代周期从九个月大幅压缩至九天，帮助医药界攻克阿尔茨海默病等传统难成药靶点并推动精准医疗。 (25:00) 公司采用平台基础设施模式，与礼来、诺华和辉瑞等药企展开深度合作，利用高精度模型生成的合成结构与真实湿实验数据构建自增强的数据飞轮。 (30:00)

Chai Discovery founders Matt McPartlon and Josh Meier argue that molecular drug design is fundamentally a computational scaling problem rather than an intractable biological mystery. By applying 'Bitter Lesson' principles—radical architectural simplicity paired with scaled compute and data—they contend biology can transition from empirical trial-and-error discovery into a predictable engineering discipline. Architectural Simplicity and the Bitter Lesson (17:35) Chai's engineering philosophy is anchored in the 'Bitter Lesson,' which posits that general scaling of compute, data, and model parameters consistently outperforms hand-engineered domain complexity. McPartlon notes that Chai's early model, Chai-1, suffered from research drag due to containing 23 distinct submodules. To accelerate iteration and uncover clean scaling laws, Chai stripped away architectural complexity, building subsequent models from scratch with unified architectures. Rather than manually engineering specialized representations for biological mechanisms, the founders argue that sufficiently scaled models learn intrinsic biological features—such as sequence motifs that signal glycosylation sites—directly as training loss decreases. Generative diffusion models facilitate this scaling by decomposing generation into incremental denoising steps. Instead of attempting single-step compression into Gaussian noise like Variational Autoencoders, diffusion networks apply simple learned shortcuts over successive iterations, reliably transforming random noise into biologically viable, complex 3D molecular structures. Zero-Shot Generation and Surging Hit Rates (15:00) Traditional computational drug discovery relies on iterative 'lab-in-the-loop' optimization, making minor adjustments to existing physical compounds. Chai challenges this paradigm by pursuing de novo zero-shot generation—designing completely new molecules computationally with predefined binding affinities, developability, and target specifications. According to Meier, state-of-the-art de novo antibody design historically yielded binding success rates around 0.1% (one hit per thousand tested), creating feedback signals too sparse to systematically optimize for drug-like properties. With their Chai-2 model, Chai reported antibody binding rates rising to roughly 15% to 16% (150 hits per thousand), surpassing the founders' original forecast of reaching 1% within three to four years. This statistical density provides researchers with rich data gradients to evaluate stability, manufacturability, and affinity without screening massive physical libraries, compressing design timelines from nine months to weeks or days. Dual Data Flywheels and Physical Verifiability (33:44) Chai's model training unifies two biological data streams: spatial crystal coordinates from the public Protein Data Bank (PDB) and sequence databases containing trillions of amino acid tokens. Meier, drawing on his work on Evolutionary Scale Modeling (ESM), argues that language models trained on massive sequence corpora implicitly internalize 3D spatial geometry to predict neighboring residues. As baseline models gain predictive accuracy, Chai runs them over large sequence datasets to generate synthetic 3D structures, expanding training datasets for future model generations. This computational loop interfaces directly with the wet lab, using physical experimental 'exhaust' as reinforcement feedback. While biological assays exhibit experimental noise of around ±5%, Meier argues that physical metrics like binding affinity and stability provide absolute physical truths, avoiding the hidden technical debt found in software code generation benchmarks. True progress, therefore, demands evaluating step-function performance leaps across broad target classes, such as difficult G-protein-coupled receptors (GPCRs), rather than cherry-picked tasks. Platform Infrastructure Versus Internal Drug Pipelines (30:00) Rather than building an internal proprietary drug pipeline, Chai chose an infrastructure-focused platform model aimed at developing a molecular Computer-Aided Design (CAD) suite. Meier explains that maintaining an internal pipeline consumes massive capital on clinical trials and risks confirmation bias, where an internal team over-fits to a few target molecules while failing to notice general model deficiencies. To enforce evaluation rigor, Chai partners directly with pharmaceutical companies including Eli Lilly, Novartis, Argenx, and Pfizer. Driven by expiring patent exclusivities, these partners require generalized, out-of-the-box reliability across hundreds of therapeutic targets. This partner-centric model monetizes Chai's platform immediately, creating a commercial loop that funds large-scale GPU infrastructure and model training while testing models against decades of refined wet-lab protocols. The founders anticipate that molecular CAD platforms could fundamentally reshape pharmaceutical economics by 2035–2040, replacing incremental 'first-in-class' compounds with fully optimized 'last-in-class' therapies and making precision medicine for rare diseases economically viable. However, realizing this vision depends on managing the operational strain of massive distributed GPU clusters, consistently outperforming legacy wet-lab protocols, and proving that zero-shot computer-generated molecules translate safely and effectively into active human clinical trials.

Chai Discovery 联合创始人 Matt McPartlon 与 Josh Meier 认为，药物设计本质上是一个模型与算力扩展问题，遵循通用人工智能的“苦涩教训”（Bitter Lesson）。该公司主张舍弃繁琐的人工设计子模块，通过构建架构极简的基础模型与分子 CAD 平台，将传统试错式的生物发现转化为可预测的工程学科。 架构极简主义与苦涩教训 (17:35) Matt McPartlon 总结早期模型 Chai-1 的教训指出，包含 23 个独立子模块的系统过于繁琐，研究人员必须单独追踪每个模块的动态，造成了严重的研发阻力。为了实现有效的扩展规律（scaling laws），Chai 团队转向剥离非核心结构、追求极致的架构极简。他们完全从零开始构建模型，而非在既有生物学模型上微调，以此加速研究迭代并准确识别扩展方向。团队坚持通用机器学习原理，认为生物学并非必须依赖高度定制化的复杂人工架构。 Josh Meier 将 Chai 定位为解决开放性重大 AI 挑战的“新型实验室”（neo-lab），其核心支柱在于坚信数据、算力与参数规模的扩展力量。Matt 观察到，随着模型损失（loss）随扩展规律持续下降，模型会自主涌现出深层生物学表征，例如无须专门设计结构模块即可自动识别蛋白质的糖基化位点基序。团队强调，在生物数据验证中必须保持严谨，因为湿实验本身存在约 ±5% 的误差波动，微小增益很容易产生误导，只有阶跃式的性能跨越和广谱靶点测试才能证明真实的算法进步。 从大海捞针到高命中率从头设计 (15:00) 传统药物发现长期依赖高通量筛选数以亿计的随机化合物，如同大海捞针，或者依赖微调既有分子的缓慢循环。Chai 致力于推动“从头设计”（de novo generation），允许研究人员直接输入理想目标分子的特性并由模型生成。Matt 指出，历史上扩散模型打破了变分自编码器（VAE）单步压缩高维生物结构的局限，通过多步去噪赋予网络逐步优化的时间，从而稳定生成复杂的生物大分子。 在抗体设计领域，行业起步阶段的业界顶尖结合成功率仅约为 0.1%（即 1000 个分子中仅 1 个命中），极度稀疏的反馈使得通过数据统计优化成药性变得不可能。随着 Chai-2 模型的推出，其抗体结合成功率提升至约 15% 到 16%（1000 个分子中产生 150 个命中）。该跃升远超创始团队最初设定的“耗时 3 至 4 年达到 1%”的预期，高密度的数据信号使研究人员得以跳过漫长的大型物理文库筛选，直接根据设定约束生成候选分子。 序列与结构融合的数据飞轮 (33:44) 在数据燃料层面，Chai 将两种范式结合起来：一种是 Matt 所代表的结构进路，专注于蛋白质数据库（PDB）中的三维空间坐标；另一种是 Josh 源自开发 ESM 模型经验的序列扩展进路，认为在大规模氨基酸序列上训练语言模型会自然涌现出对 3D 几何空间的理解。这种机制类似于大型语言模型通过上下文预测下一个词，蛋白质序列模型在预测邻近氨基酸时也会隐式编码三维构象。 由此形成的复合数据飞轮包含多层闭环：基础模型从 PDB 结构与数万亿 Token 的序列数据中学习，高精度基底模型对海量序列进行预测以生成高精度的合成 3D 结构，用于扩充下一代训练集；当生成的分子进入物理实验室测试时，湿实验产生的真实实验“废气”（exhaust data）会重新反馈给模型，形成类似于强化学习的持续优化闭环。与此同时，公司内部科学团队持续对自身模型进行极限测试，指导生成针对性的新实验数据以推动迭代。 分子 CAD 愿景与平台商业模式 (25:00) 与自建临床管线的公司（如 Isomorphic Labs）不同，Chai 选择构建分子计算机辅助设计（CAD）平台。Josh 解释，推进临床试验会占用大量可用于底层算法和算力研发的资金；且仅针对少数内部管线可能“只见树木不见森林”，导致团队误以为能泛化的分子实际上无法扩展至更广泛的靶点。通过与礼来（Eli Lilly）、诺华（Novartis）、Argenx 和辉瑞（Pfizer）等跨国药企合作，外部严格的工业化标准消除了研究人员自欺欺人的空间，所获商业收入也能直接再投资于算力和模型基础设施。 Chai 的长期愿景是将传统长达 9 个月的分子设计迭代周期缩短至 9 周甚至 9 天，从而在经济上使针对复杂罕见病的精准治疗和针对阿尔茨海默病等攻坚靶点的药物研发切实可行。创始团队展望，2035 至 2040 年间的制药行业将告别为赶进度而将“勉强及格（80%）”的候选分子送入临床的妥协做法，转向打造更具特异性、更少脱靶副作用且能彻底解决疾病状态的终极疗法（last-in-class）。 Chai Discovery 展示了将生命科学转变为确定性计算工程的路径，但其长期目标仍取决于诸多现实考验：大规模分布式 GPU 集群在长期训练中的高负荷维护稳定性、计算生成的分子在真实复杂体内环境下对成药性与毒理约束的匹配度，以及由高命中率转化出的临床候选分子能否在人体试验中真正兑现疗效。

