Google DeepMind facts for kids

Kids Encyclopedia Facts

Quick facts for kids
DeepMind Technologies Limited

Headquarters in Kings Cross, London
Trade name	Google DeepMind
Type	Subsidiary
Industry	Artificial intelligence
Founded	23 September 2010; 14 years ago (2010-09-23)
Founders	Demis Hassabis Shane Legg Mustafa Suleyman
Headquarters	London, England
Key people	Demis Hassabis (CEO) Lila Ibrahim (COO)
Products	AlphaGo, AlphaStar, AlphaFold, AlphaZero
Owner	Alphabet Inc.
Number of employees	c. 2,000 (2023)
Parent	Deepmind Holdings Limited

DeepMind Technologies Limited, also known as Google DeepMind, is a British-American research lab. It focuses on artificial intelligence (AI) and is part of Google. DeepMind started in the UK in 2010. Google bought it in 2014. In April 2023, it joined with Google AI's Google Brain to become Google DeepMind. The company's main office is in London, England. It also has research centers in Canada, France, Germany, and the United States.

DeepMind has created special computer programs called neural networks. These programs can learn to play video games and board games. In 2016, DeepMind's AlphaGo program made headlines. It beat a world champion human Go player, Lee Sedol, in a five-game match. This event was even made into a documentary film. A more advanced program, AlphaZero, learned to play go, chess, and shogi (Japanese chess). It beat the strongest computer programs in these games after just a few days of playing against itself.

In 2020, DeepMind made big progress in understanding protein folding with its program AlphaFold. Proteins are tiny building blocks of life. Knowing how they fold helps scientists understand diseases and create new medicines. By July 2022, AlphaFold had predicted the shapes of over 200 million proteins. This includes almost all known proteins. AlphaFold3 was released in May 2024. It can predict how proteins interact with other molecules.

How DeepMind Started

DeepMind was founded in September 2010. The founders were Demis Hassabis, Shane Legg, and Mustafa Suleyman. Hassabis and Legg first met at University College London (UCL).

Demis Hassabis said that DeepMind started by teaching AI to play old video games. These games were from the 1970s and 1980s, like Breakout, Pong, and Space Invaders. The AI learned each game without knowing its rules beforehand. After some time, the AI became an expert. The goal was to create a general AI that could be useful for many different tasks.

Big investment companies like Horizons Ventures and Founders Fund put money into DeepMind. Other famous people like Peter Thiel and Elon Musk also invested. In January 2014, Google bought DeepMind. The price was between $400 million and $650 million. DeepMind was then renamed Google DeepMind.

In 2014, DeepMind won the "Company of the Year" award. This was from the Cambridge Computer Laboratory.

AI Ethics and Growth

After Google bought DeepMind, they set up an ethics board. This board looks at the moral questions around AI. DeepMind also started a new group called DeepMind Ethics and Society. This group studies how AI affects society.

In December 2019, co-founder Mustafa Suleyman left DeepMind. He moved to a policy role at Google. In April 2023, DeepMind merged with Google AI's Google Brain division. This created Google DeepMind. The merger happened to speed up AI work. It was a response to new AI tools like OpenAI's ChatGPT.

DeepMind's Amazing Creations

DeepMind has published over a thousand research papers. Many of these were in top science journals like Nature and Science.

AI in Games

DeepMind's early AI programs were designed to be general. This means they could learn many different things. They used a method called reinforcement learning. This is where the AI learns from experience. It uses only the raw information it sees, like pixels on a screen.

They tested their system on old arcade games. These included Space Invaders and Breakout. The same AI code could play these games better than any human. In 2013, DeepMind showed an AI that beat humans in games like Pong and Breakout. This work helped Google decide to buy the company.

In 2020, DeepMind released Agent57. This AI agent performed better than humans on all 57 Atari 2600 games. In 2022, DeepMind developed DeepNash. This system can play the board game Stratego as well as a human expert.

AlphaGo and Its Family

In October 2015, DeepMind's computer Go program, AlphaGo, made history. It beat the European Go champion, Fan Hui. This was the first time an AI beat a professional Go player. Go is a very complex game. It has many more possible moves than chess. This makes it hard for computers to play.

In March 2016, AlphaGo beat Lee Sedol. He was one of the world's best Go players. AlphaGo won the match 4 to 1. In 2017, AlphaGo also beat Ke Jie. He was the world's top-ranked player for two years.

Later in 2017, an improved version called AlphaGo Zero was created. It beat the original AlphaGo in every game. Then came AlphaZero. This program learned to play chess and shogi. It became super-human at these games. In 2019, DeepMind released MuZero. This new model mastered Go, chess, shogi, and Atari games. It did this without needing human data or rules.

AlphaGo learned by playing against itself. It used a method called deep reinforcement learning. It started by looking at millions of human Go moves. Then, it played against itself and learned from the results. This helped it get better and better.

AlphaGo Zero was even more amazing. It learned without any human games. It just played millions of games against itself. It used less computing power than the first AlphaGo. It also learned faster. It beat its older version in just three days.

AlphaStar for Strategy Games

In January 2019, DeepMind introduced AlphaStar. This program plays the real-time strategy game StarCraft II. AlphaStar learned by watching human players. Then, it played against itself to improve. It won 10 games in a row against two professional players.

By October 2019, AlphaStar reached the top level in StarCraft II. It became the first AI to reach this level in a popular esport. It did this without any special advantages.

