Google Keynote (Google I/O ‘23)

Melhorando a vida das pessoas com IA

Nesta seção, o palestrante fala sobre como a IA pode melhorar significativamente a vida das pessoas e transformar as maneiras como podemos ajudar. Ele também destaca que embora a ambição seja ousada, a abordagem será sempre responsável.

A nova era da IA

• A IA está tendo um ano agitado e temos muito o que conversar

• Temos a oportunidade de tornar a IA ainda mais útil para as pessoas, empresas e comunidades

• Com uma abordagem ousada e responsável, estamos reformulando todos os nossos principais produtos

Exemplos de como a IA generativa nos ajuda

Gmail: Help Me Write

• O Help Me Write é um modelo generativo avançado que extrai informações do e-mail anterior para criar um rascunho completo

• Ele será implementado como parte da atualização do Workspace e melhora com o tempo

Maps: Street View

• Desde os primeiros dias do Street View, a IA uniu bilhões de imagens panorâmicas para que as pessoas explorem o mundo usando seus dispositivos

O poder da IA nos produtos Google

Visão geral da seção: Nesta seção, o palestrante fala sobre como a IA está sendo usada em vários produtos do Google para ajudar os usuários.

Google Maps: Immersive View para rotas

• O Immersive View permite que os usuários vejam toda a sua viagem com antecedência.

• Ao clicar no Immersive View, é possível ter uma visão panorâmica incrível do passeio.

• É possível verificar a qualidade do ar e o tráfego e clima nas próximas horas.

• O Immersive View para rotas será lançado em 15 cidades até o fim do ano.

Google Fotos: Magic Editor

• A Borracha mágica usa fotografia computacional com tecnologia de IA para remover distrações indesejadas.

• Com a nova experiência chamada "Magic Editor", é possível ajustar várias coisas em uma foto, como iluminação e remoção de objetos indesejados.

• O Magic Editor recria automaticamente as partes da imagem que não foram capturadas na foto original.

Conclusão

• A IA pode ajudar os usuários em momentos importantes nos produtos do Google.

• O objetivo do Google é organizar as informações do mundo e torná-las universalmente acessíveis e úteis.

Tornando a IA útil para todos

Nesta seção, o palestrante fala sobre como tornar a IA útil para todos e apresenta quatro maneiras importantes de fazer isso.

Quatro maneiras de tornar a IA útil para todos

• Melhorando o conhecimento e aprendizado da IA e aprofundando sua compreensão do mundo.

• Estimulando a criatividade e produtividade para que as pessoas possam se expressar e realizar tarefas.

• Permitindo que desenvolvedores e empresas criem produtos e serviços próprios e transformadores.

• Criando e implantando IA de forma responsável para que todos possam se beneficiar igualmente.

PaLM 2: O mais recente modelo em produção

Nesta seção, o palestrante anuncia o PaLM 2, um novo modelo em produção, explica suas características principais e como ele pode ser usado.

Características do PaLM 2

• Altamente capaz para uma ampla gama de tarefas.

• Fácil de implantar.

• Oferece excelentes recursos básicos em diversos tamanhos.

• Mais avançado em lógica e raciocínio graças ao amplo treinamento em temas científicos e matemáticos.

• Treinado em textos multilíngues abrangendo mais de 100 idiomas.

• Pode ajudar os desenvolvedores a colaborar em todo o mundo.

Exemplos de modelos PaLM 2

• Sec-PaLM: ajustado para casos de uso de segurança, usa IA para detectar melhor scripts maliciosos e pode ajudar os especialistas em segurança a entender e resolver ameaças.

• Med-PaLM 2: ajustado no conhecimento médico, alcançou uma redução de 9x no raciocínio impreciso quando comparado com o modelo, abordando o desempenho de especialistas clínicos que responderam ao mesmo conjunto de perguntas.

Pré-lançamento do PaLM 2

Overview: Nesta seção, Sundar Pichai anuncia o pré-lançamento do PaLM 2, que é o passo mais recente na jornada de uma década para levar a IA de forma responsável a bilhões de pessoas. Ele também fala sobre as equipes envolvidas no projeto e como estão trabalhando para criar sistemas mais capazes com segurança e responsabilidade.

