AI Timeline

Learn how Google has worked over the past 20 years to make AI helpful for everyone.
Our AI Journey
ML for Spell Check

Google begins using machine learning to help with spell check at scale in Search.

Google Translate

Google launches Google Translate using machine learning to automatically translate languages, starting with Arabic-English and English-Arabic.

Deep learning applied to speech recognition

A new era of AI begins when Google researchers improve speech recognition with Deep Neural Networks, which is a new machine learning architecture loosely modeled after the neural structures in the human brain.

Google trains AI at scale

In the famous cat paper, Google Research begins using large sets of “unlabeled data," like videos and photos from the internet, to significantly improve AI image classification. Roughly analogous to human learning, the neural network recognizes images (including cats!) from exposure instead of direct instruction.

DeepMind acquisition

Google acquires DeepMind, one of the leading AI research labs in the world.

Sequence to Sequence Learning with Neural Networks

Google presents Sequence To Sequence Learning With Neural Networks, a powerful machine learning technique that can learn to translate languages and summarize text by reading words one at a time and remembering what it has read before.

Generative Adversarial Networks

Google researchers help develop a new framework called Generative Adversarial Networks, or GANs, which is a type of machine learning model that can learn to create new things by playing a game with itself.
The program has two parts: one part tries to create new things, and the other part tries to tell if the new things are real or fake. The two parts play against each other, and over time the program gets better at creating new things that look real.

Google Photos

At its annual I/O developers conference, Google introduces Google Photos, a new app that uses AI with search capability to search for and access your memories by the people, places, and things that matter.

RankBrain in Search and Ads

Google deploys RankBrain in Search and Ads providing a better understanding of how words relate to concepts.

TensorFlow

Google introduces TensorFlow, a new, scalable open source machine learning framework used in speech recognition.

AlphaGo

AlphaGo, a computer program developed by DeepMind, plays the legendary Lee Sedol, winner of 18 world titles, famed for his creativity and widely considered to be one of the greatest players of the past decade. During the games, AlphaGo played several inventive winning moves. In game two, it played Move 37 — a creative move helped AlphaGo win the game and upended centuries of traditional wisdom.

Tensor Processing Units

Google publicly announces the Tensor Processing Unit (TPU), custom data center silicon built specifically for machine learning. After that announcement, the TPU continues to gain momentum: 

  • • TPU v2 is announced in 2017

  • • TPU v3 is announced at I/O 2018

  • • TPU v4 is announced at I/O 2021

  • • At I/O 2022, Sundar announces the world’s largest, publicly-available machine learning hub, powered by TPU v4 pods and based at our data center in Mayes County, Oklahoma, which runs on 90% carbon-free energy.

WaveNet

Developed by researchers at DeepMind, WaveNet is a new deep neural network for generating raw audio waveforms allowing it to model natural sounding speech. WaveNet was used to model many of the voices of the Google Assistant and other Google services.

Neural Machine Translation

Google announces the Google Neural Machine Translation system (GNMT), which uses state-of-the-art training techniques to achieve the largest improvements to date for machine translation quality.

Detection of Diabetic retinopathy

In a paper published in the Journal of the American Medical Association, Google demonstrates that a machine-learning driven system for diagnosing diabetic retinopathy from a retinal image could perform on-par with board-certified ophthalmologists.

Federated Learning

Google Research proposes a new, decentralized approach to training AI called Federated Learning that promises improved security and scalability.

Transformer

Google releases "Attention Is All You Need,” a research paper that introduces the Transformer, a novel neural network architecture particularly well suited for language understanding, among many other things.

Smart Compose

Google announcesSmart Compose, a new feature in Gmail that uses AI to help users more quickly reply to their email. Smart Compose builds on Smart Reply, another AI feature.

AI Principles

Google publishes its AI Principles – a set of guidelines that the company follows when developing and using artificial intelligence. The principles are designed to ensure that AI is used in a way that is beneficial to society and respects human rights.

BERT

Google introduces a new technique for natural language processing pre-training called Bidirectional Encoder Representations from Transformers (BERT), helping Search better understand users’ queries.

AlphaFold

DeepMind’s AlphaFold is recognized as a solution to the 50-year “protein-folding problem.” AlphaFold can accurately predict 3D models of protein structures and is accelerating research in biology.

First Quantum AI Milestone

Google’s Quantum AI demonstrates for the first time a computational task that can be executed exponentially faster on a quantum processor than on the world’s fastest classical computer -- just 200 seconds on a quantum processor compared to the 10,000 years it would take on a classical device.

Reinforcement Learning for chip design

Google Research proposes using machine learning itself to assist in creating computer chip hardware to accelerate the design process.

MUM

At I/O 2021, Google announcesMUM, multimodal models that are 1,000 times more powerful than BERT and allow people to naturally ask questions across different types of information.

