ART ARGENTUM ANALYSIS
Olfactory Intelligence and AI: Bridging the Gap
Analysis of olfactory intelligence and its implications, based on 'How AI Learns to Smell' | The TWIML AI Podcast with Sam Charrington.
OPEN SOURCESUMMARY
AI has primarily focused on digital formats like text and images, leaving the challenge of digitizing smell largely unresolved. Smell plays a crucial role in animal communication and disease detection, yet it has remained outside the reach of computing due to the lack of practical methods for large-scale digitization. Alex Wilchko, CEO of Osmo, is working to create olfactory intelligence by developing AI systems capable of modeling, predicting, and designing scents, thus bridging the physical and digital realms.
The process of enabling computers to smell involves three key steps: converting atoms into data, mapping that data for digital use, and reproducing scents. This remains a significant gap in current AI capabilities, especially given the complexity of the olfactory system, which has over 300 channels of information compared to the simpler systems for color and sound.
Olfactory sensory neurons (OSNs) are specialized cells that detect smells, extending from the brain through the skull to interact with the environment. Humans possess over 300 types of olfactory receptors, allowing them to detect odors with remarkable sensitivity. Contrary to popular belief, humans can effectively track scents, as demonstrated by experiments where individuals successfully follow scent trails.
The structure-odor relationship problem, unresolved for a century, was tackled using a graph neural network to predict molecular smells based on chemical structure. An odor touring test showed that the AI model's predictions surpassed individual human assessments, demonstrating high accuracy in olfactory predictions.
Osmo aims to create safe and affordable fragrance molecules, addressing market gaps for specific scent profiles. The company has amassed the largest olfactory dataset, digitizing over 5.4 million scent samples, which is essential for training AI models to assess the characteristics and safety of new fragrance molecules.
The future of olfactory intelligence includes applications in disease detection and consumer products, with a focus on collecting comprehensive olfactory datasets. This initiative emphasizes the complex, hands-on work required for effective olfactory data collection, contrasting with traditional AI data scraping methods.
XDETAIL
INFO
YOUTUBE2026-07-08the twiml ai podcast with sam charrington
How AI Learns to Smell
STANCE
00:00
05:00
10:00
15:00
20:00
25:00
30:00
35:00
40:00
45:00
50:00
55:00
12 intervals • swipe left
How AI Learns to Smell
Alex Wiltschko, CEO of Osmo, is developing AI systems to digitize smell, a challenge that has remained largely unresolved. The process involves converting atoms into data, mapping that data, and reproducing scents, which…
FULL
00:00–05:00
Alex Wiltschko, CEO of Osmo, is developing AI systems to digitize smell, a challenge that has remained largely unresolved. The process involves converting atoms into data, mapping that data, and reproducing scents, which could lead to significant advancements in various fields.
- AI has largely concentrated on digital formats like text and images, leaving the challenge of digitizing smell largely unresolved, despite its critical role in animal communication and disease detection
- Alex Wiltschko, CEO of Osmo, is working to create olfactory intelligence by developing AI systems capable of modeling, predicting, and designing scents, thus connecting the physical and digital realms
- The process of enabling computers to smell involves three key steps: converting atoms into data, mapping that data for digital use, and reproducing scents, which remains a significant gap in current AI capabilities
- The olfactory system is more intricate than color perception, with over 300 channels of information in the human nose, presenting a higher-dimensional challenge for AI compared to established systems like RGB for color
- Advancing the understanding of scent mapping could lead to breakthroughs in various fields, including medical detection and consumer products, underscoring the transformative potential of olfactory intelligence
Read full analysis
STANCE
STANCE MAP
Proponents of Olfactory Intelligence
- Advocate for the digitization of smell to enhance AI capabilities and applications
- Highlight the potential for olfactory intelligence in disease detection and consumer products
Skeptics of Current AI Models
- Question the ability of current AI frameworks to fully capture the complexity of olfactory perception
- Raise concerns about individual variability in scent perception affecting model accuracy
Neutral / Shared
- Acknowledge the significant advancements in olfactory intelligence and its applications
- Recognize the challenges in collecting comprehensive olfactory datasets for training AI models
FULL
05:00–10:00
Alex Wiltschko discusses the development of AI systems to digitize smell, focusing on the mapping of olfactory receptors to molecular structures. The advancements in olfactory intelligence could lead to applications in various fields, including disease detection and consumer devices.
