Creating the Future of AI: How Salesforce Research Advances AI for CRM | Salesforce

Aloha everybody. Thank you so, so much for coming. I am very excited to be here today and talk to you about the future of AI that we're creating together. And the thing we're looking forward the most is probably not this forward looking statement, but I still have to say it. This is the most forward looking of all the keynotes. So please do not make any purchasing decisions based on the things you'll see here and only based on products that are actually available. All right, now to the fun parts. AI is currently in what I call a dual state. And you can roughly categorize those two states as the electricity and the Sci Fi state. And the electricity state, AI is basically ready to have a huge impact on a variety of different applications. In agriculture, you can already use AI to weed out just the things around a lettuce head, for instance. In commerce and in the supermarket, we can already use computer vision to count different kinds of objects you see in the shelf and identify them and make sure they're placed in the right order. For driving, we already have driving assist features that help you stay in the lane. And in medicine we can classify specific types of breast cancer and histopathology images or brain bleeds and head CT scans. But in each of these applications, that is just the first one where we applied AI like electricity was applied to all the different industries out there. What is also in the other state, though, the sci Fi state, is that we still can have full general conversations with an AI. We don't have fully self driving cars that can deal with all kinds of complex situations and don't need a steering wheel anymore. We don't have fully autonomous tractors giving us unlimited free food supply. And we don't will not for a long time be able to do all the different types of things that a doctor does. And so that is exactly what we at Salesforce are also focused on. We're focused on trying to predict this future and the best way to predict this, to actually create. So in our group we actually now have four different subgroups. And this is, I think is an interesting changed to where we were a couple of years ago when we really were just on the fundamental science part. Now we still do fundamental science because that second Sci Fi bucket, we can work on making that more and more real, but we also work on applied research, things that we know we can do, but we can do them better and we can do them in a CRM context. For instance, machine translation is something we can do pretty well. You can use Google Translate, but you might want to translate Salesforce documentation and not accidentally translate Salesforce as the force of the sale into Japanese or German. And so we work on applied research and then we connect that applied research to a new product incubation, really excited. And you'll see some of the things that we had talked about last year as almost in the sci fi bucket, things that we're working on in incubation and are actually going open beta now and then. We've also seen a lot of different AI problems that actually boil down to the same kinds of algorithms. For instance, text classification. You might want to classify a tweet and know whether it's a positive or negative sentiment and whether you know a marketing campaign has changed how people perceive your product or your brand. But the same underlying algorithm can be used to do case classification and route emails into different places in your company. An email comes in, you might automatically assign it to sales or service teams. So we see these different problems across the company, have built a platform and now we have this virtuous cycle where in several cases the fundamental science influences what kinds of products we think will be possible. But sometimes also we have applied research and a platform realizing new problems that there hasn't been any research about. And so we've had a tremendous amount of patents and papers that we publish and share, we open source code. And the research community has really given us a warm welcome now that we're.

A part of it.

