Larry Ellison Keynote on Oracle’s Vision and Strategy: Oracle AI World 2025 (YouTube Video Transcript)

(00:00:00) Your YouTube transcript will appear here (00:00:00) [UPBEAT MUSIC] (00:00:03) (00:00:50) [ROCK MUSIC] (00:00:53) PRESENTER: Please welcome via live broadcast, (00:00:56) Chairman of the Board and Chief Technology Officer of Oracle, (00:01:00) Larry Ellison. (00:01:01) [ROCK MUSIC] (00:01:04) (00:01:08) LARRY ELLISON: Hi, everybody. (00:01:10) Let's see. (00:01:11) OK. (00:01:12) Says, AI changes everything. (00:01:16) That's a kind of a big statement. (00:01:18) Everything. (00:01:19) I think it's pretty close. (00:01:23) So I'm going to talk a little bit (00:01:24) about how Oracle's been responding (00:01:27) to these changes that started-- (00:01:29) well, I guess they started in earnest when ChatGPT 3.0 came (00:01:35) out and suddenly, AI models started (00:01:39) sounding a little bit like us. (00:01:40) (00:01:45) There are two big phases of this AI technology. (00:01:52) One is the dawning of the AI era, (00:01:55) which is a bunch of companies building these enormous AI (00:02:02) models. (00:02:04) They're actually an AI model right (00:02:06) now, what's called a multimodal AI model, (00:02:09) is made up of several neural networks, (00:02:12) like your brain has several parts. (00:02:16) It's actually it's kind of a perfect analogy. (00:02:19) To do vision, you use one part of your brain. (00:02:23) To do language, you use a different part of your brain. (00:02:25) When you build an AI model, you use a different neural network (00:02:32) for vision-- (00:02:34) seeing something, seeing its edges, seeing its shape, (00:02:39) seeing its color, seeing it move. (00:02:41) You use one neural network for seeing it (00:02:44) and a quite a different neural network for recognizing (00:02:46) what it is, identifying it. (00:02:49) And then a third neural network, to classify it and organize (00:02:54) it and reason with that data. (00:02:56) So very much like our brains, the modern AI system, (00:03:01) a modern AI model is a multimodal model (00:03:04) that has multiple neural networks to look (00:03:08) at different kinds of data-- (00:03:10) video data, textual data, hear, hearing, things like that. (00:03:14) Well, what's going on right now is a series (00:03:19) of companies are spending vast fortunes training these AI (00:03:27) models on publicly available data on the internet, (00:03:31) enormous amounts of data. (00:03:34) And this has become this AI training. (00:03:38) It's very apparent after a few years of it. (00:03:41) It's the largest, fastest growing business (00:03:47) in human history, bigger than the railroads, bigger (00:03:50) than the Industrial Revolution. (00:03:53) I mean, it is a whole new world that is dawning. (00:03:57) (00:04:01) There's the building of the models. (00:04:04) And then once those models are built, (00:04:07) it's the actual using those models (00:04:10) to solve very important problems, early diagnosis (00:04:15) of cancer, for example. (00:04:18) But there'll be a lot of-- (00:04:20) surgery that is more precise and more accurate than human beings (00:04:25) can do. (00:04:27) Robots will be much better surgeons than human beings can (00:04:30) be for all sorts of interesting reasons you might not guess. (00:04:34) Anyway, the big opportunity in AI training is upon us. (00:04:42) And Oracle is a major participant (00:04:44) in building data centers to do AI training. (00:04:48) But the much, much larger opportunity, (00:04:52) the one that will truly change the world (00:04:54) isn't the creation of the models themselves, (00:04:57) the training of the models. (00:05:00) What will change the world is when (00:05:02) we start using these remarkable electronic brains-- (00:05:06) and that's what they are-- (00:05:07) these remarkable electronic brains to solve humanity's (00:05:13) most difficult and enduring problems. (00:05:18) Now, there's one thing that's kind of interesting, (00:05:21) where Oracle's explicitly involved, which is, (00:05:26) as I said earlier, these AI models (00:05:28) are trained on publicly available data, (00:05:30) all the data on the internet. (00:05:32) So if you look at ChatGPT, Anthropic Claude, Llama, (00:05:36) what have you, they're all trained on all of the data (00:05:42) on the internet, in other words, publicly available data. (00:05:46) But for these models to reach their peak value, (00:05:52) you need to not train them not just on publicly available data, (00:05:56) but you need to make private, privately owned data available (00:06:02) to those models, as well. (00:06:04) And that's where Oracle plays a particularly important role (00:06:07) because most of the world's high-value data (00:06:09) is already in an Oracle Database. (00:06:12) We just had to change-- (00:06:14) and it is past tense. (00:06:15) We had to change. (00:06:16) We did change that database so that Oracle Database can (00:06:23) take the data that's already in the Oracle Database (00:06:26) and make it available to AI models for reasoning. (00:06:31) So the AI model can reason on not just public data (00:06:34) but on private data. (00:06:37) AI is an incredible tool. (00:06:42) People think some people think it's (00:06:44) going to replace all human beings and all (00:06:46) of our human endeavors. (00:06:48) I don't think that's true. (00:06:51) It will help us solve problems we couldn't solve on our own. (00:06:54) However, it will make us much better scientists and engineers (00:06:58) and teachers and chefs and bricklayers and surgeons (00:07:01) and what have you, that we've never built (00:07:05) a tool, anything like this. (00:07:06) (00:07:10) I press the button and the slide didn't move. (00:07:16) Do it again. (00:07:17) I press the button again. (00:07:19) [LAUGHTER] (00:07:22) And this is not an AI device. (00:07:24) (00:07:30) One more time. (00:07:31) And then I'm just going to say the word slide. (00:07:34) Oh, there we go. (00:07:34) OK. (00:07:35) Well, I did both. (00:07:36) So who knows what-- who knows why it moved. (00:07:39) OK. (00:07:40) (00:07:43) I remember when this wasn't called AI World, (00:07:46) remember what it was called CloudWorld (00:07:48) a long, long, long time ago? (00:07:49) (00:07:52) I had a presentation about AI. (00:07:54) Even though it was called CloudWorld, (00:07:56) I was still allowed to do a presentation on AI. (00:07:58) And I said, is AI the most important technology (00:08:01) in human history? (00:08:03) And dot dot, dot, we're going to find out soon. (00:08:08) Well, it's pretty clear. (00:08:11) The smartest people I know are working-- (00:08:16) what? (00:08:17) I didn't press the button. (00:08:21) OK, so not pressing the button. (00:08:23) You back up the slide, please. (00:08:26) Thank you. (00:08:27) I'm just going to put that down. (00:08:29) [LAUGHTER] (00:08:32) We're going to get a better one next time. (00:08:34) The smartest people I know are investing fortunes. (00:08:40) To be specific, they're investing their fortunes (00:08:44) in building and training these AI models. (00:08:49) That's how important they are. (00:08:51) That's how extraordinary they are. (00:08:53) And by the way, Elon, Mark, Sam, in alphabetical order (00:08:59) are all really smart guys, extraordinary people. (00:09:05) People say, well, this AI thing, maybe it's just a bubble. (00:09:08) Maybe it's not that big a deal. (00:09:11) Maybe it's just a bubble. (00:09:14) Well, the internet, really-- (00:09:16) I mean, the internet was a big deal. (00:09:18) I mean the most-- (00:09:20) if you look at the fortunes created of the internet, (00:09:23) I mean, certainly worked out for Google, (00:09:26) searches seemed to have paid off. (00:09:29) And Elon on that list did start PayPal, (00:09:33) and that paid off nicely. (00:09:35) But I know I have asked him. (00:09:36) Elon, he said he definitely didn't put a dime into pets.com. (00:09:41) And the thing is, when people talk about bubbles, (00:09:44) what is a bubble? (00:09:45) I mean, people get exuberant. (00:09:46) I mean, the internet was an incredible new technology, (00:09:49) remains the foundation of computing. (00:09:53) And we couldn't have AI without the internet. (00:09:57) So it's incredibly important technology. (00:09:59) But people started confusing internet companies (00:10:04) like a PayPal or even worse, internet search worse, meaning (00:10:09) better, with pets.com. (00:10:11) I mean, the fact that if I can sell (00:10:13) pet food in an e-commerce site, that suddenly (00:10:15) means I'm an internet company. (00:10:17) Not really. (00:10:20) So yes, there'll be people spending money on AI (00:10:24) because almost every tech company these days (00:10:26) call themselves an AI company, but they're not. (00:10:30) A lot of them are not. (00:10:32) But AI in terms of its value, this (00:10:36) is the highest value technology we have ever seen by far. (00:10:41) Next slide, please. (00:10:43) (00:10:46) AI. (00:10:49) It's interesting because it's called artificial intelligence, (00:10:54) but as opposed to artificial perception. (00:10:58) But it does perceive, it hears, it smells. (00:11:02) Think about smelling. (00:11:03) I mean, the idea that you can pick up chemicals that are just (00:11:07) drifting around in the atmosphere (00:11:09) and figure out what those chemicals are. (00:11:11) Dogs can smell cancer in patients. (00:11:15) We should be able to do that with AI. (00:11:18) We should be able to-- in fact, there's (00:11:21) a project I know of called the dog's nose that I'm actually (00:11:27) a part of. (00:11:28) And we're building sensors. (00:11:31) We're building sensors that can smell cancer or other illnesses. (00:11:39) But AI perceives. (00:11:42) It's got the part of the brain that hears and sees, (00:11:47) in addition to reasoning. (00:11:48) I mean, it can read street signs. (00:11:51) It can read a page on a book. (00:11:53) They take a look at you, it'll recognize you. (00:11:55) (00:11:58) Can identify the song that's playing. (00:12:01) You can talk to AI and ask it a question or type it out. (00:12:07) And AI can reason logically, very quickly, (00:12:12) using language the same way-- the same way (00:12:14) we do, and mathematics. (00:12:18) And I remember I was over at Tesla (00:12:21) looking at the Optimus robots. (00:12:24) And I was curious just how the robots were (00:12:26) going to learn, and then just thought about it for a minute (00:12:28) and said, well, how would a robot (00:12:31) learn how to clean your house or scramble eggs (00:12:36) or play