LaMDA

At I/O 2021, Google announces LaMDA, a new conversational technology short for “Language Model for Dialogue Applications.”

Google Tensor Chip

Google announces Tensor, a custom-built System on a Chip (SoC) designed to bring advanced AI experiences to Pixel users.

PaLM

At I/O 2022, Sundar announces PaLM – or Pathways Language Model – Google’s largest language model to date, trained on 540 billion parameters.

LaMDA 2

Sundar announces LaMDA 2, Google’s most advanced conversational AI model.

Imagen & Parti: text-to-image

Google announces Imagen and Parti, two models that use different techniques to generate photorealistic images from a text description.

AlphaFold 2

The AlphaFold Database--which included over 200 million proteins structures and nearly all cataloged proteins known to science--is released.

Phenaki: text-to-video

Google announces Phenaki, a model that can generate realistic videos from text prompts.

MusicLM

Google introduces MusicLM, an AI model that can generate music from text.

Second Quantum AI milestone

Google’s Quantum AI achieves the world’s first demonstration of reducing errors in a quantum processor by increasing the number of qubits.

Bard

Google releases Bard, an early experiment that lets people collaborate with generative AI, first in the US and UK — followed by other countries.

Google DeepMind

DeepMind and Google's Brain team merge to form Google DeepMind.

PaLM 2

Google launches PaLM 2, our next generation large language model, that builds on Google’s legacy of breakthrough research in machine learning and responsible AI.

Gemini, our largest and most capable AI model

Google introduces Gemini, our most capable and general model, built from the ground up to be multimodal. Gemini is able to generalize and seamlessly understand, operate across, and combine different types of information including text, code, audio, image and video.

Gemini ecosystem

Google expands the Gemini ecosystem to introduce a new generation: Gemini 1.5, and brings Gemini to more products like Gmail and Docs. Gemini Advanced launched, giving people access to Google’s most capable AI models.

Gemma

Gemma is a family of lightweight state-of-the art open models built from the same research and technology used to create the Gemini models.

AlphaFold 3

Introduced AlphaFold 3, a new AI model developed by Google DeepMind and Isomorphic Labs that predicts the structure of proteins, DNA, RNA, ligands and more. Scientists can access the majority of its capabilities, for free, through AlphaFold Server.

Connectomics

Google Research and Harvard published the first synaptic-resolution reconstruction of the human brain. This achievement, made possible by the fusion of scientific imaging and Google's AI algorithms, paves the way for discoveries about brain function.

Learn how Google has worked over the past 20 years to make AI helpful for everyone.
Our AI Journey
ML for Spell Check

Google begins using machine learning to help with spell check at scale in Search.

Google Translate

Google launches Google Translate using machine learning to automatically translate languages, starting with Arabic-English and English-Arabic.

Deep learning applied to speech recognition

A new era of AI begins when Google researchers improve speech recognition with Deep Neural Networks, which is a new machine learning architecture loosely modeled after the neural structures in the human brain.

Google trains AI at scale

In the famous cat paper, Google Research begins using large sets of “unlabeled data," like videos and photos from the internet, to significantly improve AI image classification. Roughly analogous to human learning, the neural network recognizes images (including cats!) from exposure instead of direct instruction.

DeepMind acquisition

Google acquires DeepMind, one of the leading AI research labs in the world.

Sequence to Sequence Learning with Neural Networks

Google presents Sequence To Sequence Learning With Neural Networks, a powerful machine learning technique that can learn to translate languages and summarize text by reading words one at a time and remembering what it has read before.

Generative Adversarial Networks

Google researchers help develop a new framework called Generative Adversarial Networks, or GANs, which is a type of machine learning model that can learn to create new things by playing a game with itself.
The program has two parts: one part tries to create new things, and the other part tries to tell if the new things are real or fake. The two parts play against each other, and over time the program gets better at creating new things that look real.

Google Photos

At its annual I/O developers conference, Google introduces Google Photos, a new app that uses AI with search capability to search for and access your memories by the people, places, and things that matter.

RankBrain in Search and Ads

Google deploys RankBrain in Search and Ads providing a better understanding of how words relate to concepts.

TensorFlow

Google introduces TensorFlow, a new, scalable open source machine learning framework used in speech recognition.

AlphaGo

AlphaGo, a computer program developed by DeepMind, plays the legendary Lee Sedol, winner of 18 world titles, famed for his creativity and widely considered to be one of the greatest players of the past decade. During the games, AlphaGo played several inventive winning moves. In game two, it played Move 37 — a creative move helped AlphaGo win the game and upended centuries of traditional wisdom.

Tensor Processing Units

Google publicly announces the Tensor Processing Unit (TPU), custom data center silicon built specifically for machine learning. After that announcement, the TPU continues to gain momentum: 

  • • TPU v2 is announced in 2017

  • • TPU v3 is announced at I/O 2018

  • • TPU v4 is announced at I/O 2021

  • • At I/O 2022, Sundar announces the world’s largest, publicly-available machine learning hub, powered by TPU v4 pods and based at our data center in Mayes County, Oklahoma, which runs on 90% carbon-free energy.