So for today's keynote, our main headline is the end of Data entry and the beginning of Data conversations. And you will see what we mean with that. Throughout this talk, we will actually talk about ocr, something that you think is kind of solved, but it's kind of solved in a way that wasn't until this research quite ready for you to really use in complex situations. And we'll actually make it real. And the same with voice. Voice is finally here. And voice being able to naturally interact with your CRM is one of the reasons that the most annoying part of any CRM and system of records really in general is data entry. And with voice, that can be much, much more seamless. But once you have all the data and you're able to get it into the CRM very easily, then you want to have these data conversations. And so we'll show you how you yourself don't have to know how to program databases anymore. Instead we have researchers like Victoria Lin, who worked on translating natural language English questions into database or SQL queries for you. And that is one of the data conversations. Sam Ming, one of Our senior directors of research will talk about an actual fully conversational AI using all that's available right now. And that will be a really exciting live demo of that capability. And we'll have some incredible research done by NITISH on generating whole documents, long sequences of text, which was impossible just a couple of years ago. But, but his model is one of the largest or the largest language model in the world that is publicly released. And as we're exploring and realizing, wow, the end of data entry and the beginning of these data conversations, we're going to have more and more of these conversations together with an AI and we were going to rely on it more and more. We want to make sure to know and have you know that we work on trusted AI. Also, we're thinking a lot about ethics and we'll talk about some of that throughout this keynote too. In the last chapter, Nazneen will talk about how we actually try to get these systems to be explainable so that as we rely on them more and more, we can really trust them. And what's interesting about this keynote is this is on the more technical side and all the people that are coming on stage are the actual engineers and researchers who, who built and programmed these systems. So these are not your common and wonderful marketing folks that you're used to seeing on stage. So hopefully we'll give them a round of applause when they come in and give them the benefit of the doubt. Usually they're hard at work programming every day, but I'm really excited to have them talk about their work. And so that I want to bring on stage Indira. She's our VP of Software engineering in Salesforce R and D Group and she will talk about making some of the things that we've developed in research real and bring them to you. Thanks everybody.

Thank you, Richard. It's great to be here today. Let's get started. Raise your hand if you just love manual data entry. Nobody, of course not nobody enjoys logging calls, creating reports and updating tedious records. You're professionals and you want to spend your day serving your customers well. One exciting thing about the field of AI is that it makes us all more productive. And Salesforce Research is committed to advancing AI that keeps you doing your most important work and helps you do the same for your customers. The end of data entry is the beginning of data conversations. Last Dreamforce, we announced Einstein Voice Assistant Enclosed Beta. This application showcases how how we can leverage state of the art voice and natural language understanding to intelligently automate data entry and your interactions with Salesforce CRM through a voice interface. We are very excited today to announce the open beta of Voice Assistant available to all all our customers. Please be sure to check out our voice tech in the Einstein booths and breakout sessions. One other example of using AI advancements to boost productivity is by bringing together OCR and NLP technologies to automate data and content uploads into CRM. Optical character recognition leverages advances made in computer vision technology to extract data from documents. Business card readers, table scans and uploads, identifying and extracting part and label information are some applications of ocr. Well, these are just some examples of how AI can be used to intelligently automate data entry into CRM. And now I'd like to welcome Victoria, a research scientist in our team, to tell us more about how AI can be used to extract data out of CRM. Thank you.

Thanks Indira. It's so great to be here. Well, how many of you are familiar with SQL query? Raise your hand please. I saw lots of hands. Thank you. So there are definitely lots of SQL experts among us today. For those of you who just raised your hand, great. This session is going to be really relevant to you. And for those of you who didn't, I have some good news for you in just a second. Lots of enterprise data exists in relational databases. Whether I'm a CEO who wants to hear the sales performance of flagship product, or a marketing officer who wants to learn the distribution of customers, it typically requires me to know a query language such as SQL using a specifically designed dashboard or having a team of analysts to run reports for me. But what if you could query your databases by simply asking a question using natural language? Let's imagine the following scenario together. In this scenario, I'm no longer retarded a scientist. I'm the CEO of nto, a startup company that offers sports products. As a small business owner and a trailblazer, I'm wearing multiple hats. I need to make sure the core business is going well as well. Understanding the market to gain important insights. And oh, latest delivery. Let's quickly import that shipping invoice. Yeah, I forgot to mention keeping track of shipment is also one of my hats. But fortunately this part doesn't take me much of the time. Thanks to Einstein Vision OCR here I just took my smartphone, open the Quip app and I take a picture, Make some small adjustments and I can let the algorithm do the rest. Let me take another picture. Yes, sometimes research. Okay, so make a small adjustment down spreadsheet. So here it scans the entire shipping table. Understanding the Table headers as well as the contents in the rows and columns. It does a couple of checks to make sure everything is correct and boom. In the end I got a spreadsheet. This is ready to be directly imported into my CRM and camera has become my new keyboard. Looks like we got a large section of hoverboards recently. They have become a huge hit. I want to understand more about this trendy product in terms of where we sell it, how much we sold it for and how many we have sold. So I came to my personal data scientist, Einstein, my great buddy at work, to learn what advice it has to offer. Hey Einstein.

