You.com Rolls Out GPTChat-style Function Into Search
On December 23rd You.com announced the integration of their ChatGPT-like system into their search results. I got to chat with them about it.
Created on January 12|Last edited on January 31
Comment
There's been a lot of hype around ChatGPT, and for good reason. The text generation tool, built on GPT-3.5, illustrates the massive leap forward LLMs are making and how they are increasingly solving real-world problems and tasks.
Articles discussing ChatGPT as a Google-killer flow, and now Microsoft is in discussions to invest $10 Billion in OpenAI, the company responsible for ChatGPT (a great deal in my opinion), seemingly to integrate the technology into their own business products (and search?).
Well, it seems that You.com had a similar plan, and simply rolled theirs out first.
On December 23rd, You.com announced the integration of a ChatGPT-like system into their own search results.
I had a chance to chat with Saahil Jain of You.com about the project, which you can watch below.
The big takeaway? The differentiator between ChatGPT and YouChat is the access YouChat has to realtime data.
This is pretty clearly illustrated with the query:
Prompt: what is the status of the war in the ukraine
Big credit to You.com for being the first to make this leap.
Here's my interview with Saahil Jain, where we discuss what YouChat is, the challenges of producing trustable results, and more.
Transcript (from Whisper)
Dave Davies - Hello, everyone. This is Dave Davies from Weights & Biases. I am joined by Saahil Jain from You.com. Saahil, thank you so much for jumping on this very quick and important chat. I saw the announcement from You.com, I believe from December 23rd and got super excited to talk to you as soon as I saw it. There's a lot of hype. I'm just going to dive right in. There's a lot of hype around ChatGPT. Everybody's been talking about it. Is Microsoft going to implement it? And then I saw that You.com had already integrated a ChatGPT-like system into its engine. And then, of course, I had to use it and try it out, comparing it to ChatGPT, for example. So let's start at the beginning. This came right after the ChatGPT hype. So was this something that you guys were working on already, before all this hype? How long have you been working on this project?
Saahil Jain - First of all, thanks for having us. We really appreciate it. And thanks for using YouChat and You.com in general. We have definitely enjoyed building this. And that's a great question. We have been actively integrating generative AI for around a year now, starting with YouWrite as of March last year. Generative AI has been on our radar for a while. And I think a lot of the ideas and things we had built along those lines were a precursor to YouChat.
As a search engine, we've always strived to be at the forefront of integrating generative AI into our search. We had it on our minds for a while and when the time came, we rapidly incorporated it into our search engine.
Dave Davies - Something I have to ask, and I'm going to give you a chance to promote yourself because there's no other way to word this. The question that I think needs to be asked is, why would I use YouChat and not just go straight to ChatGPT? What's the difference between these two things?
Saahil Jain - That's a great question. ChatGPT came out and has changed a lot of things, raising a lot of questions about search engines. But in general, there were a couple of reasons that we considered when we developed YouChat and that are influencing its further development. The first reason is that users can access the origins of the results of the ChatGPT-like system in the context of their search. We want to allow our users to get the best of the advances in generative AI, as well as the advances in search and information retrieval. Users can quickly benefit by viewing the web links and You.com apps to verify the information and augment the generated content they see from their chat bot and chat-like interface. ChatGPT-like systems in the context of search have unique benefits.
The second reason is more grounding. Based on web results, we can ground our system a little bit more, we can stay more up to date, we can do things like providing citations. So if you ask a question, such as who won the World Cup, we'd be able to take the most recent information into account. Whereas with the ChatGPT, you would be limited by the amount of data that it has been trained on.
And lastly, we work so that anybody can easily access it. We've worked a lot on our UI and accessibility across countries, especially for places where chat functions might be unavailable..
Dave Davies - Right. Well, that's very interesting. And you bring up one of the more interesting points - YouChat has access to the latest information, right? As opposed to when you're using ChatGPT, which is based on the last bit of information it was trained on.
I don't know how you're doing it, but I'm going to have you on a show where we'll dive into how you're pulling that off and what's involved in computing and stuff like that.
But right now, one of the questions that I want to make sure that I dive into, because I think it's a question that a lot of people have in their minds when it comes to generative AI like this. You guys covered this on your site. When I look at YouChat, it has this disclaimer, that states that you are not liable for the content generated.
Good, everybody should understand that this is the AI building this content, it’s not you, and it's a good model. But one of the questions that occurs to me is that right now, whether it's Youchat, ChatGPT, or whatever else you might be using, it can't be trusted 100%. For example, if you have a health problem, you shouldn't enter in your condition and expect the AI to diagnose you and then trust whatever it generates.
Do you think YouChat will ever get rid of that disclaimer and if so, when do you think that will happen? Will there come a time when we'll be able to use systems to access the world of information, for example where you understand that it’s the Mayo Clinic versus some affiliate site, where it’s apparent where the source of this is being generated from? Will this time ever come, and how far are we away from it? What are you guys doing to help bridge that gap and get us to a point where we have that kind of reliability, where we don't need a big disclaimer on it?