the guitar? (00:12:39) Well, it would just watch an internet video. (00:12:44) It's connected to the internet. (00:12:46) It can learn to play piano, just like we (00:12:50) would watching an internet video, (00:12:53) except it would do it a little faster because it (00:12:58) can play the internet video at very, very high speed (00:13:01) and learn to play that piece by Chopin in about (00:13:06) five seconds, which I know my kids can't learn to play piano (00:13:11) that fast because I listen to them practice every day. (00:13:16) And five seconds is out of the question. (00:13:18) (00:13:21) AI robots. (00:13:22) AI robots will be much better surgeons than the best doctors. (00:13:27) There's this very famous surgery started (00:13:29) by Dr. Mohs, who actually would take cancer lesions off patients (00:13:34) faces. (00:13:35) And he was so famous for it because he did the least damage. (00:13:40) He took the least amount of tissue off your skin. (00:13:44) So cosmetically, he had fantastic results. (00:13:47) And what he did, he would take a couple of layers of skin (00:13:50) off and then take that skin over to a microscope and look at it (00:13:53) and see. (00:13:53) Is he just taking-- (00:13:54) is he taking any healthy cells? (00:13:57) I mean, how deep does the cancer go? (00:13:59) So he's back and forth, cut a little bit of tissue, (00:14:02) look at it on microscope. (00:14:03) Cut a little bit more tissue. (00:14:04) Look at it. (00:14:06) Well, AI robots aren't just don't play fair. (00:14:10) Their vision, the vision on the robot, it is microscopic. (00:14:17) They can see what-- they don't need a microscope (00:14:21) to see individual cells. (00:14:24) They don't need a microscope to see where the cancer ends (00:14:27) and the healthy tissue begins. (00:14:31) And their coordination, and it is exactly, (00:14:36) they're better surgeons than we are. (00:14:38) Not because they're smarter than we are, (00:14:41) but because they have better hand-eye coordination. (00:14:44) Their eyes are way better than ours, (00:14:46) and the precision of their hands is way better than ours. (00:14:49) So they can cut between a layer of healthy cells (00:14:52) and a layer of cancerous cells. (00:14:54) It's truly stunning to watch and will make us all-- (00:14:59) and it'll be very reassuring when (00:15:01) we can go to a doctor who's using a robot to do the surgery. (00:15:08) The surgery will be perfect. (00:15:11) (00:15:15) I said this earlier, but it is so interesting. (00:15:20) It's built just like the brain's specialized neural networks, (00:15:25) one for vision. (00:15:26) I mean, literally, a convolutional neural network (00:15:29) simulates the visual cortex. (00:15:31) And the visual cortex has five layers. (00:15:33) It's right in the back of your head. (00:15:34) And evolution produced the very first layer of V1, (00:15:37) was just so the animal could first (00:15:40) perceive edges of something they were looking at. (00:15:44) Then it got up to four for color. (00:15:45) And the very famous V5 for motion (00:15:48) detecting, detecting motion and threats and threats (00:15:51) in the environment. (00:15:54) The ViT is the Vision Transformers (00:15:58) that then took that bitmap. (00:16:00) So the convolutional neural network (00:16:03) produces a bitmap map, an image, a bunch of pixels, if you will. (00:16:07) And then the vision transformer then compares that to things (00:16:11) that you already know. (00:16:13) And you start and you can recognize faces and things that (00:16:16) are-- things that are familiar. (00:16:18) And that's a different-- that's a transformer. (00:16:21) That's a ViT neural network for a holistic understanding (00:16:27) of the image and recording it. (00:16:29) (00:16:32) Version 3 of ChatGPT was the one that used the huge transformer (00:16:37) networks that did comprehensive language and reasoning. (00:16:42) The only drawback with that, the transformer networks-- (00:16:46) because we had facial recognition long (00:16:49) before we had the ability to converse and reason (00:16:54) using language with the GPT network, (00:16:59) the generalized pretrained transformer (00:17:02) network, which is what is doing the language and the reasoning. (00:17:06) But that requires enormous amounts of compute. (00:17:09) Thus, the requirement for fortunes to train these models, (00:17:13) these networks. (00:17:15) The transformer network is much bigger and much more (00:17:17) complex than some of the other networks, as you (00:17:20) think reasoning would be more complex than vision. (00:17:25) And then there are networks for certain types of mathematics. (00:17:29) Anyway, looks a lot like the brain. (00:17:33) (00:17:37) So the brain has a lot of-- it's amazing. (00:17:40) 20 watts human brain. (00:17:42) 20 watts. (00:17:43) Anyone screwed into 20-watt light bulb know that's not a lot (00:17:47) of light. (00:17:48) But it's enough to run 86 billion neurons (00:17:51) and give you vision and balance and reasoning and language (00:18:00) and creativity and the ability of deduction and inferencing. (00:18:04) You can do all of that with this incredible what Elon calls (00:18:09) a 20-watt meat computer. (00:18:12) (00:18:16) Sensation, recognition. (00:18:20) After recognition, the ability to reason on that. (00:18:23) Again, the visual cortex is right behind the parietal lobe. (00:18:27) Behind and below it, the prefrontal lobe, (00:18:30) as you can see on the left side, is a big language center. (00:18:37) The brain is very specialized, So are the AI models. (00:18:40) (00:18:43) But we're not building a 20-watt meat computer. (00:18:51) We're building a 1.2-billion-watt AI brain. (00:19:01) Did you ever try to do multiplication (00:19:03) as fast as an HP calculator? (00:19:07) These electronic brains, the AI, these AI models reason, and they (00:19:14) reason very quickly. (00:19:17) And they can deal with a lot of data. (00:19:19) (00:19:22) And they can get to answers that we've never gotten to. (00:19:29) So this is a picture of a data center (00:19:33) we're in the process of building. (00:19:34) Actually, it's up and running. (00:19:36) Part of it is up and running. (00:19:37) (00:19:40) Eventually, it's going to have a half a million NVIDIA (00:19:42) GPUs in it. (00:19:43) By the way, to give you an idea. (00:19:45) A 1.2 billion watts, what does that really mean? (00:19:48) That's enough to power 1 million four-bedroom homes (00:19:54) in the United States. (00:19:58) A million, that's a pretty good sized city. (00:20:02) And I think we've got a video on the construction. (00:20:07) [UPBEAT MUSIC] (00:20:10) (00:20:14) PRESENTER: Oracle is building the world's largest AI cluster (00:20:18) for OpenAI in Abilene, Texas. (00:20:20) [UPBEAT MUSIC] (00:20:23) (00:20:23) The project began as empty land in June 2024, (00:20:27) and is delivering GPUs in less than one year. (00:20:31) The cluster will contain more than 450,000 NVIDIA GB200s when (00:20:36) fully provisioned. (00:20:38) Power is provided by a combination (00:20:40) of grid power and on-site natural gas turbines. (00:20:44) Capacity is provided in eight separate buildings, (00:20:47) spanning 1,000 acres, all interconnected together (00:20:50) to support a single workload. (00:20:53) This site deploys the latest technology (00:20:55) across AI accelerators, liquid cooling, and networking. (00:21:00) More than 3,500 people work on site each day to deliver (00:21:04) capacity at an unprecedented rate. (00:21:07) Demand for AI continues to exceed supply. (00:21:10) And Oracle is committed to delivering the largest and most (00:21:13) advanced AI clusters to support our customers (00:21:16) all over the world. (00:21:18) [UPBEAT MUSIC] (00:21:21) (00:21:23) LARRY ELLISON: Well, that's a long way from writing code (00:21:26) in my bedroom in college. (00:21:28) [LAUGHTER] (00:21:31) (00:21:32) What happened? (00:21:33) [LAUGHS] I have no idea. (00:21:35) (00:21:38) OK. (00:21:39) So we're training. (00:21:41) We are in the middle-- (00:21:43) we train the very first version of Grok for Elon. (00:21:46) We're training a number of other of these multimodal AI models. (00:21:54) Almost all of these AI models are in the Oracle Cloud. (00:21:57) And I'll come back to that. (00:22:00) But yeah, we're probably involved. (00:22:04) We're certainly involved in training more multimodal AI (00:22:07) models than any other company. (00:22:10) And it's very exciting. (00:22:12) It's very exciting, and it's daunting. (00:22:15) I mean, the size of these projects that we're running, (00:22:20) it's not just building the network (00:22:22) of GPUs, the computer rooms, and the networks and the cooling. (00:22:27) And that was hard, by the way. (00:22:28) That was hard in the first place. (00:22:31) But now we have to build the power transmission plants. (00:22:35) There's a natural gas pipeline that (00:22:38) goes to the gas turbines, that fires up the gas turbines (00:22:42) and then generate electricity. (00:22:43) That electricity then has to be moved to the data center. (00:22:47) So it's power generation, it's gas pipelines, power generation, (00:22:55) power transmission, data centers, networks. (00:23:00) And those data centers are filled (00:23:02) with lots and lots and lots of complex software (00:23:05) and a lot of very smart, hardworking engineers. (00:23:09) These are enormous engineering projects, each (00:23:12) and every one of them. (00:23:14) And what we're building-- what we're (00:23:16) building-- what we're trying to build (00:23:18) are these multimodal neural networks, (00:23:22) trained on all types of data-- textual data, images (00:23:26) data, audio, video. (00:23:28) Every publicly available piece of data, plus synthetic data, (00:23:32) we train these models on. (00:23:35) Some of the models are designed to be real time. (00:23:38) Actually, Google has two models, one is Gemini, one is DeepMind. (00:23:44) The DeepMind is highly specialized (00:23:47) around molecular structures. (00:23:49) And one in DeepMind won a Nobel Prize last year. (00:23:53) Not this year, last year, won a Nobel Prize on protein folding. (00:23:57) We understand, it's taking a molecule, (00:24:01) where you understand the chemical formula (00:24:03) of that protein, a chain of amino acids. (00:24:07) And what happens-- what does it look like in 3D when you fold it (00:24:10) up, and it's no longer a string? (00:24:14) It's now folded. (00:24:15) That's a problem we've been working on for a very long time, (00:24:18) folding proteins. (00:24:19) And they solved it