 Hi everyone, it's great to be here. Thank you so much. I'm here to help paint a picture of what you're seeing at this conference and how it ties back to all the incredible innovation that seems to have burst on the scene in the last year in AIML. My goal is to help that you think about how all these innovations and related tools that you might use to solve the biggest human needs that inspire to build these impactful solutions. To truly innovate, we always need to root our innovation back into the problems that we're solving. And so great ML solutions like any other solution ideally start with design thinking. Who's your user? What are their needs and how will solving those needs impact their lives? So right now is an incredibly exciting time for machine learning whether you're just getting started or a seasoned practitioner. Large model integrations and applications have exploded from larger companies through to startups and advancements across the industry range from low code offerings to those as well as those for more advanced practitioners. Now, so many of you are here today because of large language models. So let me walk you through a super short overview. Large models and specifically large language models through generative AI have fundamentally changed technology in the last year or so. It's not just the technology, but the ease of access, the natural conversational style of many of the popular applications and several other underlying factors that have made LLM seem to burst onto the scene. Whatever the underlying magic though, this shift has helped quickly unlock solving big human problems and in a variety of ways that many of us haven't even figured out yet. And so even if large models though don't solve everything, the majority of our critical ML solutions are still running with traditional ML. It's still important to understand that all the underlying advancements have helped make LLM's possible because at the end of the day, all this machine learning is inspired by deep human needs. So let me walk through a quick path from the Google side. You know, at first traditional programming consisted of providing a specific set of instructions for a computer to execute. You had to know precisely which set of actions to follow to get from input to output. So you could teach the computer something in the old days like cat type, animal, legs four, ears two, for usually likes, yarn, catnip, naps, and ultimate power. Well-defined inputs, well-defined outputs. Then neural networks came around in roughly 2012. We started training AIs with a bunch of data, like a pile of animal pictures, teaching the AI which ones were pictures of cats and which were dogs or other animals or just not a cat. After a certain amount of training, the AI became able to figure out which pictures were cats with great quality. The model is single task model, trained to solve a specific problem. But sometimes the outputs could surprise us in good ways, and there was a lot of increased flexibility in the types of inputs we could use. Then this evolved into large language models or LLMs, which ML algorithms that can help recognize, predict, and generate human languages. LLMs are normally pre-trained for multiple tasks on a large corpus of text, followed by fine tuning on a specific task, like if you gave a whole encyclopedia to a child and then worked directly with the child to figure out how to do specifically well on questions for cats. And for this training, though, you had to think on terabyte scale, resulting in large models with tens of billions of parameters to the weights that the model learned during training and used to predict the next token in the sequence. So now we have a system where we can use really general inputs and get human-like outputs across a wide variety of questions. And so when you take a look behind the scenes of this path, you can see that this large model technology has been gaining momentum for years. And our own path is illustrative to understand where we are now. For the purposes of this chat, I'm going to zoom past a bunch of stuff from Google towards 2017 and the introduction of transformers, an architecture designed around the idea of attention. This architecture made it possible to process longer sequences by focusing on the most important part of the input solving memory issues that have been encountered in earlier models. Taking us back to everyday life, you can imagine that this state-of-the-art architecture can also be incredibly useful at solving everyday problems like translation. Or AlphaFold, developed by our DeepMind team, which allows access to over 200 million protein structures or predictions to dramatically accelerate scientific research. And several more impactful models started emerging quickly, one after another, powering incredible products here at Google from T5 to Lambda to Palm to Gemini, our newest model. Now, when thinking about LMS, it's not just thinking about the architecture. There's a number of underlying advancements that I'm going to talk a little bit about across ML software, from computational power to ML techniques to training data to access of innovation that are all relevant to making this new innovation possible. So first, advancements in computational power. So one of the most important things that you need to understand at a high level is the key infrastructure that ML runs on. Specifically, when you're trying to crunch a very, very large amount of data, you need a whole lot of computing power. Computing resources have become more powerful, such as with GPUs and our own TPUs, first launched in 2015, where you can do a much more complicated processing much faster than in traditional processing units like CPUs. And the key takeaway on this is that it takes a lot of power to run these large models and even more to train and crunch all the vast amounts of data that are required to get them to a point where they're ready for primetime consumption. And then data. So data is really key to everything. Without data, we know that models simply can't learn. And as we get better at collecting and analyzing large amounts of data, be it structured data or unstructured data like audio or video files, the model performance can improve dramatically. But for you as you develop, figuring out what type of data your customers need or have gets really complicated quickly. Do you need just text to better understand and generate natural written language? What about when it becomes dialogue and spoken conversation? Or we need to communicate with computers through code, or maybe you need image or audio? Now, I want to pause to call out that with more data becomes more responsibility. And it's important to put that responsibility at the core of all product creation that we do in AI. So next, advancements in techniques. So