Anúncio do pré-lançamento do PaLM 2

• O PaLM 2 está disponível em pré-lançamento.

• É o passo mais recente na jornada de uma década para levar IA de forma responsável a bilhões de pessoas.

• O projeto se baseia no progresso feito por duas equipes de alto nível: a Brain Team e a DeepMind.

• Essas equipes contribuíram para um número significativo de avanços decisivos da IA na última década, incluindo AlphaGo, Transformers e modelos Sequence-to-Sequence.

• As equipes foram reunidas em uma única unidade, o Google DeepMind, utilizando os recursos computacionais do Google para criar sistemas mais capazes com segurança e responsabilidade.

Gemini - Modelo de fundação multimodal

• O Gemini é um modelo de fundação multimodal que ainda está em treinamento.

• Embora esteja em etapa inicial, vemos capacidades multimodais impressionantes não vistas em modelos anteriores.

• Uma vez ajustado e rigorosamente testado quanto à segurança, o Gemini estará disponível em vários tamanhos e capacidades.

Responsabilidade da IA

• O Google investe profundamente na responsabilidade da IA.

• Duas abordagens importantes para identificar conteúdo gerado sinteticamente são marcas d'água e metadados.

• Estão sendo criados modelos para incluir marca d'água e outras técnicas desde o início.

• Garantiremos que todas as imagens geradas por IA tenham metadados.

Bard - Experimento de IA conversacional

Overview: Nesta seção, Sundar Pichai fala sobre a evolução do Bard, um experimento de IA conversacional. Ele destaca como o PaLM 2 melhorou as habilidades do Bard em matemática, lógica e raciocínio, permitindo que ele ajude os desenvolvedores com programação.

Evolução do Bard

• A equipe tem trabalhado arduamente para fazer melhorias rápidas no Bard.

• Com o PaLM 2, as habilidades de matemática, lógica e raciocínio do Bard deram um salto enorme.

• O Bard agora pode colaborar em tarefas como geração de código, depuração e explicação de snippets de código.

• O Bard já aprendeu mais de 20 linguagens de programação.

Exemplo de uso do PaLM 2

Nesta seção, é apresentado um exemplo de como o PaLM 2 pode ser usado para adicionar comentários em código em outro idioma.

Adicionando comentários em coreano ao código

• O PaLM 2 reconhece que o código é recursivo e sugere uma correção.

• Ele explica a razão por trás da correção sugerida.

• O modelo adiciona comentários em coreano ao código, conforme solicitado.

Fine-tuning do PaLM 2 para casos específicos

Nesta seção, são apresentados exemplos de como o PaLM 2 pode ser ajustado para casos específicos.

Sec-PaLM

• Versão do PaLM 2 ajustada para casos de segurança.

• Usa IA para detectar scripts maliciosos e ajudar especialistas em segurança a entender e resolver ameaças.

Med-PaLM 2

• Ajustado com conhecimento médico.

• Redução de 9x no raciocínio impreciso quando comparado ao modelo original.

• Desempenho se aproxima dos especialistas clínicos que responderam às mesmas perguntas.

• Primeiro modelo de linguagem a atingir nível especialista em exames médicos.

Gemini - Próxima geração do modelo foundation

Nesta seção, é apresentada a próxima geração do modelo foundation chamada Gemini.

Características do Gemini

• Multi-modal e altamente eficiente nas integrações API.

• Construído para permitir futuras inovações, como memória e planejamento.

• Ainda em treinamento, mas já apresenta capacidades multimodais impressionantes.

Investimento em responsabilidade de IA

• Ferramentas para identificar conteúdo gerado sinteticamente.

• Abordagens importantes incluem marca d'água e metadados.

• Modelos estão sendo construídos com técnicas de marca d'água desde o início.

• Metadados permitem que os criadores de conteúdo associem contextos adicionais aos arquivos originais.

