Name: Brainstorm AI 2023: Future of Google Bard and Generative AI
Uploaded: 2023-12-12
Duration: 19 min 6 s
Description: Sissie Hsiao, Vice President and General Manager, Google Assistant and Bard, Google Moderator: Jeremy Kahn, Senior Writer, FORTUNE; Co-chair, Fortune Brainstorm AI

Sissie Hsiao, Vice President and General Manager, Google Assistant and Bard, Google Moderator: Jeremy Kahn, Senior Writer, FORTUNE; Co-chair, Fortune Brainstorm AI

Brainstorm AI 2023: Future of Google Bard and Generative AI

Conferences

Hello. Hi, everybody. Welcome. Welcome to the 2023 Fortune Brainstorm A I conference. Great to see you all back here again in San Francisco. Um This is the third year of Fortune's brainstorm A I conference. Yes, we did get started before chat. GP T long before. Uh And many of you have been here with us from the beginning, artificial intelligence as we all know is not a new concept. In fact, it's been around since at least 1956 when Dartmouth hosted a summer research project and conference. And the concept was not all that different from ours here today to gather some of the greatest minds in the field and come to uh more or less of an agreement may be on a shared vision that computers can be made to perform intelligent tasks. I think we've established at least that much. And now with the explosion, of course, of open A is chat G BT and other technologies. Um Last November generative A I has garnered the attention of regulators as well, the general public and businesses alike with A I's newfound attention and capabilities comes a wave of both fear and excitement. On how it will be deployed in the coming years and what effect it's going to have on our society as a whole. A I as well as on its way, revolutionizing fields like health care, entertainment, manufacturing, education and a whole lot more. But with power comes great responsibility. A I has the potential to perpetuate, perpetuate biases and in the hands of bad actors spread misinformation and open the door to major security risks. Many fear it's going to replace jobs. Some have even argued that A I could even replace CEO S and journalists. Oh, no. Um With so much unknown and unregulated, we look forward to insightful conversations over these next two days as we examine new business cases for A I and explore how it's already changing the way we live work and do just about everything. We have a stellar lineup of speakers. But remember that this event was really created for your participation and is meant to be an interactive discussion among speakers and attendees. The most important conversations are those among all of you. Um As we navigate the innovations and the challenges of implementing new technologies and new ways of collaborating. Before we start some important Acknowledgments. I want to recognize the incredible work of our brainstorm A I, co-chair fortune, senior writer Jeremy Kahn Fortune, senior editor Vern Kito and co-founder and Ceo of Afecta. Uh Doctor Rana LK Calli. Thank you to our co-chair. We are also pleased to be joined by additional Fortune colleagues who are going to be moderating some of our discussions, Alan Murray, CEO of Fortune Ellie Austin, Fortune's deputy editorial director for live media and Jeff John Roberts, editor of Fortune crypto. Uh Before we get started, I also want to extend a tremendous thank you to our partners for their contributions to this event. Our presenting partner, Accenture, our partners, a ARP Amgen checkpoint hand show Sales Force Workday and our supporting partner Felicia. Before we get into our sessions, an important reminder about our conversations, you probably get sick of us saying this, but it's important. What's said on the main stage or during the strategy sessions uh is on the record unless otherwise noted. And what's said over a break or over dinner or a walk is of course private unless agreed upon otherwise by all parties. OK. Let's get started. As mentioned earlier. The release of open A is chat GP T has brought worldwide attention to the uses of generative A I chat bots and the ability to release large language models to the public. A few months later, Google released its conversational generative A I chatbot known as Bard. Since then, Google merged the Bard teams with the Google assistant team to create assistant Wood Bard, a new personal assistant chatbot powered by generative A I. Google also unveiled a generative A I search experiment that may alter how we all find information and possibly disrupt digital business models across industries. Last Thursday, Google announced the release of their newest model Gemini, which you all probably know all about. Uh And that is now going to Power Bard here to talk about the future of Gemini powered Bard. What's next for the chatbot? Please welcome General manager for Google assistant and Bard Sissy show. She's going to be interviewed by Fortune, senior writer and brainstorm A I co-chair Jeremy Kahn Sissy, thanks for joining us and thank you all for joining us. Um We're going to go to questions from the audience uh hopefully towards the end of the session. So please start thinking of your questions. Um Sissy Big News last week with the unveiling of Gemini and the announcement that Gemini was now going to be powering barred. Um There was also a lot of uh disappointment, I think over the video of the, the demonstration that came out with uh that announcement um in, in Heinz, do you regret sort of how that that video was handled? Well, maybe let me start by saying thanks for having me, Jeremy and thanks for having me for start there. Um Very excited to be here to talk about generative A I and, and bart, of course, I mean, let me start with the last part of your question. The video is completely real. All the prompts and the model responses are real. We did shorten parts for brevity which we put in the video as you know, informational on the making of the video. But we're also looking forward to releasing the API in A I studio this week so that developers can play with it. But you know, that's really addressing the video. I think the main, you know, headline story is we've entered the Gemini era. This is the best set of base models in the sort of medium size range and the large size range on a variety of industry benchmarks that all researchers benchmark their large language models off of. And so I've been really excited to take that expression of that power in those Gemini models and put them in Bard, which we saw with our side by side blind evaluations, right? So Bard is the best most preferred free chat bot now um in the