Osmo Rolls Out AI Engine to Cut Fragrance Design to Single Minute

AI & Perfumery: A New Creative Frontier With Generation by Osmo — Ai4

Ai4

By combining graph neural networks with automated liquid-handling robotics, Osmo compresses traditional 18-month perfume development cycles into instant, regulatory-compliant chemical formulas.

AI-assisted summary of AI & Perfumery: A New Creative Frontier With Generation by Osmo — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published on YouTube by Ai4 on August 24, 2025, this video features a presentation by Alex Wiltschko, CEO and Founder of Generation by Osmo. The description outlines how Olfactory Intelligence (OI) serves as a creative power tool for perfumery, transforming emotion and behavioral data into rapid, safe, and transparent fine-fragrance creation.

## TL;DR Former Google Brain neuroscientist Alex Wiltschko outlines Osmo's mission to digitize the sense of smell using a three-tier architecture of spectrometry, graph neural network scent mapping, and robotic compounding. He introduces Generation, an enterprise AI platform that condenses the traditional 18-month fragrance design cycle into minutes. Through a live demonstration, Wiltschko turns an audience prompt for a melancholic luxury perfume into 'Melancholy Bloom,' generating both creative notes and a production-ready, IFRA-compliant chemical formula. The presentation demonstrates how combining large language models with molecular prediction models allows brands to formulate and physically dispense fine fragrances at the speed of culture.

## Atmosphere & style The presentation maintains an engaging, visionary, and intellectually rigorous tone that bridges cutting-edge machine learning with the nuanced elegance of high perfumery. Wiltschko speaks with crisp, dynamic clarity, demystifying complex neurobiology and molecular structures through lucid visual analogies. The live software demonstration delivers a palpable sense of innovation and spontaneity, making the real-time translation of abstract human sadness into a physical aromatic formula feel both poetic and technologically profound.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=z9qcf46vKOM&t=4s) Digitizing olfaction lagged behind sound and vision because human smell relies on over 300 distinct receptors, compared to three color channels for sight and linear frequency for sound. - [00:04](https://www.youtube.com/watch?v=z9qcf46vKOM&t=4s) Advances in graph neural networks allowed researchers at Google Brain to map molecular structures directly to human olfactory perception, publishing the breakthrough in Science. - [00:04](https://www.youtube.com/watch?v=z9qcf46vKOM&t=4s) A complete digital olfaction stack requires three core pillars: reading molecules via spectrometry, mapping perception through Olfactory Intelligence, and writing scent using automated robotics. - [02:01](https://www.youtube.com/watch?v=z9qcf46vKOM&t=121s) Osmo spun out of Google Brain in 2022 with venture backing from Lux Capital, GV, and Two Sigma to build fine fragrances through its enterprise platform, Generation. - [02:01](https://www.youtube.com/watch?v=z9qcf46vKOM&t=121s) Traditional fragrance formulation is heavily resource-intensive, routinely taking 18 months and $250,000 to $1,000,000 in development costs per scent. - [02:01](https://www.youtube.com/watch?v=z9qcf46vKOM&t=121s) Because Osmo's models operate on universal molecular representations, the technology has cross-industry applications in defense inspection, supply chain tracking, and medical diagnostics. - [03:06](https://www.youtube.com/watch?v=z9qcf46vKOM&t=186s) Generation operates as a generative scent design interface, translating natural language prompts and emotional concepts directly into full fragrance briefs. - [03:06](https://www.youtube.com/watch?v=z9qcf46vKOM&t=186s) During a live demonstration, the platform synthesized a winter luxury fragrance brief for 'sadness' into a complete creative profile called 'Melancholy Bloom.' - [03:06](https://www.youtube.com/watch?v=z9qcf46vKOM&t=186s) The engine produces exact raw-ingredient percentages that act as technical source code, ready to be immediately compounded into physical perfume by automated fluid-handling robots. - [05:44](https://www.youtube.com/watch?v=z9qcf46vKOM&t=344s) The platform orchestrates large language models with graph neural networks trained on millions of global sensory data points to predict volatility, blending dynamics, and hedonic appeal. - [05:44](https://www.youtube.com/watch?v=z9qcf46vKOM&t=344s) Safety and regulatory compliance checks are built natively into formula generation, ensuring instant adherence to international standards such as IFRA guidelines. - [05:44](https://www.youtube.com/watch?v=z9qcf46vKOM&t=344s) Compressing formula generation from 18 months down to one minute enables brands to rapidly iterate and launch retail products aligned with dynamic cultural trends.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=z9qcf46vKOM&t=4s) The Quest to Map Olfaction and Give Computers a Sense of Smell

Alex Wiltschko introduces his life's work at the intersection of neuroscience and artificial intelligence, tracing a career that spans roughly twenty years. With a background as an olfactory neuroscientist who has always harbored a dual obsession with fragrance and computer science, Wiltschko completed his PhD at Harvard, founded several AI ventures, and eventually landed at Google Brain—famously the birthplace of the transformer model. Inside Google Brain, he quietly founded the Digital Olfaction group, a project that operated beneath the public radar until it was spun out into Osmo, an independent company. Wiltschko describes this pursuit as a mission to endow computers with an entirely new human sense, posing a fundamental question: why has sound and vision been digitized for decades, while smell remained analog?

