Ceramic AI wants to slash web search costs from $5–15 to 5 cents per thousand queries

and it should download ffmpeg do it, but who knows how good it is yet. That okay, that's our new benchmark. uh music video driven by reels. Let me tell you about console.com. Console builds AI agents that automate 70% of IT, HR, and finance support, giving employees instant resolution for access requests and password resets. Um, and without further ado, we have Anna Patterson from Ceramic AI here.

What's going on? How you doing, Anna? Good to meet you.

Great to meet you.

Yeah, you too.

Thanks so much for joining us.

Are you guys Are you guys having a good um St. Patrick's Day? I see the green.

I am. Yes, we're very

greened up.

Yes, greened up. We we we have these bright green suits from uh a show we did on Black Friday about Shopify and uh and uh I thought certainly I will not be using that until next Black Friday, but here I am on St. Patty's. Well,

happy St. Patty's Day to you. Uh since it is the first time on the show, please introduce yourself and the company.

Hi, I'm Anna Patterson. I'm a founder of Ceramic AI. I was a longtime Google engineer. I started in 2004 and I'm best known for building large search engines.

That's amazing. So, uh yeah, get you give me the pitch for ceramic AI.

So, ceramic brings the cost of search in line with the cost of inference.

As you know, inference costs have been going down and down and actually inference is faster and faster.

But search is $5 to $15 per thousand searches. But inference is maybe 50 cents per thousand searches. So the kind of analogy I like to use is tacos and salsa,

right? Tacos is kind of the me the meal and that's kind of what inference is. It's the thing that's really delivering intelligence to your application.

Um but salsa kind of makes it better, right? But search is now the dominant cost

in in tacos and salsa. you're adding search, but it's it's5 to $15 per thousand queries. So, it's really kind of time to bring um salsa in line with tacos. So, ceramic is 5 cents per thousand queries.

Amazing.

Right. So, you're saying this historically it was like you were getting a taco and it cost you $5, but then they were like, "Well, if you want salsa, it's going to cost you an extra like, you know, $500." Yeah.

And you're like, "Well, I don't not sure I really want the salsa." a weird weird exactly.

Okay.

So, help me define help me understand what search means because search can mean over the internet that's already sort of baked into LLM. It can mean web active web search like in a proprietary database or just the open web. Uh how how are you thinking about the surface area of search?

Yeah. So, we do have um a 40 billion page web search

and that is the open web. Mhm.

And then we've built proprietary systems as well.

So one of the things that we're announcing is this idea of supervised generation that as the model is generating it's double-checking what it's saying with search and it's also double-checking with search what else should I say to make a comprehensive topic. And so that way you can really enable new applications with 10x fewer hallucinations but also make the whole product affordable and fast.

Okay. So who who do you sell I sounds super valuable. Who do you sell this to? Is this going to be uh maybe similar to kind of like the data labeling market where there's like five customers that like really matter and you want to get all of them or are there a bunch of other applications that you want to actually sell to vertical specific

you know AI applications that can uh vend you in in like kind of in line with their LLM products.

Yeah, I think there are um kind of two strategies there. Um you mentioned one of getting all the big players uh that definitely would be nice but we also have a self-s serve every agentic workflow. We have one startup their uh agentic workflows do,00 searches.

So this um our search engine responds in 50 milliseconds.

So it's both affordable and um more real time than the other search engines. So um we see a number of agent workflows happening but to your idea of um a custom index or a custom application we see that as well because let's say you're a pharmaceutical company or some banks are very privacy ccentric. They don't want their searches going out and they don't even want their searches to models going out. So they kind of host their own copy of a model and here they'd host their own copy of search as well. so that they get their own proprietary environment for all their agentic flows inside their enterprise.

Um,

makes sense.

Talk about just how the open web is changing in the age of AI. I I've seen some crazy stats about how much more of the internet Googlebot sees because everyone has been indexing and been very friendly to Google for a very long time. Other publishers are getting more closed off. H how how easy is it to actually search the internet broadly these days?

So the 40 billion pages that we have are available on the open web. We do not um disobey or I should say we obey robots.ext. So we don't actually crawl news sources that um have blocked us, but we are in active talks to make deals to them

and to have a proprietary API that costs more but also um reflects back revenue to those proprietary sources.

Oh, cool. Um are are are you uh are is there any value in having like I've always been interested in the flip side of Google search versus Google alerts where the search is happening internally and then and then the information is actually getting pushed to you. That product is like probably like 0.001% as important to Google as search but it's always been interesting to me. Is that interesting to you? Is that relevant in the age of AI? Does anything change about that ratio going forward? Um well one of the interesting things um is that inference a lot of times with these models

inference has a lot of spare of compute because it's memory bound

and so with that it means that it could be thinking. So as it's inferencing, it could be getting a stream of search results, searching all all the time and actually bringing you sort of like alerts only the most interesting information or information that it doesn't think that you already know by looking at your history. So I think it's going to enable uh a lot of new um applications.

Amazing.

Uh wild card question hit me. How long until we see ads in Gemini? There's been some reporting this week. Obviously, Demis had come out and said, you know, why would we put ads in?

For the record on this show, we are extremely pro.

We're extremely pro ad.

We love Exactly. Yeah.

Uh and and I and I we both expect ads to be in Gemini. I would say this year is my guess, but

that's that would be my guess as well. Before the end of 26. Yeah. Let's go.

Cool. That gives me a lot of hope.

We're going to be very excited about

and excitement.

Faith in humanity restored.

What's What's the story of the the pink guitar on the wall there?

Yeah.

Um well, if you want to do something really humbling, learn an instrument from your children.

Oh, okay. They um they're absolutely brutal with telling you to practice everything you're doing wrong. And uh and so I kept borrowing my daughter's um electric guitar

after she taught me acoustic guitar. And so uh she decided to get me my own.

Oh wow.

Because um she saw her guitar like laying on my couch

and she said a guitar should be hanging on the wall and I said I think I've been told to clean up my room by my child.