Hi.

What is the cost of hoverboard?

$500.

Okay, $500. That's slightly above the market price. I wonder how did our sales go? Hey Einstein.

Hi.

Find me the customer orders of hoverboard and sort them by date.

There are 2,990 in total.

Okay, up and to the right. I wonder where all my hovering customers come from. Hey Einstein.

Hi.

Find me the customers who order the hoverboard and group them by location.

Sorry, I don't understand.

Let me try again.

Could you please try again?

Hey Einstein.

Hi.

Find me the customers who order the hoverboard and group them by location.

There are 2,990 in total.

Washington, Nevada, Indiana, New York and Indiana. Let's definitely get more people to ride hoverboards in California. So there you have it folks. With Einstein's Smart speaker, I can simply speak to my CRM and get insights that usually come from a highly skilled team. Isn't this awesome? Would it be even better if I tell you that all these are not just imagination? There is a real algorithm running in the back end of Einstein Smart speaker which interprets the natural language questions I asked together with the database schema in the CRM. This is the basis of some of my research. At Salesforce Research we train a type of neural network which interprets natural language question and translates them to SQL queries end to end with text to SQL. The user simply issues a question in natural language and the system will take care of the rest. This includes identifying the database structure in terms of the primary primary keys and foreign key links representing the natural language questions and database schema in a joint vector space. Decoding the SQL query from this vector space by generating the SQL keywords, copying value phrases from the natural language and table and column names from the schema. The generated query is executed on the database and finally a post processing module take the execution results, extracts the Data information to send to the smart speaker and t data attributes to be rendered into accessible realizations. I'm super excited about this work because natural language is the most important way for everyone to communicate with their CRM. Because the model takes both the natural language and the database schema as input, it can be adapted to the customer database schema. You have. As you probably know, some Salesforce recently acquired Tableau. We are looking into more collaboration between Salesforce Research and Tableau Research in terms of combining natural language processing with data analytics. This is a huge step towards democratization of data. We are looking forward to bring this piece of research out of the lab to help you and your team work faster and dream bigger. I have given you two examples in terms of the end of data entry and and the beginning of data conversation. Next, I'm going to welcome Cai Mingxion, our Senior Director of AI Research, to talk about a very compelling case of the future of conversational AI. Thank you.

Thanks Victoria. Victoria's work is 100% build based on our Salesforce research, speech recognition and text to circle. It's a real demo, right? Okay. We just heard about how the AI can make us all more productive. Now let's talk about how the future of AI is conversational. I think most of you may have this type of experience. When you make a call to a customer service center, you are told, dear customer, your call is important to us. Please continue to hold. Then they play you some random music. Well, it is pretty hard to feel important after you heard this sentence for 10th time. And the music isn't even good. Actually, that is not the worst part. The worst scenario will be hearing light. I'm sorry. Please call us back again during our business hours. Thank you. In this situation, let's imagine if you have a conversational AI agent that can help you solve your problem. Here we will present a demo of conversational AI agent that can do human authentication, handle interruptions, and process multiple tasks in a single conversation. The background story of this demo is I ordered some pizza for my team since we are working so hard for this keynote. Everyone is starving right now. Before I started, I want to emphasize again, this is a real demo. I will call it from my phone. I can't promise what will happen, so hopefully the signal is stable. Everything goes fine. Let's do it. Fingers crossed.

Hi Kane, I am the digital assistant for Northern Trail Pizza Center. What can I do for you?

Hi, I just ordered some pizza. Is my pizza on the way?

Oh, sure. I'd be happy to help you check Your order status for security. What is your email address?

Sure. My email is k I n g s ialesforce.com.