## Background 本期节目由红杉资本（Sequoia Capital）于 2026 年 8 月 4 日在其官方播客及网络平台发布，由 Pat Grady 与 Sonali Singh 共同主持。节目围绕 Chai Discovery 展开，深入探讨了联合创始人 Josh Meier 与 Matt McPartlon 如何践行“苦涩的教训”（Bitter Lesson），将原本充满偶然性的生物制药转变为基于算力、数据与模型扩展的现代工程设计问题。 ## TL;DR Chai Discovery 致力于打造分子级的“计算机辅助设计套件”（分子 CAD），打破了传统制药依赖数百万分子盲筛和漫长试错的旧范式。团队通过扩散模型与底层架构的大幅简化，将从头（De Novo）抗体设计的命中率从不足 0.1% 跃升至约 15%，并将药物设计迭代周期从九个月极限压缩至九天。与亲自下场推进自研管线的传统 Biotech 不同，Chai 坚定选择平台赋能模式，将礼来、诺华、辉瑞等跨国药企的严苛实战反馈转化为数据飞轮，推动生物学走向确定性、高精度的工程学科。 ## Atmosphere & style 整场对话洋溢着硬核工程主义与前沿科学碰撞出的理性激情。两位创始人既有从 OpenAI 与 Meta ESM 传承而来的大模型 Scaling 笃信，又有深入湿实验物理真实后的克制与谦逊，谈吐间兼具“把 GPU 跑过热”的黑客极客风趣与攻克疑难病症的厚重使命感。两位红杉投资人穿针引线的提问节奏明快，现场没有空洞的学术概念堆砌，取而代之的是面对超大规模集群工程之痛与见证生成分子走向临床人体试验时的真挚震撼，充满务实、昂扬且极富穿透力的科技创业张力。 ## Bulletin - `16:30` Chai Discovery 将“苦涩的教训”（The Bitter Lesson）引入生物制药，其第二代模型 Chai-2 将从头（De Novo）抗体设计命中率从行业标准的 0.1% 大幅推升至 16%，使分子设计由大海捞针转变为确定性工程。 - `28:37` Chai 坚持平台技术提供商（“卖铲人”）路线，拒绝自建药物管线，以此规避后期昂贵临床试验对核心算力与研发资源的稀释，并通过与礼来、诺华、辉瑞等跨国药企的实战合作倒逼模型泛化能力。 - `25:01` 团队提出“分子 CAD”愿景，致力于将传统制药长达 9 个月的试错闭环压缩至 9 天，并推动新药研发从追求“First-in-Class”演变为无脱靶、零妥协的“Last-in-Class”。 - `21:28` 创始人反直觉地提出“生物学比代码生成更具客观可验证性”，认为湿实验结合力和可制造性等物理读数提供了不可篡改的真值，使基石模型能够像爬山算法一样持续客观迭代。 ## Key takeaways - `00:00` 生物制药正经历从依赖盲筛与偶然运气的“药物发现”（Drug Discovery），向由基石模型直接按需具象化的“药物设计”（Drug Design）范式转变。 - `10:00` 蛋白质 AI 经历了结构预测（AlphaFold 2）、骨架生成（扩散模型）与功能活性设计（Chai-2）三次关键跃迁，扩散模型通过“给予模型逐步思考的时间”打破了复杂构型生成的瓶颈。 - `06:04` 创始人破除生物学“不可预测”的刻板印象，指出抗体与迷你蛋白本质上都是 20 种氨基酸序列构成的统一数学表征，如同大语言模型用相同范式处理不同自然语言一样。 - `16:30` Chai-2 将 De Novo 抗体结合成功率从千分之一推升至约 15%，远超团队原本预计需要三四年才能达成的 1% 目标，使统计学筛选和零样本（Zero-Shot）设计成为现实。 - `18:45` 恪守“极致简约”（Simplicity）准则：团队淘汰了 Chai-1 复杂的 23 个子模块架构，坚决抵制为疑难靶点叠加补丁模块，完全依靠 Scaling Law 驱动模型自发掌握糖基化等微观物理特征。 - `20:00` 生物湿实验存在 ±5% 的固有测量误差，因此 Chai 拒绝局部表面提升，设定了极高的跨越式门槛，要求模型必须展现出数量级式的能力跃迁。 - `21:28` 相比代码生成中主观模糊的代码风格与技术债务，生物学分子的结合亲和力与可制造性等物理读数具备更高的客观可验证性，能为模型提供确凿真实的优化梯度。 - `25:24` 构建分子 CAD 软件的核心价值在于将假设验证闭环从 9 个月压缩至 9 天，极速反馈让原本因试错成本过高而被称为“不可成药”的靶点变得具备工程可行性。 - `26:32` 未来制药行业将告别将“80% 妥协分子”抢推进临床的资源浪费，从争夺最早上市的“First-in-Class”演变为凭借精准设计、零脱靶锁定终极方案的“Last-in-Class”。 - `30:00` 坚持开放平台路线而非封闭管线研发：避免将模型局限在少数自研靶点上“自欺欺人”，通过跨国大药企苛刻的真实工业场景盲测倒逼模型真正的泛化性。 - `31:14` 打破对跨国药企保守的刻板偏见：在残酷的专利悬崖倒逼下，药企巨头一旦看到可靠的盲测数据，便会以极快魄力倾力押注采纳新技术。 - `33:44` 融合 Matt 的“3D 结构派”与 Josh 的“ESM 序列派”，Chai 结合 PDB 高精结构与万亿 Token 级序列数据，利用高质量模型预测与湿实验尾气数据构建加速自愈的复合飞轮。 - `40:00` 面对工程挑战，团队需解决长达数月的超大规模 GPU 集群稳定运行、通信瓶颈与自动故障转移，把工程可靠性与工业级交付标准置于纯科研发论文之上。 - `44:35` 未来 6 到 12 个月的核心目标是推动数十个零样本（Zero-Shot）设计的分子推进至人体临床试验，并实现纳摩尔级结合力精准可控的确定性分子生成。 ## Detailed walkthrough ### `00:00` 简约原则与药物研发的“工程化”愿景 Chai Discovery 的核心指导原则之一是“极致的简约”（simplicity）。在研发第一代模型 Chai-1 时，团队构建了多达 23 个独立的子模块。这种架构在实际迭代中带来了巨大的工程阻力——研究人员不得不单独去理解每一个子模块的行为模式与内部动态，导致整个系统极其难以有效扩展。因此，团队开始反思如何去芜存菁，剥离冗余部分并找出真正起决定性作用的核心组件；一旦理顺了这一逻辑，整个科研探索与寻找 Scaling（规模扩展）方向的路径便会清晰明朗许多。 红杉资本（Sequoia Capital）将 Chai Discovery 定位为“面向生物学的基石模型实验室”（foundation model lab for biology）。针对这一宏大定位，创始人解释道，团队的核心使命是让药物研发过程更贴近现代工程学。当前以大语言模型（LLM）驱动的代码生成之所以成效显著，是因为代码本身就是一种极为纯粹且简洁的抽象层，开发者可以通过它清晰地表达意图；而当下的生物学研究却截然不同，它极度依赖盲目的试错与碰运气。回顾现代生物技术的早期历史，现存的大多数重磅药物几乎都是纯粹偶然（serendipitous）的产物。Chai 正试图通过工业化这一流程，构建起类似软件工程中的多层抽象架构，从而大幅加速研发迭代。 针对主持人关于“是否意味着现实世界实验室测试的边界在逐渐缩小”的提问，团队给出了一个反直觉的视角：实验室绝非被淘汰的对象，而是验证模型真伪的核心基石。AI 领域的根本驱动力在于“可验证性”（verification）——只要能够精准评测模型的输出并予以验证，就能实现爬山算法式的持续性能迭代。因此，问题不在于消灭湿实验，而在于将大海捞针般的“药物发现”（drug discovery）彻底重塑为目标驱动的“药物设计”（drug design）。以往药企需要筛查数百万乃至数十亿个分子才能侥幸命中一个候选分子；而未来的范式是直接向模型输入心目中分子的“终极构想状态”，由模型直接将其具象化实现。团队甚至认为，未来湿实验的规模不但不会萎缩，反而会随着投资回报率（ROI）的飙升而激增，就像软件工程师生产力提升后，社会对工程师的总需求反而爆发式增长一样。 ```mermaid flowchart TD accTitle: 从药物发现到药物设计的范式转变 accDescr: 对比传统生物制药大海捞针式的筛选模式与基于基石模型的逆向定向设计模式。 subgraph Traditional["传统药物发现（Discovery）"] A["数百万/数十亿分子库"] --> B["高通量湿实验盲筛"] B --> C["小概率撞大运（偶然偶得）"] end subgraph Modern["Chai 药物设计（Design）"] D["输入目标状态 Prompt\n（形状/结合位点/治疗功能）"] --> E["生物基石模型\n（端到端生成序列与三维结构）"] E --> F["定向湿实验验证反馈\n（高 ROI 快速迭代闭环）"] end ``` --- ### `03:25` 蛋白质 AI 的演进之路：从结构预测到扩散模型生成 回顾过去十年的发展脉络，蛋白质 AI 经历了翻天覆地的代际跃迁。早期整个领域的技术标尺是每两年举办一次的国际蛋白质结构预测竞赛（CASP）。在 CASP 竞赛中，学术界会预留一批从未公开发布且无人见过的蛋白质靶点盲测各家算法。转折点首先发生在 2018 年，基于深度学习的方法首次实现了阶跃式的性能跨越；随后在 2020 年，AlphaFold 2 的诞生彻底打破了传统预测的上限。 在早期探索阶段，深度学习主要用于预测蛋白质中不同氨基酸残基之间的相对几何距离。算法输出的往往是一张噪点密布、支离破碎的距离图谱，随后由渲染器将其还原为蛋白质的三维模型。彼时的成果虽然具有开创性，但在分辨率上存在严重局限，酷似初代图像生成 AI 渲染出的模糊噪点图；而短短几年间，该领域已经跃升至如同在数秒内生成超高清细腻画作的全新阶段。 当蛋白质三维结构的“正向折叠”预测被基本攻克后，学术界开始转向其逆向问题：给定一个特定的蛋白质空间结构，能否反向设计出一条能折叠成该形状的氨基酸序列？这一方向在 2021 至 2022 年间成为研究热点，并显著拉近了计算生物学与实际药物设计之间的距离。最终，扩散模型（Diffusion Models）的引入成为了打通一切的关键技术节点。它赋予了模型同时生成蛋白质三维骨架结构与氨基酸序列的能力，研究人员得以像向生成式 AI 输入提示词一样，向模型提供更严密的现实物理约束——例如指定某个特定构型的靶标口袋，并明确指令："我想在这个具体位置实现结合"。 - **里程碑节点**：2018 年深度学习首次在 CASP 取得重大突破；2020 年 AlphaFold 2 实现颠覆性高精度折叠预测。 - **技术进阶路径**：从初期的残基间距离预测（类似低分辨率噪点图渲染）逐步演化至序列-结构端到端高精度还原。 - **逆向工程瓶颈突破**：扩散模型让蛋白质序列与空间三维构型得以协同联合生成，支持输入结合位点几何约束等强物理 Prompt。 --- ### `06:04` 破局 2024：抗体从头设计的“圣杯”与统一表征 谈及为何选择在 2024 年成立 Chai Discovery，创始团队分享了一段早期在 Matt 家中的关键讨论。当时 Matt 展示了一组关于抗体-抗原（antibody-antigen）复合物三维结构预测的最新验证结果。长期以来，制药界一直普遍悲观地认为，通过计算手段预测并设计抗体是一个过于艰难的问题。主流观点坚信蛋白质数据库（PDB）中收录的高质量抗体数据太少，纯粹的数据稀缺性会导致基于深度学习的抗体设计遥不可及。受限于此，学术界此前的蛋白质设计研究大多绕道集中在迷你蛋白（mini-proteins）等小分子量结构上；虽然迷你蛋白本身颇具科研价值，但大型制药公司的核心管线与真正具有庞大临床疗效的产品绝大多数都建立在抗体药物之上。 能够直接在计算机中从头设计具备完备治疗功能的抗体，历来被视作该领域的“圣杯”。Chai 团队坚信底层逻辑很简单：如果你在计算机上连抗体的三维结构都无法精准预测，你就根本不可能凭空设计出可用的工程抗体。当结构预测的壁垒被打破时，整个创业机会的技术窗口便彻底敞开。 面对外界对生物学“门槛极高、深不可测”的刻板印象，两位创始人坦言生物学曾经同样令他们感到畏惧。Matt 自身的学术背景完全脱胎于纯数学与理论计算机科学，直到博士三年级才半途转行切入深度学习与蛋白质预测领域。但剥开错综复杂的医学术语外壳，底层的物理与信息本质是高度统一的：无论是所谓的抗体还是迷你蛋白，归根结底都只是一串由 20 种氨基酸按特定顺序排列构成的序列，在模型眼里它们不过是不同形式的 Prompt 罢了。就像 ChatGPT 既能解答复杂的数学推理题，又能顺畅地辅导英语论文写作一样，底层的基石模型只需建立起表征氨基酸序列的统一数学空间，就能以相同的范式同时兼顾不同分子类型的预测与从头生成。 --- ### `07:22` 跨界团队拼图与 OpenAI 基因的注入 促成 Chai Discovery 诞生的不仅是架构的突破，更是多学科思维碰撞的产物。与 Matt 的纯理论数学与理论计算机科学背景互补，Josh 从小就对人工智能与生物学的交叉点充满痴迷。与传统计算机系学生不同，Josh 就读的高中就配备了高规格的干细胞实验室，这使他很早就兼具了湿实验与编程的双重視角。 Josh 的职业生涯起步于早期的 OpenAI。在那段 OpenAI 仍是纯粹非营利机构的黄金科研时期，Josh 身处核心团队，亲身参与了初代 GPT、GPT-2 以及具有里程碑意义的 Scaling Laws（模型扩展定律）的确立过程。在见证了大型模型如何通过海量数据与算力吞吐自发学会理解并输出流畅的英语、德语和法语后，Josh 脑海中挥之不去的核心假设便是：既然模型能够学会人类的自然语言，它们为什么不能学会生命底层的通用语言——DNA 与蛋白质？ 这一底层追问直接构成了 Josh 多年来的科研核心母题，并最终促成了他与 Matt 在计算生物学前沿的汇合。在他们看来，制药行业长期以来只在细枝末节的技术边缘打转，而 Chai 的真正突破点在于将最顶尖的深度学习架构、科学设计的算力基础设施规模，以及恰到好处的任务抽象精准结合。在 2024 年初，正是依托扩散模型等新一代架构，团队首次看到了在计算机中生成具有真实生物活性的复杂抗体分子的曙光，这最终促使他们迈出创立公司的决定性一步。 > “如果模型能够学会说英语、德语、法语，为什么它们就不能学会说 DNA 和蛋白质的语言？” —— Josh Meier 回忆在 OpenAI 见证 Scaling Laws 确立时的核心思考。 ### `10:00` 扩散模型的直觉：从变分自编码器（VAE）到给模型更多“思考时间” 在扩散模型（Diffusion Models）兴起之前，业界在图像生成等领域尝试过大量生成建模方法，并积累了诸多工程与算法技巧。Chai Discovery 在生物学分子设计中大量借鉴了这些源自计算机视觉的经验，但真正的质变发生在引入扩散模型之后。在扩散模型之前，蛋白质生成设计的主流范式是变分自编码器（VAE），其核心逻辑是“单步压缩”——试图将高维复杂的蛋白质结构分布一步直接压缩映射成模糊的高斯向量（Gaussian vectors）。在当时的技术条件下，这种单步压缩映射对复杂蛋白质而言极其困难；即便放到今天倾注全力尝试，或许能够攻克，但在早期阶段该方法根本无法满足实际需求。 扩散模型之所以在生物学领域表现惊艳，其底层直觉在于“它绝非一步到位的过程”，而是“给予了模型更多的时间去思考”。扩散模型不再奢求模型瞬间吐出一个完美的分子结构，而是展示一个“稍微有点破损残缺的蛋白质”，教会模型一步步修复它；进而通过不断加噪，将蛋白质逐步破坏得更加破碎、杂乱，让模型学会微小的局部修复捷径（shortcuts）。只要不断向模型下达指令——“让它稍微变好一点点，再稍微变好一点点”，将原本极其宏大的生成难题拆解为细粒度的微小修正步，这种逐步去噪的渐进机制与生物分子规律形成了极其契合的共振。 主持人在这里插了一句幽默的题外话，笑称这种“让它稍微变好一点”的过程特别像他和女儿们玩 ChatGPT 的游戏：他们先让 AI 画一只独角兽，然后不断追加提示词“把它变得更强壮”，一遍又一遍地要求“还要更强壮”，直到最后造出了“全宇宙最强壮的独角兽”。Chai 团队打趣回应：“差不多就是这么回事，就是这么简单！好吧，或许实际原理还是有点区别的。” ```mermaid flowchart LR accTitle: 蛋白质生成范式对比 accDescr: 对比变分自编码器的单步压缩与扩散模型的渐进式去噪生成机制 subgraph VAE["传统范式：变分自编码器 (VAE)"] V1["高维蛋白质分布"] -->|"单步高难度压缩"| V2["高斯潜变量空间"] V2 -->|"单步解码"| V3["结构预测 (极难收敛)"] end subgraph DIFF["现代范式：扩散模型 (Diffusion)"] D1["完全高斯噪声"] -->|"逐步修复优化"| D2["低噪破损结构"] D2 -->|"“让它稍微变好一点”"| D3["高精度蛋白质设计"] end ``` --- ### `11:41` 组建生物与 AI 的“复仇者联盟”：按研发里程碑精准引援 在计算生物学这个交叉领域，团队必须汇聚化学家、生物学家以及顶尖 AI 研究员，宛如组建一支精通多学科语言的“复仇者联盟”。