Bard - Experimento para IA conversacional

Nesta seção, é apresentado o experimento Bard para IA conversacional.

Evolução do Bard

• Suporte a uma ampla gama de capacidades de programação.

• Melhorias na resolução de problemas matemáticos e raciocínio lógico.

• Agora totalmente executado no PaLM 2.

Programming Language Options

In this section, the speaker discusses various programming languages that Bard can use.

Programming Languages

• Bard can use several programming languages including C++, Go, JavaScript, Python, and Kotlin.

Visualizing School Locations on Google Maps

In this section, the speaker demonstrates how to visualize school locations on Google Maps.

Mapping Schools

• The speaker uses Google Maps to visualize where schools are located.

Organizing School Options in a Table

In this section, the speaker shows how to organize school options in a table.

Creating a Table

• The speaker creates a table to organize school options.

Adding Public/Private Column to School Table

In this section, the speaker adds a column showing whether schools are public or private.

Adding Column to Table

• The speaker adds a column showing whether schools are public or private.

Moving School Table to Google Sheets for Collaboration

In this section, the speaker moves the school table to Google Sheets for collaboration with family members.

Moving Table

• The speaker moves the table from Bard to Google Sheets so family members can collaborate on it later.

Unlocking Creativity with Large Language Models

In this section, the speaker discusses how large language models like Bard can unlock creativity and productivity.

Advancements with Large Language Models

• Large language models like Bard can unlock creativity and productivity.

• Partnerships with companies like Instacart, Indeed, and Khan Academy will allow Bard to tap into all kinds of services from across the web.

• Adobe Firefly will allow users to generate completely new images from their imagination right in Bard.

Expanding Bard's Availability

In this section, the speaker announces that Bard is becoming available in more countries and languages.

Availability

• Bard is becoming available in over 180 countries and territories.

• Bard is becoming available in Japanese and Korean.

• Support for 40 languages is coming soon.

AI as a Collaborator

In this section, the speaker discusses how AI can act as a collaborator in Workspace.

Collaboration with AI

• AI can act as a coach, thought partner, source of inspiration, and productivity booster across all apps of Workspace.

• Help Me Write feature in Gmail and Docs allows users to collaborate with AI as a writing assistant.

• Job descriptions can be improved using Docs' helpful features.

Starting Point

In this section, the speaker talks about having a good starting point written out.

Having a Good Starting Point

• A good starting point is important.

Contextual Collaboration

In this section, the speaker talks about contextual collaboration and how it can boost productivity and creativity.

Examples of Contextual Collaboration

• Generating images for fun.

• Writing to neighbors about an upcoming potluck.

• Creating speaker notes for an important presentation.

Bringing Generative AI in Search

In this section, the speaker talks about bringing generative AI into search.

Generative AI in Search

• Using generative AI to answer complex questions.

• New integrated search results page with an AI-powered snapshot that quickly gives you the lay of the land on a topic.

• Links included in the snapshot to dig deeper and explore more details of the topic.

National Park Service

The speaker mentions the National Park Service as a source of expertise.

Expertise from the National Park Service

• The speaker mentions that experts, like the National Park Service, can provide valuable insights.

Search Generative Experience

The speaker introduces a new search generative experience called SGE and invites users to join the waitlist.

Introduction to SGE

• Google is experimenting with generative AI in Search.

• A new Search generative experience, also known as SGE, will be available in Labs.

• Users can join the waitlist today by tapping the Labs icon in the latest version of the Google app or Chrome desktop.

Hot Dog Sandwich Debate

Sundar Pichai weighs in on whether a hot dog is a sandwich.

Is a Hot Dog a Sandwich?

• Sundar Pichai answers the question "Is a hot dog a sandwich?" and says it's more like a taco because the bread goes around it.

AI for Business Transformation

The speaker discusses how AI can help businesses transform and innovate.

AI for Business Transformation

• AI is not only a powerful enabler but also a big platform shift.

• Every business and organization is thinking about how to drive transformation.

• Google is focused on making it easy and scalable for others to innovate with AI.