with the DeepMind model (00:24:23) that Google owns, when they bought DeepMind in London. (00:24:28) Elon has two AI models that are very, very different. (00:24:32) One is Grok multimodal AI model. (00:24:37) The other is Tesla. (00:24:39) And it's a real-time model. (00:24:40) And real-time models have some different characteristics (00:24:46) than, let's say, than Anthropic, which (00:24:49) generates code, or ChatGPT, which (00:24:52) is solving a legal problem or a medical problem, (00:24:55) something like that. (00:24:57) If you're driving cars, things happen very fast. (00:25:03) So yes, you have to have vision. (00:25:06) You have to have cameras all over those cars. (00:25:09) But if something happens, you might (00:25:11) be required to respond in a microsecond, (00:25:16) in a millisecond, at least. (00:25:17) A microsecond is really fast. (00:25:19) A millisecond, in the car, a thousandth of a second. (00:25:24) The ball suddenly is coming off a curb, (00:25:26) and a bike is following the ball. (00:25:29) And you have to see it, understand what's going on, (00:25:32) and take evasive action. (00:25:33) So there's no accident, and no one is injured. (00:25:35) (00:25:38) You have to build things differently (00:25:40) when you can't afford the network traffic to go back (00:25:43) across the network and talk to a computer, an AI model (00:25:47) on a network that's far away. (00:25:52) You need a very, very low latency response time. (00:25:55) That's why all the Tesla cars, all the Tesla robots (00:25:59) have to have local compute in the car, local compute (00:26:04) in the robot to make an immediate decision, a very (00:26:07) low latency decision. (00:26:09) That's not required, for example, if you're writing code. (00:26:13) I can tell you what code to write. (00:26:14) And you can take a moment to think about it (00:26:16) and then give an answer. (00:26:18) So the real-time models are a bit different (00:26:20) than the models that aren't, that don't require real time, (00:26:25) where you have some time to reason and compute your answer. (00:26:31) But both models are-- both types of models are very important, (00:26:34) and both types of models are being built. (00:26:39) These models do multistep reasoning. (00:26:43) Now, the reason-- what I'm calling reasoning not long ago (00:26:46) was called inferencing. (00:26:48) People would talk about we got to train the model. (00:26:51) There's one thing where we train the model (00:26:53) and then we're using the model when the model's reasoning. (00:26:57) We reduce that to just inferencing, (00:27:00) a type of reasoning. (00:27:01) That's no longer true. (00:27:02) In the early days, that's what models did was inferencing. (00:27:06) Not anymore. (00:27:08) They reason like we reason. (00:27:11) And there's a list. (00:27:13) They do deductions. (00:27:14) They do inferencing. (00:27:15) They do calculations. (00:27:16) They have strategy. (00:27:17) They have rules. (00:27:18) All the techniques, the reasoning techniques (00:27:21) that we use, they simulate and use, (00:27:25) but they think a lot faster than we do and solve problems (00:27:29) a lot faster than we do. (00:27:31) Or they solve really complicated problems (00:27:34) that we can't solve at all. (00:27:36) And that's what makes this so exciting (00:27:38) and makes this so enormously valuable. (00:27:43) These models can answer your questions. (00:27:46) They can generate computer code. (00:27:48) A lot of the code that Oracle is writing, Oracle isn't writing. (00:27:53) Our AI models are writing. (00:27:56) We just tell the model what we want the program to do. (00:28:03) And then the AI comes up with a step-by-step process (00:28:07) to actually do it. (00:28:09) We don't write the procedure. (00:28:11) We declare our intent, but the model (00:28:14) writes the step-by-step procedure, (00:28:17) that thing that we commonly think of as a computer program. (00:28:20) They diagnose medical images far better than we do. (00:28:23) They design drugs that we can't. (00:28:28) (00:28:31) But there's a big gotcha. (00:28:35) There's a big gotcha on these models. (00:28:37) And that is the models do not get trained on your private data (00:28:42) because for some reason, that people want to keep (00:28:48) their private data private. (00:28:51) And that's not going to change. (00:28:55) But people also want these models (00:28:57) to reason on their private data. (00:29:01) Have your cake and eat it too, whatever you want to call it. (00:29:04) I want to keep my data private. (00:29:06) I don't want to share it with anybody else. (00:29:08) However, I'd like to use this enormously powerful tool (00:29:11) to reason on my private data. (00:29:16) And that's what Oracle-- (00:29:18) was one of the big things Oracle's (00:29:19) been applying itself to in terms of solving (00:29:22) that particular problem. (00:29:24) And we have this new thing where we're talking about this week (00:29:28) here in Las Vegas is the Oracle AI Data Platform, the Oracle AI (00:29:33) Database, and the Oracle AI Data Platform. (00:29:36) And the interesting thing about the AI Data platform, (00:29:41) it includes a multimodal model of your choice, well, (00:29:47) a multimodal model of your choice. (00:29:50) That's great. (00:29:52) So if you want to use Grok in the Oracle Cloud, (00:29:55) you can use Grok. (00:29:56) If you want to use ChatGPT, you can use ChatGPT. (00:29:59) If you want to use Llama, you can use Llama. (00:30:01) You want to use Gemini, you can use Gemini. (00:30:04) And we'll attach that model, the model of your choice, (00:30:10) to not only the public data. (00:30:12) It's already-- the model is already (00:30:13) connected to the public data. (00:30:15) That's done. (00:30:16) But we give you the ability to add your private data (00:30:21) to the model's library of information and knowledge. (00:30:27) So the model can reason across not just public data, but also (00:30:33) private data, while keeping your private data private, (00:30:37) not sharing it with anybody else. (00:30:41) That's very, very important. (00:30:42) And it's not easy to do it in a highly secure way. (00:30:47) It's not easy. (00:30:48) If it was easy, a lot of people would have already done it. (00:30:51) (00:30:57) So as I said, the OCI includes all (00:31:03) of the popular multimodal models. (00:31:06) You can mix and match. (00:31:11) And we have the AI database and the AI Data (00:31:14) Platform that lets you add private data to the models. (00:31:21) In fact, I'm going to be a little more precise this time. (00:31:24) What it really does-- (00:31:26) and it's called RAG, by the way. (00:31:29) You basically take a bunch of data (00:31:33) that the model has not been trained on. (00:31:36) And by the way, that might be today's stock prices. (00:31:39) I mean, the model doesn't know today's news. (00:31:42) The model doesn't know, hasn't been trained on today's news. (00:31:45) The model hasn't been trained on today's stock prices. (00:31:49) Now, the model knows where to look for it. (00:31:52) It knows how to ask to look at the day's stock prices. (00:31:56) It knows how to look at the ticker (00:31:57) and get the very latest quote on today's stock prices. (00:32:01) And you just put that information in a database (00:32:05) that the model can access. (00:32:07) And you put your private data in an Oracle Database. (00:32:11) An Oracle Database-- the new Oracle Database (00:32:13) is called an AI database not just because AI is fashionable. (00:32:18) The new Oracle Database is called an AI database (00:32:21) because it has this RAG capability. (00:32:24) It has the ability to take any of the data in the Oracle (00:32:27) Database and make it accessible to the AI model (00:32:32) by vectorizing it. (00:32:35) So since a lot of your data is in an Oracle Database already, (00:32:40) you simply have the Oracle database (00:32:43) to make that-- put that data in a format the model (00:32:48) will understand. (00:32:49) And that's called a vector format. (00:32:50) (00:32:54) And the Oracle Database will vectorize any data (00:33:01) that you want to make available to the model. (00:33:04) And then you can reason on it. (00:33:08) By the way, but it's not just data. (00:33:11) It's not just data in an Oracle Database than an Oracle (00:33:14) Database-- the Oracle AI Database will vectorize. (00:33:17) Let's say you have a lot of data in OCI Object Store (00:33:23) or Amazon Object Store for that matter. (00:33:26) And you'd like to make that data available to the model, (00:33:31) to the Oracle AI Data Platform. (00:33:33) No problem. (00:33:34) The Oracle Database can go into OCI Object Store and vectorize (00:33:40) and create what's called a vector index to data (00:33:44) in OCI Object Store. (00:33:45) It can go into Amazon-- (00:33:47) go into Amazon Cloud storage and vectorize (00:33:51) portions of that that belong to you, (00:33:54) and make that accessible for reasoning (00:33:58) by the multimodal model. (00:34:01) So you're not restricted to data that's just in the database. (00:34:06) The Oracle Database can vectorize anything (00:34:08) that's in an Oracle Database, a different database, (00:34:12) a different cloud, and make that data easily accessible to the AI (00:34:20) model for reasoning. (00:34:22) And the reasoning is fascinating. (00:34:25) The first thing that Oracle did. (00:34:27) So the first project Oracle did in terms of taking private data (00:34:31) and making it accessible to AI models, (00:34:34) is we took all of our customer data, (00:34:38) and we vectorized it, and used RAG-- (00:34:45) and used RAG to make it available to the models. (00:34:49) So we started with customer data because we (00:34:52) think there's nothing more important to us (00:34:55) than our customers. (00:34:58) Now, some people were cynical. (00:35:00) You would say, there's nothing more (00:35:01) valuable to us than our customers. (00:35:04) But they go hand in hand. (00:35:06) So there are certain interesting questions we wanted to ask, (00:35:10) we thought were extremely high-value questions. (00:35:12) And there's a whole industry called customer relationship (00:35:16) management. (00:35:16) Actually, it's not called that anymore. (00:35:18) They changed the name to customer engagement management. (00:35:24) Whatever the name is, we know what the questions are. (00:35:27) So we ran this project inside of Oracle, (00:35:30) took our private customer data, put it in an Oracle Database, (00:35:36) vectorize it, and used RAG to make (00:35:39) it accessible