Chai Discovery 联合创始人探讨 AI 如何将药物发现从试错转变为以设计为导向的工程过程。
Chai Discovery co-founders discuss how AI is turning drug discovery from trial-and-error into a design-oriented engineering process.
要点 · TL;DR
Chai Discovery 旨在利用 AI 将药物发现变成一门工程学科，强调简洁性和缩放定律。 Chai Discovery aims to make drug discovery an engineering discipline with AI, using simplicity and scaling laws.
生物学提供可验证的反馈，使 AI 能够迭代改进，这与代码生成不同。 Biology offers verifiable feedback, enabling iterative AI improvement, unlike code generation.
AI 将因更高的投资回报率而增加实验室测试，与制药公司的合作推动采用。 AI will increase lab testing due to higher ROI, and partnerships with pharma drive adoption.
核心观点 · Key points
目标是将药物发现变得更像工程学，拥有类似于代码的抽象层。 The goal is to make drug discovery more like engineering, with abstraction layers similar to code.
简单性是指导原则；具有许多子模块的复杂模型难以扩展。 Simplicity is a guiding principle; complex models with many submodules are hard to scale.
缩放定律至关重要；识别它们以指导数据、模型和算力的投资。 Scaling laws are crucial; identify them to guide investment in data, models, and compute.
严格的评估至关重要；生物学误差棒很大，因此要对进展保持诚实。 Rigorous evaluation is essential; biology has large error bars, so be honest about progress.
与制药公司的合作推动采用；模型必须为合作伙伴带来实际价值。 Partnerships with pharma drive adoption; models must deliver real value to partners.
数据复合：更好的模型生成更多数据，从而实现进一步改进。 Data compounds: better models generate more data, enabling further improvement.
反共识 · Contrarian takes
生物学比代码生成更可验证；分子存在客观读数。 Biology is more verifiable than code generation; objective readouts exist for molecules.
随着AI改进，实验室测试可能增加而非减少，因为投资回报率更高。 Lab testing may increase, not decrease, as AI improves, due to higher ROI.
由于数据稀缺，抗体设计曾被认为不可能，但现在已可实现。 Antibody design was thought impossible due to data scarcity, but it's now achievable.
制药公司比预期更技术前沿；当看到数据时，他们会迅速采用AI。 Pharma is more tech-forward than expected; they adopt AI quickly when shown data.
最好的模型不一定是模块最多的；简单性扩展得更好。 The best model is not necessarily the one with the most modules; simplicity scales better.
成为同类中最后一名比第一名更好；在药物设计中追求最终答案。 Being last-in-class is better than first-in-class; aim for the final answer in drug design.
本期章节 · Chapters（共 24）
指导原则：简洁Guiding Principle: Simplicity
Chai Discovery 的大构想The Big Idea of Chai Discovery
工程与试错边界Engineering vs. Trial and Error Boundary
前沿技术演进State-of-the-Art Evolution
公司创立介绍Introduction and Company Founding
独角兽类比介绍Introduction and unicorn analogy
组建四语团队Building a quadrilingual team
顿悟时刻与分子质量Aha moments and molecule quality
提升命中率与扩展Improving hit rate and scaling
模型进展与硬目标Model Progress and Hard Targets
硬目标的跨学科性Interdisciplinary Nature of Hard Targets
分子计算机辅助设计套件Computer-Aided Design Suite for Molecules
行业未来Future of the Industry
药物发现未来The Future of Drug Discovery
采用与激励机制Adoption and Incentive Structure
数据来源与方法Data Sources and Approaches
范式与数据生成Paradigm and Data Generation
竞争格局Competitive Landscape
评估与模型能力提升Evaluation and pushing model capabilities
在 Chai 工作的优劣Best and worst things about working at Chai
生产挑战与代码库优先级Production challenges and codebase priorities
Chai Discovery 名称由来Origin of the name Chai Discovery