Google Cloud

The speaker discusses how Google Cloud is using the power of AI to transform the way people work.

Transforming the Way People Work with AI

• Google Cloud is using the power of AI to transform the way people work.

• Thousands of companies are already using Google's generative AI platform to create amazing content, synthesize and organize information, automate processes, and build incredible customer experiences.

• There are three ways that Google Cloud can help businesses take advantage of the massive opportunity in front of them.

Building Generative Applications with Vertex AI

• Businesses can build generative applications using Google's AI platform Vertex AI.

• With Vertex, businesses can access foundation models for chat, text, and image.

• They can select the model they want to use, create prompts to tune the model, and even fine-tune the model's weights on their own dedicated compute clusters.

Enterprise Search

• To help businesses retrieve fresh and factual information from their company's databases, corporate intranet, website, and enterprise applications, Google offers Enterprise Search.

• The platform guarantees privacy of data so that businesses have sole control over their data and costs when using generative AI models. In other words, your data is your data and no one else's.

Alternative AI Technologies

• Many leading companies are using alternative AI technologies to build super cool applications.

Conversations with our Customers

In this section, the speaker talks about PaLM 2, a language model that powers the PaLM API. The speaker also discusses how developers have used the PaLM API in their applications and how it has been integrated into Google Developer tools.

PaLM 2 and the PaLM API

• PaLM 2 is a powerful language model that powers the PaLM API.

• Developers have used the private preview of the PaLM API in their applications, including Chaptr, Game On Technology, and Wendy's.

• Leading developer tools companies like LangChain and Chroma are working with Google to support the PaLM API.

• The PaLM API has been integrated into Google Developer tools like Firebase and Colab.

Project Tailwind: An AI-first Notebook

• Project Tailwind is an AI-first notebook that helps users learn faster by creating a personalized and private AI model based on their notes and sources.

• The idea for Project Tailwind was developed with authors like Stephen Johnson and tested at universities like Arizona State and University of Oklahoma.

• A live demo of Project Tailwind shows how it can be used to quickly create study guides based on different sources.

• Users can use Project Tailwind to compile glossaries associated with all different notes and articles relating to specific topics or concepts.

• Project Tailwind is helpful for anyone synthesizing information from many different sources they choose, such as writers researching an article or lawyers preparing for a case.

Generative AI at Google

• PaLM 2 is just one example of how generative AI is changing what it means to develop new products.

• Google offers the best ML infrastructure, with powerful models like those in Vertex, and APIs and tools to quickly generate.

Innovations in AI

In this section, the speaker talks about the innovations in AI and how they are being used to improve learning comprehension. The speaker also discusses the challenges of deep fakes and how Google is working to prevent misuse of their technology.

AI for Learning Comprehension

• Google is using models to replicate the style and tone of speakers, then matching their lip movements to translate what they are saying.

• This innovation has resulted in promising results for course completion rates.

Challenges with Deep Fakes

• While this technology can be incredibly beneficial, there is an inherent tension as it could be misused by bad actors to create deep fakes.

• Google has built guardrails into their service to help prevent misuse and make it accessible only to authorized partners.

• New innovations in watermarking will also be integrated into generative models to help with the challenge of misinformation.

Responsible Use of AI

• Google's AI principles guide them on what not to do. For instance, years ago, they were the first major company not to make a general-purpose facial recognition API commercially available due to inadequate safeguards.

• They tackle challenges as they emerge by reducing the risk of problematic outputs generated by their models through automated adversarial testing using large language model technology.

• They share these innovations with others such as their perspective API which was originally created to help publishers mitigate toxicity but is now being used in large language models.

Bringing Computing Benefits To Everyone

In this section, Sameer Samat talks about how Android is bringing computing benefits to everyone through its open OS and use of AI models.

Android's Openness Enables Ecosystem

• From the beginning, Android believed that an open OS would enable a whole ecosystem and bring smartphones to everyone.

• This openness creates the freedom to choose devices that work best for you.

AI Improving Experiences at Scale

• With more than 3 billion Android devices, AI is being used to improve experiences at scale.