Okay, so K I n g s I@salesforce.com.

Yes.

Got it. What is your good code?

The zip code is 94301.

Okay, I have verified your account. One moment please. From what I can see here, it looks like your order for 10 pieces of combo pizza has been placed but not yet shipped.

Great. Oh wait. Could you please hold on second?

Sure.

I forgot to ask. Anyone want some pizza? Yeah, I think I see someone wants it. Let me do it. Okay. I'm back. Hi, I'm back.

Welcome back.

What can I do for you?

Please add 5 more pizza to my order.

Oh, sure. I'd be happy to help you add more to your order. Okay, so 15 combo pizzas in total.

Yes.

Is there anything else I can help you deal with today?

No, I don't think so.

Sounds good. I'm glad I could help. Thanks for calling and have a good one.

Yes. This is the first ever demo of a fully autonomous conversational AI agent in front of a large live audience. It is not prerecorded. These are two AI systems that we just interacted with, fully built as Salesforce Research. The entire system works today and in the future it will be even more seamless, faster, with more capabilities and you will never have to wait for customer calls again. Now let's get into how it works. Our system consists of multiple different components. Each turn the voice engine transcribe human voice into text. It is also called speech recognition. Oh, I want to add one more thing. At Salesforce, equality is one of our values. And it also applied to AI research and development. In our team, we constantly work hard to make sure our voice engine accessible to different accents. Given this unbiased and accent robust transcript, the NLU engine Natural language understanding engine analyzes the customer's feelings, the intent and the key information. Learn the dialogue management engine decides what is the next best action and generates the response. More concretely, in our demo, I told the agent I ordered some pizza and asked, is it on the way? The system detected my intent was to track the PISA order. So the dialog management engine decided to authenticate and verify the caller's identity. It triggered the authentication task and asked me for the information, my email and zip code. Of course, this is configurable for different components. After the authentication task, the system continued the uncompleted tasks. In this case checking my order status. Also, our system is not just for single task. It can easily embed new more tasks and even handle human interruption. In a demo we showed, we paused to get more feedback from you and change the order to have more pieces. Yes, this is our conversational AI system Visit. Congrats. In the future you can build your own AI agent for your customers. Well, but we know not all conversations happen with our voice. Next, let's welcome Nitish to tell us more new advancements in natural language generation. Thanks.

Have fun.

