Name: Brainstorm Tech 2023: In Search Of... A Better Search
Uploaded: 2023-07-11
Duration: 15 min 28 s
Description: Jordi Ribas, Corporate Vice President, Search and A.I., Microsoft In conversation with: Jeremy Kahn, FORTUNE

Jordi Ribas, Corporate Vice President, Search and A.I., Microsoft In conversation with: Jeremy Kahn, FORTUNE 

 Brainstorm Tech 2023: In Search Of... A Better Search 

Conferences

Brainstorm Tech 2023: In Search Of... A Better Search 

Welcome back. Please take your seats and get ready for an exciting segment ahead. Fun fact. Did you know the name NVIDIA is derived from the Latin word NVIDIA which means envy. It was chosen to symbolize the company's ambition to be the envy of the computer graphics industry and boy have they achieved it? Yes, NVIDIA stock is skyrocketing faster than you can say artificial intelligence and the company has now moved from just making the chips that power, the generative ai boom to competing with its own customers in many cases, including the company that our next guest works for when it comes to A I models and cloud services, fasten your seatbelts folks. This is going to be one wild race and the competition is only getting fiercer. Remember when just a few months ago Microsoft unleashed their A I powered search engine. They learned some tough lessons about the power and unpredictability of generative AI. But now Microsoft has set their sights on transforming bing from a product into a full fledged platform, redefining the future of Search, future of Search. However, with this reinvention comes to many questions. Our next guest, Joie Ribot has devoted 15 years, the development of AI products, mastering search, speech, vision and data output. This pioneering work has paved the way for a new era of A I powered search. Please give a warm welcome to Microsoft's Head of Search and Ai Jordie Rebus is me Jodi. Thanks so much for being with us, for those who are not familiar with the new, the new Bing and what you've done with, with Bing and the integration of generative A I into the platform. Can you walk us through what's different about it? Sure. In fact, as you know, web search has been with us for a very long time over 20 years. And if you had asked everybody in the audience six months ago, hey, do we need anything new in search? Most people would have said no, you know, if you search it work just fine. But if you actually looked at how people use search, the vast vast majority of queries were only two to three words in length in a way, we had adapted ourselves to search, using these keywords as opposed to use search in a natural way. And so what we were able to do by integrating the GPT model from opening eye into B, we were able to answer these very long complex queries that in the past or even in any traditional search engine, you'd not be able by any means to get any response that made any sense basically. And in fact, I have an example if you don't mind that I Yeah, no, let's take a look and let's walk through how being handled query is. Now, this is an example. This is a long query. You can see it here at the top. In fact, maybe we'll have to play it again. But this is imagine that you want to go to Barcelona, which is where I'm from originally and you want to spend five days and you've got a set of constraints, meaning you've got a budget, you've got a few days that you want to spend on the beach. You want to go and want to go to a mountain range nearby called Monserrat. That's very beautiful. Coincidental is where I got married and you know, you really want to come up with an itin and a good plan. And so if you play this again, I don't know if they can do it. If you type a query that's long natural. That basically describes what I said in a traditional search engine. You will get a bunch of links, vast majority and you can go ahead and tie it to live with. Your favorite place, are not going to be relevant at all. But if you do it with the new bank that incorporates the GPT technology, what we do, we reason over the data of the web and we come up with an answer that's really structured and it can really answer this very long complex wordy that again, no search engine was able to answer before you guys at Microsoft came out with this gene of AI search product. First, you got this stuff into Bing. And I think the hope was that this was, you know, people are going to gravitate to this new search and you were going to grain tons of market share. But I've been sort of watching some of the stats on this and, you know, Bing had about 3% market share before this compared to Google with about 93%. And if you look at some of the numbers today, at least the ones that are publicly available from similar web or stat counter, Bing still has about 3% market share and Google is still up there at 93%. So you know what's going on here, how come you're not gaining more share if this is so much better in experience? Well, this is a very tough space to be in. Obviously, Google is very dominant in the space, but we've actually grown faster this last year than we've ever grown before. And we're really, really pleased now, obviously, it's a long road and market share is an important metric and we obviously pay attention to it. But I think what's more important to us and the reason we're working in B or in search in general is that we feel it's such an important technology for society and it had kind of stagnated in the innovation. And one of the things that gets me out of that every morning is being able to bring competition into a space that there hasn't been a lot of competition. And what we were able to show six months ago is that search will never be the same from now on. And you saw Google Reacting right with Bart or they survey gene exactly like my CEO said, no, we're making them dance. And at the end of the day, that's what I tell the team. Look, this is what it's all about. We're bringing competition and innovation in a space that's ultimately going to benefit society. We're growing faster than we've ever grown before. And we're very happy with that. But again, ultimately, what's most important is to do this contribution? Absolutely. I want the audience to please think of, think of questions for Jodi and we'll come to you for those in a moment. So please put on your thinking caps. But in the meantime, I want to ask Jordi. Yeah, one of the, one of the issues here I think people have with a gene of AI search product is there's less room for advertising, there's potentially less room to link out to outside sites. And this has a lot of people very nervous both about how do you monetize this, but also for everyone else, including publishers who are dependent on, you know, search results to drive traffic to their sites. How are they going to sort of make money from this ecosystem and how is this ecosystem ultimately going to reward high quality information? Now, that's a very important question. And I know a lot of publishers are asking us about this and we actually had a Publisher summit not too long ago to primarily address questions like this. And it's a long complex answer. But let me try to simplify it a little bit. Basically, one thing that we're noticing in the new being, I don't know if you realize that it's actually an interactive experience where you go from the search mode to a chat mode. In a very simple interface, you can swipe up and swipe down. Sometimes we'll have that. It's like the, you know, when you are in these dating apps, you swipe left and right. So in search you go up and down but you go from the normal search mode with the links, the regular way that advertisers advertise to this chat mode. And