Keep pulling the thread on Alex Wiltschko.
Osmo uses a proprietary AI system called "OI" (olfactory intelligence) to design new molecules for customers.
Osmo is launching a new fragrance house called Generation, which combines its AI technology with human perfumers to design scents for new customers like social media creators.
Osmo has fully automated the interpretation of Gas Chromatography-Mass Spectrometry (GCMS) data using its OI system, a process traditionally done by a combination of software and human experts.
Osmo has created a library of 10,000 to 20,000 molecules that have all been designed by its AI.
Osmo uses an active learning loop where it creates new smells, turns them into data, and uses that data to train its AI, with each cycle designed to teach the system more efficiently.
The development of graph neural networks was a key breakthrough that enabled AI to work effectively on chemistry problems, which was then applied by Osmo to the domain of scent.
In a double-blind trial at Google Brain, an AI model predicted the smell of 400 novel molecules more accurately than the average human panelist, effectively passing an "odor-touring test."
Osmo has successfully digitized a human sense by creating a system that can "round trip" a smell: reading a physical scent, representing it digitally, and then recreating the physical scent.
Osmo has developed a sensor system for StockX that can determine if a shoe is real or counterfeit within 20 seconds by analyzing the smell from the shoe box.
A long-term goal for Osmo is to use its scent detection technology for human health and wellness applications by analyzing health signals in scents emitted by the human body.
Ilya Sutskever stated at a NeurIPS conference that large language models have already been trained on the entirety of the internet, and there is no more raw text data to fuel further innovation at the same pace.
Osmo has built a proprietary, internal AI ecosystem for smell, including human labelers, 24/7 molecular sensors, and scent-creating robots, to generate the data needed to train its models.