Thanks Saiming, I'm thrilled to be here today to tell you about some of the research that I've been doing. We truly are at a point where data entry is behind us and data conversations are our future. And a key component of that future is going to be automatic natural language generation. And without knowing it, many of you probably use this in your daily lives already. This could take on the form of email autocomplete, where you start typing your email and your favorite email engine completes it for you. It could be for machine translation where you convert between one natural language, say English to another, like Japanese. It could be to solve and triage your customer complaints online, or to correct your grammar both at word and phrase level. These are all great technologies, but they typically happen at sentence or snippet level. Here at Salesforce we are looking always to push the state of the art in natural language processing and set out with an ambitious goal of doing long form text generation. So that means we are not generating the next sentence, we are generating the whole article. In that respect, I'm thrilled to announce Control. It's the world's largest open source language model. It has an unprecedented 1.63 billion parameters and has been trained on 140 gigabytes of text, which includes millions of documents, thousands of books, and also all of Wikipedia. Now, Control is simply trained to predict the next word given a history of few words. And by doing this over and over again, we can generate as much as 500 words into the future. One defining feature of Control also is that it's easy and explicit to communicate to the model what your intentions and background is. And finally, we understand that this technology is not without its cause for concern. So we are working actively with our partners to ensure that we promote understanding and awareness of how this model can and should be used and to mitigate any negative consequences. Without any further delay though, I do want to show you a quick demo that highlights some of this functionality and MJ and Ankit are helping me out here. But I do want to emphasize while they get set up that this is also a rive demo. This is not simulated or recorded in any way. So let's just go off to the demo server and here we go. So imagine for a moment that I'm no longer a research scientist. I'm actually a marketer for Northern Trail Outfitters and the company is releasing a new hoverboard that I have to write a press release about. Now I could do this myself, but I'm going through a little bit of a writer's block, so I'm hoping that Control can help co create with me. So to start off, I'm going to give it two bits of inputs. One, I'm going to select the genre to be press releases. That's from a bunch of genres that we can choose from. And it's essentially telling Control to complete my sentences as if it was a press release. The next I'm going to give it some initial CTEXT so it knows what the press release is about. So I'm just going to write out. Today we are releasing the most affordable wheelless hoverboard. It's just some see text for it to understand what the context is. And just like you do in your browser or in your codecomplete editors, you'd hit tab for it to trigger a completion. This is calling a real life backend model. And it's coming back with something that I would write myself. It's emphasizing the two things that I would do myself, which is affordability, that it's not going to break your bank. And the second is that it's exciting that it's a thrill to write now with existing technology, this is as good as it can get. You got your sentence completion, there's nothing more you can do. But not with Control. You can seamlessly hit tab again for it to generate another sentence. And because it's read millions and millions of documents, it knows what the most plausible next sentence is. So it came back with something about beginners and experienced writers will enjoy this likely. And that's great. That's exactly what seems to be a good point to talk about that Now I could keep tabbing and keep getting control to generate more and more, but that's not the point of this. And I feel a little bit more confident about the press release. I'm going to intervene, I'm going to co create and I'm going to say that this board has a maximum range of to sort of trigger it that sort of completion. And I'm going to hit tab and just like before, it's going to come back with something reasonable. It comes back with 20 miles. Well, that's not exactly true. Our board is actually using a more sophisticated technology, it's 50 miles. But that's not a problem because it's not a take it or leave it situation. You can go in and seamlessly edit it. And not only are we doing this as easily, we're also influencing controls future generations. Now I'm going to get a bit technical here just for the sake of it, and we're going to type in how the board works. So we're going to say that the board works on electromagnetic propulsion. If you're like me, it's been a while since we've taken physics classes, so I don't quite remember what that is. So I can either go on and read a few Wikipedia articles and try to summarize it myself, or I could just go on to the next line and choose a new genre to completion. So I can just say from my dropdown that I want it to be like Wikipedia. I wanted to use the information that it's learned from Wikipedia and finish my sentences as if it was writing an encyclopedia article. So I'm going to write off electromagnetism is and hit tab just like before and it comes back with a reasonable one liner about what electromagnetism is. And again, we have not changed. We have not gone into a different tab, tried to do anything fancy. We're just simply using the same model in the back end to do this and it comes back with this. And we can also go into a little bit of history of electromagnetism by just saying the theory of electromagnetism and then hitting tab and having it come back with a reasonable completion about James Maxwell. Now, it's important to fact check this. It's always important to check that the models are generating things that are plausible and correct. But in this case that is correct. Maxwell did write a seminal paper in 1861 about this. Now there you have it. You can keep going in Wikipedia, you can switch back to press releases or go off into some of the other genres that we have available. But here I hope I have given you a quick glimpse of into what the future of AI looks like for writing. Next up we have Nazneen, who's going to talk about how to make these systems more interpretable and explainable and hence trustable. Thank you.

Thank you, Nitish. Explainability is crucial for us humans to have meaningful interactions with AI systems, to understand AI safety and to take appropriate actions. For example, if Einstein predicts you risk losing two of your customers their Current cost is too high. I recommend offering them a 15% discount. That's the AI systems prediction, the why behind the prediction and the suggested response. So remember that trust is our number one value with the future of data conversations. We want to be able to have trust in these AI systems. And for that we need explainability. We want AI systems to explain their reasoning. We want them to have common sense, just like people.

So?

