Chai Discovery：生物 AI「相位转移」——当模型异常其实在揭示数据/评估问题；从预测工具走向可驱动实验的系统，关键是数据质量、闭环与可信任评估。
- 评估先于叙事
- 数据质量决定相位
- 湿实验闭环
嘉宾：Matthew McPartlon & Neil Patil（Chai Discovery） · 日期：2026-08-11（PT）
播出信息
- 日期：2026-08-11（PT）
- 时长：约 95 分钟（RSS=5719s）
- 章节末段：1:32:07（约 97%）
节目结构（YouTube 章节）
- What the product actually looks like — 00:00 🎧 · ▶️
- The software-layer bet — 05:00 🎧 · ▶️
- ADCs, bispecifics, and pressing the switch — 11:22 🎧 · ▶️
- Chai-1 and the MSA detour — 17:26 🎧 · ▶️
- Chai-2: all-atom diffusion and crossing into design — 25:49 🎧 · ▶️
- Validation, cryo-EM, and the 0.33 Å result — 32:36 🎧 · ▶️
- Convincing scientists who hate AI tools — 42:32 🎧 · ▶️
- Levels of abstraction, and throwing the product away — 49:04 🎧 · ▶️
- Modalities you can't get from immunization — 57:29 🎧 · ▶️
- Triangle attention and why GPUs hate it — 1:04:18 🎧 · ▶️
- Inductive bias and data efficiency — 1:09:37 🎧 · ▶️
- Per-partner fine-tuning — 1:21:13 🎧 · ▶️
- Pharma as VC: Genentech and Eroom's law — 1:24:20 🎧 · ▶️
- Takeaways — 1:32:07 🎧 · ▶️
分节详解
1) What the product actually looks like — 00:00 🎧 · ▶️
💬 他们说了什么： 围绕「What the product actually looks like」讨论。字幕要点大意：嘉宾谈到与此相关的实践与判断（原文片段：looks a lot less like a, you know, a looks a lot less like a, you know, a looks a lot less like a, you know, a chat GBT …）。
💡 为什么重要： 产品、IP 与资本配置决定科研能否变成工业流程。
⚡ 争议点： 自动字幕非人工精校；技术细节以论文/官方材料为准。
2) The software-layer bet — 05:00 🎧 · ▶️
💬 围绕「The software-layer bet」讨论。字幕要点大意：嘉宾谈到与此相关的实践与判断（原文片段：modalities, train bigger models and ultimately just build what our our ultimately just build what our our partners and c…）。
💡 该段推进本集核心主张与证据链。
⚡ 自动字幕非人工精校；技术细节以论文/官方材料为准。
3) ADCs, bispecifics, and pressing the switch — 11:22 🎧 · ▶️
💬 围绕「ADCs, bispecifics, and pressing the switch」讨论。字幕要点大意：嘉宾谈到与此相关的实践与判断（原文片段：putting a drug on the other side or something like that and that causes the something like that and that causes the some…）。
💡 该段推进本集核心主张与证据链。
⚡ 自动字幕非人工精校；技术细节以论文/官方材料为准。
4) Chai-1 and the MSA detour — 17:26 🎧 · ▶️
💬 围绕「Chai-1 and the MSA detour」讨论。字幕要点大意：嘉宾谈到与此相关的实践与判断（原文片段：we're raising now will let us train we're raising now will let us train bigger models that can maybe be even bigger mode…）。
💡 该段推进本集核心主张与证据链。
⚡ 自动字幕非人工精校；技术细节以论文/官方材料为准。
5) Chai-2: all-atom diffusion and crossing into design — 25:49 🎧 · ▶️
💬 说明为何在基因表达上偏好扩散而非自回归等架构选择。 依据字幕：lighting and a video things like that lighting and a video things like that but at the end of the da…
💡 架构与算力约束决定方法能否规模化。
⚡ 自动字幕非人工精校；技术细节以论文/官方材料为准。
6) Validation, cryo-EM, and the 0.33 Å result — 32:36 🎧 · ▶️
💬 围绕「Validation, cryo-EM, and the 0.33 Å result」讨论。字幕要点大意：嘉宾谈到与此相关的实践与判断（原文片段：structure every time, like same sequence every time. So you also want to see every time. So you also want to see like, o…）。
💡 该段推进本集核心主张与证据链。
⚡ 自动字幕非人工精校；技术细节以论文/官方材料为准。
7) Convincing scientists who hate AI tools — 42:32 🎧 · ▶️
💬 围绕「Convincing scientists who hate AI tools」讨论。字幕要点大意：嘉宾谈到与此相关的实践与判断（原文片段：are you sitting with people who are are you sitting with people who are designing these antibodies, you know, designing …）。
💡 该段推进本集核心主张与证据链。
⚡ 自动字幕非人工精校；技术细节以论文/官方材料为准。
8) Levels of abstraction, and throwing the product away — 49:04 🎧 · ▶️
💬 围绕「Levels of abstraction, and throwing the product away」讨论。字幕要点大意：嘉宾谈到与此相关的实践与判断（原文片段：and like you need to be able to like and like you need to be able to like actually prompt the model to do this you actua…）。
💡 产品、IP 与资本配置决定科研能否变成工业流程。
⚡ 自动字幕非人工精校；技术细节以论文/官方材料为准。
9) Modalities you can't get from immunization — 57:29 🎧 · ▶️
💬 围绕「Modalities you can't get from immunization」讨论。字幕要点大意：嘉宾谈到与此相关的实践与判断（原文片段：is some of that too, right? But it's like, no, there are just like, like, no, there are just like, >> hey, how do …）。
💡 安全、伦理与治理是不可绕过的约束。
⚡ 自动字幕非人工精校；技术细节以论文/官方材料为准。
10) Triangle attention and why GPUs hate it — 1:04:18 🎧 · ▶️
💬 谈 triangle attention 等算子为何与 GPU 不友好。 依据字幕：which which changes sort of like the which which changes sort of like the >> maybe the compute…
💡 架构与算力约束决定方法能否规模化。
⚡ 自动字幕非人工精校；技术细节以论文/官方材料为准。
11) Inductive bias and data efficiency — 1:09:37 🎧 · ▶️
💬 围绕「Inductive bias and data efficiency」讨论。字幕要点大意：嘉宾谈到与此相关的实践与判断（原文片段：engines? It's like Raptor 1, has a engines? It's like Raptor 1, has a [clears throat] bunch of pipes and like [clears th…）。
💡 数据与扰动设计决定因果/预测模型能否可信落地。
⚡ 自动字幕非人工精校；技术细节以论文/官方材料为准。
12) Per-partner fine-tuning — 1:21:13 🎧 · ▶️
💬 围绕「Per-partner fine-tuning」讨论。字幕要点大意：嘉宾谈到与此相关的实践与判断（原文片段：rather than doing research in a vacuum rather than doing research in a vacuum you know based on what would you know base…）。
💡 安全、伦理与治理是不可绕过的约束。
⚡ 自动字幕非人工精校；技术细节以论文/官方材料为准。
13) Pharma as VC: Genentech and Eroom's law — 1:24:20 🎧 · ▶️
💬 在制药 IP 约束下做产品与合作的现实。 依据字幕：more than all of the AI labs put together. Yeah, together. Yeah, together. Yeah, >> I don't th…
💡 产品、IP 与资本配置决定科研能否变成工业流程。
⚡ 自动字幕非人工精校；技术细节以论文/官方材料为准。
金句
- 他们以为模型坏了——其实是现实比假设更有信息。
- 生物 AI 的瓶颈常常是标签与评估，不是又一个注意力层。