what we do if you look at chat, we actually have links that are some to advertising sites sometimes are actually links to publishers. And as we look at the overall traffic that we're sending to publishers, it's actually increased with the new bank because people are engaging more, they are spending more time rate is bigger with the click rate is bigger overall in terms of the total, the total traffic that we send. And so obviously it's early and we're working with a lot of publishers to come up with even better approaches. They're going to see us experiment with different new access that will have little cards, for example, that will attract the user attention more. But definitely for us to really be successful, we need the publisher and the advertising industry to be successful. That's how the ecosystem works. Let's kind of questions from the audience who has questions for Jodi that's looking, I don't see, we've got some mic handlers. If I don't see a question, I'll go to my own next one while you're thinking of questions, Jordi, one of the big issues with, with general over is the issue of grounding. How do you make sure that what it's saying is accurate and of course, with a search product, you really want to make sure the results are accurate and corroborated to outside information. But, you know, I think anyone who's experienced, you know, working, you know, playing with the chatbot has found that they sometimes invent information or sometimes say something that's not accurate. How have you tried to deal with this in the new Bing search? Yes, this is something that is maybe my number one worry even today, right? Because at the end of the day, if you build a search engine, you need to be responsible on providing data, that results basically that are comprehensive, that are accurate. And when I saw GPT four for the first time, that was last September, I realized we had something special. The model was more powerful than anything we had ever seen before. But it had two big problems. One is what you just said, it had a tendency to hallucinate basically with made up stuff, right? And for a search engine, you cannot do that, right. The other is that it was string with data through September 2021. So if you had a question on any event or anything that happened after that, it was not able to actually answer the question. And so then we developed this technology that we called founding, which people don't talk enough about. But anybody that works with an LM, a large language model realizes that it's fundamental. And what it does in our case, we, we basically combine the search results themselves for the query with the LM with GPT four. In this case, you can think of it as instead of relying on the data that the LM restrain, what we do, we go to our fresh index to collect the data and then we use the LM GPT model to reason on top of the and that's what provides the result. And that's why if you look at our results today, they are more accurate than anybody else's out there at this point. And it's basically because of this grounding technique that we're still perfecting, it's not perfect by any means, but we've reduced hallucinations by over 90% right now. And what about the issue you had initially with B Chat where you could sort of drive it off its guard rails. And I think you guys had to limit it. You know, the number of conversational turns you could have with the chat bot in order to prevent that from happening. What's, what's the current status there? Well, yeah, this happened, I think it was the first week that we launched. Obviously, we launched in preview and there was a report specifically from the New York Times that spent hours trying to break the model and he did and the model was starting to do crazy things to leave his wife. Yeah, I was asking him to leave his wife saying that, that, you know, he was saying that his name was Sydney was not B CHAT. And ultimately, what happened is after all those attacks, even though it was not portrayed that way. But those are what we call adversarial attacks, something that you also can refer as jail break. And there are a lot of these different types of attacks. And so one of them went through and then the model basically lost its its context and started getting out of hand. And so what we did right away, we limited the number of turn and we've since then been improving the rare and now essentially we've increased the terms to where 99 point plus percent of the people are happy. Some people are still asking us to do more. But definitely this was one of the can you do now? Now we can do 300 marks per day and 30 per session, 30 per session. OK. Great questions for Jodi who has a, who has a question again? If I don't see any, I'll just carry on. But I want to give the audience a chance to ask. See I've been so quick. I know I'm going to go on to another question then myself, you know, this is obviously all built off of open A I technology. You have a very close partnership with them. But there's been some reports of tension in that relationship, you know, they have their own product in the marketplace. Chat GPT, which does some of the same things there is now given that product access to bing to do search. So you can you talk about how that relationship is working. You are competitive in some ways and yet you're partners. Yeah, we're primarily partners, right? But you're right, we're also competitors in a way and, and any time that you have what we call a coal peat relationship, you know, there will be some tensions but there are tensions within my team, you know, and so so I think that some of what was reported, I think was, was the journal was exaggerated. I think fundamentally what we do. You notice how I talked about Opening eyes, Gpt today. So we bring them into our presidents they bring us into their presidents. I mean, I think we really collaborate whenever we can, we are partnering very, very closely. And I think in general when they do well, we do well, because all the GPT models are running on Azure and when we do, well, we also, you know, a lot of the grounding expertise you notice now that they are starting to do grounding themselves using them. Right. And so at the end of the day, I think it's a close partnership and like every partnership, there will be some ups and downs ask you sort of where this goes next. There's a lot of speculation that the next step is to give things these things agency essentially to allow them to act as digital assistants for us across the internet. So you couldn't, you wouldn't just sort of plan your trip to Barcelona. It could actually make all your reservations for you. Is that what's coming you think? And how quickly are we going to get there? I think. So, I think what we've seen so far and if you haven't tied, you know, the new bank, please go or by any of the other generative models. And you'll see that you can ask, especially in being chat, very complex questions and you'll get answers that you never thought you could get before. But where we're going next is going to be even more exciting because we're going to have all these plug ins. So we'll be able to work with all these partners and we already have, you know, dozens of partners that want to build them. And so you'll be able to, you reserve a restaurant, book tickets, do all the things that now takes multiple steps to do. You'll be able to do it in a more natural way and the chatbot basically is going to be able to do it for you. It sounds very exciting. We're out of time, but I want to thank Jordy for being here with us. It's been fantastic. Give him a round of applause, please. Thank you. Thank you so much.