Neurosymbolic AI outperforms chatbots and product search | E2327

Onto is a search and discovery engine for e-commerce powered by a neurosymbolic AI model. In a live demo, CEO Zach showed how the system can interpret a mood board of images and text—say, photos of hotels Jason likes—and return furniture that matches the aesthetic, learning what "organic burl wood" or "oversized creamy off-white chunky knit" means in real time without a separate training run. The model can explain exactly why it returned each result, something traditional LLMs cannot do. The performance is already 2.5 times greater than some of the largest search companies in the world, at one one-thousandth the cost of an average frontier model training run in the US.

Why LLMs struggle with taste

The episode draws a sharp distinction between two types of shopping: needs (toothpaste, deodorant) and wants (items above $50 driven by taste, self-expression, emotion, and imagery). LLMs are not built for the second category. "They regress to the mean and they find the most probable thing, and that's often not what you need personally," Zach explained. Neurosymbolic models, by contrast, combine an intuitive side with a logical side—analogous to system one and system two thinking in psychology. LLMs handle system one (fast, associative) well, but when you need to remove hallucination and build trust, you need to see inside the model. That is where neurosymbolic models shine: they