The first Grounded Decision Model

The recording opens mid-sentence ("It's really important to make Chat a useful and functional vision model"); the MC's introduction is not captured, and the speaker never says his name. The ASR is heavily garbled throughout. The Q&A turns are left unlabelled: the transcript does not mark who is speaking.

Transcribed locally with Qwen3-ASR 1.7B (timings per ~5 s segment), prompted with the speaker names and affiliations of overlapping sessions. "parallel" at 00:06:34 and 00:07:20 was in that prompt (the unprompted Voice Memos transcript has "apparel" and "parallel" there). "Jarred" at 00:04:34 and "Exa" at 00:07:20 are prompt artifacts naming other sessions: the Voice Memos transcript has "jazz" and no word there. Echoes of the prompt were cut from 6 segments before this text was written. Other words are kept as transcribed: the inference revolution is "evolution" and "Versailles", the model being compared is "Jeff" and "Web Jeff", and the speaker's product is a "vision model". The Voice Memos transcript is kept beside it as transcript.txt.

00:00:00 — Introduction

It's really important to make Chat a useful and functional vision model for real world. Cool. Please. That's that's the bar contest. Now we're here talking about evolution. We saw Yogo yesterday, and I think he's like awesome. He thinks of it in a very different way, but specifically, I think the thing I want to talk about is like. On Versailles, the crowd has seen the explosion, the very explosion of Versailles, and that's because it was.

So much cheaper and so much faster than anything could do before, and specifically because often what we find is we'll read code. Of L1 intelligence, it will be important, and the right point, it will be not. Like as soon as I understand the content I'm working with, I can do a lot, and I don't really have to like write that much. And so, results of the evolution, like the quantity of requests for chat GPT models, has absolutely blown up.

And I'm here to make the case that like grounding for those is extremely important.

00:00:55 — A question a model cannot answer

Okay, so maybe a question: Who won the Michigan Senate race? This should be easy. This is like well known. I mean, a couple days ago, if you do Google searches, the first thing that shows up. But if you ask Jeff this question, this is what'll give you back. Mike Rogers with fifty-eight percent. It really doesn't know. Like maybe he had some pre-trained idea about Mike Rogers. Maybe had some pre-trained idea about Julian Sanchez. But like. He doesn't have a good sense of what the anthropic question is.

And so it's kind of crazy that you would use a decision model in your workflow that has like almost no understanding of like what. And so this is where we think that like not only is it wrong, but also its confidence is too high. So I think Joe made a really good point about actually like. The thing about using jobs is not just because of crap, but ultimately because no one is content with jobs, and so we think that.

00:01:50 — Grounding unlocks calibration

Part of what this unlocks with web grounding is that you completely unlock actual calibration of the. So this is like an example of many. Where it kind of just feels unfair, like. Using Java just doesn't solve the problem, and so web grinding helps to unlock the whole portion of things here. Cool. So we understand that one sort of. Offcase example: We've been for release of open source benchmark for many questions where.

We direct our web traffic to real workloads. Where it's like it's kind of a very unique, unique decision. The AIs that have underground things don't. It doesn't really make sense to ask them. And so, sort of the same the same things you might intuit that like. There's huge accuracy differences and huge calibration differences you can see in real benchmarks too. So here's a here's an accuracy benchmark, and then here's a calibration benchmark where we show that.

For the prayers, for it's much more accurate, much better calibrated. What is correct is much more confident that's correct if you use actual grounding to ground your decision. So there's like a whole set of things here that unlocks, and specifically like. Calibration is the important thing about good decision models.

00:03:01 — How it works

Awesome. Here's kind of how it's built and how it works. Effectively, you take some state input. It's like a wrap. It looks exactly like Jeff in terms of state. We do a live search, web search call, and we get information from our proprietary index, which is like. On the size, I think that's at least their size, which is about 100 billion pages, and we find the most relevant content for that decision. So. We take a bunch of results and we include that part of the state.

It runs under a second in response time. It's one dollar per thousand requests. So. Two orders of magnitude, maybe an order of magnitude cheaper than the LLM workflow we were using to be back in April. And then it's also like an order of magnitude faster. For that one call, you can also ask the same way project. Ask how much quantum state they want to parallelize. And so the same search content can be reused across the whole.

And like like we show here, like there's a bunch of different sort of use cases. There's true false questions. Employee benefits, all normal things.