over the last several years, significant advancements in some of these techniques to train the models have also led to some of these large leaps in performance. We covered more power, which means more data crunching and the ability to handle more diverse types of data. But it's not just about training on more data, but the ability to train faster so that we can have faster feedback loops, meaning the ability to incorporate higher level and higher judgment evaluations of the outputs of the models. Specifically, techniques like reinforcement learning with human feedback or RLHF has been critical as models are learning from vast amounts of unstructured data like image audio, etc., which require a much higher judgment in knowing how to evaluate the outcomes. Computers in the traditional ML world have only been able to handle simple questions like, is this a hot dog? Yes or no? But with faster iterations of humans providing real time feedback, we can create more nuanced, expert opinions to help guide the models to better judgment like determining which corgi is the cutest, most dramatic and corgi-tastic. And with the increased availability of open-source models, as we heard from our last speaker, on all sizes of platforms from Kaggle, the hugging face, etc., and open-source frameworks like TensorFlow, Jack's, Keras, and PyTorch, more and more developers have been able to learn more quickly from each other, innovating faster on the models themselves, the data, the techniques, and the fine-tuning for it to support a wide variety of use cases solving deep customer needs. This innovation, the overall community expertise has been critical in supporting the acceleration of innovation around large models. And easier access comes through no-code platforms like LMChat interfaces like Bard or ChatGPT, which dramatically open up access to everyday customers to start using and being inspired by large models without needing to know how to code. It made it as simple as typing whatever you're thinking in whatever language you think in to be able to interact with these LLMs. This has meant that the large models that were previously only available to researchers with deep technical experience and access to large amounts of compute are now available to almost anyone. Because it's been so much easier to access LLMs and build with them, we've seen rapid integration of these LLMs in applications solving an adverse set of customer problems across a wide variety of tasks critical to our lives, ranging from assistance to coding to translation, summarization, generation, and more. And this means that ML running behind the scenes is just that much more approachable for people. Adding to this a proliferation of smart devices over the last decade has also meant that customers now have devices like phones, wearables, and IoT with them almost all the time anywhere in the world. This popularity has been, in part, powered by increased advancements in device machine learning, including critical techniques around model optimization and quantization. These capabilities are often baked into the frameworks themselves, whether it's like TFLite or MediaPipe, so the developers can meet end users' demands for smarter and smarter applications. On the device that they have them with them all the time means there's greater need for advanced ML capabilities than ever. And with diverse applications and services bring diverse responsibility as you look to build. As we get better at collecting and analyzing large amounts of data, think about how these practical applications can be centered around key human needs and how can we be responsible. We'll run into a diverse set of opportunities to develop this responsibility, each with their own nuance. And many of the answers to what responsible needs are still being weighed. So it's critical for all of us to keep educating ourselves, staying informed on the latest guidelines, and I know I'm always asking if it's good enough. So what are some of the specific applications out there using AIML? I'm sure you all know of a whole bunch. I'll give you a few from my day job. We've launched all sorts of products that are powered by LMS from Bard and Duet AI, which is your AI-powered collaborator across Google Cloud to help you get more done faster. And this technology also powers a lot of consumer features you probably know and love, like real-time editing features and photos all the way through to the incredible effects in meat. And developer products like Vertex AI, Kaggle, Visual Blocks, and Keras 3.0, which is something I'm very excited about, a new rewrite of the Keras APIs that let you run their workflows on either Jax, TensorFlow, or PyTorch, unlocking brand new large-scale model training and deployment capabilities enabling multi-framework machine learning. We've leaned in to make sure that RML works great on all your devices so that you have that little bit of magic in your pocket, on your wrists and your ears, and more. And as of last week, of course, and even more being announced today, Gemini. Gemini is the next phase in our journey around making AI more helpful for everyone. It's the most capable and general model that we've built, and it's the result of a large-scale collaborative effort by teams across Google. And it's state-of-the-art capabilities will significantly enhance, we believe, the ways that developers and enterprise customers build and scale with AI across multiple different model sizes. And while I love all that, Mel, I'm particularly passionate about devices, which is why I'm actually really excited for the smallest-sized model, the Gemini Nano, optimized for mobile, bringing that power of large models with you anywhere, anytime. And the reason this matters is running on device enables features where the data should not leave the device, such as suggesting replies to messages and to encrypted messaging app, and it enables consistent experiences with deterministic latency. So features are always available, even when there's no network. Anyway, bringing it back to you. We're excited about the amazing possibilities of a world responsibly powered by AI, a future of innovation that will enhance creativity, extend knowledge, advance science, and transform the way billions of people live and work throughout the world. And so one parting thought, just don't forget to think about all the cool and core human needs that you're building for and you're building these solutions to address. Whether that's how you might use AI ML to find a safer path to Mars, or figure out what's for dinner, or just bring a little more fun into your daily life. And we're in a great world of possibility where you can make a brainy day a little bit brighter, whether that's literally or figuratively. And I can't wait to see what you all build. Thank you so much.