Neurosymbolic AI outperforms chatbots and product search | E2327

**Zach Hudson** (0:00)
LLMs are not good at this. They regress to the mean, and they find the most probable thing, and that's often not what you need personally. When you need to remove hallucination, and you need trust, you need to be able to see inside the model, that's where neuro-symbolic models can really shine.

**Jason Calacanis** (0:13)
You really have built something between Houzz and Pinterest that is more powerful than both of them.

**Zach Hudson** (0:19)
The performance of it is already 2.5x greater than some of the largest search companies in the world. It learns that new information, so you don't have to go do another training run. It's literally updating in real time. It's one-one-thousandth the cost of an average front-tier model training run in the US.

**SPEAKER_3** (0:33)
This Week in Startups is brought to you by YSecurity, the on-demand security team for startups. Need enterprise-grade security without hiring a $400,000 CISO? Security gives you 40-plus expert engineers matched to exactly what you need by the hour, with your first six hours completely free. Go to ysecurity.io/twist.
agree.com Stop chasing invoices and automate your entire contract to cash stack. Go to agree.com and tell them Jason sent you to get 50% off for life and Vanta. Compliance and security shouldn't be a deal breaker for startups to win new business. Vanta makes it easy for companies to get a SOC2 report fast. Get $1,000 off for a limited time at www.vanta.com/twist.

**Lon Harris** (1:13)
We got Zach Hudson joining us. He's the CEO of Onton. They're a search and discovery engine for e-commerce. It's powered, Jason, by a neuro-symbolic AI model. What does that mean? You can see it for yourself at onton.com. You're going to find out right now.
Zach, thank you so much for joining us.

**Zach Hudson** (1:29)
Thanks for having me.

**Lon Harris** (1:29)
Pleasure.

**Jason Calacanis** (1:30)
Yeah, and you're going to give us a demo, so let's show Don't Tell.

**Zach Hudson** (1:33)
Oh, yeah, yeah, yeah.

**Jason Calacanis** (1:34)
Don't tell me what have you built, yeah.

**Zach Hudson** (1:35)
We can dive right in.

**Lon Harris** (1:36)
Let's do it.

**Jason Calacanis** (1:37)
Oh, and we're investors in the company, just as a full disclosure here, yeah.

**Zach Hudson** (1:42)
I'll give your audience a chance to go and do some basic searches for themselves, but why don't I jump in and do some more of the mind-blowing searches.
And a little bit of background about Onton, we have these things called surfaces within our product. If you've ever had Lon or Jason, I've either of you used Notion before, I'm sure you have.

**Lon Harris** (2:00)
Plenty of times.

**Jason Calacanis** (2:02)
More than once.

**Zach Hudson** (2:02)
So they have these things called blocks there, they let you organize information in different ways. People do this quite often in shopping journeys too, where they create lists, they create mood boards, these types of things, these are surfaces within Onton. And so what I've prepared for you was I put together some of your favorite things, Jason, or at least some of the hotels that you've mentioned, some of your other things you mentioned were...

**Lon Harris** (2:23)
We sent him some cheat codes, Jason. Monocle Magazine, a few hotels we know you like. We helped Zach build this for you.

**Zach Hudson** (2:29)
I went back through some of the podcasts and grabbed a few others. But we put this into a canvas. But what we've realized recently is it wasn't possible to search these without this model. And so I want to use this to demonstrate what Ontology 1 can do in the amount of context these models can take. So I'm just going to go search with it. I'm going to use this just to go, hey, try to find some stuff. OK.

**Jason Calacanis** (2:54)
So you put images on a mood board of hotels I've stayed in and the design of those hotels. So I saw one of them there look like the room at the Artisan in Singapore, yeah?

**Zach Hudson** (3:06)
And I put even more than that. I put actual products in here. I put some text and other context in here. And Ontology is able to suck all of that up. So it is a lot like a mood board. But Ontology 1 will look at that whole thing. And it will try to understand it. The thing that I want to call out here is that it's learning in real time. We don't have any tags for these things.
We don't have any, this wasn't like human labeled. It is going off and learning them.
But more importantly, it's storing that understanding in the model. So every next search that's happening with Onton is getting a tiny bit smarter and a little bit faster, which is incredible. We don't have to go back and do another training run with this model to get better accuracy.