So I trained a language model similar to what NITISH demonstrated to generate common sense explanations. Let me give you an example. Say we have a question like where would you not ride a hoverboard? And the options are road, grass or ocean. Imagine that this language model has seen FAQ pages along with all of Wikipedia and several hundred books. So it generates an explanation that looks something like a hoverboard can be ridden anywhere on land. In the next step, the AI classifier takes this explanation along with the original question and the answer choices and selects the right answer. Because the AI system can explain its reasoning, it can make better predictions. To learn more about my work on explainable AI, please join me in my session on Thursday. The details are on the slide. Once you understand the why behind your AI system's prediction, you'll be able to trust it even more. Thank you. Now back to you, Richard.

So I personally was pretty nervous because again, those were all real life demos and you saw that sometimes they fail. And what you also saw is that we built them so they're robust and they can recover from failure. And those were not scripted. So I want to give a big round of applause to our demo team and all the researchers for their research. What you could also see there were our values. Trust is our number one value. So that bot actually announced itself as a bot. We're not trying to fool people into thinking they're talking to people, even though right now that might not look like that realistic because there's a beep and so on. But as we're making this better and better, we'll continue to have that trust and be clear about what are your customers are communicating with. We also think about customer success trying to make it seamless to do data entry. And that OCR system was indeed given an extra tough challenge because I'm sure you're a little more nervous when you take that picture on a big stage. And when it comes to equality, I'm really excited that this also works, no matter what your linguistic background is and where you come from, and eventually also for different languages. Now, we also realized that AI is basically an omni use technology, you can use it for a lot of different things in the world and we want it to be a force for good. So we have a special subgroup inside Salesforce research that works on applications that have actually zero product and monetary impact necessarily. But we feel like the world will be a better place if they are successful and supported and researched. We work on histopathology, trying to classify breast cancer and you'll see next year some of those papers and research results coming out and hopefully also actual products and collaborations with universities and hospitals. We also work with the Global Fund for Women on helping to continue to make the grant making process that are using natural language processing to understand for a large organization that gives out millions and millions of grants and grant money to support girls and women's causes all over the world. How they can continue to have an unbiased, equitable and even more efficient process in supporting different groups all over the world. And I'm very excited about this. We also work on counting sharks. You might be like what I thought I meant. Salesforce and Dreamforce. Where's the customer story? But there are a lot of customers. A lot of our customers are surfing. And sharks are incredibly important for our ecosystem. People are obviously even more important and they're both wonderful, but it's better when we keep them a little separate. So I'm going to play this film first and then we're going to talk to some of the main researchers from the Benioff Ocean Institute and a wonderful friend of mine right after play the film.

The fact of the matter is that like sharks are not super predators in the oceans. People are the super predators in the oceans. In terms of what we're doing to change ocean ecosystems, life would be very different without sharks and very clearly would not be better for us. What do you think is the most dangerous shark in the ocean?

The bull shark, the tiger shark and.

The great white shark are like the top three most dangerous sharks. But like all things, consider not very dangerous. Not very dangerous. A selfie is more dangerous. A selfie.

A hot dog's more dangerous.

It's true. We have pretty frank conversations, he and I, about sharks. And it's not something that we ignore. It's something we talk about directly because I want a pretty smart kid to process that there are things out there that he needs to respect. I think the thing that creates fear is a lack of understanding. If you don't understand and have data about what a shark is and when it's around, you start to fear that. So this is why the AI is cool to me because it's creating this new stream of new data which helps us get to that point of thoughtful respect for these creatures.

It's an interesting time in the United States.

There's now more great white sharks off the coast than we've ever seen before. And there's not a great resource for understanding those populations. You can use cool technology like drones to look for sharks, look for wildlife. You can use really cool software like artificial intelligence to be able to search all that video. And if you're a seal that's up close to the surface, if you're a dolphin that's close to the surface, if you're a shark that's close to the surface, it'll be able to see it. With this particular project, we're using two different technologies that are available at Salesforce. What we've done is we've taken the field service lightning platform and we've coupled it with Einstein vision. And so we've built a special model, special AI model that understands sharks, specifically.