- Olfactory sensory neurons (OSNs) are specialized cells that detect smells, extending from the brain through the skull to interact with the environment
- Humans have over 300 types of olfactory receptors, enabling them to detect odors with remarkable sensitivity, akin to identifying a single drop of a substance in an Olympic-sized swimming pool
- Contrary to popular belief, humans can track scents effectively, as demonstrated by experiments where individuals successfully follow scent trails similar to dogs
- Digitizing smell presents challenges, particularly in mapping the intricate structure of olfactory receptors to their corresponding molecular structures for AI representation
- Developing olfactory intelligence requires training neural networks on large datasets that connect molecular structures to their associated smells, improving AIs predictive and generative capabilities
FULL
10:00–15:00
Alex Wiltschko is developing AI systems to digitize smell by mapping olfactory receptors to molecular structures. This advancement could lead to applications in disease detection and consumer devices.
- The structure-odor relationship problem, unresolved for a century, was tackled using a graph neural network to predict molecular smells based on chemical structure
- An odor touring test showed that the AI models predictions surpassed individual human assessments, demonstrating high accuracy in olfactory predictions
- The graph neural network analyzes molecular graphs, where nodes represent atoms and edges represent bonds, to derive insights into potential smells of molecules
- A principal odor map was created from a nearly 300-dimensional vector, visualized in two dimensions to group molecules by similar scents, such as sweet or cucumber
- The research underscores the potential of olfactory intelligence for applications beyond fragrance, including medical detection and emotional sensing, utilizing extensive datasets and advanced machine learning
METRICS
OTHER
300-dimensional vectordimensions
details
CONTEXT: the dimensionality of the embedding used for smell prediction
WHY: This dimensionality is crucial for accurately mapping and predicting scents
EVIDENCE: that embedding turned out, if you do the engineering right, it just kind of needs to be around 300 dimensions to work really well
OTHER
5,000 dotsmolecules
details
CONTEXT: the number of molecules plotted in the principal odor map
WHY: This dataset size is significant for training the model effectively
EVIDENCE: there were about 5,000 dots on that plot
FULL
15:00–20:00
Alex Wiltschko is developing AI systems to digitize smell by mapping olfactory receptors to molecular structures. This advancement could lead to applications in disease detection and consumer devices.
- The mapping of scents reveals a complex hierarchy among odors, with floral scents having specific sub-regions for flowers like jasmine and rose, highlighting a biological link in scent production
- Osmo aims to create safe and affordable fragrance molecules, addressing market gaps for specific scent profiles, such as long-lasting citrus and clear vanilla notes
- The companys olfactory dataset has expanded dramatically from 5,000 to 6 billion molecules, leveraging innovative data licensing and industry expertise to improve fragrance development
- The identification and production of new scents are driven by clear industry needs, enabling the team to focus on developing molecules that meet specific market demands
METRICS
OTHER
6 billionmolecules
details
CONTEXT: the size of Osmo's olfactory dataset
WHY: A larger dataset enhances the accuracy and reliability of AI models in scent prediction
EVIDENCE: we've digitized 6 billion molecules at this point.
OTHER
5,000molecules
details
CONTEXT: the initial size of Osmo's olfactory dataset
WHY: The dramatic increase in dataset size indicates significant progress in data collection capabilities
EVIDENCE: we had I think 5,000 molecules in our first data set.
FULL
20:00–25:00
Alex Wiltschko discusses the creation of AI systems that digitize smell by mapping olfactory receptors to molecular structures. This technology has potential applications in various fields, including disease detection and consumer products.