谈及如何吸引并选拔这些跨学科明星人才，Chai 展现出了极强的实用主义导向：团队不会盲目扩张，而是根据每一代模型的演进里程碑，极其务实地调整并扩展所需的人才画像。在创立之初，Chai 的核心团队几乎清一色是专注于模型规模扩展与落地算法的顶级 AI 研究员；而到了研发推进至 Chai-2 这一代时，团队才开始有针对性地招募全球顶尖的抗体工程师与生物学家。 团队引入的明星科学家包括业内知名学者 Andy Yang。当 Andy 宣布加盟 Chai 时，外界甚至误以为 Chai 已经转型为一家全栈型的自主药物管线研发公司，因为业内人士纷纷感叹：“如果 Andy 都加入你们了，你们不顺势去做全流程新药研发简直是疯了。”然而在 Chai 强大的计算设计平台赋能下，Andy 在过去短短几个月内实际运行的抗体设计项目数量，已经远超他此前整个职业生涯的总和。团队中还有像 Nathan Rollins 这样的青年天才：他从小接受居家教育（homeschooled），极早便进入大学，在年仅 14 岁时就加入了刚刚斩获诺贝尔化学奖的 David Baker 蛋白质设计实验室，并在 18 岁开启博士生涯，脑海中有着源源不断的突破性创意。 - **阶段性按需招募**：绝不在抗体设计模型具备可行性之前盲目招聘抗体工程师；人才的进驻与底层模型的工程成熟度严格挂钩。 - **冷启动的技术爬坡**：在抗体科学家刚进驻 Chai 时，Chai-2 起初甚至还无法支持多特异性抗体（multi-specifics）等新一代复杂分子格式，工程师团队在数周内以极快速度迭代算法，才使前沿实验案例顺利上线。 --- ### `13:30` 极致的工程文化与黑客精神：把 GPU 跑到过热与精简团队 随着模型能力的爆发，Chai 迅速构建了一支兼具顶尖产品构筑能力与底层系统工程攻坚力的工程团队。为了确保强大的底层模型能够转化为药企客户真正可用的产品界面，Chai 招募了来自 Stripe 等头部科技公司的骨干人才——联合创始人 Jack 曾深耕于 Stripe，同时引入了曾是 Stripe 核心代码贡献主力之一的 Munaz；鉴于模型未来将部署至大型制药跨国巨头内部，数据资产的安全性至关重要，团队还吸纳了曾自主创办网络安全公司的 Neil。在底层基础设施方面，Chai 聚集了一批精通“GPU 底层魔改与算力优化”（GPU hacking）的硬核工程师，确保大规模任务并发运行时系统坚如磐石。 Josh Meier 在此分享了一个有趣的幕后故事：前些天团队收到了来自某家算力超算云厂商（Hyperscaler）发来的 Slack 告警消息，对方称 Chai 的计算集群出现了故障，原因是“你们的 GPU 温度过高彻底过热了，因为你们一次性跑了太多任务”。Chai 团队对此幽默地回应：“GPU 跑满算力难道不正是我们的目的所在吗？这恰恰证明我们尽职尽责地把硬件性能压榨到了极致。” Chai 始终坚持维持极其精简的小型团队规模，让全员始终保持在“轻微超负荷工作”（slightly over capacity）的状态中。这种精干的团队结构能够倒逼所有人聚焦于优先级最高、真正核心的要务，杜绝冗余内耗。与此同时，团队贯彻了极致的代码与执行文化：联合创始工程师 Kevin Wu 曾主导开发了全球首个蛋白质扩散模型，其惊人的工程执行速度正是整个团队的缩影——在 Chai，即使是最纯粹的 AI 研究员，骨子里也必须是极其过硬的顶级软件工程师。 ### `15:00` 软件工程标准与蛋白质 AI 的三次“顿悟时刻” 在团队建设和研发底座上，Chai Discovery 有着极其严苛的工程偏执。联合创始人明确指出，尽管团队汇聚了顶尖的 AI 研究员与蛋白质设计师，但从本质上讲，Chai Discovery 是一家软件公司，最终交付的是高水准的软件产品。因此，团队在招募能够拓展前沿技术的研究天才的同时，始终将代码库的质量与工程标准置于核心地位。 主持人回顾了蛋白质 AI 领域经历的几次标志性“顿悟时刻”（Aha moments）：首先是 AlphaFold 证明了 AI 可以精准预测蛋白质如何折叠；紧接着，扩散模型（Diffusion Models）带来了第二次顿悟，证明 AI 能够无中生有地生成蛋白质结构；而 Chai Discovery 最新模型所引发的第三次顿悟，则是生成的分子不仅“看起来像蛋白质”，更具备切实的治疗潜力——能够紧密结合靶点，并展现出极高的命中率。 ```mermaid flowchart LR accTitle: 蛋白质AI技术的三次跃迁演进 accDescr: 展现从AlphaFold的折叠预测，到扩散模型的结构生成，再到Chai-2具备高活性治疗特性的演化路线。 subgraph S1["结构预测时代"] A["AlphaFold<br/>蛋白质折叠预测"] end subgraph S2["从头生成时代"] B["扩散模型<br/>生成蛋白质骨架/外观"] end subgraph S3["功能设计时代"] C["Chai-2 等前沿模型<br/>高亲和力结合与成药特性"] end A --> B B --> C ``` --- ### `16:30` 从 0.1% 到 15%：从头分子设计（De Novo）的命中率飞跃 谈及模型生成的分子质量时，创始人表示，这契合了公司成立之初的核心假设：专注“从头生成”（De Novo generation）。当时行业内普遍的做法是拿已有分子做微量渐进式修饰，并依赖“实验在环”（lab-in-the-loop）在湿实验与计算之间缓慢倒腾。Chai Discovery 提出的核心问题则是：我们能否直接在计算机中完成全部工作？是否拥有或能够收集足够的数据，从而直接“零样本”（Zero-shot）生成具备优良成药属性的分子？ 要实现这一愿景，首要攻克的便是成功率极低的瓶颈。公司初创时，抗体设计的业内顶尖水准（SOTA）结合率仅为 0.1% 左右，意味着设计出的分子在湿实验中每一千个仅能结合一个。这种极低的结合率带来了双重困境：不仅需要极大规模的筛选，而且算法在迭代中获得的反馈梯度极其微弱，对许多难成药靶点更是颗粒无收。而 Chai-2 模型将这一命中率大幅提升至约 15%（即每筛选 1,000 个分子便能斩获 150 个阳性命中）。这彻底改变了药物研发的游戏规则，使研究人员能够从统计学层面深入剖析分子属性、搭建实验室评估基准，并在算法引擎中直接融入各类药物特性。 - 行业早期基线：抗体结合成功率仅约 0.1%（1/1000 结合率），筛选成本高昂且数据反馈极度匮乏。 - Chai-2 性能跃升：抗体结合成功率推高至约 15%，单批次千量级筛选可获得约 150 个命中分子。 - 预期差距：创始团队在早期规划预算时，曾以为达到 1% 的命中率就需要 3 到 4 年，原本认为“百分之一的结合率就已经是破天荒的突破”，实际进展的速度连他们自己都大感意外。 --- ### `18:45` 坚守“苦涩的教训”：以极致简约寻找生物学 Scaling Laws 这种超预期进展的背后，是 Chai Discovery 对通用机器学习成熟法则的坚定套用。许多人误以为生物学是一个高度定制化、特异性的杂乱领域，但创始团队坚信生物学问题与自动驾驶、大语言模型（LLM）底层遵循着相同的逻辑。Chai Discovery 是一家极其深信“苦涩的教训”（Bitter Lesson-pilled）的公司，坚信通过扩展数据、扩展模型容量和扩展算力才能带来本质突破，而这一切的前提是必须找到生物学维度的扩展定律（Scaling Laws）。 在应对生物学分子的 Token 化和架构设计时，团队的核心准则是“极致简约”（Simplicity）。例如在早期模型 Chai-1 中，架构内包含了 23 个独立的子模块。在研发迭代中，研究人员必须去理解每一个子模块各自的运行动态与协同行为，导致系统极难扩展。Chai 团队通过大刀阔斧地剔除冗余、聚焦最本质的机制，使得模型架构大幅简化，从而能够清晰识别算力与数据扩展的方向。 与直接微调现成大语言模型（如 GLM）处理蛋白质序列的做法截然不同，Chai Discovery 的模型全都是从底层（ground up）全新构建的。创始人将公司定位为一个“新形态实验室”（Neo Lab）式的探索性前沿机构：面对生物学这一开放式的 AI 难题，唯有依靠清晰、坚定的底层工程哲学作为航标，在无人区中持续迭代。 > “我们是一家极其深信‘苦涩的教训’（bitter lesson-pilled）的公司……当你在 Chai-1 这样的模型里塞进 23 个独立子模块时，迭代就会变得无比艰难，因为你必须挨个理解它们的行为，这根本无法有效扩展。” ### `20:00` 坚持苦涩的教训与湿实验的高标准验证 在探讨公司的核心指导原则时，Chai Discovery 的创始人将其归结为两大支柱：极致的“简洁性”（Simplicity）与严苛的科学求真态度。他们深受人工智能领域“苦涩的教训”（The Bitter Lesson）的启发，坚信通过持续扩展计算资源、数据集以及模型参数规模，远胜于手工搭建繁复脆弱的专家规则系统。这种对简洁的执念在生物学领域尤为珍贵，因为生物实验极易让人“自欺欺人”——在湿实验（Wet Lab）环境下，测量本身的误差范围（Error bars）通常就相当大。 与代码生成领域有着鲜明对比。在 SWE-bench 这类基准测试中，算法水平从 16% 逐步微调提升至 17%、18%，通常就代表着可见的技术优化；但在生物实验室中，由于本身就存在正负 5% 的实验波动，如果算法带来的提升只有几个百分点，往往根本无法区分这究竟是模型的真实突破，还是实验室的背景噪音。因此，Chai 对模型设定了极高的跨越式门槛，要求模型必须展现出数量级式的飞跃（Step changes）。他们坚决避免为了在一两个特定新任务上“看起来很漂亮”而硬凑出复杂的专门模型，因为唯有具备普遍泛化能力（Generalization）的模型，才能真正推动整个行业向前发展。 - 湿实验误差范围通常在正负 5% 左右，这意味着生物基础模型微小的表面提升往往只是实验噪音，必须依靠大幅度的阶段性跃升来证明能力。 - 拒绝为了个别边缘任务量身打造定制化小模型，Chai 的核心原则是将资源集中于可规模化扩展、能泛化到广泛分子结构的通用架构。 --- ### `21:28` 为什么生物学比代码生成更具“可验证性” 主持人提出，生物学向来被外界视为一个极其模糊、难以验证的黑盒领域，而 Chai 却频频向客户展示其模型的客观进展。对此，联合创始人 Josh Meier 给出了一个反直觉的论断：“生物学其实比代码生成拥有更加客观且可验证的衡量标准。” 在代码生成（CodeGen）场景下，验证代码能够编译通过、甚至跑通特定的单元测试固然直接，但代码风格与架构“品味”（Taste）却极度主观——模型是否写出了冗长臃肿、难以长期维护的烂代码？其潜在的技术债务极难量化。相反，在分子设计的湿实验中，每一个关键物化指标都可以被极其精确地量化。例如，候选分子是否能与靶点结合？它是否具备可制造性（Manufacturability）？在生物学视角中，生产制造难度就是它的“技术债务”，而这种债务可以通过实验室数据确凿地测量出来。尽管湿实验的反馈回路并不像运行代码单元测试那样只需五秒钟，而是需要几天甚至几周的时间，但这种等待换来的是无可辩驳的真实反馈，让研究人员能够诚实地评估模型的真实水平。 ```mermaid flowchart TD accTitle: 代码生成与分子设计的可验证性对比 accDescr: 展示了代码生成在品味与技术债务上的主观性，与生物分子设计在结合力和可制造性上的客观实验验证对比。 subgraph CodeGen["代码生成 (CodeGen)"] C1["单元测试与编译运行"] --> C2["快速反馈 (数秒内)"] C2 --> C3{"代码品味与技术债务评估"} C3 -->|主观模糊| C4["难以量化维护成本"] end subgraph BioDesign["分子设计 (Bio Design)"] B1["靶点结合实验与理化测量"] --> B2["实验室反馈 (数天至数周)"] B2 --> B3{"可制造性与结合亲和力"} B3 -->|客观精确| B4["确凿的物理实验读数"] end ``` --- ### `22:42` 拒绝拼凑模块：依靠 Scaling Law 攻克“不可成药”靶点 在讨论如何攻克传统筛选方法无法企及的“不可成药”（Undruggable）靶点时，Chai 再次展现了其模型哲学的定力。在研发过程中，团队经常会面临一种诱惑：“如果我们再专门增加第 24 个定制模块，是不是就能立即搞定这个疑难靶点？”但创始团队每次都会审慎地反思：这是否是值得长期维护的架构？这究竟是局部的渐进式修补，还是一种能够持续复利的技术积累？它能否帮助模型泛化到整个一类原本无法成药的靶点？ Chai 坚定地站在了 Scaling Law（缩放定律）这一边。他们深信，只要不断将训练损失（Loss）推向更低，模型为了更好地拟合全局规律，就必然会被迫自发理解靶点最深层的内在物理与生物规律。例如，团队在探讨蛋白质糖基化位点（Glycosylation sites）时，曾纠结是否需要手工设计一套特征表示来显式标注。最终他们意识到，只要模型持续学习，特定的一级序列基序（Sequence motifs）本身就会向模型揭示此处存在糖基化修饰。为了持续压低 Loss，模型自然会学会解析这些隐性特征。因此，攻克原本棘手的生物学靶点，底层的解法不是给系统打补丁，而是坚决通过更大、更好的基础模型去全面解锁。 - 抵制“第 24 个专用模块”的技术债诱惑，优先选择具有复利效应和通用泛化潜力的模型演进路径。 - 理论上隐含在氨基酸序列中的生物物理线索（如糖基化位点），在 Scaling Law 的驱动下会被模型自主习得，无需人工硬编码规则。 --- ### `24:05` 跨学科协同：各团队对“难题”的不同定义与并行演进 Chai 内部跨学科协作的一个有趣现象是，不同团队对究竟什么是“硬核难题”（Hard target）的定义截然不同。然而在创始人眼中，这种视角的碰撞非但不是认知分歧的缺陷，反而是系统协同推进的独特优势： * **湿实验科学团队（Science Team）**：他们眼中的难题是经典意义上的“不可成药”结构，例如某种至今无人能够通过功能性分子对其进行变构或调控的 GPCR（G蛋白偶联受体）靶点。 * **机器学习算法团队（ML Research Team）**：他们定义的难题完全是表征与计算层面的——例如某种结构异常复杂、空间构象极其罕见、模型根本无法正确折叠和预测的蛋白质构型。 * **产品与工程团队（Product Team）**：他们面对的挑战则是交互与数据表达层面的——面对带有复杂翻译后修饰（如糖基化）的多跨膜蛋白，如何向药物研发科学家直观、精准地在用户界面上渲染和呈现这些结构信息？ 这种多维度的难题定义被 Chai 视为一种“特性而非缺陷”（Feature rather than a bug）。这意味着公司不需要将所有精力赌在一个单一的技术死结上，也不存在“只要没有某个额外模块系统就彻底瘫痪”的单一瓶颈。科学团队、算法团队与产品团队各自沿着平行的维度持续向前推进，共同促成了平台整体平稳、连贯且不可逆的技术进化。 ### `25:01` 分子级 CAD 愿景：将生物学重塑为工程学科 Chai Discovery 的雄心并非成为一家手握三五个临床管线、押注一两款新药上市的传统生物科技公司，而是要彻底重构人类发现药物的范式。Josh Meier 强调，公司的核心目标是打造一套针对分子的“计算机辅助设计套件”（Computer-Aided Design Suite，简称分子 CAD）。在药物研发的历史长河中，人们无论面对多么棘手的靶点，往往只能依赖极其笨拙的手段；而 Chai 试图通过构建统一的底层计算解决方案，从根本上攻克这些难以成药的硬核靶点。 所谓的分子 CAD，本质上是将充满不确定性与偶然性的生物学研究，彻底转变为一门标准化的工程学科。传统药物研发如同在大海中捞针，科学家们需要构建极其庞大的物理化合物库，进行成千上万次盲目的试错与筛选。而分子 CAD 的逻辑完全颠倒：研发人员在项目伊始，就可以在软件中明确输入分子的各项设计原则与理化约束，由背后的 AI 引擎直接生成符合要求的分子结构，再交付湿实验室进行物理合成与验证。 尽管在真实的研发过程中，由于初始假设的偏差，科研人员依然需要在湿实验室与计算模型之间进行几轮迭代，但 CAD 软件彻底颠覆了从“提出构想”到“形成可检验假设”的速度。正如 Josh 所指出的，工程化的核心不在于一次性消灭所有湿实验，而在于让每一次试错都高度受控且迅速。 - **愿景对比**：摒弃传统 Biotech 押注 5 个候选药进入临床的商业路径，转向打造通用的分子级计算工程底座。 - **范式转移**：从依赖大型文库“盲筛捕捞”（fishing out of a library）与海量试错，转变为“前置指定设计原则，AI 直接生成分子结构”。 --- ### `25:24` 缩短设计闭环：从九个月压缩至九天 分子 CAD 套件带来的核心变革在于迭代周期的断崖式压缩。目前在制药行业中，从锁定一个靶点概念到真正合成出可验证的候选分子，一个循环通常需要耗费整整九个月。