to models, to a multimodal model, an AI model. (00:35:44) And then we asked the question, what Oracle customers (00:35:50) are likely to buy another Oracle product in the next six months? (00:35:55) Now why should that be important to us? (00:35:58) And specifically, what product each and every customer that's (00:36:05) going to buy something in the next six months? (00:36:07) Do you mind telling me what product they're going to buy? (00:36:11) Now, did you answer that question, (00:36:13) they're most likely to buy. (00:36:15) And then one thing-- and by the way, (00:36:16) it's not just questions that this thing does. (00:36:20) You can ask questions, prompt. (00:36:22) You can prompt it and get answers. (00:36:24) But you can also ask it to do things via agents. (00:36:28) You can create little computer programs, (00:36:30) sometimes not so little, and ask the AI to actually do something (00:36:35) to orchestrate some process. (00:36:37) And then we said, OK, let's send a mail to all (00:36:41) of our prospective buyers with the three best customer (00:36:45) references encouraging them to buy. (00:36:48) Now, that request required the generation of a computer program (00:36:56) called an AI agent, that had to figure out, OK, you (00:37:01) were going to buy this product. (00:37:02) You're a bank in Switzerland. (00:37:06) So we think the best references would (00:37:10) be the banks in Switzerland that have already (00:37:13) bought that product for you. (00:37:16) Those would be the best references for you. (00:37:18) So all of the references would be customized (00:37:21) based on what we know about you as a customer (00:37:24) and the exact situation you're in, the business you're in, (00:37:28) the products you have, the other banks you (00:37:31) have good relationships with, and you (00:37:33) can call for a reference. (00:37:34) Anyway, it's extremely interesting (00:37:38) that it can solve a problem like this so quickly (00:37:44) and tell us what the sales force should be concentrating on (00:37:52) at Oracle over the next six months. (00:37:56) It's kind of amazing. (00:37:56) (00:38:00) So that application, that AI agent-- (00:38:07) I can just back this up one. (00:38:09) I'm going to have to back up my slide once. (00:38:12) The last thing, the last line, send an email (00:38:16) to prospective buyers with the three best-- (00:38:19) the three best references. (00:38:21) From that single line, we can generate the AI agent (00:38:26) to actually do that properly. (00:38:30) You can generate the AI agent. (00:38:32) Or if you wanted to do a little bit more, (00:38:34) you could get even more precise. (00:38:36) You could add more things to it, what (00:38:38) exactly what you want to do. (00:38:39) What kind of letter do you want to send them? (00:38:42) What make the agent even more capable? (00:38:45) And that's actually what we did. (00:38:47) (00:38:50) And by the way, I don't know if you've heard this term. (00:38:53) I mean, I thought it was a little strange the first time (00:38:55) I heard it-- vibe coding. (00:38:56) Sounds very Gen-- what is the latest one? (00:38:59) Z? (00:39:00) It sounds very Gen Z, which is just (00:39:04) say what you want the program to do, generate the prototype, (00:39:08) and try it out. (00:39:11) Don't think about it too hard. (00:39:13) Just get a feeling for it. (00:39:15) And feel the vibe, I guess. (00:39:19) I mean, you can use English. (00:39:21) You can generate computer programs directly from English. (00:39:24) Personally, I've had debates with other engineers here (00:39:27) at Oracle about whether using English as a programming (00:39:32) language is a good idea because English (00:39:35) is notoriously imprecise. (00:39:38) And wouldn't we be better off if we (00:39:41) want to generate programs to create a custom, highly precise, (00:39:45) declarative language for computer programming? (00:39:47) Well, that's what we did at Oracle using APEX. (00:39:52) We added declarative AI generation language to APEX (00:39:56) for generating applications. (00:39:58) But there are plenty of people out there still (00:39:59) working with English. (00:40:01) And that's fine. (00:40:01) It's up to you. (00:40:02) We don't make those decisions for you. (00:40:04) We just make sure that you have options. (00:40:07) But most of the new applications that Oracle's creating now (00:40:14) are AI agents that were generated, (00:40:17) not handwritten, that were generated. (00:40:19) And they're connected by workflows. (00:40:23) And the interesting thing when we generate these applications-- (00:40:26) there are no security holes in these applications, (00:40:31) because the application generator (00:40:34) doesn't forget things and leave things out (00:40:37) and doesn't make those kinds of mistakes. (00:40:41) Every application that we generate (00:40:45) is stateless and reliable. (00:40:47) In other words, if the computer that application (00:40:51) was running on suddenly blows up, (00:40:54) loses power, whatever happens, someone (00:40:58) catches fire, that application can immediately (00:41:02) restart in a different data center because it is stateless. (00:41:05) And even though it stopped running in location A, (00:41:09) it will pick up running in location B (00:41:11) without missing a beat, without losing any data, (00:41:13) without the customer ever perceiving it. (00:41:16) So when you're generating these applications, (00:41:18) they have built-in backup, no single point of failure, (00:41:25) built-in reliability, built-in security, (00:41:28) and built-in scalability. (00:41:30) All the applications are written. (00:41:34) A lot of people-- (00:41:36) these low code application programming languages (00:41:38) are designed to write departmental things. (00:41:40) Maybe they work for 20, 30, 40 users. (00:41:43) But after that, they start to slow down (00:41:45) because they're really not designed (00:41:47) to scale to millions of users. (00:41:50) Well, because we generated, the design is always the same. (00:41:53) We always design it for millions of users. (00:41:55) Even if there are only five, it will run faster that way (00:41:58) and use fewer resources. (00:42:01) The productivity gains we're getting from this (00:42:07) is one of the reasons we feel so good about our efforts (00:42:11) in health care, that we can build, rebuild the Cerner code (00:42:23) base. (00:42:24) We can rebuild the entire Cerner code base, (00:42:26) modernize it using AI, build a modern version of Cerner (00:42:32) by generating it. (00:42:34) And we got all of the code for clinics operating already. (00:42:42) And next year, we'll have all acute hospitals. (00:42:45) We have rewritten everything that Cerner wrote (00:42:49) over a quarter of a century-- (00:42:50) we'll have rewritten in three years. (00:42:54) But what ours does is much more than theirs ever did. (00:42:58) (00:43:01) We attack the problem not just as automating a hospital (00:43:04) or clinic but automating the entire ecosystem. (00:43:08) Those are the kind of enormous productivity gains (00:43:11) you get when you use these incredible AI tools. (00:43:14) (00:43:18) The example of rebuilding Cerner is fascinating (00:43:25) because it's really not what we're doing. (00:43:27) We're not just, yes, we're rebuilding Cerner. (00:43:30) But we're also building accounting systems (00:43:34) for hospitals designed for hospitals, HR systems designed (00:43:39) for hospitals. (00:43:41) And hospitals are very unusual. (00:43:43) They're kind of 50/50 gig economy in and out. (00:43:46) A lot of nurses, they'll work for one hospital. (00:43:49) They'll work for private patients. (00:43:51) They'll have schedules. (00:43:54) You don't how many nurses you need or doctors for that matter (00:43:57) you need on Monday. (00:43:59) It depends what you're doing, how many patients you're (00:44:01) seeing, how many operating theaters are available. (00:44:05) So an HR system for a hospital is very, very different (00:44:08) and complicated. (00:44:10) There's a lot of certifications that doctors and nurses (00:44:14) and other health professionals, technicians have to get in order (00:44:17) to do certain tests, in order to do certain procedures, (00:44:20) in order to handle certain patients. (00:44:24) And our HR system has to deal with those certifications, (00:44:30) schedule the training, schedule when they're working. (00:44:35) They trade shifts a lot, be flexible about doing (00:44:39) all of that, paying them properly when they're (00:44:42) working a lot of overtime, but also understanding when they're (00:44:46) only working two days a week here and four days that week (00:44:52) at another hospital across town. (00:44:54) So we're building HR systems and accounting systems and banking (00:44:59) systems. (00:45:00) And this will be the one that maybe surprises you. (00:45:02) And then I'll go into my example. (00:45:04) And banking systems that cater to hospitals, (00:45:11) making hospital loans based on their receivables. (00:45:17) So I'm going to describe an AI agent. (00:45:21) So our goal was to not just automate (00:45:26) hospitals, like Cerner did or other competitors of ours (00:45:31) do automate hospitals and automate clinics. (00:45:34) We thought, following Elon Musk's rule, (00:45:38) that if we really want to be successful in health care, (00:45:42) we can't just automate hospitals and clinics. (00:45:46) We have to automate the entire ecosystem. (00:45:50) Like, Elon had to build a worldwide charging network. (00:45:56) Or electric cars weren't going to work. (00:45:59) He couldn't just make the cars and assume (00:46:02) that Standard Oil would provide the fuel, which (00:46:07) is what Ford did. (00:46:12) To build electric cars, he had to not only design (00:46:15) an electric car and manufacture batteries and put robots (00:46:21) in the manufacturing plant and figure out how (00:46:24) to sell cars on the internet. (00:46:26) He had to build a worldwide network of charging stations. (00:46:30) He had to build a complete ecosystem for electric cars. (00:46:35) If we want to automate hospitals and clinics, (00:46:41) those hospitals and clinics are not (00:46:43) going to be very efficient if the people who (00:46:46) regulate those hospitals and clinics are not also automated. (00:46:51) (00:46:56) If the patients who are making appointments or receiving (00:47:03) the results of a blood test and all of it-- if the patients (00:47:07) not also have access to that automation technology, (00:47:13) you have to automate the patient, the