market in blind evaluations with third party Raiders. Great. And what has been the most interesting use cases you've seen so far with Bard or, or things that have really surprised you that people are doing with it? I think what surprises me is the just the breath of things that people do uh with the chat bots. It's really hard to say it does just one thing or the other thing because it does so many things. So I'll give you an example um you know, for my personal use um at my home life, like my daughter just got her ears pierced. And so I was looking for a particular type of earring that she was looking for that wouldn't cause allergies. I brainstormed with Bart about that. Uh Bard gave me some brands. Bart also told me how to take care of your ears to not get an infection. That's a very like personal use case that was able to, you know, take that problem and, and solve it for me. But um at work in productivity cases, we have people using it for, we have, I have a coworker who's uh English is a second language uh for him. And so he uses Bard to rewrite his emails to be more professional. So really these a is have just such a broad, you know, use set of uses because they can role play, they can generate content, they can brainstorm with people. So I'm really excited to see just like the breath of uh things that people are trying and also, you know, what really resonates with users in the world. And when we look ahead to Gemini, what really excites you about what you're going to be able to do with Bard now powered by Gemini that maybe you couldn't do six months ago. Yeah. Well, I mean, first of all, Gemini is just a stronger base model which means like everything that comes out of it is more high quality. So we see it in text generation, we see it in code generation, we see it in instruction following. So uh typically complex prompts have many, many instructions embedded and the model is able to follow those instructions. Um But I think the most like sort of visually stunning is of course multimodality, right? This is where the model can ingest, you know, not just text but image and audio and video and also communicate back out to the user um in not only text but also, you know, in other sort of means of communication like video and image and, and audio. Um a very simple example. Um I just used bar to uh I went to a dinner, I had a menu that was like, choose a course from the 1st, 2nd, 3rd course. And then I had a wine list and I took a picture of the two menus side by side and I said, choose me a dinner for tonight, one course from each section with a wine pairing to go with it. And it actually did fabulously. And so like even something like that as simple as just taking a picture of something and asking a question or for the bot to do something for you um will just become second nature as like these multimodal models uh become, you know, as easy to access as your phone. Yeah, interesting. Um And a lot of people are looking forward today when, when these models, um you know, you can prompt them to go out and do something for you to take an action on the internet on your behalf. Uh How far away is that? Kind of agency do you think and will Bard be for Google the thing that, that carries out that your mission in the world for you? Yeah, I mean uh getting things done for people is definitely within reach of these uh agents or we call them agents because they have this agent of capability. So already Bard can summarize your emails, right? We've connected Bard to gmail, youtube. I can summarize youtube videos for you. You can ask questions. So already like this action ability is starting to emerge when you connect different tools that exist in the world into the large language model and foundation. The reason why it's able to do this is you're teaching the model to speak another language, which is essentially code, right? When it can seek API S, it can actually translate the user requests and convert it into an a a set of API calls with reasoning on top. So I'll give you an example. Um One thing I want Bard to be able to do for me personally is uh book my entire kids summer camps. Yeah, all the, all the parents are happy. OK? I have two kids, they have different interests. I need to find camps every week of the summer that are within driving distance and actually I can pick up and drop them off, right? And so this is really hard, right? So you have to think about all the reasoning that goes into that. And so I've gotten barred to the point where it's giving me three options per child per week. Um But wouldn't it be awesome if I could just say, please now book this. We're almost there. How, how far away do we think? Mm. Things move really fast in this industry. I can't say exactly when, but I think using tools and not only using tools but the logical reasoning on top is very powerful and we're definitely working on it. One of the things that people have been concerned about with all of these LM powered chat interface has been hallucination. And I remember being here last year, maybe some of you were also here last year. And there were definitely people saying, don't worry, hallucination is going to be solved within nine months. It's going to come back next year, it'll be solved, it's not solved really. But how are we doing on making progress on hallucination? And do you think that this is a solvable problem? I uh look, so hallucinations are inherent to all large language models. Now, we even know this word in the lingo of this industry, we all are aware of this. And before we didn't, you know, use this word very much, um It's two pronged approach. One is, can you actually just improve the issue, right? Can you make the bot hallucinate? Yes, uh uh uh hallucinate less the answer is absolutely yes. And that takes hard work, fine tuning evaluation. And we've seen the hallucination rate um decrease in bard over the last eight months. We've been working on it really hard and definitely improving. And I would also say the smarter the base model, the more it tends to understand how to use search as a tool to really reduce its hallucination. So that's like sort of one step. The other step is we're, we're first to market to launch something we call double check. Um in the industry, it's called an entailment model. But really what it is is checking all the sentences that come out of the bot with search. And if you think about it, if you're writing an essay or doing something important as a human, you're going to be doing a lot of searching to corroborate your facts, right? You're going to be reading all of the