Josh 指出，如果能将这个周期压缩到九周，甚至进一步缩减至九天，那么研发团队能够探索的假设空间和创意数量将迎来指数级的跨越。 研发速度与靶点难度在本质上是相互收敛的。行业内常常争论究竟应该优先关注“研发速度”还是“攻克靶点的难度”，但在 Josh 看来这两者密不可分——一个靶点之所以被定义为“不可成药”或极其困难，往往是因为现有的试错成本太高，人们在九个月一次的漫长迭代中耗尽了资源与耐心；一旦假设的验证闭环缩短数个数量级，原本不可逾越的高难度靶点就会在快速试错中被迅速攻破。 ```mermaid flowchart LR accTitle: 药物设计闭环效率对比 accDescr: 对比传统制药长达9个月的试错周期与CAD套件驱动的9天超快迭代循环 subgraph TRAD["传统发现流程（~9个月/循环）"] direction TB T1["模糊假说"] --> T2["文库盲筛试错"] T2 --> T3["漫长湿实验验证"] T3 --> T4["周期漫长/吞吐受限"] end subgraph CAD["分子 CAD 引擎（~9天/循环）"] direction TB C1["精准参数约束"] --> C2["AI 生成定向分子"] C2 --> C3["靶向湿实验验证"] C3 --> C4["高频迭代/攻克硬靶点"] end TRAD -.->|"效率跨数量级提升"| CAD ``` - **数量级跃迁**：假设验证周期从常规的 9 个月，压减至 9 周甚至 9 天，每一次数量级的提升都会成倍放大可探索的科学构想空间。 - **速度与难度的统一**：速度本身就是解决高难度靶点的利器，极速反馈闭环让原本无法落地的科学假说具备了工程可行性。 --- ### `26:32` 站在未来看药物研发：从“首创”走向“终极（Last-in-Class）” 在主持人的引导下，对话跳出眼前的技术局限，采用“站在未来看现在”（future-back）的视角，畅想 2035 年甚至更远时期、当分子 CAD 成为行业标配时的制药景象。Josh 描摹了一幅极具变革性的行业图景：未来的核心变化首先是上市药物质量的质的飞跃。在今天的临床管线中，充斥着大量妥协的产物——很多分子可能只达到了理想状态的 80%，但药企为了赶在进度节点前交差、尽早进入市场收回成本，依然选择将其推进临床。这类药物虽能使部分患者获益，但往往在一年后就被竞品击溃，造成了行业研发资源的巨大浪费。 随着分子计算设计的普及，过去那些因经济效益低下或技术难度过高而被放弃的疾病领域将迎来曙光。阿尔茨海默病等久攻不克的神经退行性疾病有望被真正攻破；而受限于患者群体规模过小、传统研发模式下算不过账的罕见病与高度定制化的个性化药物，也将因研发成本的大幅摊薄而具备商业可行性。 更深刻的行业范式转变在于研发目标从追逐“首创”（First-in-Class）或“同类最优”（Best-in-Class），演变为追求成为“同类终极”（Last-in-Class）。未来的药物分子将具备前所未有的“精准设计意图”（intentionality），在计算阶段就彻底消除脱靶效应与非预期相互作用，成为针对特定病灶的最完美、不可被超越的“最终答案”。 - **终结 80% 的妥协**：解决行业当前因抢夺研发进度而将半成品分子推入临床的资源浪费现象。 - **激活被遗忘的赛道**：攻克阿尔茨海默病等高难度靶点，并让极低患者基数的罕见病和个性化疗法在经济上变得可行。 - **全新竞争维度**：从争夺最早上市的“First-in-Class”，演进为凭借极高特异性与零副作用锁定终极解决方案的“Last-in-Class”。 --- ### `28:37` 平台商业模式之辨：为何不做药企，而是为药企“造铲子”？ 谈及 Chai Discovery 与同样由顶级 AI 人才创立的 Isomorphic Labs（Alphabet 旗下从 DeepMind 拆分出的药物研发公司）之间的战略差异，主持人指出：Isomorphic 选择了一条亲自下场开发药物管线的全栈道路，而 Chai 却选择了赋能现有制药产业链、提供底层基础设施的技术平台路线。Josh 对此剖析了 Chai 做出这一决断背后的核心底层逻辑与团队基因。 Josh 坦言，自己和联合创始人 Matt McPartlon 以及创始团队的多位核心成员，过去都曾在全栈型 AI 制药企业打拼过，亲历过从零训练模型、筛选分子并一步步送进临床试验的全流程。正是这段经历让他们彻底成为了坚定的“苦涩教训信徒”（Bitter Lesson-pilled）。在传统全栈模式下，随着项目推进，企业必须将绝大部分资金和资源抽离核心算法与模型研发，砸向耗资极其巨大的湿实验临床试验之中。 Chai 的核心信念是：计算机辅助设计必将成为未来绝大多数药物诞生的必由之路，因此行业最紧缺的是能够支撑整个时代的底层技术基础设施。Chai 坚持平台模式，能够构建一个极其健康的自我强化飞轮：随着 AI 模型能力的不断提升，Chai 能够赢得更多顶级药企的商业合作，这些真金白银的技术授权与服务收入能够被立刻反哺回底层模型的算力与数据迭代中，而非被临床研究所吞噬，从而在工程与商业上形成极具可扩张性的正向循环。 - **商业分野**：Isomorphic 押注自研药物管线资产，Chai 坚定定位为整个制药工业的计算底层基础设施（卖铲人）。 - **避开资源泥潭**：恪守“苦涩的教训”（The Bitter Lesson），避免因开展高成本临床试验而稀释核心模型研发的算力与资金。 - **正向飞轮效应**：更强大的基础模型赢得更多药企商业订单，商业回报全额反哺算力与模型迭代，确立长期的模型领先优势。 ### `30:00` 拒绝自欺欺人：为何坚持平台合作而非自建管线 Chai Discovery 从创立之初就确立了与整个生物医药生态广泛合作的路线，而不是把技术关在门内去跑几条自己的专有药物管线。创始人解释道，这一抉择的底层逻辑在于避免在生物学领域“自欺欺人”。如果一家公司只押注于少数几个药物靶点，即使针对这几个靶点打磨出了极为精妙的分子、创造了局部的巨大价值，团队也很可能“一叶障目，不见泰山”，根本无法察觉自己的模型是否能泛化到行业每年合成的其余 500 个分子上。 走向开放合作则彻底断绝了侥幸心理。Chai 的合作伙伴包括礼来（Eli Lilly）、诺华（Novartis）、argenx、辉瑞（Pfizer）等顶尖跨国药企，这些巨头绝不会轻易把技术宣称当作既成事实，只有实打实地交付价值，合作才能推进。如果走自研药物管线的路子，模型预测哪怕有瑕疵，往往还能指望内部化学家团队后续“缝缝补补”，直到进入临床阶段才见真章；但在平台合作模式下，交付的模型必须直接管用，没有打掩护的余地。这种模式在商业上极其苛刻，这也是此前极少有团队敢于践行的原因，但 Chai 坚信，如果能把这条难走的路彻底走通，它所创造的行业影响力与回报将无可比拟。 - **泛化性检验**：模型如果仅针对少数靶点优化，极易掩盖其无法泛化到成百上千种复杂分子的本质缺陷。 - **合作倒逼严谨**：面对礼来、诺华、argenx 与辉瑞等挑剔的顶级药企，交付的模型必须稳定输出价值，不能寄希望于“化学家手动擦屁股”。 - **商业路径差异**：自建管线允许在实验后期修补缺陷，而平台赋能模式要求算法在最前沿直接面对盲测与真实工业场景的检验。 ```mermaid flowchart TD accTitle: Platform Partnership vs Closed Pipeline accDescr: 比较封闭药物管线与开放平台合作两种模式在模型泛化性与外部反馈机制上的差异 subgraph Closed["封闭药物管线模式 (Closed Pipeline)"] A["仅聚焦少数内部靶点"] --> B["设计出精巧分子"] B --> C["化学家手动修补算法缺陷"] C --> D["极易自欺欺人，缺乏泛化性"] end subgraph Open["Chai 开放平台合作模式 (Broad Partnering)"] E["跨国大药企盲测验证<br/>(Eli Lilly, Novartis, Pfizer 等)"] --> F["算法必须开箱即用，真实见效"] F --> G["多场景数据与边界反馈"] G --> H["驱动模型具备通用泛化能力"] end ``` ### `31:14` 药企认知的反差与专利悬崖下的创新动力 当 Chai 走出学术实验室、深入制药工业一线时，外界普遍流传着“大型药企保守笨拙、根本不懂如何运用前沿 AI 技术”的刻板印象。然而实际接触后，创始团队发现事实恰恰相反：跨国大药企不仅具备极高的专业严谨度，甚至在部署新技术前会对算法的每一项技术宣称进行地毯式审查与复测；一旦盲测数据得到确证，药企高层的拥抱速度与魄力往往是“倾力押注（go all in）”。这种敏锐嗅觉本质上是由制药产业残酷的商业逻辑驱动的——新药受到专利独占期的严格限制，一旦遭遇专利悬崖，现金流便会骤降。即便是如今市值达到万亿美元级别的礼来，如果无法持续产出新的重磅炸弹（blockbuster）药物，其地位也绝不可能长久。 这种生死存亡的紧迫感迫使跨国药企保持高度竞争态势，一旦 Chai 的模型跨过准入门槛，药企团队便会迅速扩大技术采纳规模，并展现出极具创造力的新用途。同时，Chai 内部设立了专职的科学家团队进行高强度的“自产自销（dogfooding）”，全天候挖掘模型在结合剂（binders）与抗体生成中的潜力与边界。Chai 从未抱有“打造一个终极模型就能解决所有生物学问题”的幻想，而是将研发拆解为类似其他现代工业的渐进式迭代：通过前线合作与内部实战解锁新的应用场景，催生高质量的全新数据，进而训练出性能更优的下一代模型。 - **打破行业偏见**：大型药企绝非拒绝技术的前沿落后者，而是一旦看到坚实数据就会迅速全线押注的敏锐买家。 - **专利悬崖机制**：药物专利独占期终会到期，万亿市值巨头若没有源源不断的重磅产品储备便难以为继，从而构成了采纳前沿技术的绝对推力。 - **内部实战与外部反馈**：Chai 科学团队与外部客户同步进行算法“吃狗粮”式测试，重点覆盖结合剂与功能性抗体的实际生成与验证。 ### `33:44` 数据飞轮与技术分野：结构派与序列语言模型 关于底层算法迭代的核心燃料——数据，生物医药 AI 面临着与其他深度学习领域截然不同的困境：团队无法像训练大语言模型（LLM）那样从互联网抓取海量网页，生物数据的沉淀方式极其特殊。生物物理领域的“黄金标准”数据源当属蛋白质数据库（Protein Data Bank, PDB）。自 1970 年代以来，几代实验科学家在实验室中日积月累，耗费心血测定并上传蛋白质及各类分子的三维晶体结构数据；可以说，没有半个世纪以来 PDB 的无私积累，现代结构预测与从头药物设计根本不可能诞生。 在对这些数据的利用思路上，Chai 的两位联合创始人早年分别代表了行业内两条截然不同却高度互补的技术流派：联合创始人兼 CTO Matt McPartlon 是坚定的“结构派（structure-pilled）”，执着于如何让机器学习模型直接输出三维空间坐标，这是一套完全不同于通用语言模型或计算机视觉的独特架构体系；而联合创始人兼 CEO Josh Meier 当年则走向了完全相反的“序列语言模型派”——他曾是 Meta 划时代蛋白质语言模型 ESM（Evolutionary Scale Modeling）的早期核心作者之一。将语言模型直接施加在蛋白质一维氨基酸序列上的研究表明，仅凭纯粹的序列理解，模型同样能够捕捉深层的生物进化法则，而 Chai 正是融合了这两派认知集大成的产物。 - **生物数据无法网络爬取**：模型输入严重依赖高精度的实验物理测定数据，而非低质量公开文本。 - **PDB 的奠基性意义**：自 1970 年代起由全球结构生物学家以实验晶体结构沉淀构建，是结构预测与分子设计的基石。 - **技术路线的双剑合璧**： - **Matt 的结构路线（Structure-pilled）**：专注于让深度学习直接理解并生成 3D 空间原子坐标，跳出传统 CV/LLM 的框架。 - **Josh 的序列路线（ESM）**：将蛋白质氨基酸序列视作生命语言，利用语言模型框架从海量自然演化序列中捕获生物机理。 ### `35:00` 序列与结构的融合：生物数据的飞轮效应 Matt McPartlon 详细解释了模型如何从纯序列中隐式学习到三维物理结构。就像大语言模型预测句子中的下一个单词一样，蛋白质语言模型预测的是序列中的下一个氨基酸。为了精准预测缺失的氨基酸，模型必须深刻理解该位置所处的微观局部环境（microenvironment），识别哪些氨基酸与周围环境兼容，而要做到这一点，模型内部就必须构建出蛋白质的三维形态表征。Matt 最初倾向于基于结构的建模路径，而 Josh Meier 则更加彻底地践行“苦涩的教训”（bitter lesson），坚信只要喂入海量序列，结构感知能力便会自然涌现（emerge）。 当前驱动模型训练的两大数据源是：PDB（蛋白质数据库）中有限的高精结构数据，以及规模达到数万亿 Token（trillions of tokens）的庞大生物序列数据库。更有趣的是一种正向滚雪球效应：高质量的基础模型可以直接对海量序列数据库进行推断生成新的预测结构，从而为下一代模型制造更多训练数据。Josh 强调，Chai 内部从不拘泥于“序列学派”还是“结构学派”的学术教条，而是极其务实地将两者融为一体。事实上，互联网上存在的生物序列 Token 总量可能比英语语言 Token 还要多，尽管其中伴随着大量噪声和非冗余清洗的挑战。 ```mermaid flowchart TD accTitle: Data Flywheel in Protein AI accDescr: The compounding cycle of sequence data, structure prediction, wet lab validation, and model refinement. subgraph DataSources["数据底座"] SeqData["万亿级序列数据库"] PDB["PDB 3D结构数据"] end subgraph Modeling["模型迭代"] BaseModel["Chai 基础模型"] SyntheticStruct["高精度预测结构数据"] end subgraph LabValidation["湿实验闭环"] LabExp["靶点设计与湿实验验证"] DataExhaust["实验尾气数据 (Data Exhaust)"] end SeqData --> BaseModel PDB --> BaseModel BaseModel --> SyntheticStruct SyntheticStruct --> BaseModel BaseModel --> LabExp LabExp --> DataExhaust DataExhaust --> BaseModel ``` 随着模型精准度突破临界点，Chai 能够直接针对全新靶点进行从头设计，湿实验中产生的大量“实验尾气数据”（data exhaust）再次回流训练集。Josh 回忆起在 OpenAI 早期研发的经历：团队曾尝试在初代 GPT-1 架构上引入强化学习（RL），但当时由于基础模型太弱而彻底失败；只有当 Base Model 足够强大时，RL 实验与数据飞轮才能真正起飞。现在的生物分子 AI 正迎来这种大模型时代的“起飞时刻”。 - 生物序列数据规模极为庞大，潜在 Token 数量甚至可能超越互联网上的全部英文文本。 - 优质模型生成合成结构（Synthetic Data）与湿实验尾气数据相结合，构成了加速自愈的复合飞轮。 - 算法路线遵循极端实用主义，不纠结于单纯的序列模型或结构模型，以解决现实靶点问题为唯一标准。 --- ### `37:16` 对标大自然与湿实验巅峰：跨越制药界的技术门槛 面对制药行业与 AI 生成领域日益激烈的竞争，主持人询问 Chai 如何在巨头林立与新兴竞品层出不穷的环境中持续领跑。Josh 明确表示，Chai 的第一准则是“看清客观结果，绝不自欺欺人（not fooling yourself）并保持极度严谨”。团队进行海量评估的目的不是为了公关跑分，而是为了让模型像“爬山算法（hill climb）”一样不断突破物理界限。从 Chai-2 彻底告别大型文库筛选（Large Library Screening）、实现超高成功率，到随后几个月将可成药性（Developability）与可制造性（Manufacturability）引入设计，Chai 始终在以极快节奏将涌现的新能力产品化并交付给药企客户。 Josh 进一步指出，Chai 真正的竞争对手并不是其他 AI 模型提供商，而是“大自然”与“传统湿实验百年来积累的极限方法”。传统生物学家在既定管线中不断微调打补丁，其复杂程度好比在实验协议上叠加了“第 240 号模块”。例如 Chai 团队的资深科学家 Andy Young，二十年前在麻省理工学院（MIT）便是酵母展示（Yeast Display）技术的早期开拓者，后在辉瑞（Pfizer）和基因泰克（Genentech）深耕二十载，手握上市抗体药物。传统领域专家凭借现有技术已能制造出顶级抗体，而 Chai 必须正面超越这套极高水准的行业基准。 > “我们创办这家公司，绝不是为了造出一只快十倍的传统筛选小鼠（10 times faster mouse），而是为了去实现前所未有的突破性药物。” 