provider, (00:47:17) the payer, the regulator, the pharma companies, banks (00:47:21) who finance the hospitals, and governments (00:47:23) who regulate the hospitals and collect information (00:47:27) from the hospitals. (00:47:28) You have to automate the entire ecosystem. (00:47:31) That then, you will get a truly modern, efficient health care (00:47:35) system. (00:47:35) And that's what we were after when we bought Cerner (00:47:39) as a first step. (00:47:41) Anyway, one of the most interesting AI agents (00:47:44) we've ever built connects providers to payers (00:47:51) because this is a very interesting problem. (00:47:54) And it took me a while to fully grasp this problem (00:47:57) when we were working on this. (00:47:59) And the best possible care-- (00:48:04) what do we want the hospital to do? (00:48:06) The hospital has to figure out, what (00:48:09) is the best possible care I can give this patient? (00:48:12) (00:48:15) Well, that's kind of true. (00:48:20) But let's say you're in the UK. (00:48:23) And the best possible care said that you have high blood sugar. (00:48:27) And I've got to put you on Ozempic or another GLP-1. (00:48:32) Well, guess what. (00:48:35) The NHS in the UK doesn't pay for Ozempic. (00:48:38) They won't reimburse you for it. (00:48:40) And it's very expensive. (00:48:44) So are there any other drugs that (00:48:45) will help you manage your blood sugar levels? (00:48:47) Yes, there are. (00:48:49) And are they pretty good? (00:48:51) Yes, those drugs are pretty good. (00:48:53) And will the NHS reimburse you for those? (00:48:56) Yes, they will. (00:48:59) So what you're really doing when you're automating (00:49:02) a hospital in the UK-- (00:49:05) what you're doing is you're trying to work with the doctor (00:49:11) to come up with the best possible quality of care that (00:49:16) is fully reimbursable if the patient can't afford (00:49:21) to pay themselves. (00:49:24) So those two things are tightly coupled together. (00:49:26) (00:49:29) So it's pointless to prescribe Ozempic to someone (00:49:34) in the UK who can't afford it because the government is (00:49:39) the insurance company in the UK. (00:49:40) And NHS doesn't pay for Ozempic. (00:49:43) It's true today. (00:49:45) So this is what we had to build. (00:49:46) And we had to build something that worked in the United States (00:49:49) and in the UK and all over the world and solve this problem. (00:49:52) The problem was the best possible care (00:49:55) that's fully reimbursable. (00:49:57) That's what our goal was. (00:49:58) So the AI model that we built first (00:50:02) used RAG to access the latest medical literature (00:50:07) and your latest test results in the EHR, vital signs, and all (00:50:13) of that information, all your blood tests, (00:50:16) to assist the doctor to come up with the best possible care. (00:50:21) And we had to things like, well, there's (00:50:23) a new clinical trial for this particular type of cancer (00:50:26) that applies to this patient, that the doctor should consider (00:50:29) putting this patient in that clinical trial. (00:50:32) So the AI model, not surprisingly, (00:50:37) will have all of the latest information (00:50:39) about clinical trials, which drug is working better (00:50:46) than the other drugs for this particular patient the doctor is (00:50:50) looking at. (00:50:51) So we'll provide information to doctors. (00:50:55) The AI model will provide information to doctors (00:50:58) as the doctor tries to figure out the best possible care (00:51:01) for the patient. (00:51:03) Then the AI model is also trained, (00:51:07) uses RAG to access the latest rules and policies. (00:51:10) Now, in the United States, those would (00:51:12) be insurance policies and rules depending (00:51:14) what insurance do you have. (00:51:16) Do you have Medicare? (00:51:16) Or do you have Medicare or Medicaid? (00:51:18) Do you have supplementary insurance? (00:51:20) What are all the different things you have? (00:51:22) So I got to figure out what is covered, (00:51:24) what do you get reimbursed And it's (00:51:27) really those intersecting sets. (00:51:30) What's the best care? (00:51:32) What's fully embraceable? (00:51:34) So I have to train the model on all of the insurance rules (00:51:40) to make sure that what the doctor is prescribing (00:51:43) is fully reimbursable. (00:51:46) And I've got to catch little snags along the way. (00:51:49) Well, actually I do reimburse for Ozempic in the UK (00:51:55) if your body mass index is beyond this point. (00:52:00) And I've got to make sure that the doctor knows that. (00:52:04) And I can let the doctor know-- actually, (00:52:08) this case is an exception. (00:52:10) This patient is eligible for Ozempic (00:52:13) because they're overweight to past a certain threshold. (00:52:17) And the rule that was just changed (00:52:20) says that they now can get Ozempic. (00:52:23) I had to do that. (00:52:26) So the AI agent then reasons with all of this data (00:52:32) to propose the best possible care at the highest (00:52:38) reimbursement level achievable. (00:52:41) That's the goal of it in most places (00:52:46) in the world where the government is (00:52:48) the payer of health care. (00:52:52) And the one last thing that we also did-- (00:52:55) and we have examples of this-- (00:52:57) that we've experienced where a lot of clinics, (00:53:02) a lot of hospitals in the world, including in the United States, (00:53:08) don't have lots of cash on hand. (00:53:10) And if they haven't gotten the reinsurance reimbursements (00:53:14) on time, sometimes they can't provide care to new patients. (00:53:21) They're just running short of cash all the time. (00:53:24) And what the AI agent can do here (00:53:29) is give the bank all of the information (00:53:33) about a particular collection of reimbursements, (00:53:37) assuring the bank that those reimbursements will (00:53:40) be adhered to all of the reimbursement rules. (00:53:44) And the clinic and the hospital will, in fact, be reimbursed. (00:53:48) 99% chance, 95% chance, they'll be reimbursed. (00:53:52) You can discount it a little bit. (00:53:54) And the bank will then loan on those receivables. (00:54:01) So it's a fascinating set of problems. (00:54:07) When you look at the health care ecosystem, the financial aspects (00:54:11) of the health care ecosystem, it's very expensive to run. (00:54:15) There's a lot of administrative duties and administrative tasks (00:54:19) that we can automate away using AI (00:54:23) and let patients spend more time with their doctors who (00:54:27) are worried about care. (00:54:29) And we can figure out how to get the highest achievable (00:54:33) reimbursement, how to get the hospital the cash (00:54:37) that they need to continue operating. (00:54:39) But that's all done via automation. (00:54:42) And the doctor's time and the nurse's time (00:54:44) is spent much, much more efficiently with patients. (00:54:46) As I say, AI will make things so much better for all of us. (00:54:52) (00:54:55) So Oracle Cloud is very unusual. (00:54:59) (00:55:04) In the simplest sense, Oracle does infrastructure (00:55:06) and applications. (00:55:08) We do scaled enterprise applications. (00:55:11) And we do scaled AI infrastructure. (00:55:15) And we're the only cloud that does that. (00:55:19) The other big clouds-- (00:55:21) Microsoft, Amazon, and Google-- (00:55:24) really do not do health care applications, (00:55:27) enterprise applications, big financial applications. (00:55:29) They don't do that. (00:55:31) In other words, they develop AI technology. (00:55:34) They may or may not develop AI technology. (00:55:36) Google does. (00:55:38) The other two don't. (00:55:40) They may or may not develop AI technology. (00:55:43) But also, they are not building large, scaled applications, (00:55:47) where they're trying to automate industries or automate (00:55:52) ecosystems using this technology. (00:55:57) So our goals are different than those other clouds. (00:56:00) We're a participant in creating AI technology. (00:56:05) And we're also a participant in using that technology (00:56:08) to solve problems in different ecosystems (00:56:12) and different industries. (00:56:14) And we're obviously very large in training the AI models. (00:56:23) But we have those models, a bunch of those models, (00:56:27) some of which we trained, some of which we didn't. (00:56:29) We have those models in our cloud for you (00:56:35) to use to solve your problems, for you (00:56:39) to do AI reasoning on your private data, (00:56:43) to solve the problems you want to solve at your company. (00:56:46) (00:56:53) We have AI code generators. (00:56:55) Anthropic is-- it's the thing they're most famous for, (00:56:58) Anthropic, is code generation. (00:57:01) We've been doing this for a long time. (00:57:03) We think we have our new APEX code generator. (00:57:09) (00:57:14) One thing I can say about APEX-- (00:57:17) every application it generates is scalable, secure, reliable, (00:57:23) everyone. (00:57:25) And we've been doing that for a long time. (00:57:27) Now, we're doing complete code generation using AI and APEX. (00:57:32) We are the only ones that are building suites of applications (00:57:38) to modernize not just industries but complete ecosystems. (00:57:44) And health care is one example. (00:57:46) But utilities is another. (00:57:49) And we're taking on entire ecosystems, which makes things (00:57:56) work much more efficiently. (00:57:59) I mean, you're only as strong as the weakest link in the chain. (00:58:04) If you have to interact with, let's say, (00:58:08) a regulator that does clinical trials and the clinical trial (00:58:12) regulator says, OK, once you finish your clinical trial, (00:58:16) print out all the results, and send it (00:58:18) to us in boxes of paper-- (00:58:21) and I won't mention any names. (00:58:23) But that happens all over the world. (00:58:29) It makes new drugs incredibly expensive (00:58:32) and take forever to come out. (00:58:34) It's a huge problem. (00:58:37) So you have to automate these entire ecosystems as a goal. (00:58:43) And then agents, you have to build these complex processes, (00:58:49) these robotic pieces of software called AI agents, (00:58:53) that not only automate processes within a company (00:59:00) but also automate processes between companies-- (00:59:03) how one company talks to another company, how (00:59:06) a hospital talks to a bank. (00:59:08) (00:59:13) That's phase 1 of my presentation. (00:59:19) We'll be serving dinner. (00:59:20) [LAUGHTER] (00:59:24) That's