- Osmo has created a detailed catalog of potential scent molecules, filtering them based on physical properties and manufacturability to guide scent development
- The company has amassed the largest olfactory dataset, digitizing over 5.4 million scent samples, which is essential for training AI models to assess the characteristics and safety of new fragrance molecules
- Safety testing for new scents involves rigorous empirical validation and predictive modeling to ensure compliance with regulations related to skin, inhalation, and environmental safety
- To quickly address client needs for products like air fresheners and shampoos, Osmo primarily blends existing approved molecules while also innovating entirely new scents
FULL
25:00–30:00
Alex Wiltschko is developing AI systems to digitize smell by mapping olfactory receptors to molecular structures. This technology has potential applications in various fields, including disease detection and consumer products.
- Osmos olfactory intelligence system enables customers to create scents tailored to specific demographics, such as Gen Z men, by combining AI technology with the expertise of master perfumers
- The scent design process incorporates customer descriptions into a perceptual space, which is then translated into a fragrance formula, supported by a factory that can produce a new scent every 100 seconds
- Each new fragrance not only meets customer preferences but also contributes data that improves Osmos predictive models, fostering a self-reinforcing cycle of enhancement and business growth
- By integrating multimodal inputs like text and images into the scent development process, Osmo expands creative opportunities for both customers and perfumers
FULL
30:00–35:00
Alex Wiltschko is developing AI systems to digitize smell by mapping olfactory receptors to molecular structures, which could lead to applications in disease detection and consumer devices. The company Osmo has created the largest proprietary olfactory dataset to train predictive models for scent creation.
- Osmo utilizes advanced neural networks to establish connections between molecular structures and their associated scents, incorporating diverse input types such as chemical sensor data
- The company has created the largest proprietary olfactory dataset, enabling swift data generation and model training, which surpasses traditional fragrance companies that depend on outdated data collection techniques
- Osmo employs a fleet of predictive models for scent creation, akin to the architecture used in autonomous vehicles, facilitating nuanced scent development tailored to specific applications
- A significant challenge for Osmo is understanding the interactions between different molecules, as refining their models to accurately predict scent mixtures is a primary goal in advancing olfactory intelligence
FULL
35:00–40:00
Alex Wiltschko is developing AI systems to digitize smell by mapping olfactory receptors to molecular structures, which could lead to applications in various fields. The technology aims to predict scents while addressing safety, manufacturability, and cost through distinct predictive models.
- Osmos olfactory intelligence aims to predict the scent of substances while addressing safety, manufacturability, and cost through distinct predictive models
- Smell significantly influences flavor perception, with approximately 90% of flavor derived from olfactory input, underscoring the close relationship between taste and smell
- Cultural background plays a crucial role in shaping individual scent and flavor preferences, as evidenced by differing views on foods like kimchi and strawberries across cultures
- While the current focus is on smell, Osmo recognizes the intricate relationship between taste and smell, with potential plans to explore taste in the future
- The predictive models created by Osmo are adaptable, enabling various outputs that meet regulatory standards and align with consumer preferences
FULL
40:00–45:00
Alex Wiltschko is developing AI systems to digitize smell by mapping olfactory receptors to molecular structures, which could lead to applications in various fields. The technology aims to predict scents while addressing safety, manufacturability, and cost through distinct predictive models.
- The development of olfactory intelligence focuses on mapping, reading, and writing smells to enhance scent detection capabilities through a foundation model
- Osmo is working to collect a comprehensive olfactory dataset, including scent profiles from diverse sources, to support predictive models for applications such as early disease detection
- A significant challenge in olfactory data collection is quantifying subtle scent changes related to health conditions, which is more complex than data collection in other AI domains
- Alex highlights the importance of collaboration in gathering a large olfactory dataset, which must include samples from both healthy and sick individuals, as well as everyday items
- The initiative contrasts with traditional AI data scraping, emphasizing the complex, hands-on work required for effective olfactory data collection
METRICS
OTHER
10,000 peoplepeople
details
CONTEXT: of individuals to collect scent data from
WHY: A large dataset is crucial for building effective predictive models
EVIDENCE: let's go sniff 10,000 people.
FULL
45:00–50:00
Alex Wiltschko is developing AI systems to digitize smell by mapping olfactory receptors to molecular structures, which could lead to applications in disease detection and consumer devices. The technology aims to predict scents while addressing safety, manufacturability, and cost through distinct predictive models.