• For example, Android used AI models to protect users from more than 100 billion suspected spam messages and calls.

Bringing Computing Benefits To Everyone

• There are so many opportunities where AI can make things better.

• Sameer Samat talks about two big ways Android is bringing computing benefits to everyone: continuing to connect people to the most complete ecosystem of devices where everything works better together and using AI models in new ways.

Introducing a New System for Multitasking

In this section, the speaker introduces a new system for multitasking.

New System for Multitasking

• A new system for multitasking is introduced.

Applause

This section contains applause from the audience.

Personalization of Android Experience

In this section, the speaker talks about how personalization has always been at the core of the Android experience and how they are combining guided customization with Google's advances in generative AI to make phones feel even more personal.

Combining Guided Customization with Generative AI

• Personalization has always been at the core of the Android experience.

• They are combining guided customization with Google's advances in generative AI to make phones feel even more personal.

Magic Compose Feature

• Magic Compose is a new feature coming to Google messages powered by generative AI that helps you add that extra spark of personality to your conversation.

• Users can type their message like they normally would and then choose how they want to sound. Magic Compose will do the rest so messages give off more positivity, more rhymes, more professionalism or if you want, in the style of a certain playwright.

New Personalizations Coming to OS Layer

• Material You is a design system that combines user inspiration with dynamic color science for a fully personalized experience. It was introduced two years ago at Google I/O and it's being expanded on in Android 14 with all-new customization options coming to your lock screen.

• Users can add their own personalized style to the lock screen clock so it looks just the way they want.

• With new customizable lock screen shortcuts, users can instantly jump into their most frequent activities.

Emoji Wallpapers

• Emojis are such a fun and simple way of expressing yourself. So Google thought it would be cool to bring them to your wallpaper.

• With emoji wallpapers, users can choose their favorite combination of emoji, pick the perfect pattern, and then find just the right color to bring them all together.

Cinematic Wallpaper Feature

• With a new cinematic wallpaper feature, users can create a stunning 3D image from any regular photo and then use it as their wallpaper.

New Wallpaper Features

In this section, the speaker introduces new wallpaper features for Pixel devices. Users can choose from cinematic wallpapers, emoji wallpapers, and generative AI wallpapers that are created based on prompts selected by the user.

New Wallpaper Features

• Cinematic and emoji wallpapers will be available first to Pixel devices next month.

• Generative AI wallpapers allow users to select prompts such as "City by the bay in a post-impressionist style cool" and then create a unique wallpaper based on their selection.

• Google's text-to-image diffusion models generate completely new and original wallpapers behind the scenes.

• Material You automatically adapts the system's color palette to match the wallpaper created by the user.

• Generative AI wallpapers will be coming this fall.

AI-Powered Pixel Devices

In this section, the speaker discusses how Pixel devices are leading in AI-driven hardware experiences. The Google Tensor G2 chip is custom designed to put Google's leading-edge AI research to work in Pixel devices.

AI-Powered Pixel Devices

• Pixel was conceived as an AI-first mobile computer that brings together all of Google's breakthroughs across the company into one device.

• The Google Tensor G2 chip is custom designed to put Google's leading-edge AI research to work in our Pixel devices.

• Personalized AI enables helpful experiences that aren't available on any other mobile device, like Pixel Call Assist and Pixel Safe.

• On-device machine learning translates different languages for you, transcribes conversations in real-time, and understands how you talk and type.

• Pixel Camera understands faces, expressions, and skin tones to better depict you and the people you care about so your photos will always look amazing.

• Pixel experiences continue to be completely unique in mobile computing because it is the only phone engineered end-to-end by Google that combines Google Tensor, Android, and AI.

Conclusion

In this section, the speaker concludes by stating that Pixel devices allow users to experience all of the incredible new AI-powered features seen in the presentation in one place.

Conclusion

• The combination of hardware and software in Pixel devices allows users to experience all of the incredible new AI-powered features seen in the presentation in one place.

