Chai Discovery：药物设计也是规模化问题Chai Discovery: Drug Design Is Another Scaling Problem
The Takeaway：药物发现不必永远是大海捞针，用足够简单、可规模化的模型，就能把“试错”变成真正的分子设计。
Chai Discovery 联合创始人 Josh 与 Matt 正在把 AI 变成生物学的“工程语言”。他们的核心信念是 bitter lesson：不要堆复杂模块，而是靠数据、算力与模型规模把问题推到极限。早期抗体设计成功率只有 0.1%，Chai-2 已做到约 15%，足以开始统计分子性质并持续爬坡。他们坚持从零训练模型，而不是微调通用大模型，并强调验证必须诚实——湿实验误差很大，只有显著跨步才算进步。
公司选择与辉瑞、诺华、礼来等药企合作，而不是自建管线，因为“模型越好，伙伴用得越多，数据飞轮越强”。Josh 说：“我们不是在和别的模型公司竞争，而是在和自然本身竞争。”目标是打造分子的计算机辅助设计套件，让从想法到可测试假设的周期从九个月缩短到九周甚至九天。
他们最兴奋的是：零样本设计的分子已进入真实项目，未来几年可能有成百上千个 Chai 分子进入人体试验。
Chai Discovery 联合创始人 Josh 与 Matt 正在把 AI 变成生物学的“工程语言”。他们的核心信念是 bitter lesson：不要堆复杂模块，而是靠数据、算力与模型规模把问题推到极限。早期抗体设计成功率只有 0.1%，Chai-2 已做到约 15%，足以开始统计分子性质并持续爬坡。他们坚持从零训练模型，而不是微调通用大模型，并强调验证必须诚实——湿实验误差很大，只有显著跨步才算进步。
公司选择与辉瑞、诺华、礼来等药企合作，而不是自建管线，因为“模型越好，伙伴用得越多，数据飞轮越强”。Josh 说：“我们不是在和别的模型公司竞争，而是在和自然本身竞争。”目标是打造分子的计算机辅助设计套件，让从想法到可测试假设的周期从九个月缩短到九周甚至九天。
他们最兴奋的是：零样本设计的分子已进入真实项目，未来几年可能有成百上千个 Chai 分子进入人体试验。
The Takeaway: Drug discovery does not have to stay a needle-in-a-haystack problem. With sufficiently simple, scalable models, trial-and-error can become true molecular design.
Chai Discovery co-founders Josh and Matt are turning AI into an engineering language for biology. Their guiding principle is the bitter lesson: skip elaborate submodules and push data, compute, and model scale instead. Early antibody design hit rates sat at 0.1%; Chai-2 reaches roughly 15%, enough to start measuring drug-like properties and climb further. They train models from scratch rather than fine-tuning general LLMs, and insist on rigorous wet-lab verification because error bars are large—only clear step-changes count as progress.
They partner with Pfizer, Novartis, Lilly and others instead of building their own pipeline, because better models drive more usage, more data, and a stronger flywheel. Josh notes: “We are not competing against other model providers; we are competing against nature.” The vision is a computer-aided design suite for molecules that compresses the idea-to-hypothesis loop from nine months to nine weeks or nine days.
What excites them most: zero-shot molecules are already entering real programs, and within a few years hundreds of Chai-designed molecules could reach patients.
查看原文 →
Chai Discovery co-founders Josh and Matt are turning AI into an engineering language for biology. Their guiding principle is the bitter lesson: skip elaborate submodules and push data, compute, and model scale instead. Early antibody design hit rates sat at 0.1%; Chai-2 reaches roughly 15%, enough to start measuring drug-like properties and climb further. They train models from scratch rather than fine-tuning general LLMs, and insist on rigorous wet-lab verification because error bars are large—only clear step-changes count as progress.
They partner with Pfizer, Novartis, Lilly and others instead of building their own pipeline, because better models drive more usage, more data, and a stronger flywheel. Josh notes: “We are not competing against other model providers; we are competing against nature.” The vision is a computer-aided design suite for molecules that compresses the idea-to-hypothesis loop from nine months to nine weeks or nine days.
What excites them most: zero-shot molecules are already entering real programs, and within a few years hundreds of Chai-designed molecules could reach patients.