正是因为跨过了与传统黄金标准平齐乃至超越的拐点，各大跨国药企才在近期相继与 Chai 达成深度战略合作。随着主持人抛出关于团队文化的问题，Josh 与 Matt 开始复盘加入 Chai 所面临的硬核挑战与非凡回报。 ### `40:00` 团队使命感与“成长的痛苦”：从8张GPU到大规模集群工程 谈到在 Chai Discovery 工作中最享受与最痛苦的事情时，Matt McPartlon 毫不犹豫地指出，最棒的部分在于整个团队纯粹的使命感驱动。团队里的每个人都对自己的工作极度投入、拼命工作，并拥有明确且罕见的共同目标。Matt 回忆说，自己此前在其他公司工作过，唯二见过类似极高专注度与动力的场景，就只有在读博士期间的实验室伙伴；Chai 从创立之初就确立了清晰的哲学和提升路径，全体员工对此深信不疑。 至于最不满意的一点，Matt 先是开玩笑说新办公室“没能直接建在 Dandelion Chocolate 巧克力的楼上”，随后坦言，真正的痛点在于如何让一切在大规模算力下稳定运转。在创立 Chai 之初，团队拥有 128 张 GPU，当时 Matt 甚至觉得“对一个实验室来说这已经是难以置信的极限规模了”——毕竟读博时是四个人挤着用 8 张 GPU，能有 128 张卡简直感觉“暴富”（GPU rich）。然而现在的 Chai 维护着极其庞大的算力基础设施：虽然 GPU 具备天然的并行优势，但也带来了严峻的系统级挑战——如何长达数月保持集群的健康状态？当超大规模训练任务意外中断时，如何实现自动故障转移与任务恢复？如何将卡间通信开销降至最低？又如何极致优化模型以压榨硬件效能？主持人 Pat Grady 随即引用了知名科技高管 Frank Slootman 的名言来宽慰：“你只有两种选择，要么承受失败的痛苦，要么承受成长的痛苦”，而这显然是令人欣喜的成长之痛。 - **读博时期的算力落差**：4 人共享 8 张 GPU，对比 Chai 创业之初拥有 128 张 GPU 时感到前所未有的“富足”。 - **集群工程核心挑战**：持续数月的超长训练稳定性、故障自动断点续训（automatic resume）、通信瓶颈压减及硬件资源的高效利用。 ### `42:10` 从“发表好论文”到“拯救真实患者”：产品化工程的严苛标准 Josh Meier 认为最振奋人心的部分则是实打实的成果（Results）——“它真的起作用了”（It's working）。团队中许多人在 AI 领域已经深耕多年，Josh 和 Matt 在这个方向上也钻研了整整十年。Josh 反思道，几年前大家也曾写出过很多很酷的论文，业界一片赞誉，但他总忍不住拷问自己：“我们真的让世界变得更好了吗？这项技术真的能影响到现实中的患者吗？”而今天，答案终于变成了斩钉截铁的“能”。从实验室科学团队传回的验证数据、新模型在基准测试中的突破，到产品端推出的每一个切实减轻客户负担的功能，这种真切感受到技术正在深刻改变世界的体验令人无比兴奋。 然而，这也带来了 Josh 认为最艰难的挑战：走出纯粹的“象牙塔研究”（blue sky research），承担起工业级交付的严苛责任。全球顶尖的制药巨头正依托 Chai 的模型推进核心管线，模型绝不能存在致命 Bug 或性能回退（regressions）。Chai 的另一位联合创始人 Jack 经常强调：“如果你想在长期跑得快，就必须在短期内学会慢下来。”团队坚决摒弃“为了让眼前的指标好看而强行塞入第 25 个临时模块”的短视做法，转而致力于构建健壮、易于多人协同的高水准生产级代码库，将每一次突破沉淀为可持续复利的工程系统。 ```mermaid flowchart TD accTitle: 从前沿研究到临床落地的工程飞轮 accDescr: 展示 Chai Discovery 如何从学术蓝天研究转向严苛的工业级与临床落地体系。 subgraph Research["科研探索模式 (以往)"] R1["单点突破与发表高引论文"] --> R2["临时堆砌模型补丁 (如Module 25)"] R2 --> R3["止步于学术演示，脱离真实患者"] end subgraph Engineering["Chai 工业化交付体系 (当前)"] E1["扎实规范的生产级代码库"] --> E2["杜绝性能回退与高频 Bug"] E2 --> E3["长期复利架构设计 (短期审慎/长期高速)"] end subgraph Impact["现实商业与临床应用"] P1["全球顶级药企管线全面接入"] P2["数月内完成分子 Zero-Shot 临床推进"] end Engineering --> Impact ``` ### `43:45` 名字背后的寓意：“Chemistry and AI”与化繁为简的初心 在被问及公司为何命名为“Chai Discovery”时，Josh 揭晓了这个名字背后的双重含义：它不仅代表着“**Ch**emistry（化学）”与“**AI**”的直接缩写组合，同时也取自他们喜爱的印度奶茶（Chai tea）——团队办公室里甚至摆满了各种奶茶主题的周边物品。 Josh 进一步解释，许多传统生物技术公司的命名往往晦涩冗长、充满复杂的专业术语，而 Chai Discovery 希望传达一种亲民且直观的感受。团队的终极使命就是让原本极度繁琐的药物研发工程变得极其简单，因此必须配上一个同样极致简洁、平易近人的名字。 ### `44:35` 未来展望：Zero-Shot从电脑屏幕走向人体临床试验 谈及未来 6 到 12 个月最令人期待的技术与业务进展，Josh 重点指出了与全球战略伙伴合作管线的实际部署。目前签署的巨额合作绝不是让模型躺在硬盘里吃灰，而是真正部署在攻克致命恶性疾病的实际研发管线中；AI 设计出的一枚有效分子，承载着拯救无数患者生命的希望。Josh 指出，技术演进的速度堪称惊人：仅仅一年前，业界还根本无法想象通过“零样本”（Zero-Shot）方式直接在计算机上生成分子并对其成药性抱有充足信心；而今天，客户已经可以拿着 Zero-Shot 生成的全新分子坚定地宣布“我们将把这个推进到临床”。从电脑屏幕上的数字分子到注入患者体内的候选新药，整个周期被压缩到了前所未有的程度。 Matt 过去在上一家公司申请的专利中，已经有在计算机上直接设计的分子走入了人体临床试验阶段；而一想到未来几年内可能会有几十甚至上百个由 Chai 设计的分子进入临床试验、走向患者，两人都感到“激动到起鸡皮疙瘩”。面对外界关于长时间高强度工作是否会导致职业倦怠（burnout）的疑问，Matt 坦言在看到实质性进展时不仅完全没有倦怠感，反而每天都在被“下一代模型一定会比这一代更强大”的确定性所鼓舞。接下来，团队正致力于攻克分子设计的终极“可控性”（Controllability）——不再是盲目祈祷碰运气拿到高亲和力分子，而是能够精准设定参数，按需定制出纳摩尔级（如 10 nM）结合力的靶向分子，将生物学真正转化为精密的确定性工程。 - **设计范式跃迁**：从一年前无法实现可靠的 Zero-Shot 分子设计，跨越至直接生成可推进临床阶段的高置信度分子。 - **下一代核心攻坚目标**：从概率性筛选转向确定性参数控制，例如精准定向生成特定结合力（如 10 nM 结合活性）的分子。 ## Notable quotes > `07:22` 如果模型能够学会说英语、德语、法语，为什么它们就不能学会说 DNA 和蛋白质的语言？ — Josh Meier > `13:30` GPU 跑满算力难道不正是我们的目的所在吗？这恰恰证明我们尽职尽责地把硬件性能压榨到了极致。 — Chai 团队 > `18:45` 我们是一家极其深信‘苦涩的教训’（bitter lesson-pilled）的公司……当你在 Chai-1 这样的模型里塞进 23 个独立子模块时，迭代就会变得无比艰难，因为你必须挨个理解它们的行为，这根本无法有效扩展。 — Josh Meier > `21:28` 生物学其实比代码生成拥有更加客观且可验证的衡量标准。 — Josh Meier > `37:16` 看清客观结果，绝不自欺欺人（not fooling yourself）并保持极度严谨 — Josh Meier > `37:16` 我们创办这家公司，绝不是为了造出一只快十倍的传统筛选小鼠（10 times faster mouse），而是为了去实现前所未有的突破性药物。 — Josh Meier > `40:00` 你只有两种选择，要么承受失败的痛苦，要么承受成长的痛苦 — Frank Slootman > `42:10` 我们真的让世界变得更好了吗？这项技术真的能影响到现实中的患者吗？ — Josh Meier > `42:10` 如果你想在长期跑得快，就必须在短期内学会慢下来。 — Jack > `44:35` 我们将把这个推进到临床 — 客户 ## Mentioned Chai Discovery, Josh Meier, Matt McPartlon, Pat Grady, Sonali Singh, Sequoia Capital, Chai-1, Chai-2, AlphaFold, AlphaFold 2, OpenAI, GPT, GPT-2, ChatGPT, Andy Yang, Nathan Rollins, David Baker, Stripe, Jack, Munaz, Neil, Kevin Wu, SWE-bench, Isomorphic Labs, Alphabet, DeepMind, Eli Lilly, Novartis, argenx, Pfizer, Genentech, PDB (Protein Data Bank), CASP, ESM, Frank Slootman ## Video description Most people treat biology as a bespoke, messy science. Josh Meier and Matt McPartlon, co-founders of Chai Discovery, treat it as an engineering problem. They make the case that drug design obeys the bitter lesson: scale data, models, and compute, and the model can learn what a hand-built pipeline simply couldn't capture. The results are concrete: Chai-2 pushed de novo antibody design from a sub 0.1% hit rate to 16%, turning a needle-in-a-haystack search into something more like designing a key to fit a lock. Josh argues, counterintuitively, that biology is more verifiable than code, and explains why the goal should be more lab experiments, not fewer. Their bet: a design suite that collapses drug discovery from nine months to nine days, and arms the pharma industry rather than competing with it. Hosted by Pat Grady and Sonali Singh, Sequoia Capital 00:00 Introduction 01:52 From Discovery to Design 03:25 Protein AI Breakthroughs Timeline 06:04 Why Start in 2024 10:13 Diffusion Models Intuition 11:41 Building the Avengers Team 15:22 Hit Rates and Scaling Laws 25:01 Molecular CAD Vision 25:24 Faster Design Loops 26:32 Future Drug Discovery 28:37 Platform Business Model 31:14 Partnering Reality Check 33:44 Data Flywheel Explained 37:16 Staying Ahead at Scale 39:44 Culture and What's Next [Source](https://pscrb.fm/rss/p/traffic.megaphone.fm/CPUAI9886655614.mp3)

[7:22] Josh Meier: 如果模型能够学会说英语、德语、法语，为什么它们就不能学会说 DNA 和蛋白质的语言？ [13:30] Chai 团队: GPU 跑满算力难道不正是我们的目的所在吗？这恰恰证明我们尽职尽责地把硬件性能压榨到了极致。 [18:45] Josh Meier: 我们是一家极其深信‘苦涩的教训’（bitter lesson-pilled）的公司……当你在 Chai-1 这样的模型里塞进 23 个独立子模块时，迭代就会变得无比艰难，因为你必须挨个理解它们的行为，这根本无法有效扩展。 [21:28] Josh Meier: 生物学其实比代码生成拥有更加客观且可验证的衡量标准。 [37:16] Josh Meier: 看清客观结果，绝不自欺欺人（not fooling yourself）并保持极度严谨 [37:16] Josh Meier: 我们创办这家公司，绝不是为了造出一只快十倍的传统筛选小鼠（10 times faster mouse），而是为了去实现前所未有的突破性药物。 [40:00] Frank Slootman: 你只有两种选择，要么承受失败的痛苦，要么承受成长的痛苦 [42:10] Josh Meier: 我们真的让世界变得更好了吗？这项技术真的能影响到现实中的患者吗？ [42:10] Jack: 如果你想在长期跑得快，就必须在短期内学会慢下来。 [44:35] 客户: 我们将把这个推进到临床

[1:00] Josh Meier: One of the reasons why we call this field drug discovery is we're often looking for a needle in a haystack... If we can instead put in what is like the dream state of the molecule you want, and then the model can materialize that, that's going to be really powerful. [10:00] Matt McPartlon: You can kind of break that protein more and more and make it look more and more noisy... and teach the model just quick little shortcuts... You can just keep asking over and over again: make it slightly better, make it slightly better. [17:35] Matt McPartlon: When you look at a model like, let's say, Chai-1, there are 23 distinct sub modules... when you're trying to iterate on something like that, it gets really hard... simplicity allows the research process and identifying scaling directions to become a lot simpler. [20:00] Josh Meier: In biology, if you're like plus or minus 5% in your lab, that might all be the same... at least with physical readouts, you can be honest with yourself. [25:00] Josh Meier: The goal of the company is to build a computer-aided design suite for molecules... to change the way that medicines are discovered. If that loop takes nine months versus nine days, each order of magnitude scales in a very big way the number of ideas you can really sort through.