different maybe dissenting views and really, you know, deciding for yourself, what is the right way to write any particular piece of content? So we put double check into Bar to help people really quickly get access to like search corroboration on top of the model's output. And that's yet another way that if you're, you know, doing something like really, really important that it must be factual you can use double check is another tool to help you um really understand what Bard is saying? Interesting. And do you think that uh Bard and search, which right now are very distinct products? Will they remain distinct or are we going to see increasing kind of merger between the two. You know, I think of um you know, search and ste which is the application of large language models inside search and Bard is really complementary. Um A I is a, a raw technology that has application in lots of products, not just search and a stand alone chatbots, but I'm sure many of you in the room are thinking about how can this transform like my business operations and the way we do work. And, and so I think that um the way I think about it very simply is you go to search because you mostly are wanting information, right? And I think she does a fantastic job sort of summarizing what uh the web is really saying about your question. Whereas Bard is like you are directly tunneled into the A I, right? And it has this full amplitude of expression, you can have it write poems, you can have it to critique your code. You can, you know brainstorm ideas for a birthday party with the bot, you can have a conversation just for fun. And so it's just kind of a very different expression um of what underlying is, is basically a similar uh base model technology, right? I wanna try to get questions. Uh I want to have a question from the audience is uh if we can get a mike to the gentleman down front uh here at the first table. Um please. And if you could please state your name and where you're from? Hi, uh I'm Edward Hall. Uh I teach A I at NYU and I run in TABS and I'm uh behind the big parts of data commons project. Um So, uh question is uh how does this transform Google's business model? Opening eyes, business model is charge subscriptions to consumers. And so, um while of course, the internet is still advertising, um how does consumer get clean information out of this? And what's the transformation of the model business model? You know, um really, we're really focused and barred on making the A I very, very helpful for people. You know, monetization is something that, you know, we're, you know, we're always looking at, but it's not really our focus right now. It's really about like, how do you make this technology really great and help people with their everyday tasks and other questions. Um Who else has a question? Raise your hand. Oh, the gentleman here at this, there we go, wait till the mic gets in. There you go. Yeah, hi. This is Sar I go on with fresh works. And um you know, one of my question is around security for the models because people are doing prompt hijacking uh and being able to understand what's coming out. So is there going to be the need for firewalls for these models? Well, I mean, security has like different facets, right? One is, for example, if you have um specific data that you're grounding the model against, you don't wanna leak that data, right? So you need to make sure that any um uh data that's, you know, used to flavor your model in any particular way is protected. There's also and uh there's also ways, for example, when you use tools to make sure that you know, the the chat bot is operating tools in a sandbox, right? So there's different like layers of security that we add to uh avoid prompt hijacking, avoid leaking grounding data or training data or um make sure that like its execution is safe. So at in Bard, we do a lot of red teaming, we have adversarial testing to make sure that like we're constantly, you know, looking at the model and making sure that it has different layers of protection around all these potential exploits that um might occur. Next question just here, stand up on a table cloth. I am Michael, the American Institute of Physics Publishing. And um so I, I wanna go back to what you were saying about the distinction between search and the use of an A I because fundamentally we searched because we had to break down what we wanted to do into these obscure pieces to find something. But if we could have just asked, we never would have searched, we would have just gone right to, you know, it's almost, it's kind of a, it's a choppy way and I guess I'm thinking back about the gentleman's question about business model. Can you see a time when we don't search, we ask? And then how do you get ad revenue from that? Um, like I said earlier, I think, you know, there's the two parts of the question one is as the A I gets smarter. Can it take tasks off your plate? I think the our endeavor is to make that. Yes. Right. Because we think that a is have such incredible aptitude like solve my summer camp problem, right. That's a very complicated problem, right? With a lot of steps to that. But um in terms of, you know, is it ads is other forms of monetization, it's just very early and it's not something that like I could say, you think we're gonna see, are we gonna see with Bard a kind of store? I mean, uh open eyes, talked about trying to allow people to create their own models and maybe have some kind of a, a kind of app store for a model store. Are, are you thinking about that for Bard? You know, um I think it's great that there's so much innovation, right? Like customizing uh uh an LLM into a particular type of behavior is something that I think some people wanna do, especially power users. So I'm interested in how people wanna use that, you know, right now we're really focused on making Bard really, really helpful. Um And why I have a distinction is because Bard can play act many different flavors of behavior, right? Because you can just prompt it into being an interior designer or a statistician and you can actually prompt your way to many different uh flavors of it if you will. So I think um you know, whether you need to really split that bot into many little sub bots or whether the main bot is really helpful at many things like these are just different models of expressing the, the core technology. And I'm very interested in how this, you know, will, will work for people. Well, sissy, we're out of time. Thank you so much for joining us. Uh And uh thank you all for listening. All right. Thank you so much.