why I arrived. (00:59:25) I arrived a little late because this way, (00:59:27) we can go straight to dinner when we're done. (00:59:30) So this is looking at-- (00:59:34) I went into how the AI models work, how they're built, (00:59:39) how Oracle is different. (00:59:40) And I'd like to just take a look at the world (00:59:44) as I think it's going to be because of AI. (00:59:48) And I think by and large, we are going (00:59:51) to live much better lives, healthier, longer lives, (00:59:59) eat better food, live in better houses. (01:00:04) It should be a much better world because these tools (01:00:07) are so enormously powerful. (01:00:09) (01:00:12) But some of the things they'll do is a little bit shocking. (01:00:14) (01:00:19) So these are some of the things we're working on. (01:00:21) I can go through them. (01:00:22) On the line, we're working on biometric. (01:00:26) We can prevent identity theft using AI. (01:00:30) Just stop it. (01:00:31) So no more logging on. (01:00:34) No more passwords that get stolen. (01:00:37) No more intrusions. (01:00:38) No more data that gets stolen. (01:00:41) No more credit card. (01:00:42) No more you have to send in your credit card and get a new one. (01:00:45) (01:00:48) We can make them all credit proof if that's what you want (01:00:53) or fraud proof if that's the kind of credit card you want. (01:00:57) (01:01:01) I don't know of anyone who likes spending time in the hospital. (01:01:04) And the hospitals have figured out, (01:01:06) the sooner they can get you out of the hospital, (01:01:08) the better it is for them also because some of the nastiest (01:01:12) bugs, some of the nastiest pathogens (01:01:14) are lurking in the halls of hospitals. (01:01:16) And the quicker you home, the patient's happier. (01:01:19) And you're safer at home. (01:01:23) So we can build these IoT medical devices (01:01:28) where we can monitor you at home as well as we can monitor you (01:01:31) in the hospital. (01:01:32) And even if you're in an emergency, (01:01:33) you're being transferred back and forth, (01:01:35) the ambulance is also always connected. (01:01:38) So your home, if you had a patient at home, (01:01:42) they're always being monitored by hospital staff. (01:01:44) You've got a patient being transported in an ambulance. (01:01:47) The hospital staff-- there's an audio-video digital connection (01:01:52) between the ambulance and the emergency room. (01:01:57) Diagnostic images-- when AI reads them-- (01:02:02) I remember one time, I flipped my motorcycle upside down. (01:02:04) Don't ask what was I doing. (01:02:08) And I wasn't that young either. (01:02:10) I don't even have that as an excuse. (01:02:13) Anyway, I landed on my right side. (01:02:16) And I broke eight ribs. (01:02:18) I remember going into an MRI. (01:02:20) And they were counting 1, 2, 3, 4. (01:02:22) What are you doing? (01:02:23) I'm counting your broken ribs. (01:02:24) Oh, great. (01:02:26) But I was having an MRI. (01:02:27) But the only thing they did was count my broken ribs. (01:02:31) There was all this other data that that MRI produced. (01:02:34) No one looked at it. (01:02:36) That's always the case when you get one of these scans. (01:02:39) You're looking for one or two things. (01:02:41) And the rest of the stuff, you just ignore. (01:02:44) AI will find it. (01:02:45) We'll find things that no one was looking for. (01:02:48) And plus, it's just more precise and more accurate. (01:02:50) (01:02:53) Actually, if I do this, I'll finish all the slides (01:02:56) on this one page. (01:02:57) So I'm going to just do this. (01:02:59) Identity theft-- we said earlier in the early slides, (01:03:06) AI knows who you are. (01:03:07) We recognize your face, your voice, your fingerprint. (01:03:10) When you log in, sit down at the computer, say, hi, Safra. (01:03:16) What do you want to do today? (01:03:18) (01:03:23) There's no-- passwords are insane. (01:03:26) That's what get stolen. (01:03:27) People write them down. (01:03:28) The fact that your password has to be 17 characters long (01:03:32) with at least two underscores next to each other-- (01:03:37) are you out of your mind? (01:03:39) You think this is a good idea? (01:03:41) The only way I'll ever remember this (01:03:43) is I write it down and put it on a sticky note (01:03:46) right next to my computer. (01:03:49) Why? (01:03:50) This is just idiotic. (01:03:52) So no password. (01:03:53) No passwords. (01:03:54) It's all biometric. (01:03:57) Better for everybody. (01:03:58) Better data privacy. (01:04:00) Credit cards-- if you want them, we (01:04:04) will have optional credit cards that are biometric. (01:04:08) So it's very hard to imitate people. (01:04:14) So this dramatically reduces credit card fraud. (01:04:18) The banks pay for all the credit card fraud. (01:04:21) The banks don't have to pay that. (01:04:22) Your interest rates are going to go down. (01:04:24) It's going to be better for everybody. (01:04:26) It's going to save a lot of money (01:04:28) and keep your data private. (01:04:29) (01:04:31) Patient monitoring-- I mentioned this. (01:04:33) (01:04:36) We're going to have these low cost. (01:04:38) They're going to be so low cost. (01:04:40) We're going to have these fabulous medical devices (01:04:42) that we can mass produce that are higher quality. (01:04:45) But all medical devices should be attached to the internet. (01:04:50) And they should go into a secure database, where only (01:04:54) you and it's your data. (01:04:57) And you can decide who gets to see it, (01:04:59) your doctor or a health professional (01:05:02) who's monitoring your care. (01:05:04) And you keep it private. (01:05:05) But that data is immediately accessible by your doc. (01:05:11) And if your doc has set an alarm, (01:05:12) if your blood pressure drops below a certain threshold (01:05:15) or goes above a certain threshold, (01:05:18) they want to be immediately notified. (01:05:20) You can do all of that. (01:05:21) You're going to get much better health monitoring-- (01:05:24) home, in the ambulance, wherever. (01:05:26) (01:05:30) And as I say, when moving between your home (01:05:36) and the emergency room, the ER doctors (01:05:40) are talking to the EMTs and the ambulance. (01:05:44) And believe it or not, we're building one. (01:05:48) We're actually building these prototypes. (01:05:51) Will we mass produce an ambulance? (01:05:53) I have no idea. (01:05:54) If you told me a couple of years ago (01:05:57) we'd be building billion watt power plants, I would have said, (01:06:05) you need to get more rest. (01:06:07) That's not going to happen. (01:06:08) But yeah, now we're looking at doing this because-- (01:06:12) and the thing is the ambulance is connected and is (01:06:14) loaded with AI. (01:06:15) And it's just a much safer way to transport patients. (01:06:18) (01:06:22) The diagnostic imaging-- my wife was pregnant. (01:06:29) We were living in Hawaii at the time. (01:06:31) And she went in for a sonogram. (01:06:34) (01:06:38) Two things were crazy. (01:06:39) One is the tech took a ruler and was measuring fetal development (01:06:44) with a ruler, measuring how big the skull was (01:06:48) and how long the spinal cord was on the screen of the sonogram. (01:06:52) And I said, whoa, whoa, whoa, whoa, whoa. (01:06:54) That's like a two-dimensional ruler (01:06:56) measuring a three-dimensional shape inside, (01:06:59) floating in a fluid. (01:07:02) Are you kidding? (01:07:03) Who thinks this is a good idea? (01:07:05) (01:07:08) We can do that with AI. (01:07:09) We can do this very accurately with the computer. (01:07:12) Even with primitive AI, we should (01:07:13) have been able to do that. (01:07:15) It then got worse. (01:07:18) We were on the island of Lanai. (01:07:19) And the dock was actually in Honolulu. (01:07:24) And she held up her iPhone to the sonogram screen (01:07:27) so that the doc could see the fetal image on the sonogram. (01:07:31) I'm like, oh my God. (01:07:34) Oh my God, (01:07:36) You can't record this in high resolution (01:07:38) and transmit it digitally. (01:07:41) You're FaceTiming the image over. (01:07:43) What the hell is? (01:07:44) No. (01:07:45) And actually, I remember her saying one thing. (01:07:47) I said to the tech, look, I promise to fix this. (01:07:50) I promise to fix. (01:07:53) This is awful. (01:07:54) I can't believe this is going on. (01:07:58) But of course, AI is 3D vision. (01:08:01) We can measure accurately fetal development on the sonogram. (01:08:07) We, again, find things doctors aren't looking for. (01:08:10) Imaging-- right now, one of our partners (01:08:14) looks at tumor biopsy slides and can (01:08:17) diagnose the cancer from the image in a few minutes. (01:08:23) We're going through the entire process. (01:08:27) Do all the genetic testing and all of these other things (01:08:30) might take a week or two, a week or two of worry (01:08:34) and a week or two without treatment. (01:08:36) And AI is going to allow us to get a response very quickly, (01:08:40) either say you're fine, you're clean, everything is good, (01:08:43) or, no, you need to start this drug right away. (01:08:46) (01:08:50) In both cases, we get better outcomes. (01:08:51) (01:08:56) This is very interesting. (01:08:58) This is a device that we're working on, (01:09:03) which is called a metagenomic testing device. (01:09:06) (01:09:12) Our ability to identify pathogens-- (01:09:15) when someone gets sick, we have a testing methodology (01:09:19) called PCR. (01:09:21) If we suspect, well, you have influenza A or influenza (01:09:24) B or this coronavirus or COVID-19, (01:09:29) we can test for a panel of some number of known respiratory (01:09:33) viruses. (01:09:34) But if you have something that's odd, (01:09:38) it comes up just as PCR negative. (01:09:40) We don't what it is. (01:09:42) (01:09:46) And what we really want to do is genomic testing on that. (01:09:49) But before we can do genomic testing on it, (01:09:51) we have to culture it. (01:09:53) We have to culture it and wait several days. (01:09:56) And it could take a week or two weeks (01:09:58) before we know what you had. (01:10:01) Either it went away or you did. (01:10:03) It was particularly bad. (01:10:05) (01:10:07) This is a new sensor that will simply do gene sequencing. (01:10:15) It will do gene sequencing of everything in the sample. (01:10:18) So you take blood. (01:10:20) And obviously, in your blood are your own genes. (01:10:26) Well, included in your own genes are something called ctDNA, (01:10:32) circulating tumor DNA. (01:10:36) So