[0:00] One of our big guiding principles is just simplicity. So like when you look at a model like let's say CHI 1, I think there are 23 distinct submodules in CHI 1. And like when you're trying to iterate on something like that, it gets really hard because you're like, I kind of need to understand each of these submodules independently. I need to understand all of their behaviors, their dynamics. And like that doesn't actually scale that well. So like you can start to think, okay, how do I simplify this? How do I identify like what's really important? And once you have that, like kind of the whole research process and like identifying these types of scaling directions becomes a lot simpler. We're here with Josh and Matt, two of the co-founders of Chai Discovery. Chai is engineering molecules with AI. So it's something of a foundation model lab for biology. Sounds like a big idea, but let me just start there. What is the big idea? Thanks for having us on the show. One of the exciting things we're trying to do at CHI is to make the drug discovery process look a little bit more like engineering And we seen all the work happening with LLM for code generation for instance right And that works really well because code is a very simple abstraction right I'd like to kind of get across what you want to do. Biology doesn't look like that today. It's a lot of like trial and error. I think if we go back to the early days of just like modern biotech, right? A lot of the medicines that we have today were kind of discovered quite randomly. It was a bit serendipitous. And what we're trying to do is allow us to industrialize that process a little bit more and try to come up with the tools that we'd need in order to have abstraction layers. The same way we have that in code and in modern engineering allows us to iterate really quickly and try to bring that into into biology and into drug discovery. So let me ask you a question on that then. So there's and tell me if this is a reasonable way to frame it. There's almost this boundary between that which can be engineered and that which needs to be tested in the real world. and it feels like that boundary has been moving over time. Like the proportion of the drug development process that can be engineered as opposed to trial and error seems to be increasing. Is that a reasonable way to think about it? If so, can you talk about like what specific developments have pushed that boundary over time? I think that's one way to think about it. The way that we often look at this is like the lab is an important part to verify that what you doing is correct And actually verification is a very big theme in AI as well If you can evaluate that your model works you can verify it then you can start to hill climb that and you can make progress on it. So I think the lab is a very important part of this. And the question is, how do we take drug discovery and make it look a lot more like drug design? So one of the reasons why we call this field drug discovery is we're often looking for a needle in a haystack. We'll screen millions, billions of molecules, try to find one that works. And if we can instead put in what is like the dream state of the molecule you want, and then the model can materialize that, that's going to be really powerful. So it's not even a question of like reducing lab testing. I mean, maybe that happens as a result. I actually might even take the opposite side of the coin and we can talk about that where maybe we'd actually do even more lab testing because the ROI will increase. The same way there's more software engineers or there's more demand for software engineers now that they become more productive. There might be more demand for the lab. But I think the key part is how do we just change the paradigm here and how do we make it more design oriented? Can we talk about what is state of the art today? And then maybe let's take a little trip down memory lane five years ago, three years ago, one year ago. Like what have been some of the major breakthroughs and like how has state of the art changed over the recent years Yeah I think the field has like it really evolved especially like over the last decade So it really wasn until like you can actually look back there like this biannual protein folding competition. So every two years, usually a bunch of academic groups who compete in this protein folding competition, what would happen is like, you kind of hold out these protein targets that no one's ever seen before, they never get deposited publicly. And then all these groups compete so you can predict the proteins the best. And like, really, it wasn't until 2018, where you started to see this like big step change in performance. And then finally, again, in like 2020 with Outfold 2. But really, like it was kind of like the advent of deep learning that really like kicked off this whole field. So at first what you do, these are kind of like bring me back to my original days in my advisor's lab. So we worked on like one of the first systems to do this with deep learning. What we would try to do is predict like the distances between amino acids and a protein. And then we'd have some render, which then took in like these kind of noisy and complete looking distances. and then emitted some protein from that. And what's pretty remarkable is that with relatively minimal information, like you could build a machine learning system that could actually produce something that really looks like a protein, and for the most part was correct. Like there was still pretty big gaps that didn't quite have the resolution of like what we have today, almost like [5:00] early image generation models where it was like a little bit grainy, a little bit blurry, and now it's like, oh my gosh, like this is crazy high definition images like in seconds. So we kind of like saw that same evolution happen over time. So it started with protein folding, and then I guess once that became more realizable, there were kind of these other subproblems that people wanted to solve. So now given a protein structure, can I design, say like a sequence which might fold to that? And that's getting closer and closer to like the drug design problem. So you're kind of like now thinking, okay, how do I start engineering proteins that like match a certain shape or might perform a specific function? And so people kind of study that problem independently. And that was kind of like 2021, 2022 time. And then kind of these ideas began to merge together. It was really like the advent of diffusion models where we started to be able to like, okay, I can now generate a protein structure and a sequence kind simultaneously. And I can start making these like what we'd call prompts more and more realistic. So I can now prompt these models on a target that might have a certain shape and I can say, hey, I want to bind over here. Kind of like add more real world constraints on the problem. What did you guys see in 2024 that made you think that was the right time to start the company? Yeah there were a lot of discussions that went into it I remember one of these early discussions actually back in Matt house when we were you know Matt was showing me some of these results on like antibody antigen like structure prediction. Matt was just talking about protein folding. But for a long time, people thought that protein folding for antibodies was just like too hard of a problem. People were like, there was not enough data for antibodies, like in the protein databank, for instance, to like solve this problem. And people as a result thought that like antibody design was going to be out of reach. A lot of the work that people even did on protein design in the early days with deep learning, it wasn't antibodies. It was these different class of proteins called mini proteins, which are actually really interesting in their own right, but they're not what most of the drug industry is looking at. So the holy grail was whether we could design antibody proteins on the computer, especially ones that had all that therapeutic function. And our thinking was that if you couldn't predict what an antibody looks like, how are you ever going to design one? Traditionally, people have thought like, like biology is scary. Like there's so much to know. I'm terrified of biology. I am as well, honestly. Like my background was was never biology. I studied like pure math and started my PhD in theoretical computer science. And it was only like after my third year that I ended up switching into like deep learning, protein structure prediction all of this stuff So it was like totally new field to me Seemed insane But like at the end of the day it much simpler And like the problems are much more like interconnected than one might think So like I think people like when they start going into the field they like oh man what an antibody What's a mini protein? These are all just like sequences of amino acids. At the end of the day, like these are just like different types of prompts for the model. But like in the same way where you might like have a math problem that goes into chat GPT, chat GPT can both answer your math problem and like help you with your English homework. So really, we have the same type of thing going on with our models. We just have some way of representing these sequences of amino acids. Then we have a way of designing, predicting those as well. And I think in that lens, things become a lot more clear. And Matt, you mentioned your background a little bit. Josh, talk a bit about your background. And then more broadly, in order to pull this off, you have a bunch of different disciplines that kind of come together. So can you just talk a bit about your background and that of some of the other core members of the team and how these things all fit together? Yeah, I've been excited about AI and biology since I was a kid. So I guess I'm like, Matt, Matt, I didn't start with like theoretical CS and get into that way, but I went to a high school with a stem cell lab. So I was just always excited about biology as a kid and I grew up as a programmer. I I really started my career at OpenAI so it was on the early team there It was a nonprofit back then so it was a pretty good time to be there We did GPT GPT scaling laws And the question was, if the models can learn to speak English, German, French, why can't they learn to speak DNA and protein? And that was kind of my research agenda since then. And I think that sort of intersects with around the time like Matt got into the field as well. And it was a I think it was a pretty important time as well. Right. Because if you look at the kind of methods that we are bringing in, like there's been a lot of these changes on the edges, if you will. Right. And, you know, Matt talked a bit about the history of what's happened in the fields here. But, you know, it's all about like how do we find like the right deep learning architectures with the right compute configuration and the right model architectures to make this happen? What are the right tasks to apply it to? we're talking about how we even knew that like you know 2024 is the right time to start the company as matt was saying you know these are all different like kinds of amino acid sequences and people thought that you know antibody class of problem was going to be too hard and we started to see the first signs of life that actually this was starting to work i think a lot of it fueled by um some of the new architectures we're bringing to the problems where i think it was diffusion models back then the first time like anyone was able to generate reasonable looking proteins uh was the advent of diffusion models. So it was pretty crazy. [10:00] There were a bunch of generative modeling approaches that would kind of work if you had a bunch of data. So people got these working for images. There were a bunch of tricks to make this better and better along the way. We've definitely borrowed a lot of those ideas in our domain as well. But it was really like once diffusion models came around. Is there an intuition for why diffusion models work? Yeah. So diffusion is not a one-step process. So like I think kind of up until this point, like the the main generative design paradigm was like called variational auto encoders. And in that case, you're like saying, I want to just like compress my input distribution. So like you have some like proteins, you want to make these look like kind of like fuzzy Gaussian vectors. That task is just like really hard. And maybe today if we tried like super hard, I think we'd crack it. But at the time, like I think like it wasn't really as developed enough to work for our problems. What turned out working really well was just kind of giving the model more time to think and showing it like more examples like here's like a slightly broken looking protein. How do you make it better? You can kind of break that protein more and more and more and you can make it look more and more noisy, more and more broken and teach the model just quick little shortcuts. All right here how I make it slightly better You can just keep asking over and over again make it slightly better make it slightly better and kind of like breaking the problem down to that scale worked really really well for for biology yeah the make it slightly better reminds me of a game that i like to play with my daughters where we have chat gpt give us a unicorn and then we make it stronger and we just keep telling it to make it stronger and by the time we're done we have the strongest unicorn in the world. So about the same, right? Yeah. It's that easy. Okay. Maybe not the same. Yeah. So I feel like in this domain, you need to assemble a kind of like a quadrilingual group of people, like an Avengers squad of chemistry people, biologists, AI folks. And so that's a challenge. How have you guys gone about finding people, convincing people to join the team and who are your superstars? So we've been really pragmatic about this at CHOP. You look at the founding team, it was mostly AI researchers. So people who had worked on either scaling models or getting them to work in this domain. But really with each generation of model, the kind of people we've needed for the next milestone has changed, or I'd say probably has expanded, right? So if you look at Chai 2, right, that's the point when we started to bring in some of the most incredible like you know antibody engineers and scientists in the world One of the scientists Andy Young actually when we hired him people asked us if we had pivoted into building a full stack drug pipeline because they like you be crazy not to do that if Andy joined your team But I think Andy has enjoyed running more antibody campaigns in the past couple of months than he's probably run in his whole career, which is very cool to see. You have folks like Nathan Rollins on the team. Nathan was actually homeschooled and then started college very early on. So you joined David Baker's lab, who won the Nobel Prize for protein design, like when he was 14, started his PhD when he was 18, and has so many creative ideas. As the model started to get better, we needed to build up a product team because while the researchers might get the models to the point that they're very powerful, you need to build the right product interfaces so that the models are actually useful. And that's where we started to bring on people who have built some of the most exciting products we know about today. Like our co-founder, Jack, worked at Stripe. Munaz, who was one of the top and code contributors at Stripe. Neil, who ran his own cybersecurity company before security started to become very important as we deploy this to our big partners. As we started to scale up, we brought in people who've really done a lot of the GPU hacking, if you will, in order to scale up our systems that they don't break when we're running them at scale. We had an email from or a Slack message from one of our hyperscalers the other day where we had a cluster I think that an issue too And they like oh the GPUs got too hot I think you guys are running too many And we like isn that the point right That's probably, I was like, good, we're doing our job at least. Yeah, it's like, how do you convince those people to join? I think, again, a lot of it comes back to the results and like a clear need. We didn't hire antibody engineers before in an antibody design model. Like, what are those folks going to do? We were even, I think, worried when we started that trend because for some of the next generation formats, the complicated antibodies, like multi-specifics, like they didn't even work with CHI-2. So it actually took a couple of weeks when some of those people showed up before the models could work at a point that they could work on some of these interesting case studies. But fortunately, progress was fast enough to kind of bring that online. So I think we're always like evolving that team and going for, you know, that next milestone. We've got the team very small as a result, too. So this way, you know, everyone is a little bit like slightly over capacity, I think, which means we have to prioritize. It forces us to work on the things that really matter most. I think on the research side as well, one of the founding engineers, Kevin Wu, he had the first, I think it was the first protein diffusion model ever. And that speaks to Kevin's speed of execution. He is a heck of an engineer. And I think engineering has just always been important since day one. So really, even our researchers are all excellent engineers. And we really care about... [15:00] a high quality code base. At the end of the day, we are technically a software company. We're AI researchers, we're protein designers, we're a lot of things, but our deliverable is some piece of software. So we've kept that bar really high while also trying to level that with great research talent and people that can actually push the frontiers of what's possible. And you've had a number of, it seems like, aha moments in the field. AlphaFold was an aha, we can figure out how a protein folds. And then the diffusion models, aha, like we can generate proteins. And it seems like your latest models are a kind of another aha moment. We're not only generating molecules that look like proteins, but they also have therapeutic properties, like they can bind really tightly. They have really high hit rates. Can you talk sort of about the quality of the molecules that your models are producing now and everything that sort of went in to those models to make them able to do that? Yeah. If we look at the quality of the molecules that's coming out. It goes back to, you know, one of the theses when we started the company that we really wanted to focus in on this like de novo generation of the molecules. If you look at what a lot of the drug discovery, I work was at the time was about how do I take a molecule and just make it a little bit better Right Which we just talked about in a sense but it was doing it with a lab in the loop style right where I take some data try to make it better that way And the question we started with is well can we actually just