00:04:05 — Since launch

Okay, so we talked about the industrial revolution. I think that the second phase of this is that there's a web industrial revolution. Where since we launched two weeks ago, we've seen absolutely. Like explosion of number of amazing AI community guys, both on the proprietary and on the proprietary, on our first-rate. And so we think this is just like the beginning of the exponential. People have only begun to realize.

First, what Jeff can do, but also how Web Jeff enables you to do other things around.

00:04:34 — Customer use cases

Cool. Okay, we have a bunch of amazing customers. I want to highlight some of the use cases of where Jarred. And web running. So some of our customers are using. You have a restaurant. They have a list of like a hundred thousand restaurants. California, and you search every single one of them up here on the website, and you can actually classify for every single one of them.

Is a reservation platform like phone only? Is OpenTable? Is a Resy and. What is the website platform? So I'm using Squarespace. I'm using Wix. What are they doing? What is the website? So if you were if you're a sales engine. And you want to go find for every single for every single question in all of California, this is a hundred dollars to classify, and then you build this model for every single question.

Second case, you have some sort of business Kyc flow. You search a business like Anchor Holdings Ltd. And you want to contemplate: What is the business model? Is it currently operating? Is there any sort of real threat against it? You can do that for a dollar per thousand epochs, which is like. And then a third example: If you have some news, you can do like instant categorization. That's not just a form of Theorized later scenes, but also all the other views.

So it shouldn't crash and extremely. These are some examples.

00:05:45 — Live demo

I'm going to go and do a really quick live demo. And hopefully, it's going to land well. Always a risk. So here is sort of an example that we cooked up. So, if you ask something like, "Who is the CEO?" Will give you true with a 96% probability. So I will run this live. Just so you guys know that this is nice. And seven people. And it's again kind of unfair. It doesn't. It's trained on like years worth of information.

We're all seeing is that Justin Wong is CEO of the video. Justin Wong is the CEO of the video for 30 years. And so it's highly confident, and it's. It's like truly uncalibrated. If you use a parallel decision API, which asks the same question. Not only will we give you the correct answer. Not only do we give you the answer, we also give you the citation, so you can go fact check the source yourself.

And so it all comes to like one dollar per thousand requests. It's super easy. It's super fast.

00:07:03 — Pricing

Yeah, that's what I wanted to talk about. I have some other slides here that we can sort of look through if you guys are interested about like. The pricing difference between what people currently do, which is that they use some sort of like open AI model. Go to research. Find your outfit scheme that you attend. Versus like the parallel system, which is using the Exa. Anyways, recorded back like two or three gigabits.

So something that we're seeing is like people are doing this live in every single hotel. As you're typing your business name. We can go search it up, find out what institutes you're in, what institutions you're in. How it's like. Sub-second latency, so like you—you don't even get to the paid part of the information.

00:07:45 — Q&A

Yeah. As We have our own deep models. And you can imagine, like anyone looking at that, would say, Wow! The size of our web index is. Is we'll get more specific, but it's roughly the size of the exa. We talked about a little bit. Which is more than the size of pretraining, like something like AI does. And so, if you have OpenAI's WebMDX, if you have LLMs, you should probably.

So I was actually playing around with this. Yesterday, it works really well if you don't. The cards are fun. Parallelize questions. Yes. So, like, if you ask two questions in parallel, they obviously will share context, the state that you. You all retrieve. So, in your election example, you might be like, "Who won?" What's their position on immigration? What party are they a member of? Because the same search will back those results, it actually degrades as opposed to asking each of those as.

Separate questions. Do you all have plans to address that? How do you? Where is this my user area? Yeah, I think the intention behind the product shape is that we want. A single sort of grounding or single search to answer a question. We think that there's a lot we can do to the project. It's very early days in terms of object, and so we're collecting product feedback.

And so I think that is somewhat out of scope, but also I think we're going to do a lot of work to make it much better than today. Yeah. So the question was, do we use it for calibration techniques to? Also, the also the calibration. Again, like I think that the thing is, is that when you have the weather context information. It's almost very clear what the answer is. Part of the reason why you needed this complex calibration, you didn't have a complex calibration.

Is because you were kind of guessing, like you had to sort of guess some probability, like this percentage of the time is used, and that percentage. Awesome. Thank you.