Isomorphic Labs 首席 AI 官 Max Jaderberg 讨论他们构建通用 AI 药物设计引擎的愿景，将其与 AlphaStar 和夺旗游戏相类比，并描述了药物发现领域可能出现的 GPT-3 时刻。
Max Jaderberg, Chief AI Officer of Isomorphic Labs, discusses their vision for a general AI drug design engine, drawing parallels to AlphaStar and Capture the Flag, and describes what a GPT-3 moment for drug discovery might look like.
要点 · TL;DR
像 AlphaFold 3 这样的通用 AI 模型在药物设计中优于局部模型。 General AI models like AlphaFold 3 outperform local models in drug design.
强化学习在人类知识有限的问题上表现出色。 Reinforcement learning excels where human knowledge is limited.
生成模型和智能体需要探索广阔的化学空间。 Generative models and agents are needed to explore vast chemical space.
核心观点 · Key points
药物设计需要像 AlphaFold 3 这样的通用模型，而不是针对每个靶点的局部模型。 Drug design needs general models like AlphaFold 3, not local models for each target.
强化学习对于人类不知道答案的问题至关重要。 Reinforcement learning is key for problems where humans don't know the answer.
多人游戏为训练可泛化的 AI 智能体提供了多样化的任务。 Multiplayer games provide diverse tasks for training generalizable AI agents.
AlphaFold 3 能够建模所有分子间的相互作用，而不仅仅是蛋白质。 AlphaFold 3 enables modeling of all molecule interactions, not just proteins.
需要生成模型和智能体来探索 10^60 分子的广阔化学空间。 Generative models and agents are needed to explore the vast chemical space of 10^60 molecules.
反共识 · Contrarian takes
生物学并不受数据限制；历史数据存在，但并非为机器学习而创建。 Biology is not data-constrained; historical data exists but wasn't created for ML.
招聘没有生物学背景的机器学习专家可能是一种资产。 Hiring machine learning experts with no biology background can be an asset.
生物学中的 GPT-3 时刻将更像 AlphaGo 的第 37 手，而非类人文本。 A GPT-3 moment in biology will look like AlphaGo's Move 37, not human-like text.
来自物理模拟的合成数据为化学领域提供了巨大机遇。 Synthetic data from physics simulations is a massive opportunity for chemistry.
随着 AI 预测毒性和疗效，未来的临床试验可能发生根本性变化。 Future clinical trials may change radically as AI predicts toxicity and efficacy.
本期章节 · Chapters（共 14）
引言与公司愿景Introduction and Company Vision
通过多人游戏实现 RL 泛化Generalization in RL via Multiplayer Games
早期抱负与 Isomorphic LabsEarly Ambition and Isomorphic Labs
药物设计中的通用与局部模型General vs. Local Models in Drug Design
半打 AlphaFoldHalf a Dozen AlphaFolds
科学圣杯模型与智能体Holy Grail Models and Agents for Science
超越预测模型：分子空间的生成模型与智能体Beyond Predictive Models: Generative Models and Agents for Molecular Space
AlphaFold 3 对药物设计的影响AlphaFold 3's impact on drug design
训练 AlphaFold 3 与扩散架构Training AlphaFold 3 and diffusion architecture
生物学中的数据约束Data constraints in biology
数据生成的挑战与机遇Data generation challenges and opportunities
团队建设与跨学科合作Team building and interdisciplinary collaboration
AlphaFold 服务器发布与未来方向AlphaFold server release and future directions
超越人类理解的生物学 AIAI in biology beyond human understanding