The fundamental obstacle has been the lack of a coordinate system. Vision relies on RGB, a simple three-number representation matching the three color photoreceptor channels in the human eye, while sound is mapped linearly along low-to-high frequencies. In contrast, the human olfactory system relies on more than 300 distinct olfactory receptors, demanding an exponentially more complex map. To resolve this, researchers had to wait for AI and graph neural networks to mature enough to extract latent structures from molecular data. The breakthrough came while at Google Brain, where Wiltschko's team succeeded in mapping scent molecules, publishing their findings in the journal *Science*. Translating olfaction into the digital world requires three interconnected pillars: sensing, mapping, and synthesizing.

```mermaid flowchart LR accTitle: The Three Pillars of Digital Olfaction subgraph Sense ["1. Sensing"] A["Spectrometry (Read)"] end subgraph Brain ["2. Cognition"] B["Olfactory Intelligence (Map)"] end subgraph Act ["3. Actuation"] C["Fluid Handling & Robotics (Write)"] end A --> B B --> C ```

- **Sensory mapping disparity:** The human eye relies on three color channels to construct vision, whereas the human nose utilizes over 300 distinct receptor channels. - **Three-tier technology stack:** - **Reading the physical world:** Uses spectrometry to capture volatile molecular signatures. - **Mapping scent:** Powered by Osmo's proprietary Olfactory Intelligence (OI), predicting perceptual scent qualities from molecular structures. - **Writing scent:** Utilizes precise digital fluid handling and automated compounding hardware to dispense target formulas.

---

[02:01](https://www.youtube.com/watch?v=z9qcf46vKOM&t=121s) Launching Osmo and Generation: Disrupting the Fragrance Industry

Osmo officially spun out of Google Brain in 2022, securing venture capital backing from prominent firms including Lux Capital, GV (Google Ventures), and Two Sigma Ventures. After spending two and a half years engineering the underlying AI and robotics stack, Osmo unveiled its commercial enterprise unit, Generation, designed to craft fine fragrances for consumer brands "at the speed of culture." Historically, the fragrance creation process has been painfully slow and prohibitive, routinely taking brands 18 months and between $250,000 and $1,000,000 in upfront development capital simply to settle on a final scent formula.

While Generation focuses initially on the commercial fragrance and beauty sector, Wiltschko highlights that the underlying scent-digitization technology is already demonstrating cross-industry utility. Because the models operate on fundamental molecular representations, the same core tech stack is being deployed in supply chain authentication, defense and security inspection, and diagnostic human health and wellness applications.

---

[03:06](https://www.youtube.com/watch?v=z9qcf46vKOM&t=186s) Live Demonstration: Formulating "Melancholy Bloom" in Real Time

To demonstrate the platform's agility, Wiltschko presents Generation’s design interface, describing it as an interactive "ChatGPT for scent" currently operating in private beta with commercial clients. He solicits a prompt from the audience, receiving an abstract conceptual challenge: create a perfume that captures the essence of "sadness." Embracing the ambiguity of an emotional vibe, he guides the system through specific fine-fragrance parameters: positioning it as a winter, nighttime luxury fragrance evoking "the highest end of depression and melancholy."

``` [ User Brief ] "Luxury sadness perfume for a winter night" │ ▼ [ Olfactory Intelligence Engine ] (LLMs + Graph Neural Networks + Scent Datasets) │ ┌───────────────┴───────────────┐ ▼ ▼ [ Creative Profile ] [ Technical Formula ] "Melancholy Bloom" (Source code of ingredients) • Top: Metallic, Aldehydic • Material A: xx.x% • Mid: Muguet, Orris, Violet • Material B: xx.x% • Base: Sandalwood, Musk • Material C: xx.x% │ ▼ [ Compounding Robot ] ```

The system generates a bespoke scent concept titled **"Melancholy Bloom."** The description opens with "a whisper of cool metallic air" designed to evoke a solitary, introspective chill: - **Top Notes:** Metallic and aldehydic facets, delivering a crisp, sparkling, soapy, and cooling atmospheric opening. - **Heart Notes:** Muguet (lily of the valley), orris, and wet violet, creating what Wiltschko characterizes as "a bouquet of sad flowers." - **Base Notes:** A grounding dry-down composed of powdery sandalwood and musk.

Beyond the narrative description, Wiltschko displays the actual output: a complete, production-ready fragrance formula that functions as the "assembly code or the source code for a fragrance." The screen reveals a detailed, line-by-line breakdown of raw perfumery materials and exact numerical proportions. Rather than waiting weeks for manual compounding, these instructions can be immediately transmitted over the network to an automated liquid-handling robot that dispenses the concentrate into physical vials within minutes.

---

[05:44](https://www.youtube.com/watch?v=z9qcf46vKOM&t=344s) The Underlying AI Tech Stack and Speeding Up Product Launches

Building a system capable of moving from raw linguistic prompts to stable physical perfumes requires orchestrating dozens of specialized sub-disciplines. The architecture combines large language models (LLMs) trained across digital conversations with proprietary graph neural networks (GNNs) trained on vast scent databases—representing millions of individual sensory evaluations collected across global markets.

Wiltschko explains that turning abstract emotions into compliant consumer goods demands answering intricate physical and chemical questions simultaneously: - **Molecular Structure & Blending Dynamics:** Predicting how disparate organic molecules interact, volatilize, and bind together on skin. - **Perceptual Scent Prediction:** Determining not only how strong or diffusive the compound will be, but also forecasting hedonic appeal—whether consumers will find it pleasant and emotionally resonant. - **Regulatory Compliance:** Automatically verifying that every suggested formulation complies with strict international cosmetic safety and allergen regulations (such as IFRA standards).