like do all that on the computer right Is there a way that the, maybe there would be enough data or we could collect enough data so that we could just zero shot a molecule that has a lot of these properties. So the first thing we needed to do to get there was to design molecules with really high success rates. When we started the company, the state of the art for any body design was about like a 0.1% binding rate. So one in a thousand of the molecules you design would actually bind in the lab. So first of all, that means you have to screen a lot of molecules to find some good ones. It also means that like the gradient you get on your process is actually quite weak as well. So for many targets, you won't get any hits. For the ones that, you know, you do get some hits, you won't have enough to actually see whether you're having like the drug-like properties. So we really focused in on how do we just make this process more accurate. We got to, with our CHI-2 model, about like a 15% success rate. So now if you screen a thousand molecules, you're getting 150 back. Now you can start to get some interesting statistics on the properties of the molecules, right? And a lot allowed us to iterate on that, build the right evaluations around that in the lab, and really try to hill climb that as well. And we getting to a point now where we can actually bake in a lot of these different properties from the start into this engine And then maybe if you think about how does how does this happen Or like how are we approaching this as well And like why do we think it going to continue improving Yeah so like honestly I think Josh and I were both surprised at how quickly this works. Like when we were originally budgeting this, we're like, ah, maybe like 20% hit rate, like three or four years, we're like really shooting for like a 1%. We thought one in 100 would be amazing. We're like, this is going to be a groundbreaking thing. We had a philosophy on like an approach that we wanted to take. And it just like ended up working really well. And that approach is like very similar to what's worked in the rest of machine learning. So people kind of treat biology as this like bespoke problem or like bespoke field. But really, it's like the same principles as like self-driving LLMs. Well, same principles. But one of the things we've talked about before is how you manage to find scaling laws. And it's one thing to tokenize a string of text. It's another thing to tokenize biology. Can you say a couple words about without giving away any of the magic? You can you just say a couple words about that challenge and how you guys solve that yeah so one thing that i really like about chai is like we're we're a very bitter lesson-pilled company so like we really believe in like scaling data scaling models scaling compute in order to do that obviously you need to identify scaling laws otherwise you just kind of like wasting time and resources um and i think like without giving away too much like one of our big guiding principles is just simplicity So like when you look at a model like let say Chai 1 so like, they're like, I think there are 23 distinct sub modules in Chai 1. And like, when you're trying to iterate on something like that, it gets really hard. Because you're like, I kind of need to understand each of these sub modules independently, I need to understand all their behaviors, their dynamics, and like that doesn't actually scale that well. So like, you can start to think, okay, how do I simplify this? How do I identify what's really important? And once you have that, the whole research process and identifying these types of scaling directions becomes a lot simpler. One of the other things too, I think it's interesting about this, is we take a lot of these lessons from what's worked in the rest of the deep learning space. But as you're pointing out, the data itself is different. The models at Chai are completely built from scratch. We're not fine-tuning GLM or something like that on some protein data. We build everything from the ground up. I think a lot of the company building process, though, is taking a philosophy and actually sticking with it and iterating on that and just having some guiding principles. When you're building like a blue sky research company, if you will, you know, Chai is almost like one of these neo labs, right? Like we have this big AI problem we're going after where as we make progress on it, you know, that opens up opportunity for our customers. But if you're gonna work on something so open-ended that way, you need some principles to guide you. And I think we've done a very good job [20:00] I'm like tracking those principles in the company, working on it. So a lot of the things that Matt's saying- Yeah, can you share them? What are the guiding principles? So I think simplicity was one of the ones that Matt mentioned. It's this like bitter lesson pilness of like, you know, scaling compute and data and models. It's being really rigorous. That's something that's so important in this space. You can fool yourself so easily in biology. Like the error bars in the wet lab are actually quite large as well. So it's actually a little bit different than if you look at like cogeneration, for instance. If you look at Sweebench, people were like, oh, this is maybe like a year ago. People were like, oh, I got 16%, then 17%, then 18%. I mean, in biology, if you're like plus or minus 5% in your lab, that might all be the same. So it actually just means the bar is really high in terms of the step changes that you want to see with the models. But you also need to be really honest with yourself about whether you're making progress or not. So you could come up with some fancy model that looks like it works well in like one or two new tasks. but it's very important to show that that works more generally if you're actually trying to build a product that can bring the field forward. Yeah, and that's, I think, pretty interesting because biology is one of those inherently not so verifiable domains. And you guys have been really good at sort of showing your progress to customers and to people like us who know very little about biology And so can we talk a little bit about the evals and the verifiable part of the model progress Like how do you guys know that your models are getting better Well, I would say, first of all, that I actually think this is one of the more verifiable domains. It's actually a very objective readout. If you look at something even like CodeGen, right? Like, maybe it's a verifiable task, like, did my code compile? Did it solve these unit tests? But how do you think about the taste, right? Like, did I write some really sloppy code that can't be maintained? Like, what does that look like? When we think about designing a molecule in the lab, we can actually be quite specific about many of these properties, right? So maybe we get a molecule that binds the target, but can I manufacture it, right? That might be your version of like some tech debt, but you can measure that. And I think those evals, again, actually make this domain more verifiable. Maybe it takes a little bit longer to validate it, right? It's not like five seconds to like get a readout and run a unit test. You might have to spend a couple of days, a couple of weeks in the lab to get that readout, but at least you can be honest with yourself. Yeah. When we talk about progress and how good the models are, there's like a domain of targets in biology that you can, as you mentioned, sort of address with traditional screening methods. They take a long time. They're very slow and rudimentary. And then there are targets that just aren't addressable with existing methods. They not druggable for whatever reason And so where are we in terms of model progress in terms of you know working on existing targets and generating molecules faster That one end of the spectrum And then the other end of the spectrum is unlocking novel biology new targets and things that we couldn't drug before. This kind of goes back to like, kind of like our core modeling philosophy. So like, there have actually been like plenty of times where like, man, if we had a 24th module, like we can actually like unlock that new target. and we're like, is that really something that we want to maintain long term? Is this incremental or is this like actually a compounding improvement? Will this actually like help us generalize to the broader class of like these, this whole class of targets that we really can't hit? And so like our philosophy has been like, all right, let's just like continue to focus, like identify your scaling laws. Like there comes a point where like if the model is able to push loss down even further, it has to understand like something like very intrinsic about the target that it's operating on. So like one example we were talking about the other day, Paul and I, maybe the way that we're looking at like certain glycosylation sites on proteins were like, oh, we might want to represent them differently or something like this. And we're like, well, even if we didn't represent them, like there are certain sequence motifs that will tell the model like there should be a glycosylation site here. And in order to drive like loss down further the model should just like have to learn that So there are all these hidden features of targets where if you really believe in scaling laws you can believe the models will get there Like these types of targets like should just unlock with better models And of course like you still have to take this very seriously. And like, you still need all the proper validation. You need to like really challenge yourself and like make sure that this is truly working. But I think our approach has always been like, you know, we with better models, like we should be able to unlock a lot of these targets. One of the other interesting things is if we look at we talked about like the interdisciplinary nature of this, if we look at the different teams at CHI, what people will call a hard target is actually different in literally every team. So on the science team, it might be, you know, like a undruggable GPCR target. No one's gotten something that has like, you know, modulated that in a functional way. On the ML research team, it'll be something like, oh, there's like the model just can't fold this thing up. It doesn't know what it looks like. And then on the product team, it might look like, oh, I've got all these like, you know, modifications, like my glycosylations, and it's a membrane protein? How do I represent that to the user? And actually the fact that it's different for each of these groups, I think is a feature rather than a bug. And it means that if we want to make like broad progress over here, everyone is kind of pushing in parallel on these different ways. And that means that there's very smooth progress that we can make all the time. And there's usually not one bottleneck at chai. It's not like, oh, if we only had that one extra module, things would work. Or if we only, you know, like try to push this into the product a [25:00] some way we can unlock it. We're trying to build this unified solution because at the end of the day, the goal of the company is to build a computer-aided design suite for molecules. It's not to make one or two molecules. It's not to get a pipeline of like five interesting therapies that we bring to market. It's the change of the way that medicines are discovered. And we're going to do that. We need to work on all the hard targets, regardless of how you define hard. Say more about this idea of computer-aided design suite for molecules. What does that mean? So at its core, it goes back to this point about making biology look more like an engineering discipline. So we're not going and fishing something out of a large library or doing a ton of trial and error. You want to be able to specify upfront the principles that go into designing your molecule and then have an engine that can actually realize that into some molecule that we're going to go and create in the lab. And look, you still might do some iteration on the lab and on your model because maybe your hypothesis was wrong. But what we want to speed up is actually, again, have that computer-aided design suite so that you can go from idea to testable hypothesis very quickly. And if that loop right now takes something like nine months, I don't know, to go and discover your molecule versus if it takes nine weeks or it takes nine days, each order of magnitude just scales in a very big way the number of ideas you can really sort through I think that ultimately how the field is going to converge on better medicines It really comes back to like people sometimes talk about do we care and you kind of noted on it not like do we care about like speed or do we care about like the difficulty of the targets. At some point they converge in this way as well because a hard target if we can like iterate through hypotheses a lot faster then maybe it'll be easier to crack it. So if we can maybe detach ourselves from the present reality and go far enough into the future that we're not thinking present forward. We're actually thinking future back, 2035, 2040, 2100, whatever you want. What does the industry look like? Let's imagine that computer-aided design suite for molecules has become a standard. Let's imagine a lot of innovation has flowed downstream into some of the wet lab parts of the process. Can you paint the picture of what the industry might look like 10 years from now? Yeah, I think it's going to be a really exciting time. And we can look at this on a couple of angles. So first of all, the quality of the medicines that we develop will hopefully go up. There's a lot of molecules we put into the clinic today that there is really hard to discover a molecule. You find something that's like 80% of the way there. Maybe I'm just gonna advance it anyways, like hit my timelines. It's probably gonna benefit some patients. But then you get beat a year later by someone else And it really just not the most efficient spend of resources in the industry You got the kind of diseases that are just too hard to go after today People have been trying to drug Alzheimer forever and unfortunately haven made as much progress as we'd like. You have things that just aren't economical to go after today. Think about like personalized medicines, rare diseases, things where maybe the patient population is going to be smaller. But again, if we can iterate through those ideas faster, if we can launch something faster, if we can do it in a less expensive way, then those probably come into reach as well. So I think there are just so many different axes that we're able to push on. And I think that means that the future is really bright. It used to be like, yeah, they want to be first in class or best in class. Now it's like you want to be last in class because you actually just want to be the final answer. So I think a lot of what you'll see is just way more intentionality in the types of drugs that you're designing. Like this drug will be super specific to the disease of interest. It won't have like certain interactions, like certain negative interactions that a lot of drugs today do. A lot of this stuff is actually like you're able to model most of this computationally, like maybe not today, but there's definitely a path towards getting there. And I think that's like one of the most exciting things for me. Very cool. Can we talk about a business model decision that you guys made Because I think a lot of times folks think about chai and Isomorphic and the same neighborhood Isomorphic is developing drugs You guys are enabling the existing industry to develop drugs more efficiently, better, faster, cheaper than they had before. Why did you decide to go down that path versus the Isomorphic path? Yeah, first of all, I think both of these paths are great and they can create tremendous value. We've always been really excited about building infrastructure for the industry. our bet when we started to go back to those results in Matt's house, you know, a few years ago, was that this is how most future drugs are going to be discovered. And if that's the case, somebody needs to go and like build that infrastructure and make it happen. I think part of this too, is a lot of our founding team, like including Matt and I had worked in companies before where we had built these full stack drug pipelines, right? And we were building AI models. We're putting those drugs into the clinic. Again, we're pretty bitter lesson-pilled as well. And our thinking was, as the models get better, we want to be spending more of the money making better models as opposed to diverting those resources into clinical trials and things like that. And one of the interesting things about our business models, as the models get better, actually wins us the right to continue investing more in them. Right. And you have partnerships with with these pharma companies that are paying off today and allow us to invest in it. So it's a much more scalable business in that way. And I just think about the old. [30:00] impact that we can create for the world is a lot larger. We've always wanted to just partner broadly with the ecosystem. It goes back to this point about fooling yourself in biology. If you work on a small number of drug targets, you might come up with the most exquisite molecules, creating a ton of value for the world by doing that. But you might miss the forest for the trees because maybe your model doesn't generalize to the other 500 molecules that people are going to make that year. And if you go and partner with people, you just can't fool yourself. Like you look at our partners, Eli Lilly, Novartis, Argenix, Pfizer, like these are not companies that are, you know, they they take this stuff for granted. Like you have to really deliver on these partnerships for them to take you seriously. And that means it can't just work like some of the time. Like when we ship models at Chai, they really have to work. They have to deliver value to our partners. And it's not like, oh, we made some molecule doesn't fully work. We're going to have our chemists like clean it up a little bit. So it's it's almost a harder business to pull off. I think that's one of the reasons why you haven't seen it pursued many times. If you work on a drug pipeline, again, there might be ways to fix things up later. There will be the proof of what happens in the clinic, of course. But when you have this partnering-based model, you have to be really rigorous about your models. Your models have to work really well because otherwise those partners are not going to come easily. So it made our life harder I think in many ways But I think it also the more rewarding path if we can get it to work Well and what have you learned I mean you know you out of the lab so to speak and in the real world you know delivering real value for actual pharma companies What have you learned as you started to work with these partners in terms of any surprises in terms of how their needs might have differed from what you expected or how their level of sophistication