Saahil Jain - Yeah, that's a great question. It's also a very tough question.
It depends on a lot of advances and improvements that we're making along the way. I know that this is an annoying answer, but it probably depends on the context and the use cases, as well as the user expectations of the tool.
So I think in some areas, we're already there, but in some areas, we're still very far off. So in general, I guess when you talk about a disclaimer, I think maybe I'll just start off with thinking about the different use cases.
So if we think about three different groups of use cases, we can see that the need for a disclaimer might be different. I think in one use case, people often use it for creative outputs. This would be in the case of developing images, writing poetry and brainstorming ideas. In these areas, I think we're already in a state where you may not need disclaimers.
I think it's important to consider the use case.
The second group of use cases would be low to medium stakes information gathering. In these cases, we currently do need to show a disclaimer. And there are probably going to be a couple of advances necessary here.
But I think the main point here is that because we're a search engine, we have a retrieval-based approach. We're more likely to be able to reuse the possibility of users having to need the disclaimer by grounding our chat in live information, as well as information that we've retrieved.
And then I think the last group might be taking high stakes action. For example, providing medical advice or legal advice, my sense is that disclaimers will continue to be very important for the foreseeable future in that domain. Another dimension to think about this question might be around user expectations.
So disclaimers are necessary when user expectations aren't clear. But I think what's going to happen is that users are going to understand how to leverage generative AI better in this chat-bot like interface, which will become one of their habits.
And as that happens, I think it'll become clearer and a disclaimer will be less necessary, even if the tool remains imperfect.
Dave Davies - Right. You'll be understanding that it's the same as if you were asking me, right?
Like you have to understand that I don't know everything. And so if you're asking me, if you're asking me for health advice, you have to understand that I'm a marketer and this is the level of information you're about to get.
Whereas yeah, the generative modeling would be based on everything that You.com is collecting, right? So it depends on how much you understand You.com, especially when asking it for health or finance advice.
But for fun and creativity, it’s okay for it to go off the rails. We've all played with DALL-E or DALL-E mini or whatever, and just let it generate the worst kind of nightmare fuel, but we understand that it doesn't matter at the end of the day.
So it must be a very, very big challenge for You.com.
So you know what? Well, I'll say what is the approach you take to that? I'm going to grab you on one of these episodes and we're going to really dive into that sort of area of things.
But before we go, one of the questions I really want to know the answer to, what's You.com working on now?
Like you just announced that you clearly have a road map. Like, here's the next thing for what you can tell us. Obviously you're not going to put all your cards on the table.
But is there something that we should be keeping our eyes open for? For example, what you're doing to maybe push the boundaries of AI at ML in search? You guys beat a lot of companies to the punch here on this one. And you've hit the ground running because you have the advantage of being the first or, at the very least, way ahead of a lot of people.
What is the next thing that we should be on the lookout for over the coming months?
Saahil Jain - That's a great question.
And yes, I think there's probably a couple of different kinds of dimensions of work that we're doing. One dimension is based on our values of trust, facts and kindness. We're going to continue working on improving the trustworthiness of YouChat results. So there's definitely a lot of research we have to do there.
As you mentioned, the topic of the scammers, for now, we have a disclaimer, and we want to promote very responsible use of YouChat.
We're going to continue looking into ways of grounding it, improving the results, citing sources, etc, so that users can really trust YouChat in ways that they have not been able to trust chatbots in the past. And then another thing we'll broadly do is really think about how to more cohesively integrate YouChat into search. So on You.com right now, we have a rich interface, with a lot of these concepts of You.com apps.
We're definitely going to think about how to integrate these two types of things and how to push the boundaries of what a chatbot can do. I can't go into too much detail, but definitely stay tuned.
I would just say to keep watching YouChat. We have a lot of people on our team working on it. And it's a really collective effort. So just keep an eye out, stay tuned. And yes, we're going to have a lot of exciting advancements coming out soon.
Dave Davies - Okay. Well, and then that leads just so wonderfully to the last question that I have for you here today, which is, as you know, I'm launching a podcast on ML and search, I would love to have you on again in a couple of months and to dive into some of the stuff that you learned from having YouChat running for a couple months at that point, as well as some updates on what you guys are doing. So would it be okay for me to drag you on again and dive a little further into how this thing is progressing?
Saahil Jain - Yes, of course. Thanks for having us on. Yes, we really appreciate it.
Dave Davies - Wonderful. Well, thank you so much. And I'll talk to you again in a couple months then.
Saahil Jain - Okay. Thanks, Dave.
Add a comment
Tags: ML News
Iterate on AI agents and models faster. Try Weights & Biases today.