The TWIML AI Podcast

The episode contains substantial technical and scientific content about olfactory intelligence, including the structure-odor relationship problem, the principal odor map, graph neural networks applied to molecules, and data collection at scale (543 million sniffs, 6 billion molecules digitized). However, significant portions are devoted to foundational biology explanations (olfactory receptors, human smell sensitivity) that are educational but not novel to informed operators, and considerable time is spent on business positioning and fragrance applications that are somewhat tangential to the core technical substance.

“over 300 channels of olfactory information in the nose and it's still a mystery exactly what they code for”

“we've digitized 6 billion molecules at this point”

The framing of smell as a modality for AI is genuinely novel and underexplored in the AI discourse. The principal odor map discovery, the odor Turing test, and the observation that embedding dimensionality (~300) mirrors biological olfactory receptor count show creative thinking. However, the technical approaches (graph neural networks, embeddings, multimodal learning) are standard ML patterns applied to a new domain rather than methodologically novel. The business pivot to fragrance, while pragmatic, is presented as opportunistic rather than generating breakthrough insights.

“we've Passed an Odor Turing test. Like, our model predictions were human quality”

“this region of the map is vanilla, this region of the map is Red Barry, et cetera. Um, without that, you actually can't do that classification problem. So that embedding turned out, if you do the engineering right, it just kind of needs to be around 300 dimensions to work really well, which is, like, suspicious, but, you know, just suggestive”

Alex Wiltschko is a highly credible practitioner: former Google DeepMind researcher who founded Osmo and is actively building production systems for olfactory AI. He demonstrates deep domain expertise spanning neurobiology, chemistry, machine learning, and commercial scale-up. He has shipped products (fragrance design), collected proprietary datasets at scale, and operates a factory with deployment infrastructure. This is not a thought leader or career podcast guest - he's an operator with skin in the game.

“Alex Wolchko, founder and CEO of Osmo and a former Google DeepMind researcher”

“we have a factory where we make it”

The episode is rich with concrete numbers and specific examples: 5,000 molecules in initial dataset, 6 billion molecules enumerated, 543 million human sniffs collected, 300-dimensional embedding space, 50 descriptive odor terms used in training, graph sizes of 3-20 atoms, odor Turing test with Joel Mainland at Monell, specific chemical names (mercaptans, rhodopsins), regulatory frameworks (EU/US/worldwide), and production capacity (one fragrance per 100 seconds). Trade-offs: less specificity on model architectures post-evolution, regulatory timelines, and business metrics (revenue, customer count).

“we've digitized 5,000 molecules in our first dataset. We've digitized 6 billion molecules at this point”

“we've digitized five 43 million sniffs”

Sam asks technically competent follow-up questions (graph nodes/edges, embedding structure, multimodal inputs, decode mechanisms) and occasionally pushes back or seeks clarification (e.g., asking about confounding examples, business model viability). However, many exchanges feel exploratory rather than pressuring - Sam often accepts high-level answers without drilling into contradictions or limitations. For example, when Wiltschko claims models 'fell out' from good data but then mentions dozens of specialized models, Sam accepts this without fully reconciling the tension. The host does not challenge vague claims about aromatherapy, emotion detection, or the business model difficulties in healthcare.

“Before we get to that structure, you mentioned a graph neural net was the fundamental architecture here. What did the nodes and the edges in the graph represent?”

“But of course you're simplifying a lot because for each of those other modalities, there's lots of different maps”