around this might be different than what you expected? Have you any surprises or any learnings from working with these partners? Yeah, so when we went into these partnerships, I think a lot of people told us that like pharma doesn't know how to use AI. It's not like a tech forward industry and things like that. And to be honest, that hasn't really been our experience. I think that these are, again, they're very rigorous partners, very rigorous customers, right? And they're going to test every one of our claims right before they start to deploy these things. But when they see the data, they go all in, right? Because pharma is an innovation industry. And it's interesting just to think about even the whole economics of the pharma industry. If you build a product in pharma, like a drug, right? You only have exclusivity on that drug for a certain period of time, right? And you have to continue to reinvest. Eli Lilly is a trillion dollar pharma company right now. If they don't get more blockbuster drugs, they will not be a trillion dollar pharma company forever And I think that forces these companies to really be on their game of adopting new technologies and deploying them and trying to stay ahead Pharma is a very competitive arena There a lot of people trying to bring these drugs to patients By the way it great for patients but it means that you have to be on top of your game here And I think that means that once you're through that door and your models are working, we've seen an upscale this adoption very quickly and people thinking about how to use the models in incredibly creative ways. I really like the incentive structure that it creates as well, like taking more of the partnership model. As our models get better, we get better results to our customers and so on. So I think that's just a nice side effect for us as well. Chai has been incredibly focused. And really, as the models get better, you can kind of iterate on a better model with better data and so on. So a lot of what we see and a lot of what our partners are using the model for, like generating binders, antibodies, so on. We're also doing a lot of dogfooding in-house. We have a whole science team that's using the models and just trying to understand what they can do. and a lot of that comes back to like okay now that like the model now that we've unlocked use case x like what type of new data can we generate how can we make the model better that way um and really like that's been the long-term vision of chai is like we never thought of like we're going to build this one model that's going to solve all of this like we know that they're just like uh in other fields like they going to be multiple iterations of this model and as you get better and better models like you get this flywheel effect i guess um well and you mentioned what data can we generate Can we touch on data for a minute You guys can exactly just go scrape the internet and have all the data you need to build your models. Where does the data come from? How does it kind of compound over time? Can you just say a word about data as an input to your models? Yeah. So I think like the primary source of data or like the gold standard source of data is this protein database. So this is like, it's actually just like legit lab scientists who since 1970 have just been depositing crystal structures of proteins and other molecules over time. And really without that, structure prediction and design wouldn't have been a thing. So that's one source of structural data. When I started in the field, I really took a structure-pilled approach. I was super interested in predicting structure. How do you think about a machine learning model that can output 3D coordinates? That doesn't look like an LLM, it doesn't look like an image model. This is really in its own class. Josh, interestingly, was taking the exact opposite approach. So he was one of the original authors on ESM, and that was really a seminal work in understanding how to apply language models to protein sequences. What's really cool about that work is if you can train a language model to understand protein sequences, [35:00] What ends up happening is it ends up kind of like representing the 3D structure internally. And there's like a really interesting reason why this happens. So like in order to predict like missing amino acids, so like same way that you'd predict like next word in a sentence, I want to predict next amino acid in a protein. In order to do that effectively, like you really need to understand like, okay, what does amino acids like immediate microenvironment look like? Because that kind of tells you, okay, what are like the compatible amino acids with everything surrounding it? And in order to do that well, you need to understand the protein's 3D shape. So I was taking this really structure-based approach, and Josh was even more bitter lesson-pilled to me. He's like, we're just going to look at the sequences, and this is just going to emerge. So yeah, the two main sources of this data, again, protein database for structures, and then these massive, massive, maybe even order of trillions of tokens sequence databases. And what you can do once you have really good models is just run them on the sequence databases to get new structures out. So again, you have this compounding effect. As your models get better, they get more and more accurate at predicting these structures. and then you have more and more training data for the next series of models. Sometimes people ask us at CHI about, these days at CHI, like which paradigm are you actually going after? I think one of the things I love about our team is that it actually just like neither we very pragmatic Like we wanna solve the problem We don really care is it a sequence approach is it a structure approach In practice it going to end up being like both of course I think you be surprised but there might even be more like biological sequence tokens on the internet than like English language tokens. And now a lot of that data is not that useful. It might not be redone. It might be very noisy, but there's a lot of data out there. There's a lot of art in like bringing it together. And I think also the exciting thing is as the models have gone into a point now where we can design things in the lab, write to new targets, for instance, we can actually use the models to generate data as well. So there's a lot of exhaust from like all the experiments that we're doing at CHI, which also helped to make the models even better. So it's a similar kind of takeoff that we saw with LLMs. Like when I was at OpenAI, we worked on reinforcement learning of like a GPT-1 architecture. Like did not work because the models weren't like good enough. But once the base model got good enough, then you could start to do those kinds of experiments. And I think there's a similar analogy that's starting to happen in our world now as well, where the models have reached a point where there's actually a renewed interest in data. And like, how do we actually bring the models into the loop on like making that happen? And then that creates another really like interesting cycle on compounding improvement of the models. Yeah. We're talking a bit about compounding improvement in the models. And you said something about how pharma as an industry is an incredibly competitive landscape. And it interesting because I think there been a renowned interest in using AI and using ML to generate proteins and molecules And so your arena has actually become quite competitive And you guys have obviously done an amazing job at staying at the frontier It's been the year of deployment for you. You've locked up a number of pharma partnerships that are making your models better. But how do you guys think about the competitive landscape and staying at the front of the frontier? Well, first of all, I think it goes back to looking at the results and not fooling yourself and being rigorous. So one of the reasons why we do a lot of evaluation of the models, we mainly do it just a hill climb in the models itself. I think if you look at many of the capabilities that we've brought in have kind of been like first in the field, you look at our CHI-2 model, right? Like getting to success rate of the models that you didn't have to do large library screening anymore to see results. A couple of months later, showing how we could bring in a lot of these developability properties we've talked about before, like the manufacturability of the molecules. So in many cases, we are pushing the model forward and trying to see these capabilities emerge. And then we try to quickly like lock those in. Like how do we make those capabilities like even more pronounced that they become production ready and we can like ship them to our customers. I think one of the things I like to tell the team is that, you know, it's not like we are head to head like with other model providers or something like that, where all of us I think are working against nature. Like nature has actually been a pretty good baseline People have done drug discovery a certain way for a long time And you know Matt talked about how like maybe you don want to add like module 24 to the CHI model but people have added like module 240 to like the existing wet lab protocols And they have been like tuned quite considerably. One of the scientists on our team, Andy Young, was he was one of the first people working on yeast display at MIT actually like two decades ago. And he's got 20 years experience in like Pfizer and Genentech, like really honing in these methods, has a drug approval to his name and antibodies. And I think you look at someone like that and like, you know, Andy knows how to make a good antibody with existing tools. And that is actually the bar that we need to clear. Now, of course, I think the ceiling on AI is going to be a lot higher than what we've managed to do before. Otherwise, what would be the point of doing this? We didn't start the company just to make, you know, a 10 times faster mouse, right? We started this to make breakthrough medicines that weren't possible before. But ultimately, that is the bar that we needed to clear in order to get adoption. I think we hit that inflection point a couple months ago. That's why you've seen a lot of these big pharma announcements. But now we just need to continue to hone in on like making these things even better. For somebody who's listening who thinks, wow, this sounds pretty cool. I wonder what it would be like to work at CHI. What is the best thing about working at CHI? And what is the worst thing about working at CHI? Yeah, I can speak to some of this. I'll think of this. [40:00] the fly the worst thing. Yeah, so I think like the best thing is just like how actually like mission driven everybody is like everyone is so dedicated to what they're doing. I've worked at other companies like the closest I've ever seen to this is like maybe some of the guys in my PhD lab. But like everyone is just like incredibly, incredibly motivated. We all work really hard. There's like an obvious shared goal. And I think that's really rare to see. And I think this goes back to just like the focus that we've had since the beginning and like we've always had like kind of a clear philosophy a clear plan on how things are going to get there how things are going to get better and really like everyone at chai is very bought into this it's pretty amazing just to see like the amount of dedication that that everyone's putting in uh least favorite thing about chai uh not directly on top of dandelion chocolate maybe no maybe the next office yeah uh actually so probably the least favorite part now uh is just like i guess it's getting things to work at scale um and really i i didn't even know what that meant actually when we started chai we had 128 gpus and i was like this is like the most scale possible for a lab like this is crazy i came from like you know my phd group where we there were four of us sharing eight GPUs And I was like I just felt GPU rich It was crazy So now like you know I try like we have a lot more infrastructure to maintain We have a lot more compute resources. Like luckily, GPUs are parallelizable. But that also kind of brings up its own set of problems. So just like, how do you keep a cluster healthy over time? How do you get that large training run? So like, how do you keep that running for months on end? And even when it does die, like, how do you automatically resume these things? How do you keep all of your communication down? How do you optimize the models and make the best use of the resources that you have? So I think these are a lot of problems that, you know, they continuously pop up. They're good problems to have, but I think they're also really difficult to solve. Yeah, and I'm just excited to work on this. Yeah, I remember one of the CEOs that we've worked with a couple of times, Frank Slootman, had this line about you either have the pain of failure or the pain of growth. You'd much rather have the pain of growth. Yeah, that's exactly right. I think my favorite part is probably the results. And that sounds a bit cliche, but there's nothing. It's working. It's working, right? And just knowing that you're, I think many of us in the company, we've been working in AI for a long time. There's all this experience you've built up. And to know that you're applying it to something that really matters just even I think just take Matt We been working on this problem for 10 years And a couple of years ago I looking at something like yeah we writing some cool papers like everyone is like celebrating this Are we actually making the world better Like is this actually going to impact some patients And I think now the answer is, like, actually, yes. Like, we have reached the point where this is going to make a big difference in the world. And every time you get one of these breakthrough results, anytime there's a new feature on the product that makes lives of our customers easier, whenever there's, like, new lab results coming back from the science team, it's just always so, honestly, it's a good idea. exhilarating to realize like this is actually going to change the world in a pretty profound way i think that's also then uh maybe it comes to the least favorite side right like you know we're running the company and it's like we have real partners that are relying on this and like things have to work right and you ship a new model generation how do you make sure that there's no bugs in that how do you make sure you don't have regressions right this is no longer just like a again the blue sky research problem of like oh we got some cool results and move on uh we've had to have really high priorities on like you know having production level code bases as the team How do we make sure that the code base is in a state that more people can contribute to this? So something that one of our other co-founders, Jack, likes to say is that if you want to move fast in the long term, you sometimes have to just move a little bit slower in the short term, right? And make sure that you are building something again that goes back to that compounding idea It goes back to we don add module 25 to make the next thing work So sometimes you know you like so excited to get to the next result and you just want to jump into it But we have real partners, some of the biggest companies in the world that are now relying on us. And it's important that we realize that we take that responsibility to heart and we make sure we're building systems that, you know, continue to work. I have a burning question. Why are you guys called Chai Discovery? Chemistry and AI. But we love Chai Chi as well. There's a lot of chai-themed stuff in the office. That's right. That's a good question. I didn't know that either. Josh is the visionary. That was all him. It is a very user-friendly name. Yeah, we also wanted a name that, like, biotech companies have such complicated names. Yeah. We wanted something that's going to be much simpler. Yes. We're trying to make this whole thing simpler, right? Yeah. So we need a simple name to go along with it. Awesome. What are you guys most excited about in the next six to 12 months? I think for me, it's just the deployments that are happening. So we've announced a couple of these partnerships, and I'm really excited just to hear about the results that our partners are bringing online. It goes back to this point of making a real difference and also why I'm so happy to see how these partnerships are going, even posting the agreement. And as we're working with these folks, the models are not just sitting on a shelf somewhere. They're actually being used on real programs, people trying to get a lot of money. [45:00] approach devastating diseases where, you know, if Chai could give them a molecule that works, it could really change the lives of patients. So I'm really excited to see how that goes. Just the pace of progress here is incredible, but also just like the pace of the models. Like a year ago, you could not zero shot a molecule and like, you know, have a good sense that your program was going to work. Like now that's changed. Someone might zero shot a molecule and be like, I think we're going to bring this program to the clinic now. And then a year later, you know, they might even have some of those first molecules going into patients. And just the speed of that that is is just incredible and it's uh you know sometimes you get some shivers thinking about this that like okay my model is gonna like matt has some patents from our last company about you know like just generating the molecule on the computer and these things are now in patients and just to think about the scale like i don't know a few years from now do we have dozens do we have hundreds of like chai molecules going into people it's a bit mind-blowing to think about what that might might look like for patients i think one of the things that like again like what makes chai unique like the dedication, like people are like, man, you work a lot. Um, and like, you like, aren't you burnout or whatever? It's like, it's actually really easy. And like it very motivating when you making progress seeing the progress that we making and just like thinking man the next model is going to be even better than the last We identified this new thing So on like that so incredibly motivating for me It more of just like what what can we unlock next And like how do we make these things more controllable? And like when someone comes to us with a certain target, instead of just like hoping we get good affinity or something like that, you know, can we actually control this? Can we say like we want exactly a 10 nanomolar binder or things like that? Like there are a lot of technical things that I think are like kind of right on the brink of solving and for me it's like it's really motivating just to like pin those things down and just like get all of this over the line and see kind of where that leads to next awesome matt josh thank you for engineering biology and thank you for sharing your story with us today thank you guys thanks for having us Thank you.