这家嗅觉初创公司称其AI鼻子媲美人类

在新泽西州纽瓦克机场附近的一间实验室里,一支约 30 人、经验丰富的人类闻香师团队正在逐一梳理地球上的每一种气味。他们仔细记录的笔记——"坚果味""桃子味""海洋味"——会与化学分析结果交叉比对,为 AI 初创公司 Osmo 所称的全球首个气味基础模型构建一套丰富的训练数据。创始人 Alex Wiltschko 说,公司的目标是"让计算机拥有嗅觉"。
"气味分子实际上是这个星球上大多数物种唯一能说的语言,"他说。"大多数物种无法用声音或视觉交流,所以它确实是地球最原始的语言。而且从记忆和情感的角度看,它对我们的影响极其强大。有些气味你只需闻一次,就能记一辈子。你在街上闻到某种气味,就可能被瞬间拉回高中时代。"
Wiltschko 声称,过去几个月里,Osmo 的模型已经达到"人类水平的嗅觉智能"。公司对此的定义是:该模型在检测和区分气味方面与普通人水平相当,在创造气味方面与初级香水设计师水平相当。公司称,其技术已被香精香料行业用于帮助品牌设计和制造新的香型。该公司表示,上个月,Osmo 的 AI 模型在一家全球最大的消费品公司中表现超过了其内部调香师。Wiltschko 相信,这项技术最终可以部署到机场用于威胁检测,或者用于疾病诊断等场景。
2023年,Wiltschko 基于自己在 Google Brain(这家搜索巨头此前的 AI 研究实验室)的工作创立了这家初创公司,他正在沿用其他领域 AI 初创公司的同一套打法:把大量数据投向难题,看看会发生什么。Wiltschko 说,气味领域的挑战在于缺乏优质数据。据他统计,在 Osmo 之前,现有数据库中人类对气味的描述不足 20 万条。此外,采集数据所需的化学传感器设备价格昂贵,而香氛行业对其嗅觉科学方面的工作讳莫如深。
公司的大量工作在于构建自己的数据集,其中包含超过 1000 万条人工标注和 8 亿个“气味 Token”,这些 Token 描述了用于营造特定气味的确切分子与浓度。Osmo 让光谱仪和其他化学传感器全天候运行,以拆解从水果、花朵到主食和器官等各类事物的分子构成。一支约 70 人、分布在美国和海外的专业数据标注团队每天闻各种东西,并用一套标准化的香氛词汇记录下它们闻起来是什么味道。
以这种方式绘制气味图谱,揭示出了一些看似不太可能的相似之处。正如改变词语顺序能大幅影响句子的含义,按不同比例重新排列气味分子,也能大幅改变人类感知到的气味。“定义新生儿气味的那些气味 Token,与一盘刚煮好的意大利面中的气味 Token 存在大量重叠,”Wiltschko 说。(他说,具体是 Barilla 品牌的意大利面,这或许会让意面工厂的人产生一些疑问。)
该公司还自行创造香氛,使用一台校车大小的气味打印机器人,按精确配方混合原料,既为 Osmo 自身的研究制作样品,也为付费客户制作香氛。它可以根据类 ChatGPT 式的提示词按需设计和打印气味。Wiltschko 翻开笔记本电脑,向我演示这个 AI 模型如何运作。他在屏幕上打开一个文本框,我请他生成一款闻起来像葡萄牙蛋挞的香氛(我们当时在里斯本参加 Web Summit),目标人群是 18 至 23 岁的男性。系统“思考”了几分钟,然后给出了三组符合这一要求的潜在香型方案(奶香、椰子香、果香),以及调配它们所需的分子。如果我继续走完下单流程,我的签名款“蛋挞之水”(L'Eau de Custard)大约一周后就能做好。
Osmo 的技术已被用于开发零售香氛,客户包括 Procter & Gamble 这样的巨头。Gates Foundation 在 Osmo 创立时投资了 500 万美元,后来又向该公司提供了 350 万美元的资助,用于评估驱避带病昆虫的分子,Osmo 找出了八种驱避效果强于 DEET 的物质。Osmo 还与线上服饰交易平台 StockX 合作,通过检测鞋类所用材料化学成分的差异,利用气味来识别假冒运动鞋。
但并非所有人都相信这项技术,也不都相信该公司的大话。嗅觉科学家、《What the Nose Knows》作者 Avery Gilbert 一直在他的 Substack 上尖锐批评 Osmo,认为其部分营销宣称缺乏可查证的记录,而且其专利申请仍依赖人类感官评审小组和调香师,而非 AI。“对于一家以超酷 AI 为立身之本的公司来说,如此严重依赖人类感官判断似乎很奇怪,”他在 3 月写道。Osmo 的一位代表未能在发稿前及时回应。
曾任赫特福德大学神经计算教授、此前曾为 Osmo 提供咨询的 Michael Schmuker 表示,该公司的技术是一项成就。他与同事一起制作了 Osmo 2023 年模型的开源逆向工程版本,但它缺少该公司此后收集的大量数据。
但 Schmuker 说,即便拥有所有这些气味数据,Osmo 也可能遗漏了让人类嗅觉独一无二的一些因素。“气味极其复杂。语境对气味非常重要。同一个分子可能因语境不同而唤起不同的联想,”他说。“卡芒贝尔这类法国奶酪的气味,部分与用过的尿布是同一个分子。你不会把两者混淆,因为语境不同。”
在香精实验室这样的受控环境中,这种语境缺失可能没那么重要。但如果像 Wiltschko 所说,Osmo 的技术成为 AI 世界模型的“嗅觉通道”,它就可能产生影响。想象一台机器,能在生产线上嗅出食物变质,或像狗在试点研究中被训练的那样,通过气味检测肿瘤。
Schmuker 表示,要实现这一点,传感器技术需要大幅提升。模型也需要不断进步。医疗诊断是最大的目标。