By unifying these capabilities into a turnkey engine, brands have already developed products that are currently sitting on commercial retail shelves. Instead of a traditional 18-month timeline, scent architecture and formula generation are condensed into a single minute, giving brands the ability to respond immediately to emergent digital trends and cultural moments.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=z9qcf46vKOM&t=4s) why has sound and vision been digitized for decades, while smell remained analog? — Alex Wiltschko

> [02:01](https://www.youtube.com/watch?v=z9qcf46vKOM&t=121s) at the speed of culture — Alex Wiltschko

> [03:06](https://www.youtube.com/watch?v=z9qcf46vKOM&t=186s) ChatGPT for scent — Alex Wiltschko

> [03:06](https://www.youtube.com/watch?v=z9qcf46vKOM&t=186s) the highest end of depression and melancholy — Alex Wiltschko

> [03:06](https://www.youtube.com/watch?v=z9qcf46vKOM&t=186s) Melancholy Bloom — Alex Wiltschko

> [03:06](https://www.youtube.com/watch?v=z9qcf46vKOM&t=186s) a whisper of cool metallic air — Alex Wiltschko

> [03:06](https://www.youtube.com/watch?v=z9qcf46vKOM&t=186s) a bouquet of sad flowers — Alex Wiltschko

> [03:06](https://www.youtube.com/watch?v=z9qcf46vKOM&t=186s) assembly code or the source code for a fragrance — Alex Wiltschko

## Mentioned Alex Wiltschko, Ai4, Generation, Osmo, Google Brain, Harvard University, Science, Lux Capital, GV (Google Ventures), Two Sigma Ventures, Melancholy Bloom, IFRA

## Video description Alex Wiltschko, CEO and Founder, Generation by Osmo

Abstracts When carpenters gained power tools, they spent less time on manual toil, and more time on developing new creations, with added precision and speed. Olfactory Intelligence (OI) is poised to do the same in perfumery, and will help brand builders push boundaries, explore new ideas, and launch with more confidence. This talk will explore Osmo’s advancements in OI, which allow us to turn behavior and emotion into data and turn that data into fragrance. It will share how OI unlocks deep consumer insights, accelerates fragrance creation without compromising quality or safety, and speeds up development cycles to compete in a world that is only becoming faster and more complex. The future of fragrance is creative, rapid, safe, transparent, and data-driven. The future of fragrance is OI.

[Source](https://www.youtube.com/watch?v=z9qcf46vKOM)

中文简报

Osmo推出AI引擎 将香水设计缩短至一分钟 通过将图神经网络与自动化液体处理机器人相结合，Osmo将传统的18个月香水开发周期压缩为即时生成、符合监管合规的化学配方。 前Google Brain神经科学家Alex Wiltschko阐述了Osmo利用光谱测定、图神经网络气味映射与机器人调配的三层架构来实现嗅觉数字化的使命。他介绍了企业级AI平台Generation，该平台将传统上长达18个月的香氛设计周期缩短至数分钟。在现场演示中，Wiltschko根据观众提出的“忧郁奢华香水”提示词现场生成了“Melancholy Bloom”，不仅给出了创意香调，还生成了可直接投产且符合IFRA规范的化学配方。该演示展现了将大语言模型与分子预测模型相结合，如何让品牌能够紧跟文化潮流的速度配制并实际分发高级香水。 嗅觉数字化此前长期落后于视觉与听觉，核心障碍在于人类拥有超 300 种嗅觉受体；前 Google Brain 团队利用图神经网络成功将分子结构映射到嗅觉感知，奠定了数字嗅觉的基础。 (0:04) 从 Google Brain 拆分并获得 Lux Capital 与 GV 投资的 Osmo 推出了商业平台 Generation，试图打破传统香氛研发需耗时 18 个月、花费 25 万至 100 万美元的高门槛瓶颈。 (2:01) 在现场演示中，Generation 宛如“气味版 ChatGPT”，根据观众提出的“冬夜奢华悲伤感”提示词，瞬间生成名为“Melancholy Bloom”的创意图谱与可由机器人直接调配的精确配方源指令。 (3:06) 该技术栈融合了大语言模型与专有图神经网络，在预测分子挥发与情感共鸣的同时自动确保符合法规安全标准，使品牌开发周期从 18 个月大幅压缩至仅需 1 分钟。 (5:44)

Reading summary

Osmo founder Alex Wiltschko presents a digital olfaction framework that combines graph neural networks and robotics to condense fragrance creation from an 18-month development cycle into minutes. Originating from research at Google Brain, the technology maps molecular chemistry to human scent perception to generate production-ready, physically compounded formulas from natural language briefs. The Coordinate Problem and the Olfactory Stack (0:04) Digitizing smell has lagged behind vision and hearing due to the complexity of the human olfactory system. While vision relies on three color photoreceptor channels and sound follows linear frequencies, human olfaction utilizes more than 300 distinct receptors, requiring graph neural networks (GNNs) to extract latent structures from molecular data. Wiltschko's team published this molecular mapping breakthrough in Science before spinning Osmo out of Google Brain. The resulting technology stack operates across three interconnected pillars: sensing volatile molecular signatures through spectrometry (reading), predicting perceptual scent qualities via Osmo's proprietary Olfactory Intelligence models (mapping), and synthesizing target formulations using digital fluid-handling robotics (writing). Fragrance Economics and Cross-Industry Deployment (2:01) Osmo spun out from Google Brain in 2022 with venture capital from Lux Capital, GV, and Two Sigma Ventures. Its enterprise unit, Generation, addresses traditional fragrance development, which Wiltschko reports typically requires 18 months and $250,000 to $1,000,000 in upfront capital per formula. Because Osmo's models operate on foundational molecular representations rather than industry-specific heuristics, Wiltschko notes that the core stack is also being deployed beyond commercial beauty. Active non-cosmetic applications include supply chain authentication, defense and security inspection, and diagnostic health monitoring. Text-to-Formula Generation and Robotic Compounding (3:06) In a live demonstration of Generation's design interface—described by Wiltschko as an interactive private-beta system for fragrance—an audience prompt requesting a winter luxury perfume evoking sadness yielded a bespoke concept titled 'Melancholy Bloom.' The system structured the scent with metallic and aldehydic top notes, a heart of muguet, orris, and wet violet, and a base of sandalwood and musk. Rather than stopping at creative descriptions, the platform generated a full production recipe detailing raw materials and exact proportions. These instructions are transmitted over a network directly to an automated liquid-handling robot, which compounds the physical liquid concentrate into vials within minutes. Integrating Chemistry, Perception, and Safety Regulations (5:44) Translating subjective prompts into viable perfumes requires coordinating large language models with GNNs trained on millions of sensory evaluations across international markets. The architecture must simultaneously resolve organic molecular blending dynamics on human skin, forecast hedonic appeal and diffusion, and verify compliance with international safety and allergen restrictions such as IFRA standards. Wiltschko states that this integrated pipeline has already produced fragrances currently available on commercial retail shelves, allowing brands to respond near-instantly to digital trends rather than waiting on multi-month manual sampling cycles. While Osmo has demonstrated that generative formulation can place physical perfumes onto retail shelves rapidly, the broader realization of digital olfaction will depend on validating its molecular predictive stack across more demanding, non-aesthetic environments such as medical diagnostics and defense screening.