Great white sharks and their sizes. Nature creates, in theory, so much data in practice. Nobody has the time to look at the ocean for hours and hours and then count that one chart. And so AI can kind of automate things that no human could do.

If ever there was a time in ocean history where we needed these tools, we needed this intelligence. We needed more AI and smart people and new communities trying to put tools together, it's now.

We want to understand how is climate change affecting them? Why are they coming to the coast in larger and larger numbers and for longer periods of time?

We can have really current, you know, up to the minute information, hopefully, and make a really good educated assessment on whether this is a good day to swim or not. We're getting texts every day that tell.

Us what the AI is picking up.

And what it's not picking up. We can use AI to identify and.

Educate us on the movements and the sizes and the activity of some of these juvenile great white sharks.

I didn't really know how this was gonna go down. You know, sharks blowing up here on the pond. I'm gonna ignore that. Hang on a sec. Do you want me to talk about this suit? We saw one great white shark just north of the creek mouth. It looked around 10ft smaller than the individual we saw yesterday. Started close to shore, swam around 100 meters out. Sounds like there's another shark down there.

Your son's down there.

Yeah. Right. So Finn just dropped him off at surf camp at 9. Right. Which is just minutes before the text Came in. It changes the way I'm thinking about my day. So this is sort of what this is about. It's about trying to understand a little bit more about what's going on down there at the beach so his camera camp counselors can make a smarter decision about how much time they're spending in the water.

We all communicate with the other surf.

Camps and other surf instructors on the beach. It's just better additional information that lets us, in theory, maybe get, like, a step ahead.

I think that's the value of this project, is really using the AI to create this data stream so that they can do what they love, which is, in this case, connect to the oceans, fall in love with the oceans, but do it more safely than they were doing before we started this.

It's up to us as researchers in both AI and collaborating with people in other fields to find the most positive use cases and actually double down and work on those.

You want to go, sir? Let's do it. Okay. Come on. My goal is to catch 20, 20 ways. I want to catch 30 waves. Okay. Here's a big tech company, but they care about the oceans. They care about making sure that these same tools are useful for people in a community like mine. This is, like, exactly what our Bunny out the ocean initiative is about. Take some smart people, some smart tools, maybe originally built for another purpose, but you bring it in to help with something that's, you might say, more important, which is about community safety, about a future of our ocean of kids.

Right.

That, for me, is like, the essence of what AI for good is.

Wow.

That is our part in thinking about all the stakeholders and not just the shareholders. And I'm really happy that we have the support from the rest of Salesforce to be able to do that kind of work. I'm really stoked to bring Doug on stage. He's a good friend of mine, and he is the main driver behind this project. Welcome, Doug.

Hey.

So Doug and I, I met him first and on one of our first scuba dives together ever. As my first scientific dive, the goal was to count giant sea bass. It's apparently one of the closest to extinct species of California. There are only 500 or so left. And I was very excited to go with a real ocean marine biologist and have a scientific dive. And after our first two dives, the unfortunate count was zero. It was really just a normal dive with no giant sea bass. But then on the third dive, you saw 10 of them, and that was a pretty large fraction of the entire species. That's Left of giant sea bass. And so across all our trips, I never asked you what I actually love most about scuba diving.

Right. Thanks, Richard. And thanks for having me here. Probably an answer that will resonate with you. I think probably a lot of, you know, Dr. Socher's superpowers in the computer science domain. I've learned a little bit about that in our collaboration together. But he has a series of superpowers. One of those is being a very proficient diver and also being a very good underwater photographer. And there's a lot of beauty in your pictures, Richard. And it's that beauty, I think, that really resonates for me when I get a chance to be underwater. And it's the unexpected nature of that beauty. As you know, you can be diving, you go around the edge of a reef and you really never know what you're going to see. You might run into this intricate, beautiful anemone and anemone fish. Another day you could go around the edge of the same reef and you could run into a car sized shark. It's that unexpected nature and beauty that's so special. It's like having your favorite art museum and every time you go, the exhibition is different. Everything is changing. You know you're going to love it, you know, to fall in love with it, you know, to be inspired, but you never know what you're going to see.