“有很多公司曾经直接去攀登那座山峰,但遗憾的是都没能成功,因为代价非常高昂,难度也非常大,”Wiltschko 说。“所以我们所做的是,在所有这些山脉的山脚下打好基础。我们派出探险队上山,进行探索,基本上是在闻遍整个世界。”
译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。
查看英文原文(This smell startup says its AI nose is as good as a human’s)
In a lab near Newark Airport in New Jersey, a team of around 30 experienced human sniffers are slowly working their way through every odor on Earth. Their careful notes—“nutty,” “peachy,” “marine”—are cross-referenced with a chemical analysis to create a rich set of training data for what AI startup Osmo calls the world’s first foundation model for smell. The company’s goal, says founder Alex Wiltschko, is “to give computers a sense of smell.”
“Scent molecules are really the only language most species on this planet can speak,” he says. “Most species can’t communicate with sound or visually, so it truly is Earth’s original language. And it’s incredibly powerful for us, from a memory and an emotional standpoint. You can smell something once and remember it for the rest of your life. You can smell something on the street and be pulled back to high school.”
In the last few months, Wiltschko claims, Osmo’s model has achieved “human-level olfactory intelligence.” The company’s definition of this is that the model is about as good as the average person at detecting and discriminating between smells, and about as good as a junior perfume designer at creating them. It says its technology is already being used in the fragrance industry to help brands design and manufacture new scents. Last month, the company says, Osmo’s AI model outperformed in-house perfumers at one of the world’s largest consumer goods companies. Wiltschko believes it could be deployed eventually for threat detection at airports, for example, or diagnosing diseases.
Wiltschko spun the startup out of his work at Google Brain, the search giant’s former AI research lab, in 2023, and he is running the same playbook as AI startups in other domains: Throw a lot of data at difficult problems, and see what happens. The challenge with scent, Wiltschko says, is that there was a dearth of good data. Before Osmo, there were fewer than 200,000 human descriptions of smells in existing databases, by his count. Plus, the chemical sensor devices needed to collect data are expensive, and the fragrance industry is secretive about its work on olfactory science.