• The new Magic Editor and Google Photos will be available for early access to select Pixel phones later this year.

Pixel Tablet

In this section, the speaker introduces the new Pixel Tablet and its features.

Features of Pixel Tablet

• The tablet has room-filling sound from the charging speaker dock.

• The redesigned Google Home app brings together over 80,000 supported smart home devices, including Matter-enabled devices. It also allows for easy access to smart home controls directly from Hub mode.

• The tablet supports multiple users and makes switching between users super easy while maintaining privacy.

• The tablet is the first with Chromecast built-in, making it easy to move content between devices.

• A new type of case for Pixel Tablet has a built-in stand that provides continuous flexibility and is sturdy at all angles so you can comfortably use your tablet anywhere.

Availability and Pricing

• The new Pixel Tablet comes in three colors and is available for pre-order starting at just $499.

• Every Pixel Tablet comes bundled with the 129 charging speaker dock for free.

Foldable Devices

In this section, the speaker talks about foldable devices and what users want in a dream foldable device.

Dream Foldable Device Features

• Users want a versatile form factor that makes it great to use both folded and unfolded.

• They also want a flagship-level camera system that truly takes advantage of the unique design.

• An app experience that's fluid and seamless across both screens is also desired.

Pixel Fold and Android Large Screen Devices

In this section, Rick Osterloh talks about the new Pixel Fold device and its features. He demonstrates how it can be used for productivity, entertainment, and photography. He also announces that pre-orders for the device are now open.

Features of Pixel Fold

• The multitasking windows are paired in Overview mode.

• The split screen can be adjusted to suit the content being viewed.

• The "help me write" feature in Gmail can be used to quickly compose messages.

• Tabletop mode allows for a single-handed lean-back experience while watching videos.

• Support for switching displays from within an app is added, which is demonstrated using the camera app.

Photography with Pixel Fold

• The viewfinder can be moved to the outside screen, allowing for high-quality selfies using the rear camera system.

Pre-ordering Pixel Fold

• Pre-orders for Pixel Fold are now open and will ship next month.

• Pre-ordering a Pixel Fold comes with a free Pixel Watch.

Conclusion

• Rick concludes by stating that these large-screen devices represent several years of work on Android OS and popular apps on Play Store.

EN Access to Speech Therapy

This section discusses the global problem of access to speech therapy, particularly in developing countries.

The Need for an App

• Access to speech therapy is a global problem, but it's particularly acute in developing countries.

• A team realized they needed to code this into an app so that people could have access to speech therapy.

Controlling Stuttering

• With the right tools and training, stuttering can be controlled.

• Speech therapy is one such tool that can help control stuttering.

EN I/O Flip Card Game Demo

This section provides a demo of the I/O Flip card game and details some of the technology used in its development.

Introduction

• Chloe and Craig introduce I/O Flip, a classic card game with a modern AI twist.

• They know the developers of the game and will play against each other for a real-time demo.

Gameplay

• Players build their own team by customizing cards with classes and special powers.

• Each player gets 12 cards in a pack and chooses three to compete against their opponent's team.

• The winner is determined by whose cards are more powerful than their opponent's.

Technology Used

• Flutter was used to create silky smooth animations for the game.

• Google AI tools were used to create all images seen in I/O Flip.

I/O Flip Game

In this section, Craig and Chloe play the I/O Flip game. They discuss the game's features and how it is easy to customize cards and characters.

Playing the Game

• Craig and Chloe play a game of I/O Flip.

• Chloe wins the game by beating Craig's fire with her water elemental power.

• Chloe talks about how fun and easy it is to play I/O Flip. She mentions that she loves being able to customize her cards and characters, as well as playing quick games if she's short on time or extending her winning streak by playing again.

How to Play

• The objective of the game is to flip all your opponent's cards before they flip yours.

• Each card has an element (fire, water, or metal), which beats one element but loses to another.

• Players can customize their cards and characters in the game.

Conclusion

I/O Flip is a fun and customizable card game that can be played quickly or for extended periods. The objective of the game is to flip all your opponent's cards before they flip yours.