in everyone's blood, if you have cancer, even (01:10:40) a stage 1, early stage 2 cancer, you (01:10:44) have small fragments of circulating tumor DNA (01:10:49) that we can discover by gene sequencing (01:10:55) everything alive in your blood. (01:10:58) The problem with the circulating tumor DNA-- (01:11:01) and people have been trying to work with it in the past-- (01:11:04) is your immune system will cure a lot of cancers (01:11:09) without you ever knowing you have them. (01:11:14) The immune system clears up a lot of cancers (01:11:16) before you're ever symptomatic. (01:11:19) And if we keep telling you, oh my God, we found this cancer, (01:11:23) we need to start treating you-- in fact, no, we don't-- (01:11:25) your immune system is going to clean (01:11:27) that up, do absolutely nothing. (01:11:30) So the false positives are deadly in this. (01:11:32) However, with AI now, we can look at the fragments (01:11:36) and distinguish between false positives (01:11:39) and a real serious problem, that you should start (01:11:43) treating immediately early. (01:11:45) So this has the promise of giving us (01:11:48) very, very early cancer diagnosis, which everyone knows (01:11:52) leads to a much higher likelihood of a positive outcome (01:11:55) with the cancer. (01:11:57) It also will allow us to find any bacteria, any fungus, (01:12:02) any virus, any living organism that you're (01:12:06) infected with-- any pathogen that you're infected with (01:12:09) and tell you exactly what that pathogen is even if it's novel. (01:12:13) Like, COVID-19 was novel. (01:12:16) So we know how to treat it. (01:12:19) Well, it'll tell if that pathogen is resistant (01:12:24) to certain antibiotics and specifically (01:12:27) which antibiotics it's resistant to (01:12:30) and which antibiotics we should treat you with. (01:12:32) Now, we actually have a partner here (01:12:36) that went on earlier, that talked (01:12:38) about working on that same exact problem, which is very, very (01:12:44) important. (01:12:46) If you imagine this device being a low-cost device (01:12:50) that's in the pathology departments in hospitals all (01:12:53) over the world-- (01:12:54) so we can do this one blood test and find whatever (01:12:57) pathogen you're infected with. (01:12:59) If we had that, we never would have been caught off guard (01:13:03) with COVID 19. (01:13:05) We would have had early warning. (01:13:06) We would have discovered it far before we discovered it. (01:13:13) Those metagenomic sequencers would (01:13:15) be the perfect early warning system for pandemics. (01:13:20) And that's why we're working on them. (01:13:21) And that's why we need them. (01:13:22) (01:13:28) Building all of these medical devices, (01:13:31) building them reliably-- (01:13:33) if you want to put these metagenomic sequencer (01:13:36) in every hospital all over the world or most of the hospitals (01:13:42) all over the world, they can't cost $1 million. (01:13:47) They can't cost $100,000. (01:13:49) You have to make them cost effectively. (01:13:53) You have to mass produce them. (01:13:55) You have to make them in robot factories. (01:13:58) If you make them in robot factories, (01:14:00) you get much higher quality and dramatically lower costs. (01:14:03) (01:14:05) I think we have a video. (01:14:06) (01:14:18) This is a disk where the test-- (01:14:24) you actually put the sample into the disk, spin the disk, (01:14:29) and run all of these tests on the disk. (01:14:31) (01:14:40) Actually, I think that video, when I saw it, (01:14:43) lasted three minutes. (01:14:45) And Maddie told me, no way am I putting that whole video (01:14:48) in your presentation. (01:14:49) [LAUGHTER] (01:14:52) But it is remarkable. (01:14:55) There are no people in the room when the device and the disk (01:15:01) and the disk is being built. Here's another one. (01:15:07) You'll be happy we don't have a video. (01:15:09) We just have a couple of pictures. (01:15:10) (01:15:13) Growing inside reduces the amount of water (01:15:18) that we use to grow food by 90% That in itself (01:15:21) is essential because we are running out of food, by the way. (01:15:24) We're running out of food in the world. (01:15:28) I think in 2050, Africa will be our most populous continent. (01:15:34) Think about that. (01:15:35) Asia is by far. (01:15:37) Asia has India, China. (01:15:38) (01:15:41) Those are big countries with a lot of people. (01:15:45) Africa will be larger. (01:15:46) We need to produce much more food than we currently do. (01:15:50) We're going to run out of water. (01:15:54) We're going to run out of arable land. (01:15:56) We can't keep taking habitat and converting it to farmland. (01:16:00) We have to be much more efficient. (01:16:01) And by growing in greenhouses and moving plants around, (01:16:05) plants only need a lot of room a few weeks (01:16:09) before they're harvested. (01:16:10) Otherwise, they can grow in much more confined areas. (01:16:13) If you can move the plants around, (01:16:15) you use up much less water, much less space. (01:16:18) You save habitat. (01:16:20) If you're growing indoors, you can grow by urban centers. (01:16:24) I mean, I don't suggest you put a greenhouse right (01:16:27) in the middle of New York. (01:16:28) But you can put it 50 miles away from New York. (01:16:32) And you're growing near population centers. (01:16:35) So the CO2 output for transporting the food (01:16:39) to population centers is greatly reduced. (01:16:41) The food is much fresher. (01:16:43) Again, in a greenhouse, there's a harvest every morning. (01:16:48) And it's to deliver the grocery that afternoon. (01:16:51) And it can be eaten that evening. (01:16:54) So the food is much fresher. (01:16:57) It's lower cost. (01:16:58) It's more nutritious. (01:16:59) It's tastier. (01:17:02) And we're actually building these things (01:17:05) in these robotic greenhouses. (01:17:08) And there should be a picture coming up. (01:17:11) Yeah. (01:17:11) That's real. (01:17:12) You just hold that. (01:17:15) As I pointed out to Elon, this is also a martian habitat. (01:17:18) (01:17:21) This building, which is very large, (01:17:25) you can imagine as a greenhouse. (01:17:28) And that yellow thing kind on the lower part is an overbought. (01:17:33) That's a rail system that moves the plants around (01:17:36) from one location to the other. (01:17:38) No human beings are allowed in the growing (01:17:41) area because human beings contaminate the growing area. (01:17:45) We literally lift the plants up and move them into a harvesting (01:17:48) area where people are allowed. (01:17:52) But also, the growing area is very, very high in CO2. (01:17:57) It's very humid. (01:17:58) It's very unpleasant for people. (01:18:00) It's very, very high in CO2, which is good for plants, (01:18:03) not so good for human beings. (01:18:07) But if you took that same building-- and the building, (01:18:10) by the way, there's no structure. (01:18:13) It is an air pressure building. (01:18:14) So the atmosphere-- it's a positive air pressure. (01:18:19) So basically, think of fans keeping (01:18:22) the pressure inside the building as higher (01:18:24) than the pressure outside the building. (01:18:25) And that's what holds up the roof, which (01:18:28) is made of ETFE, which is the most sunlight (01:18:33) transparent material known to man, also quite strong. (01:18:38) (01:18:42) And those are steel cables. (01:18:43) Those are steel cables in the arches (01:18:47) anchored to a concrete footing around the base. (01:18:52) So literally, you have a robot dig the footing. (01:18:56) You snap the steel cables onto the fiducials on the footing. (01:19:01) And then you turn the fan on. (01:19:03) And you inflate the building. (01:19:06) You fold the building up. (01:19:08) The building is fabric with steel cables. (01:19:13) You fold it up in nice packages. (01:19:16) And you transport it to where you're building it. (01:19:18) Or you transport it to Mars on one of those big rockets. (01:19:21) And then Elon can build his house right (01:19:26) in the middle of that and have beautiful (01:19:28) rose gardens and all of that other stuff. (01:19:31) It'll be lovely. (01:19:34) But I'm not going. (01:19:35) [LAUGHTER] (01:19:38) I will to go to this one, which is-- (01:19:41) the first ones are in California and Texas, (01:19:45) which is way closer than Mars. (01:19:48) Here's another picture of the same building. (01:19:50) They're big. (01:19:51) And then the green areas are the harvesting areas. (01:19:54) And the walls lift up where the trucks (01:19:57) arrive to deliver the food. (01:20:00) (01:20:04) This is going to be shocking. (01:20:05) (01:20:13) The first thing we did-- and we've actually done this. (01:20:16) We've actually done this. (01:20:17) It's actually a company that I'm involved with called Wild Bio. (01:20:23) It's part of the Oxford company. (01:20:28) I've got an institute at Oxford called (01:20:30) EIT, the first time I've ever put my family name on something. (01:20:35) EIT. (01:20:36) And one of the companies we have is this company called Wild Bio. (01:20:40) And the first thing they did was they modified (01:20:46) wheat plant, which is a grass. (01:20:49) They modified wheat to have it produce 20% more food per acre, (01:20:54) more grain per acre, which seems like we're running out of food. (01:20:57) That seems like a good idea. (01:20:59) Now, it's really interesting. (01:21:02) If you produce 20% more grain per acre, (01:21:06) what wheat does basically, it takes CO2 and sunlight, (01:21:11) mixes them together to create food. (01:21:13) (01:21:16) So if you're growing more grain, you're consuming more CO2. (01:21:19) (01:21:21) Now, where that CO2 ends up is really-- (01:21:26) if you have AI designing the wheat-- (01:21:29) is really up to us. (01:21:32) So we built this wheat that's much more (01:21:36) efficient with photosynthesis than conventional wheat. (01:21:41) Once we've absorbed the CO2 into the wheat, (01:21:46) we could choose to take that CO2 and convert it (01:21:53) into calcium carbonate. (01:21:56) By the way, that's exactly how coral reefs get built. (01:22:00) A coral reef is converting CO2 and sunlight into a structure, (01:22:10) into an inert mineral called calcium carbonate. (01:22:15) So we grow a lot of wheat around the world every spring. (01:22:22) We plant several Amazon rainforests worth of wheat. (01:22:28) And if you want to, you can not only produce more grain, (01:22:33) you can convert more CO2 directly into calcium carbonate, (01:22:39) therefore removing it from