- The future of olfactory intelligence aims to set benchmarks for applications such as malaria and cancer detection, as well as consumer products, to assess the effectiveness of AI models in predicting odors
- Expanding the understanding of intelligence to include chemical communication from various species could enhance AI models and broaden their applications
- Current chemical sensing technologies are specialized and primarily used in laboratories, but there is potential for miniaturization that could allow integration of olfactory sensors into everyday devices like smartphones
- Introducing olfactory sensors into consumer technology could lead to innovative applications, similar to the unexpected uses that emerged with camera phones, showcasing the transformative potential of incorporating smell into digital experiences
FULL
50:00–55:00
Alex Wiltschko is developing AI systems to digitize smell by mapping olfactory receptors to molecular structures, which could lead to applications in various fields. The technology aims to predict scents while addressing safety, manufacturability, and cost through distinct predictive models.
- Olfactory intelligence has potential applications in detecting spoiled food, tracking scents, and diagnosing diseases like cancer, though developing viable business models for these technologies is challenging
- There is significant interest in creating devices that mimic a dogs sense of smell, which could lead to innovative consumer applications in health monitoring and environmental sensing
- Advancements in olfactory sensors could enable miniaturization, allowing integration into everyday devices like smartphones, thus making scent detection more accessible
- Mapping scents to emotions and other sensory experiences is complex, indicating that enhanced labeling systems could improve our understanding of chemical interactions
- The development of olfactory technology is closely linked to the fragrance industry, suggesting that advancements in one area could positively impact the other
FULL
55:00–60:00
Alex Wiltschko is developing AI systems to digitize smell by mapping olfactory receptors to molecular structures. This technology aims to predict scents and has potential applications in various fields, including disease detection and consumer devices.
- Integrating olfactory information with emotional responses is a promising research area, though quantifying emotions remains complex
- Current neuroscience techniques, like EEGs, face criticism for their inability to accurately measure emotional states, underscoring the need for rigorous methods in olfactory studies
- Certain scents are believed to positively affect mood and cognitive functions, indicating a potential physiological link that could be utilized for therapeutic applications
- Traditional practices such as Ayurveda and aromatherapy may have scientific validity, suggesting that further exploration could reveal significant insights into the effects of scents on well-being
- The necessity for comprehensive research and validation in olfactory analytics to fully realize its potential in enhancing human experiences
CRITICAL ANALYSIS
The assumption that olfactory intelligence can be effectively modeled by current AI frameworks overlooks the complexity of scent perception and the potential for unaccounted variables, such as individual differences in olfactory receptors. Inference: The success of this technology hinges on the ability to accurately map and reproduce the intricate relationships between molecular structures and perceived odors, which remains untested in practical applications.
METRICS
other
300-dimensional vector dimensions
the dimensionality of the embedding used for smell prediction
This dimensionality is crucial for accurately mapping and predicting scents
that embedding turned out, if you do the engineering right, it just kind of needs to be around 300 dimensions to work really well
other
5,000 dots molecules
the number of molecules plotted in the principal odor map
This dataset size is significant for training the model effectively
there were about 5,000 dots on that plot
other
6 billion molecules
the size of Osmo's olfactory dataset
A larger dataset enhances the accuracy and reliability of AI models in scent prediction
we've digitized 6 billion molecules at this point.
other
5,000 molecules
the initial size of Osmo's olfactory dataset
The dramatic increase in dataset size indicates significant progress in data collection capabilities
we had I think 5,000 molecules in our first data set.
other
10,000 people people
of individuals to collect scent data from
A large dataset is crucial for building effective predictive models
let's go sniff 10,000 people.
THEMES
#olfactory_intelligence#AI#scent_detection#ai_in_healthcare#ai_perfumery#data_collection#science#ai_development#digitizing_smell#scent_technology#ai_research#data_generation#digital_scent#emotional_science#fragrance_development#osmo#scent_prediction#smell_ai#smell_technologyAI smell technology
DISCLAIMER
This analysis is an original interpretation prepared by Art Argentum based on the transcript of the source video. The original video content remains the property of the respective YouTube channel. Art Argentum is not responsible for the accuracy or intent of the original material.