Neurosymbolic AI outperforms chatbots and product search | E2327

Antan, founded by Zach Hudson, is a search and discovery engine for e-commerce powered by a proprietary neurosymbolic AI model called Ontology One. Unlike large language models, which Hudson says regress to the mean and return the most statistically probable result rather than what a specific user needs, Ontology One builds a separate understanding for each individual user, updates in real time without requiring a new training run, and cost one one-thousandth the cost of an average frontier model training run in the United States.

Hudson argues that LLMs applied to e-commerce data hallucinate frequently and return irrelevant products, a problem compounded by years of SEO gaming that has degraded data quality. Because Ontology One is interpretable rather than a black box, it can explain definitively why any result was returned, removing hallucination and enabling user trust. The model can perform multi-category and multi-object searches and can learn concepts such as organic, burlwood, brass, and floral directly from a mood board without human labeling or tags.

Antan's search performance is already 2.5 times greater than some of the largest search companies in the world, though Hudson notes the company currently has a smaller index size than those benchmarks. The company is on track for hundreds of millions of searches over the next year within its current single product category, home decor and furniture. Current revenue comes from affiliate commissions, but Hudson identifies the largest revenue opportunity as an API that would allow the ontology to be plugged directly into stores.

Apparel is the next product category planned after home decor and furniture, followed by electronics. As the ontology learns one category it makes the next smarter, for example learning what polyester means in home decor improves the model's understanding of polyester in apparel, contrasting with traditional e-commerce expansion that requires relabeling all products and rebuilding understanding from scratch. Hudson also noted that products priced above roughly fifty dollars are driven by taste, self-expression, emotion, and imagery rather than commodity need, which is the market segment Ontology One is designed to serve.

Hudson drew a distinction between use cases best suited to neurosymbolic models versus LLMs, arguing that LLMs are well suited to verifiable domains like coding and mathematics where outputs can be externally checked, while neurosymbolic models excel when hallucination removal, trust, and interpretability are required. Jason Calacanis disclosed that Antan is a portfolio investment made in 2021 or 2022 during COVID, when the company was a search engine predating ChatGPT.

Spaceium, a seven-person startup based in San Francisco that went through Y Combinator in summer 2024, is building in-space refueling infrastructure. The company built its first payload in five months with two people, does 98 percent of testing in-house, and has already put a payload into space. In February, Spaceium announced it had flown the most precise robotic actuator ever tested in orbit, achieving end-of-arm docking precision within 0.5 millimeters, a level required because spacecraft are built with very low tolerance for physical stress and there are no humans on board to correct errors.

Spaceium's zero boil-off technology allows propellants including Xenon to be stored in space for extended periods. Current commercial customers are asking for two to five metric tons of fuel, which is achievable with existing launch vehicles, meaning lower launch costs from vehicles like Starship or New Glenn are a tailwind for Spaceium's economics but not a prerequisite for current operations. A Xenon propellant transfer mission in orbit is planned within the next year, followed by a free-flying spacecraft docking and Xenon transfer mission within one and a half to two years, and both upcoming missions are fully funded.

Spaceium has raised close to 13 million dollars across all funding rounds, with that capital covering all three missions including the one already flown. The company also has close to 100 million dollars in commercial contracts and over 2 billion dollars in letters of intent. Ashi identified a chicken-and-egg dynamic where larger space companies are waiting for refueling stations to exist before designing technology around refueling, though some customers are already committing to purchase fuel once available. Spacecraft design changes that take advantage of refueling are not expected to be visible in the next three to five years, but companies are already designing future missions around the assumption that refueling will be available, and Ashi said groundbreaking news about refueling standardization is expected within one to two months of the recording date.