the atmosphere forever. (01:22:42) (01:22:44) So if you want to manage-- (01:22:47) there are all these interesting ideas (01:22:48) on how to manage the climate and manage the atmosphere (01:22:53) and manage atmospheric CO2. (01:22:56) But in this particular case, you can remove-- (01:22:58) if you wanted to go from the current level of 440 parts (01:23:02) per million of CO2 in the atmosphere, which (01:23:06) some people think is too high, and reduce it (01:23:09) to 400 parts per million, you can do that simply (01:23:12) by having the wheat and the corn and the soybeans and whatever, (01:23:21) converting CO2 into calcium carbonate. (01:23:25) And you can manage the CO2 level in the atmosphere to whatever (01:23:30) level you deem appropriate. (01:23:33) And if you think the sweet spot is 400 parts per million, (01:23:37) that's right. (01:23:38) Now, someone will say, no, no, we (01:23:39) want to get rid of all the CO2 in the atmosphere. (01:23:42) Well, pack a lunch because if you get rid (01:23:47) of all the CO2 in the atmosphere, (01:23:49) all the plants will die on the planet. (01:23:52) So don't go to zero. (01:23:53) That's a really bad idea. (01:23:56) But the sweet spot in terms of stabilizing the climate probably (01:24:04) is going from 440 to 400. (01:24:06) And it's something we can do. (01:24:08) And it's basically free. (01:24:10) Basically, there's no cost in doing it. (01:24:14) And it's just a natural process called biomineralization. (01:24:20) And we could use our food crops. (01:24:22) We could actually increase the food yield while lowering CO2. (01:24:26) This is what I mean by AI. (01:24:28) AI is a pretty amazing tool. (01:24:31) There are a lot of problems we can (01:24:33) tackle that we've been unable to solve (01:24:37) for a very, very long time. (01:24:38) And it's very, very contentious within our society. (01:24:43) But you absolutely have the ability to do this. (01:24:45) (01:24:49) Corn-- we're also working on corn. (01:24:54) Another huge problem with agriculture (01:24:56) is nitrogen fertilizer. (01:24:59) You fertilize all these crops to increase the yield. (01:25:03) The problem is fertilizers are made up of nitrogen. (01:25:08) And it rains. (01:25:09) And you've got huge nitrogen runoffs into river basins (01:25:12) and into the ocean. (01:25:14) And that pollution does a lot of damage in our environment. (01:25:21) Rather than using nitrogen fertilizer to nourish the plant, (01:25:29) the atmosphere has got a huge amount of nitrogen in it. (01:25:33) Why don't you simply engineer the plant (01:25:39) to take the nitrogen directly out of the atmosphere? (01:25:42) And we know how to do that. (01:25:44) There's an enzyme in the world called nitrogenase. (01:25:47) And nitrogenase quite literally takes atmospheric nitrogen. (01:25:53) Does it with soybeans, for example. (01:25:55) It's unique to soybeans. (01:25:57) Takes atmospheric nitrogen and uses it (01:26:00) as a nutrient for the plants. (01:26:03) And you don't have to use nitrogen fertilizer. (01:26:05) You can get rid of all the nitrogen fertilizer. (01:26:08) In Africa, no one can afford-- (01:26:09) I shouldn't say nobody. (01:26:10) A lot of farms can't afford to use nitrogen fertilizer. (01:26:13) But even the ones that can afford used nitrogen fertilizer, (01:26:17) it's a waste of money. (01:26:18) And it is damaging to the environment. (01:26:20) So you can engineer the plant to get the nitrogen directly (01:26:25) from the atmosphere. (01:26:27) And the plant is just as tasty and just as nutritious (01:26:30) and just as healthy. (01:26:31) Getting the nitrogen from the atmosphere (01:26:33) is getting the nitrogen from fertilizers (01:26:36) that's been added to the soil. (01:26:41) Another problem AI makes it easy for us to solve. (01:26:44) (01:26:46) You're going to be very happy that last slide-- this is (01:26:48) my last slide with words on it. (01:26:53) [LAUGHTER] (01:26:56) I have one more video, one more picture. (01:26:58) And then the three of you who are going to stay (01:27:01) can ask questions. (01:27:02) [LAUGHTER] (01:27:05) So autonomous drones-- well, anyone who's looked, (01:27:12) we've seen the drones have been developed in Ukraine (01:27:16) for military purposes. (01:27:17) Fortunately, drones have very wonderful uses (01:27:21) beyond how they're being used in Ukraine, the war in Europe, (01:27:28) which is just terrible. (01:27:29) (01:27:32) We built an air traffic control system for drones. (01:27:35) And we're actually using drones to deliver blood samples (01:27:41) from clinics and taking the blood sample by drone (01:27:46) to testing laboratories. (01:27:48) And we built what we call an RFID specimen vault, which (01:27:52) we put an RFID tag on, which identifies-- (01:27:55) so no one knows this is Larry Ellison's blood or whatever. (01:28:00) They just know there's an RFID tag on the blood. (01:28:04) And then the test results go into the cloud. (01:28:08) And eventually, they make it back to my doctor (01:28:11) and to me, the results. (01:28:13) But otherwise, in the chain of custody, (01:28:16) no one can distinguish-- (01:28:19) my personal privacy is not compromised at all (01:28:22) by doing this. (01:28:23) But also, the other problem is sometimes (01:28:26) they do a great job of protecting your personal privacy (01:28:29) by losing your blood sample or thinking it was somebody else's (01:28:35) blood sample. (01:28:36) That's not a great way to protect our personal privacy. (01:28:39) So we built this specimen. (01:28:41) Another thing we built are these specimen vaults (01:28:46) to take samples from the hospital, (01:28:48) from the clinic to the lab, where the results then (01:28:52) go into the cloud. (01:28:54) But the other thing that drones can do (01:28:58) is they can detect forest fires immediately (01:29:00) with infrared cameras. (01:29:03) They can even figure out who set the forest fires. (01:29:05) Tragically, the Palisades fire-- a number (01:29:08) of the fires in California were set by arsonists. (01:29:11) I mean, unbelievable tragedies. (01:29:14) But we can detect the immediately (01:29:18) and start to fight the fire immediately. (01:29:21) And if someone set the fire, we can figure that out too. (01:29:27) And we shouldn't have police cars chasing other cars (01:29:31) around those high-speed chases. (01:29:32) While the videos look kind of cool, (01:29:35) they are very dangerous for not just the police. (01:29:39) But for civilians and cars nearby, we (01:29:43) can have drones follow those cars. (01:29:45) It's way better. (01:29:47) I'm going to now go to my last picture. (01:29:50) That's the RFID specimen vault over there. (01:29:54) And last video will be coming up. (01:30:00) There it is. (01:30:03) Sure enough. (01:30:03) (01:30:11) So you can deploy these. (01:30:12) It'd be great in the Palisades. (01:30:15) It's the dry season. (01:30:17) You send the drones up. (01:30:18) You can have a series of these cars. (01:30:20) You've got a lost hiker out in the wilderness, something (01:30:25) like that. (01:30:25) They're portable. (01:30:26) I think it's going to now land. (01:30:28) And then if it gets down safely, I will take my first question. (01:30:32) (01:30:37) It's a video. (01:30:38) It's going to get down safely. (01:30:39) [LAUGHTER] (01:30:41) (01:30:46) Awesome. (01:30:46) And that's a charging station. (01:30:48) (01:30:51) [MUSIC PLAYING] (01:30:55) (01:31:07) DAWN TITTENSOR: Wood PLC is a large global consulting (01:31:11) and engineering company. (01:31:13) We operate across 60 countries. (01:31:15) And we have an employee workforce of 36,000. (01:31:19) As a function, we were inefficient and costly. (01:31:22) And that was across our IT, our HR, and our finance functions. (01:31:26) We had dozens of HR systems, very disparate applications (01:31:32) landscape. (01:31:33) We had lots of different ways of working, different processes. (01:31:39) And we found it very, very difficult to pull together (01:31:42) our employee data and reporting. (01:31:45) One of the key impacts was the ability (01:31:49) to make it easy to work at Wood, to engage with our employees (01:31:55) in a more unified way. (01:31:57) And also, as I say, that self-service capability (01:32:01) in the flow of work wasn't there. (01:32:03) So we chose Oracle Fusion Cloud HCM because it covered (01:32:07) across our people processes. (01:32:10) It also really focused on that employee and candidate (01:32:14) experience, which was important to us. (01:32:17) Moving to SaaS software allows Oracle to do all the hard work (01:32:21) and for us to leverage and move at pace with innovation (01:32:25) across our people processes. (01:32:26) We are live with Code HR, Talent and also (01:32:31) Workforce Comp and Learn. (01:32:33) So those were the first modules we rolled out. (01:32:36) We have since also brought in Oracle Recruit (01:32:40) with Recruit Booster. (01:32:41) We're leveraging journeys for onboarding. (01:32:44) We've also just recently implemented (01:32:47) Help Desk and Digital Assistant, which really gives us a full end (01:32:51) to end for our people processes. (01:32:55) We have enabled our AI features within Oracle Recruit and also (01:33:00) our performance management. (01:33:02) We have recently gone live with AI assist in performance goals. (01:33:07) I'm delighted to say the feedback has been incredible. (01:33:11) We have reduced our time to hire. (01:33:14) So pre-AI, for our trading craft population, we were 45 days. (01:33:22) Post-AI, we are sitting at 21 days. (01:33:26) That time to hire for our reimbursable roles (01:33:30) is a key metric and a key target. (01:33:33) If we don't have a role filled, we can't then build our clients. (01:33:37) AI agents are going to help us really drive (01:33:41) the adoption of self-service. (01:33:43) It will support our employees and managers, (01:33:46) finding the answers and giving them (01:33:48) guidance in the flow of work. (01:33:51) So no longer do they need to send an email to HR, (01:33:54) reach out to our shared services, wait for a response. (01:33:58) They are in the application. (01:34:00) They have the guidance and what they need there. (01:34:03) And they can, again, focus on the value (01:34:07) add activity of their work. (01:34:10) The support that we get from Oracle and the partnership (01:34:12) is second to none. (01:34:13) The journey never ends with digital transformation. (01:34:16) It's continuously evolving. (01:34:18) And I think partnering with Oracle, (01:34:22) they put the customer first. (01:34:25)