Nice. So most people, when they think about AI, they might think CRM and service or self driving cars. Few people think about the oceans when it comes to AI. How did you have those ideas and where do you see that connection?

Yeah, well, I'm really grateful to be able to have this experiment to try to power AI or bring AI into ocean problem solving. I know there's a lot of more traditional domains for doing AI for good. I'm glad you guys are working a lot of those spaces. But I think the oceans desperately needs more intellect and tool and help and tools like AI. So the beauty that I just referenced, the beauty that I talked about, is a great peril. There's a lot of risk now for ocean life, for the services that we get from the oceans. We have traditional threats like overfishing, but we're also facing off in the oceans against a whole new set of mega threats we've never seen before. Things like climate change, scourge of plastic pollutions. And so we need smarter tools, better tools, bigger tools, bigger brains to face off against those challenges. And in the oceans, I think maybe I'm a little biased, probably share this bias with folks Here in the room, that data is our strongest ally. It's perhaps the thing that will help us face off successfully against those kinds of challenges that I mentioned. And our problem in oceans is not having a paucity, not having enough data, it's actually having too much data and inability to actually find the signal for change in all that data. We're being hit in the face by this fire hose of data. We've got sensors on submarines and penguins and reefs, and we have data coming at us in, in mass quantities. We don't always have the capability to figure out what the solution is for the problems we're facing off against. It's like the oceans. We're a sick patient. Got a lot of sophisticated machines plugged in, a lot of data reading out here, but not necessarily a good diagnosis of that data. AI comes in to that ocean data helps us figure out exactly what the illness is, then importantly helps us diagnose and apply a cure. So it's a really important moment in ocean history when we absolutely need these new higher IQ tools like AI.

I love the projects that you're working on. In some ways, if he's successful, we might have to have the ability to do natural language processing on dolphin language or whale language. For this particular project, what was the most meaningful part for you?

Yeah, well, as you saw, there's obviously a sort of personal component. It's a rare opportunity to actually have a study system that I immerse myself in in my family, immerses themselves into. This beach is for our community, our gathering place. It's like what Trafalgar Square might be for London. That's it for our community. The beaches, it's where we go to, you know, fall in love with nature, to fall in love with each other, to actually have a good, happy social experience, to recreate, do sport. Right. But it's a place that we want to be wild, but we also want to be safe at the same time and find that sort of win, win. And I know there's sort of a lot of conversation about AI driving people apart, maybe being an isolating force when you have a machine and a network of friends or decision makers. For us, what was really surprising, was really special, was that you had this community and we're actually brought together more as a result of the AI. We had good data that helped us sit around the table and a whole bunch of strange species that don't always sit around the table together. You know, surf school instructors and marine biologists, computer scientists and computer engineers and lifeguards. And whatever, all sitting together that felt more empowered because we could see more and understand more about this space we can cared about. It's a moment in ocean history where we absolutely need to have more people falling in love with the oceans rather than the other way around. And I felt like AI brought us together to help do more of that together. Wow.

I love it.

Thank you so much and thanks again.

Thank you.

Wow.

So with that, I hope we could inspire you to join us in the AI revolution. If you want to learn more about all the different research projects and the things that are coming out now, you can check out the Einstein zones. We also have Einstein AI, a website where we have all the papers and research that we publish available. Of course, you can take trailheads. And one thing that's quite interesting, if you work in a company and you want to do some joint research and develop joint IP and patents, we're excited to collaborate with you on both the short term and the long term AI research. And with that, thanks everybody for coming. Have a great dream force.

Bye.

Bye.