精炼阅读摘要

数字嗅觉初创公司Osmo创始人Alex Wiltschko展示了基于图神经网络与大语言模型的数字化调香系统，旨在突破人类300多种嗅觉受体带来的高维度感知映射壁垒。据Wiltschko介绍，其商业平台Generation将传统耗时18个月且成本高昂的香精开发流程缩短至数分钟，实现从抽象概念提示到合规物理配方的实时自动化生成。 破译嗅觉坐标与数字嗅觉三支柱 (0:04) Wiltschko指出，视觉依赖RGB三通道映射，听觉基于线性的频率谱，而人类嗅觉系统包含300多种不同受体通道，导致气味数字化长期受限于坐标体系的缺失。随着图神经网络（GNN）的发展，研究团队在Google Brain期间成功将分子结构与气味感知建立数字关联，成果发表于《Science》期刊。该体系建立在三大支柱之上：第一是感知，通过质谱分析捕获物理世界的挥发性分子特征；第二是认知与映射，依托自研嗅觉智能（OI）模型预测分子结构的嗅觉品质；第三是执行合成，借助数字化流体处理硬件与自动化调配机器人输出物理香氛。 缩短香水研发周期与跨行业应用 (2:01) Osmo于2022年从Google Brain分拆独立，获得Lux Capital、GV（Google Ventures）和Two Sigma Ventures等机构投资。Wiltschko表示，传统香氛行业开发一款香气通常需要耗费18个月以及25万至100万美元的前期研发成本。为改变高门槛现状，Osmo推出了面向消费品牌的企业级商业单元Generation，旨在加快高级香氛的研发上市节奏。此外，由于该技术以底层分子表征为运算核心，Wiltschko声称该底层技术栈目前也已被拓展应用至供应链正品防伪验证、国防安全检测以及人体健康与医疗诊断等跨行业领域。 抽象情绪的实时调香与物理输出 (3:06) 在现场演示中，Wiltschko展示了Generation交互设计界面，并将其比作正在进行客户私测的“气味版ChatGPT”。针对现场提出的抽象需求——调制一款表达“悲伤”、适用于冬夜的奢华香水，系统实时生成了名为“忧郁之花”（Melancholy Bloom）的方案：前调采用金属与醛香营造冷冽皂感，中调融合铃兰、鸢尾根与湿润紫罗兰构成花香主体，尾调辅以粉质檀香木与麝香。除创意描述外，系统同时输出了包含具体原料成分及精确数值比例的生产级配方源码，可直接经由网络传输至自动液体处理机器人，在数分钟内将原液配制注入实体测试瓶。 多模型架构与合规化工程体系 (5:44) 从自然语言提示过渡到物理香水需要多学科AI技术的融合。Wiltschko介绍，该平台架构整合了通用大语言模型以及基于全球数百万条感官评估数据训练的图神经网络。系统在生成配方时需同步计算分子结构与混合动态，预测化合物在皮肤上的挥发与吸附行为，评估扩散强度与大众感官喜好度，并自动校验国际日用香料香精协会（IFRA）等化妆品安全与过敏原法规标准。Wiltschko强调，借助这一交钥匙系统，部分由AI辅助设计的产品已进入商业零售货架销售。 Wiltschko展示了数字嗅觉从基础神经科学研究向商业自动化落地的演进路径，但技术在更广泛商业场景中的普及程度、消费者对AI生成香气的长期接受度，以及物理调配硬件在全球供应链体系中的规模化部署仍有待市场检验。

