Osmo 创始人、前谷歌 DeepMind 研究员 Alex Wiltschko 探讨 AI 如何建模、预测和设计气味，将嗅觉带入数字世界。
Alex Wiltschko, founder of Osmo and former Google DeepMind researcher, discusses how AI can model, predict, and design scents, bringing smell into the digital world.
要点 · TL;DR
嗅觉 AI 利用图神经网络从分子结构预测气味，解决了一个百年难题。 Olfactory AI uses graph neural networks to predict smell from molecular structure, solving a century-old problem.
主气味图是气味的连续预测图，实现了数字气味设计。 The Principal Odor Map is a continuous predictive map of smell, enabling digital scent design.
香水设计是嗅觉 AI 首个可行的业务，为后续研究提供资金。 Fragrance design is the first viable business for olfactory AI, funding further research.
核心观点 · Key points
嗅觉是一个高维问题，有超过 300 种嗅觉受体类型，远超颜色的三个通道。 Smell is a high-dimensional problem with over 300 olfactory receptor types, far more than color's three channels.
主气味图是嗅觉的连续预测地图，是嗅觉 AI 的基础。 The Principal Odor Map is a continuous predictive map of smell, foundational for olfactory AI.
构建嗅觉 AI 需要大规模、专门构建的数据集；抓取互联网是不够的。 Building olfactory AI requires massive, purpose-built datasets; scraping the internet is insufficient.
香氛设计是嗅觉 AI 可行的首个业务，为后续研究和数据收集提供资金。 Fragrance design is a viable first business for olfactory AI, funding further research and data collection.
智能的哥白尼观点认为 AI 应纳入非人类形式，如化学通信。 A Copernican view of intelligence suggests AI should incorporate non-human forms like chemical communication.
反共识 · Contrarian takes
人类嗅觉极佳，能像狗一样追踪气味，与普遍认知相反。 Humans are excellent at smelling, capable of tracking scents like dogs, contrary to common belief.
结构-气味关系被认为百年无解，但图神经网络解决了它。 The structure-odor relationship was considered unsolvable for 100 years, but graph neural networks solved it.
嗅觉 AI 模型在预测气味上可超越单个人类评审员，通过了图灵测试。 Olfactory AI models can outperform individual human panelists in predicting smell, passing a Turing test.
脑电图无法测量对气味的情绪反应；只能检测睡眠或癫痫。 EEGs are not useful for measuring emotional response to scent; they only detect sleep or seizures.
通过气味进行癌症检测等医疗诊断并非好生意，因监管和数据挑战。 Medical diagnostics like cancer detection via smell are not good businesses due to regulatory and data challenges.
本期章节 · Chapters（共 23）
嗅觉智能简介Introduction to Olfactory Intelligence
嗅觉细胞与受体类型Olfactory sensory cells and receptor types
构建模型：结构-气味预测Building a model: structure-odor prediction
分析神经网络嵌入Analyzing the neural network embedding
主气味图与图神经网络Principal Odor Map and Graph Neural Network
嵌入空间中的结构Structure in the Embedding Space
混淆示例与商业应用Confounding Examples and Business Application
数据收集与规模化Data Collection and Scaling
气味设计即服务Scent Design as a Service
多模态嵌入与解码Multimodal Embedding and Decoding
从简单模型与数据驱动入手Starting with simple models and data-centric approach
模型集群vs单一基础模型Fleet of models vs. single foundation model
端到端vs模块化：类比自动驾驶End-to-end vs. modular approach in AV analogy
核心预测：气味、好感度与强度Core prediction: smell, goodness, and strength
味觉vs嗅觉，聚焦首个垂直领域Taste vs. smell and focus on first vertical
风味主要靠嗅觉Flavor is mostly smell
气味的数据收集与基础模型Data Collection and Foundation Model for Smell
物理AI与化学传感的进展Progress in Physical AI and Chemical Sensing