Pushmeet Kohli discusses how AlphaFold 3 expands beyond proteins to model interactions with DNA, RNA, and small molecules, giving scientists a superpower to understand biology.
要点 · TL;DR
AlphaFold 3 不仅能预测蛋白质结构，还能预测生物分子间的相互作用。 AlphaFold 3 predicts biomolecular interactions, not just protein structures.
AI 发现了 220 万种新型稳定材料，极大拓展了已知可能性。 AI discovered 2.2 million new stable materials, vastly expanding known possibilities.
大语言模型的幻觉可以被利用来进行创造性科学发现。 LLM hallucinations can be harnessed for creative scientific discovery.
核心观点 · Key points
AlphaFold 从氨基酸序列预测蛋白质结构，解决了一个 50 年的重大挑战。 AlphaFold predicts protein structures from amino acid sequences, solving a 50-year grand challenge.
AlphaFold 3 扩展到生物分子相互作用，包括蛋白质、DNA、RNA 和小分子。 AlphaFold 3 extends to biomolecular interactions including proteins, DNA, RNA, and small molecules.
GraphCast 在 10 天预报中优于传统超级计算机天气模型，提供更早预警。 GraphCast outperforms traditional supercomputer weather models in 10-day forecasts, giving earlier warnings.
GNOME 发现了 220 万种新的稳定无机材料，极大扩展了已知可能性。 GNOME discovered 2.2 million new stable inorganic materials, vastly expanding known possibilities.
FunSearch 使用大语言模型发现新算法，解决了如 capset 问题等难题。 FunSearch uses LLMs to discover novel algorithms, solving problems like the capset problem.
AlphaGeometry 解决了国际数学奥林匹克几何问题，是人工智能的首次。 AlphaGeometry solves International Math Olympiad geometry problems, a first for AI.
反共识 · Contrarian takes
科学中的人工智能专注于根节点问题而非狭窄应用，以产生广泛影响。 AI in science focuses on root-node problems, not narrow applications, for broad impact.
大语言模型的幻觉可与真相检测器配对，用于创造性发现。 LLM hallucinations can be harnessed for creative discovery when paired with truth detectors.
材料科学比生物学更实验性；人工智能的影响需要 5-10 年。 Material science is more experimental than biology; AI impact there will take 5-10 years.
人工智能系统必须通过严格的评估指标被迫泛化，而非记忆。 AI systems must be forced to generalize, not memorize, by using hard evaluation metrics.
DeepMind 避免顶尖大学可解决的问题，专注于需要规模的问题。 DeepMind avoids problems solvable by top universities, focusing on those requiring scale.
从 2 万到 220 万种材料的转变改变了游戏规则，但实用性仍未知。 The shift from 20,000 to 2.2 million materials changes the game, but utility remains unknown.
本期章节 · Chapters（共 17）
引言与 AlphaFold 概述Introduction and AlphaFold overview
AlphaFold 3 改进AlphaFold 3 improvements
科学家反应Reaction from scientists
AI 对科学领域的影响Impact of AI across scientific disciplines
AI 前的材料科学Material science before AI
电池与新材料探索实例Example of batteries and the search for better materials
发现 220 万种稳定材料Discovery of 2.2 million new stable materials
AI 材料科学影响时间线AI in Material Science: Impact Timeline
从计算机科学到科学团队领导Background: From Computer Science to Leading Science Team
通才路径：共同技术主线Generalist Approach: Common Technical Threads
关键要素：定义成功Crucial Element: Defining Success
研究项目选择Choosing Research Projects
将 LLM 融入科学研究Incorporating LLMs into Scientific Research
善用幻觉Harnessing hallucinations for good
AlphaGeometry 与 IMOAlphaGeometry and the International Math Olympiad