Detailed notes — zh

## Background 本内容由 Ai4 频道于 2025 年 8 月 24 日发布在 YouTube 平台上。视频记录了 Osmo 创始人兼 CEO Alex Wiltschko 的现场演讲，探讨了嗅觉智能（Olfactory Intelligence, OI）如何将人类情绪转化为数据并生成实体香氛，从而颠覆传统香水工业的研发周期与创意边界。 ## TL;DR Alex Wiltschko 分享了他从哈佛神经科学博士到 Google Brain 创立数字嗅觉团队、再到独立创办 Osmo 的历程，阐释了利用图神经网络突破人类 300 多种嗅觉受体复杂维度的技术原理。演讲重点展示了 Osmo 旗下 Generation 业务推出的“气味版 ChatGPT”系统，现场演示了如何将“冬季夜晚的奢华悲伤”这一抽象情绪秒级转化为精准的化学配方代码。该技术配合自动化调配机器人，成功将传统香水长达 18 个月、耗资数十万美元的开发流程压缩至分钟级，并已实现端到端实体货架交付。 ## Atmosphere & style 整个演讲节奏明快且极富科技前沿感，Alex 展现出跨界科学家的深厚底蕴与创业者的敏锐自信。现场演示从抽象概念到具体配方代码的瞬间生成，带来了强烈的技术震撼力。整体氛围既有硬核神经科学与 AI 架构的理性严谨，又洋溢着将诗意情绪物化为实体香气的浪漫与前卫探索感。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=z9qcf46vKOM&t=4s) 嗅觉数字化的最大瓶颈在于人类鼻腔拥有超 300 种受体通道，远比视觉的三原色和听觉的连续频率轴复杂，长期缺乏底层的结构化图谱。 - [00:04](https://www.youtube.com/watch?v=z9qcf46vKOM&t=4s) Osmo 源自 Google Brain 的数字嗅觉研究团队，创始人团队曾在顶级期刊《Science》发表嗅觉图谱论文，为机器理解气味奠定了学术根基。 - [00:04](https://www.youtube.com/watch?v=z9qcf46vKOM&t=4s) 嗅觉数字化的完整技术闭环由三大要素构成：质谱读取（Read）、嗅觉智能图谱构建（Map）以及高精度流体自动写入（Write）。 - [02:03](https://www.youtube.com/watch?v=z9qcf46vKOM&t=123s) 传统香水研发周期长达 18 个月且需耗资数十万至百万美元，Osmo 创立 Generation 业务旨在打破成本与时间瓶颈，以分钟级速度打造实体香氛。 - [02:03](https://www.youtube.com/watch?v=z9qcf46vKOM&t=123s) Osmo 研发的“气味版 ChatGPT”能够将“冬季夜晚的悲伤”等抽象情绪提示词，直接转化为具象的香气金字塔调性与商业企划。 - [03:51](https://www.youtube.com/watch?v=z9qcf46vKOM&t=231s) 生成结果不仅停留在文字层面，系统能同步输出精确到化学原料与百分比浓度的物理配方源码，充当香水工业的底层代码。 - [03:51](https://www.youtube.com/watch?v=z9qcf46vKOM&t=231s) 配方数据可直接对接自动调配机器人（Compounding Robot），实现由 AI 提示词到物理液体喷射与样品制作的无缝自动化。 - [05:44](https://www.youtube.com/watch?v=z9qcf46vKOM&t=344s) 将 AI 落地香氛产业需多模态协同，涵盖分子结构解析、混合反应预测、气味扩散度计算及全球香料安全合规审查等多维硬性指标。 - [05:44](https://www.youtube.com/watch?v=z9qcf46vKOM&t=344s) Osmo 已打通端到端交钥匙（Turnkey）商业交付，使合作品牌能够跟上互联网文化传播速度，将设计成果直接摆上实体零售货架。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=z9qcf46vKOM&t=4s) 数字化嗅觉的起源与核心难题 Alex Wiltschko 回顾了自己过去 20 年的跨界历程：他本科与博士期间专注于嗅觉神经科学（在哈佛大学获得博士学位），同时对香水和计算机怀有极大的热情。后来他创立了数家 AI 初创公司，并最终加入 Transformer 架构的诞生地——Google Brain。在 Google Brain 期间，他创立了鲜为人知的“数字嗅觉”（Digital Olfaction）研究小组，并在那里确立了自己的使命：赋予计算机嗅觉能力。该团队后来独立拆分为一家商业公司 Osmo，迄今 Alex 已经在嗅觉数字化上探索了 18 年，在 Osmo 专注深耕了 3 年。 人们经常好奇，既然人类早已将声音和视觉完成了数字化，为什么唯独没有“数字化嗅觉”？核心痛点在于嗅觉长期缺乏一张结构化的“地图”（Map）。在视觉中，RGB 色彩模型仅用 3 个通道的数值就能指导如何混合颜色以及创造新色彩；在听觉中，从低频到高频的频率轴构成了声音的图谱。然而，人类眼睛仅有 3 个颜色通道，鼻子内部却有超过 300 种各不相同的嗅觉受体通道。因此，这项研究必须等待 AI 技术和图神经网络等模式提取工具的成熟。在 Google Brain 期间，团队首次绘制出这张嗅觉图谱并将成果发表于顶级期刊 *Science*，自此奠定了 Osmo 的核心技术根基。 ```mermaid flowchart TD accTitle: 数字化感官对比与 Osmo 技术三要素 subgraph Senses["人类感官数字化对比"] V["视觉 (Vision)<br>3 种色彩通道 (RGB)"] A["听觉 (Sound)<br>频率轴 (低频至高频)"] S["嗅觉 (Smell)<br>300+ 种受体通道 (高维复杂)"] end subgraph OsmoTech["Osmo 核心技术栈"] R["读取 (Read)<br>质谱分析法 Spectrometry"] M["制图 (Map)<br>嗅觉智能 Olfactory Intelligence"] W["输出 (Write)<br>微流控与液体处理 Fluid Handling"] end S --> M R --> M --> W ``` - **学术与研究背景**：哈佛大学博士毕业，曾在 Transformer 诞生之地 Google Brain 组建 Digital Olfaction 团队并在 *Science* 杂志发表奠基性论文。 - **感官维度的不对称性**：人眼仅 3 种颜色感知通道，人类鼻腔则拥有超过 300 种嗅觉受体通道，导致传统数学模型无法直接拟合。 - **嗅觉数字化的三大支柱**： - **读取（Read）**：利用光谱/质谱分析法（Spectrometry）从物理世界采样。 - **地图（Map）**：独家自研的“嗅觉智能”（Olfactory Intelligence, OI）。 - **输出/写入（Write）**：利用高精度液体配比与流体处理技术（Fluid Handling）进行实体调制。 --- ### [02:03](https://www.youtube.com/watch?v=z9qcf46vKOM&t=123s) Osmo 的创立、Generation 业务线与“气味版 ChatGPT” Osmo 于 2022 年正式从 Google Brain 拆分独立，获得了 Lux Capital、GV（Google Ventures）以及 Two Sigma Ventures 等顶级风险投资机构的注资。经过两年半的技术积累，Osmo 于 3 月推出了首个面向消费品牌的业务部门——Generation。传统香水工业开发一款气味通常需要耗费 18 个月，仅前期的调香研发费用就高达 25 万至 100 万美元不等。Generation 的核心目标就是打破这一漫长且昂贵的瓶颈，帮助品牌方“以文化传播的速度”（at the speed of culture）打造高品质香氛。虽然该技术已被应用于安全防护、供应链以及大健康与人类福祉（Health and Wellness）领域，但香氛消费领域是首个完全落地的场景。 Alex 现场演示了他们与客户日常使用的产品设计系统——一个被称为“气味版 ChatGPT”的生成式 AI 工具。在现场互动中，一位观众提出了一个极具抽象感和情绪氛围的需求：“做一款闻起来像‘悲伤’（Sadness）的香水”。Alex 随即与现场观众将其细化为一个真实的商业品牌 Brief：一款定位为高端奢侈沙龙香、适合“冬日夜晚”、充满抑郁与忧郁（melancholy）氛围的香水。 - **资本与时间线**：2022 年分拆，获 Lux、GV、Two Sigma 投资；历时 2.5 年搭建底层模型与数据集，2025 年 3 月推出 Generation 业务。 - **传统制香痛点 vs 现代需求**：传统模式耗时 18 个月、成本 25 万~100 万美元；AI 模式支持实时秒级生成与精准配方输出。 - **现场实测 Brief**：抽象主题“悲伤”（Sadness）+ 场景细化“冬季夜晚（Winter Night）”+ 市场定位“高端奢侈沙龙香（Luxury Fine Fragrance）”。 --- ### [03:51](https://www.youtube.com/watch?v=z9qcf46vKOM&t=231s) 配方级生成：从抽象情绪到实体机器代码 系统在接收到“冬季夜晚的奢华悲伤”这一提示词后，迅速生成了一款名为“Melancholy Bloom（忧郁绽放）”的香水方案。Alex 解释道，底层系统不仅仅结合了大型语言模型（LLM）和图神经网络（GNN），更重要的是融合了海量的全球气味专有数据（包括团队在世界各地采集并闻过的数百万种不同气味配方）与互联网语料。在生成结果中，前调（Top notes）呈现出清冷金属与醛香气息（Aldehydic，具有肥皂感和闪烁微光的冷冽感），勾勒出寂寥空旷的氛围；中调（Middle notes）则是铃兰（Muguet）、鸢尾根（Orris）和湿润紫罗兰，犹如一束“悲伤的花束”；基调（Base notes）则落脚在粉质檀香与麝香的沉静余韵。 更关键的是，这不仅仅是一段文字描述。Alex 放大展示了系统同步生成的**香水化学配方（Fragrance Formula）**。这是香氛工业的“汇编语言/源代码”，其中逐行精确列出了调香师所需的原材料名称及其严格的浓度配比。该指令集可以直接发送给全自动香精调配机器人（Compounding Robot），由机器人精准喷射对应液体原液，瞬间完成实体香水的混配与打样。目前该工具已开启定向公测（Private Beta），供品牌方与创作者自主设计气味。 ```mermaid flowchart LR accTitle: Generation 气味生成与物理制造流程 subgraph Input["多模态输入"] B["情绪/意向 Prompt<br>如：冬季夜晚的奢华悲伤"] D["底层海量气味图谱<br>全球数百万种气味数据"] end subgraph AI["嗅觉智能 OI 推理"] M["LLM + 图神经网络 GNN"] F["香气三调金字塔描述<br>Melancholy Bloom"] C["物理配方源码<br>原料品名与精确配比"] end subgraph Output["实体制造交付"] R["自动调配机器人<br>Compounding Robot"] P["货架实体香水成品"] end B & D --> M M --> F & C C --> R --> P ``` - **Melancholy Bloom 气味构成**： - **前调**：冷金属感、醛香（Aldehydic），营造冰冷、清脆、独处的氛围。 - **中调**：铃兰（Muguet）、湿润紫罗兰、鸢尾（Orris），构成幽暗凄美的“悲伤花束”。 - **后调/基调**：粉质檀香木（Powdery Sandalwood）、麝香（Musk）。 - **物理交付能力**：直接输出包含原材料名与精确百分比的配方代码，与后端自动化流体机器人无缝对接。 --- ### [05:44](https://www.youtube.com/watch?v=z9qcf46vKOM&t=344s) 现实世界落地：多模态技术栈与即时端到端交付 Alex 指出，要让 AI 在物理世界的香氛产业真正运转，必须在一个系统中协同解决数十项复杂的子能力。这不仅需要理解微观层面的分子结构，还必须准确预测多种化合物相互反应混合后的物理表现、气味的扩散强度（Intensity）、人类对该气味的情感喜好度，同时必须严格符合全球各项化妆品与香精安全法规要求（Regulatory Compliance）。 Osmo 不仅在 1 分钟内完成了一款新气味从概念到配方的设计，更已经打通了全流程交钥匙（Turnkey）的商业闭环，成功帮助合作品牌将实体产品快速推向了零售货架。相比过去耗费一年半的繁冗流程，现在品牌能够跟上互联网的实时节奏和文化热潮，为这个古老的传统香氛行业注入了全新的生命力与创造力。Alex 在此结束演讲，并留出时间接受现场观众提问。 - **工业落地必须解决的复合维度**： - 分子化学结构表征与多元成分混合反应。 - 气味浓度/扩散强度（Intensity）预测。 - 人群主观感知与偏好度建模。 - 全球香精香料安全法规（Regulatory Compliance）硬性约束。 - **商业交付效果**：从 18 个月压缩至分钟级设计，实现交钥匙（Turnkey）模式全流程落地实体货架。 ## Notable quotes > When carpenters gained power tools, they spent less time on manual toil, and more time on developing new creations, with added precision and speed. — Alex Wiltschko > The future of fragrance is creative, rapid, safe, transparent, and data-driven. The future of fragrance is OI. — Alex Wiltschko > [00:04](https://www.youtube.com/watch?v=z9qcf46vKOM&t=4s) 赋予计算机嗅觉能力 — Alex Wiltschko > [02:03](https://www.youtube.com/watch?v=z9qcf46vKOM&t=123s) 以文化传播的速度 — Alex Wiltschko > [02:03](https://www.youtube.com/watch?v=z9qcf46vKOM&t=123s) 做一款闻起来像‘悲伤’（Sadness）的香水 — 现场观众 > [02:03](https://www.youtube.com/watch?v=z9qcf46vKOM&t=123s) 气味版 ChatGPT — Alex Wiltschko > [03:51](https://www.youtube.com/watch?v=z9qcf46vKOM&t=231s) Melancholy Bloom — Alex Wiltschko > [03:51](https://www.youtube.com/watch?v=z9qcf46vKOM&t=231s) 悲伤的花束 — Alex Wiltschko > [03:51](https://www.youtube.com/watch?v=z9qcf46vKOM&t=231s) 香氛工业的‘汇编语言/源代码’ — Alex Wiltschko ## Mentioned Alex Wiltschko, Osmo, Generation by Osmo, Google Brain, Lux Capital, GV (Google Ventures), Two Sigma Ventures, Science, 哈佛大学 (Harvard University), Transformer, ChatGPT, 嗅觉智能 (Olfactory Intelligence, OI), Melancholy Bloom, Ai4 ## Video description Alex Wiltschko, CEO and Founder, Generation by Osmo Abstracts When carpenters gained power tools, they spent less time on manual toil, and more time on developing new creations, with added precision and speed. Olfactory Intelligence (OI) is poised to do the same in perfumery, and will help brand builders push boundaries, explore new ideas, and launch with more confidence. This talk will explore Osmo’s advancements in OI, which allow us to turn behavior and emotion into data and turn that data into fragrance. It will share how OI unlocks deep consumer insights, accelerates fragrance creation without compromising quality or safety, and speeds up development cycles to compete in a world that is only becoming faster and more complex. The future of fragrance is creative, rapid, safe, transparent, and data-driven. The future of fragrance is OI. [Source](https://www.youtube.com/watch?v=z9qcf46vKOM)