WaveNet

Developed by researchers at DeepMind, WaveNet is a new deep neural network for generating raw audio waveforms allowing it to model natural sounding speech. WaveNet was used to model many of the voices of the Google Assistant and other Google services.

Neural Machine Translation

Google announces the Google Neural Machine Translation system (GNMT), which uses state-of-the-art training techniques to achieve the largest improvements to date for machine translation quality.

Detection of Diabetic retinopathy

In a paper published in the Journal of the American Medical Association, Google demonstrates that a machine-learning driven system for diagnosing diabetic retinopathy from a retinal image could perform on-par with board-certified ophthalmologists.

Federated Learning

Google Research proposes a new, decentralized approach to training AI called Federated Learning that promises improved security and scalability.

Transformer

Google releases "Attention Is All You Need,” a research paper that introduces the Transformer, a novel neural network architecture particularly well suited for language understanding, among many other things.

Smart Compose

Google announcesSmart Compose, a new feature in Gmail that uses AI to help users more quickly reply to their email. Smart Compose builds on Smart Reply, another AI feature.

AI Principles

Google publishes its AI Principles – a set of guidelines that the company follows when developing and using artificial intelligence. The principles are designed to ensure that AI is used in a way that is beneficial to society and respects human rights.

BERT

Google introduces a new technique for natural language processing pre-training called Bidirectional Encoder Representations from Transformers (BERT), helping Search better understand users’ queries.

AlphaFold

DeepMind’s AlphaFold is recognized as a solution to the 50-year “protein-folding problem.” AlphaFold can accurately predict 3D models of protein structures and is accelerating research in biology.

First Quantum AI Milestone

Google’s Quantum AI demonstrates for the first time a computational task that can be executed exponentially faster on a quantum processor than on the world’s fastest classical computer -- just 200 seconds on a quantum processor compared to the 10,000 years it would take on a classical device.

Reinforcement Learning for chip design

Google Research proposes using machine learning itself to assist in creating computer chip hardware to accelerate the design process.

MUM

At I/O 2021, Google announcesMUM, multimodal models that are 1,000 times more powerful than BERT and allow people to naturally ask questions across different types of information.

LaMDA

At I/O 2021, Google announces LaMDA, a new conversational technology short for “Language Model for Dialogue Applications.”

Google Tensor Chip

Google announces Tensor, a custom-built System on a Chip (SoC) designed to bring advanced AI experiences to Pixel users.

PaLM

At I/O 2022, Sundar announces PaLM – or Pathways Language Model – Google’s largest language model to date, trained on 540 billion parameters.

LaMDA 2

Sundar announces LaMDA 2, Google’s most advanced conversational AI model.

Imagen & Parti: text-to-image

Google announces Imagen and Parti, two models that use different techniques to generate photorealistic images from a text description.

AlphaFold 2

The AlphaFold Database--which included over 200 million proteins structures and nearly all cataloged proteins known to science--is released.

Phenaki: text-to-video

Google announces Phenaki, a model that can generate realistic videos from text prompts.

MusicLM

Google introduces MusicLM, an AI model that can generate music from text.

Second Quantum AI milestone

Google’s Quantum AI achieves the world’s first demonstration of reducing errors in a quantum processor by increasing the number of qubits.

Bard

Google releases Bard, an early experiment that lets people collaborate with generative AI, first in the US and UK — followed by other countries.

Google DeepMind

DeepMind and Google's Brain team merge to form Google DeepMind.

PaLM 2

Google launches PaLM 2, our next generation large language model, that builds on Google’s legacy of breakthrough research in machine learning and responsible AI.

Gemini, our largest and most capable AI model

Google introduces Gemini, our most capable and general model, built from the ground up to be multimodal. Gemini is able to generalize and seamlessly understand, operate across, and combine different types of information including text, code, audio, image and video.

Gemini ecosystem

Google expands the Gemini ecosystem to introduce a new generation: Gemini 1.5, and brings Gemini to more products like Gmail and Docs. Gemini Advanced launched, giving people access to Google’s most capable AI models.

Gemma

Gemma is a family of lightweight state-of-the art open models built from the same research and technology used to create the Gemini models.

AlphaFold 3

Introduced AlphaFold 3, a new AI model developed by Google DeepMind and Isomorphic Labs that predicts the structure of proteins, DNA, RNA, ligands and more. Scientists can access the majority of its capabilities, for free, through AlphaFold Server.

Connectomics

Google Research and Harvard published the first synaptic-resolution reconstruction of the human brain. This achievement, made possible by the fusion of scientific imaging and Google's AI algorithms, paves the way for discoveries about brain function.