A lot of the company’s work has been in creating its own dataset of more than 10 million human labels and 800 million “scent tokens,” which describe the exact molecules and concentrations that go into creating specific odors. Osmo uses spectrometers and other chemical sensors, running around the clock, to break down the molecular makeup of everything from fruit and flowers to pantry staples and organs. A team of around 70 trained data labelers in the US and overseas spend their days sniffing things and writing down what they smell like, using a standardized fragrance vocabulary.
Mapping out scents in this way is throwing out some unlikely parallels. In the same way that changing the order of words can drastically influence the meaning of a sentence, rearranging scent molecules in different proportions can drastically alter the smells humans perceive. “The scent tokens that define the smell of a newborn baby have massive overlap with the scent tokens in a fresh plate of cooked spaghetti,” Wiltschko says. (Specifically Barilla brand spaghetti, he says, which perhaps raises some questions for the people at the pasta factory.)
The company also creates its own scents, using a smell-printing robot the size of a school bus that mixes ingredients according to precise formulas to make samples for Osmo’s own research and fragrances for paying customers. It can design and print scents to order based on a ChatGPT-style prompt. Wiltschko flipped open his laptop to show me how the AI model worked. He opened a text box on screen, and I asked him to generate a fragrance that smelled like a Portuguese custard tart (we were in Lisbon for Web Summit), aimed at 18- to 23-year-old men. The system spent a few minutes cogitating, then generated three options for potential scent profiles that matched the brief (creamy, coconutty, fruity) and the molecules needed to create them. If I’d continued with the ordering process, my signature L’Eau de Custard would have been ready about a week later.
Osmo’s technology has already been used to create retail fragrances, including for giants like Procter & Gamble. The Gates Foundation, which invested $5 million when Osmo launched, later gave the company a $3.5 million grant to assess molecules that repel disease-carrying insects, and Osmo identified eight substances that were more repellent than DEET. Osmo has also worked with StockX, the online clothing marketplace, on using smell to spot counterfeit sneakers by detecting differences in the chemical composition of the materials used in the shoes.
But not everyone is convinced by the technology or the company’s big claims. Avery Gilbert, a smell scientist and author of What the Nose Knows, has been sharply critical of Osmo on his Substack, arguing that some of its marketing claims are undocumented, and that its patent filings still rely on human sensory panels and perfumers, rather than AI. “This heavy reliance on human sensory judgments seems odd for a company premised on super-cool AI,” he wrote in March. A representative for Osmo did not respond in time for publication.
Michael Schmuker, a former professor of neural computation at the University of Hertfordshire who has previously consulted for Osmo, says the company’s technology is an achievement. With colleagues, he created an open-source, reverse-engineered version of Osmo’s 2023 model, but it lacks the reams of data the company has collected since.
But even with all this smell data, Osmo may be missing some of the factors that make human olfaction unique, Schmuker says. “Smell is super complex. Context is very important for smell. The same molecule may evoke a different association depending on context,” he says. “[The smell of] a French cheese like Camembert is partly the same molecule as in used nappies. You don’t confuse the two because of the context.”
In a controlled setting like a fragrance lab, this lack of context may not matter so much. But it could have an impact if Osmo’s technology becomes, as Wiltschko suggests, the “olfactive channel” for AI world models. Imagine a machine that could sniff out spoilage in food on a production line, or detect tumors through smell the way dogs have been trained to do in pilot studies.
For that to happen, sensor technology will need to get a lot better, Schmuker says. The models will need to keep improving. Medical diagnostics is the biggest prize.
“There’s a lot of history of companies that have gone directly at that mountain peak and, unfortunately, not been successful, because it is very expensive and it is very difficult,” Wiltschko says. “So what we’re doing is building this foundation at the foothills of all these mountains. We’re sending expeditions up to the mountains, doing explorations, and basically smelling the whole world.”
正文保留英文原文,标题与摘要由机器翻译。
原文发布于 MIT Technology Review(This smell startup says its AI nose is as good as a human’s),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。