Source quotes

Alex Wiltschko: When carpenters gained power tools, they spent less time on manual toil, and more time on developing new creations, with added precision and speed. Alex Wiltschko: The future of fragrance is creative, rapid, safe, transparent, and data-driven. The future of fragrance is OI. [0:04] Alex Wiltschko: 赋予计算机嗅觉能力 [2:03] Alex Wiltschko: 以文化传播的速度 [2:03] 现场观众: 做一款闻起来像‘悲伤’（Sadness）的香水 [2:03] Alex Wiltschko: 气味版 ChatGPT [3:51] Alex Wiltschko: Melancholy Bloom [3:51] Alex Wiltschko: 悲伤的花束 [3:51] Alex Wiltschko: 香氛工业的‘汇编语言/源代码’

Source quotes

[0:04] Alex Wiltschko: why has sound and vision been digitized for decades, while smell remained analog? [2:01] Alex Wiltschko: at the speed of culture [3:06] Alex Wiltschko: ChatGPT for scent [3:06] Alex Wiltschko: the highest end of depression and melancholy [3:06] Alex Wiltschko: Melancholy Bloom [3:06] Alex Wiltschko: a whisper of cool metallic air [3:06] Alex Wiltschko: a bouquet of sad flowers [3:06] Alex Wiltschko: assembly code or the source code for a fragrance

Original transcript

[0:04] Well, I want to share a little bit about [0:06] what I've been doing for the last 20 [0:08] years. Um, so my background is as an [0:12] alactory neuroscientist who's always [0:14] loved fragrance and loved computers. So [0:16] I did my PhD at Harvard. I ended up [0:18] starting uh a few AI companies and ended [0:20] up at Google Brain. Uh, which is the [0:22] birthplace of what we're all here for, I [0:25] guess. So that's where the transformer [0:26] was born. and I started the digital [0:28] alaction group at Google brain which not [0:31] many people knew existed and we turned [0:32] it into an independent company and I've [0:35] been on a mission to give computers a [0:36] sense of smell for roughly the last 18 [0:38] years and for the last three I've been [0:40] doing this at Osmo. So, I'm going to [0:42] talk about AI and perfumery, but really [0:44] this is one step in a much larger [0:47] mission, which is to give computers a [0:48] new human sense. And um like why hasn't [0:52] this happened is a great question to [0:55] ask? Because we have digitized sound, we [0:57] have digitized vision, but why don't we [0:59] have digitized smell. And the problem is [1:00] is that sound and sight have had what v [1:03] what scent has been lacking for a long [1:05] time, which is a map, right? So RGB is a [1:08] map of color. It's three numbers that [1:10] tell you how to combine colors, how to [1:11] create new ones, and low to high [1:13] frequency is a map of sound, but what's [1:15] the map of scent? There's three channels [1:17] of color information in your eye, but [1:19] there's over 300 separate channels of [1:21] information in your nose. So, in a way, [1:24] we needed to wait for AI to mature. We [1:25] needed to wait for data sets uh for AI [1:28] to extract these patterns from in order [1:29] to find this map. And that's what we [1:31] first did at Google Brain and we [1:33] published this in the journal Science. [1:35] Then we then took that and we've been [1:36] building a company around that insight. [1:38] So you have to read have a sensor to [1:41] read the world. You have to map that's [1:42] been the missing piece and then you have [1:44] to write that sense back out again like [1:47] with a printer or with a display. So we [1:50] are building each of these text stacks [1:51] at Osmo. So reading uses something [1:54] called spectrometry. Writing uses [1:55] something called fluid handling. And the [1:58] map is something that's completely [1:59] proprietary and we call that alactory [2:01] intelligence. So we spun out of Google [2:03] brain in 2022. Uh we raised venture [2:05] money from funds such as Lux and GV two [2:08] sigma ventures and we spent two and a [2:10] half years building the technology and [2:12] then in March we launched our first [2:13] business unit called generation and the [2:15] goal of that company is to do one thing [2:18] which is to build beautiful fragrances [2:19] for brands at the speed of culture. It's [2:22] a very slow process today. Normally 18 [2:24] months to design a scent usually [2:26] anywhere between a quart million and a [2:27] million dollars just to get that scent [2:29] in place. Um, and I'm going to skip over [2:31] the technical details because we don't [2:34] have time. I just want to show you how [2:36] it works. So, we're starting in the [2:39] fragrance industry and we've had to [2:41] build an entire tech stack for creating [2:43] an enormous amount of data, but the same [2:46] technology is already working in [2:48] security and supply chain and we're [2:49] already beginning to apply it uh in the [2:51] domain of human health and wellness. Um, [2:54] Generation is taking this first piece of [2:56] technology and it's actually applying it [2:58] in the realm of fragrance. And what I [3:00] want to do is just show you what it [3:03] looks like to actually use this [3:04] technology if you're building a brand. [3:06] So, this is a tool that we use with our [3:08] customers. And what we can do is right [3:11] now we can just make a smell. So, we can [3:12] This is like a chat GBT for scent where [3:15] we can actually design new fragrances. [3:17] Um, does anybody want to make a smell? [3:19] Anybody have an idea for what they want [3:20] to create? Yeah. What do you want to [3:22] make? [3:27] a perfume that smells like sadness. Um, [3:31] I like that because it's abstract. It's [3:33] a vibe, right? Who knows what it's [3:35] supposed to smell like? But what I want [3:38] to do is refine it. Like, let's say [3:39] let's say we're actually launching a [3:40] real brand. So, let's say it's a fine [3:42] fragrance. And let's say this is a [3:46] luxury sadness perfume, right? It's the [3:48] highest end of depression and [3:49] melancholy. [3:51] And um what you think this is this is [3:53] definitely like a dusk like Yeah. night. [3:55] Okay. Night. Great. In the winter, [3:56] right? Okay. There we go. All right. [4:00] I've not made a nighttime luxury sadness [4:03] perfume before, but let's try it and see [4:05] what it comes up with. Um and so what's [4:08] going on is yes, we're using LLMs. Yes, [4:10] we're using graph neural networks, but [4:11] the most important part is we've [4:12] integrated a ton of data on every [4:14] conversation on the internet and an [4:16] enormous amount of proprietary data [4:18] where we make I mean we've smelled [4:20] millions of different fragrances across [4:22] the globe and integrated all this data [4:24] together in order to make you your [4:26] fragrance melancholy bloom. Let's see. [4:28] It opens with a whisper of cool metallic [4:30] air. I think we're getting something [4:32] here. This is kind of interesting. Uh [4:33] evoking a sense of quiet solitude. So [4:36] the top notes are metallic and alahitic. [4:38] Aldahitic is like sparkly and soapy. So, [4:40] it's kind of like I don't know cooling. [4:44] Um, and then we've got mug orus wet [4:46] violety. So, these are kind of like a [4:48] bouquet of sad flowers. And then we've [4:50] got a base of powdery sandalwood musk. [4:53] And so these are top, middle, base [4:55] notes. It's just words. But what's [4:56] actually going on behind the scenes is [4:58] we've created a fragrance formula. And [5:01] this is kind of like the assembly code [5:03] or the source code for a fragrance. [5:05] Let's see if I can zoom in here. There [5:06] we go. So, we've actually created, you [5:08] know, line by line, here's the raw [5:10] materials that a perfumer would work [5:12] with, and here's the proportions. And [5:14] this is a set of instructions that can [5:15] be sent to a big fragrance compounding [5:17] robot, and you squirt the juice in the [5:19] right proportions, and you have your [5:21] scent ready to go. And so, this is [5:23] something that we use with our clients. [5:24] We also have a private beta right now. [5:26] So, if you want to use and drive this [5:28] tool and create your own fragrances, [5:29] you're welcome to sign up. But, um, [5:31] that's what it looks like in practice, [5:33] right? So that's a new kind of AI on a [5:36] new human sense. Uh and it's really just [5:38] the first step on a much larger uh [5:40] journey that we're on. Um [5:44] let me [5:47] just pull back here. The the way that [5:51] this is working and I'll give you just a [5:53] little bit of detail and then if there's [5:54] questions we'd be happy to chat. We've [5:56] had to integrate an enormous amount of [5:58] different kinds of modalities of sense. [6:00] So, we have to know about molecular [6:02] structure. We have to know about how [6:04] multiple ingredients combine together. [6:07] We have to be able to predict not just [6:08] the intensity of the formula, but what [6:10] it's going to smell like and whether [6:11] you'll like it. It has to be regulatory [6:13] compliant. And so, as you begin to [6:15] double click into solving real world [6:17] problems in AI for scent, you end up [6:19] having to assemble uh dozens of [6:21] different sub capabilities that all have [6:23] to work perfectly together. And that's [6:25] what we do. So, we already have products [6:27] that we've helped people put on shelves. [6:28] And so instead of launching brands in 18 [6:30] months, well, we've first of all just [6:32] designed our scent in one minute, uh, [6:34] but we've able to actually go end to end [6:36] in turnkey and help people launch, uh, [6:38] their products really at the speed of [6:40] the internet and the speed of culture. [6:42] So that's what's possible now. It's a [6:45] very old industry. Uh, and we're [6:47] bringing some new life and uh, [6:49] perspective into it. So I'll pause there [6:50] if there's I think we have two minutes [6:52] for questions. So if anybody has a [6:53] question, happy to to chat.