Understanding Proteins with AlphaFold

In 2016, DeepMind started using AI to study protein folding. This is a big challenge in molecular biology. Proteins are like tiny machines in our bodies. Their shape determines what they do.

In December 2018, DeepMind's AlphaFold won a major competition. It correctly predicted the shapes of many proteins. In 2020, AlphaFold's predictions were as accurate as lab experiments. Scientists said the problem of protein folding was "largely solved."

In July 2021, DeepMind made AlphaFold available to everyone. A week later, they announced that AlphaFold had predicted almost all human proteins. It also predicted proteins for 20 other common organisms. These shapes are now in the AlphaFold Protein Structure Database. By July 2022, over 200 million protein predictions were added.

The newest version, AlphaFold3, came out in May 2024. It can predict how proteins interact with DNA, RNA, and other molecules. It greatly improved accuracy in predicting DNA interactions.

Language Models and Chatbots

In 2016, DeepMind created WaveNet. This system turns text into speech. It sounds very natural, almost like a human. Google now uses WaveNet for its commercial text-to-speech products.

In May 2022, DeepMind released Gato. This is a very flexible AI model. It can do many different tasks, like describing images or having conversations. Gato doesn't need to be retrained for each new task.

Sparrow is a chatbot created by DeepMind. It aims to build safer AI systems. It uses feedback from humans and Google search results.

AlphaCode for Programming

In 2022, DeepMind showed off AlphaCode. This AI can write computer programs. It performs as well as an average human programmer. DeepMind tested AlphaCode on coding challenges. AlphaCode earned a rank similar to 54% of human programmers.

Gemini: A Powerful New AI

Gemini is a very large and powerful AI model. It was released in December 2023. Gemini can understand and work with different types of information. This includes text, images, and more. It comes in three sizes: Nano, Pro, and Ultra. Gemini is also the name of Google's chatbot that uses this AI.

Gemma: Open Source AI

Gemma is a family of smaller, open-source AI models. They were released in February 2024. These models are designed to be used on different devices. They were trained using similar methods as the Gemini model.

SIMA for Virtual Worlds

In March 2024, DeepMind introduced SIMA. This AI agent can understand and follow instructions in games. It can complete tasks in many different 3D virtual worlds. SIMA learned from nine video games and research environments. It can adapt to new tasks without needing the game's code.

AI in Robotics

In June 2023, DeepMind released RoboCat. This AI model can control robotic arms. It can adapt to new types of robots and new tasks.

AI in Sports

DeepMind researchers have used AI for football (soccer). They model how players behave during games. This includes goalkeepers, defenders, and strikers. AI can help understand how players make decisions.

AI models could also help in the football industry. They could automatically pick out interesting video clips. This would create highlights of games. This is possible because AI can analyze video and player data.

AI in Archaeology

Google has a new program called Ithaca. It helps researchers restore damaged ancient Greek texts. It can also guess when and where these texts came from. Ithaca is based on an earlier DeepMind text analysis network called Pythia. Ithaca can restore damaged texts with 62% accuracy. It can also guess the location with 71% accuracy.

The team is working to use this model for other ancient languages. These include Demotic, Akkadian, Hebrew, and Mayan.

AI for New Materials

In November 2023, Google DeepMind announced GNoME. This tool uses AI to suggest millions of new materials. These materials were not known to chemistry before. It found hundreds of thousands of stable crystal structures. Some of these have already been made in labs.

AI in Mathematics

AlphaTensor for Math Problems

In October 2022, DeepMind released AlphaTensor. This AI uses reinforcement learning, like AlphaGo. It finds new ways to do matrix multiplication. For example, it found a faster way to multiply two 4x4 matrices. This was an improvement on a method known since 1969.

AlphaGeometry for Geometry

AlphaGeometry is an AI that combines different methods. It solved 25 out of 30 geometry problems from the International Mathematical Olympiad. This is a performance similar to a gold medalist.

Traditional geometry programs use strict rules. AlphaGeometry combines these rules with a special AI model. This model learns from many geometry proofs. If the rule-based system gets stuck, the AI model suggests new ideas.

How DeepMind Helps Google

Google says DeepMind's AI has made its data centers more efficient. The AI automatically balances cooling costs with hardware issues. DeepMind also helps Google Play suggest apps to users.

DeepMind has worked with the Android team at Google. They created two new features for Android Pie phones. These are Adaptive Battery and Adaptive Brightness. They use AI to save energy and make phones easier to use. This was the first time DeepMind used AI on such a small scale.

DeepMind and Healthcare

In July 2016, DeepMind started working with Moorfields Eye Hospital. They wanted to use AI to analyze eye scans. The goal was to find early signs of diseases that cause blindness.

In August 2016, DeepMind began a project with University College Hospital. They aimed to create an algorithm. This algorithm would tell the difference between healthy and cancerous tissues.

DeepMind also worked with other hospitals. They developed mobile apps linked to patient records. These apps send alerts to doctors about patients at risk. For example, they can warn about acute kidney injury.

In November 2017, DeepMind partnered with Cancer Research UK. They wanted to improve breast cancer detection. They used AI for mammography. In February 2018, DeepMind worked with the United States Department of Veterans Affairs. They used AI to predict kidney injury and patient health changes.

DeepMind created an app called Streams. It sends alerts to doctors about patients. In November 2018, DeepMind announced that its health division and the Streams app would join Google Health.

DeepMind's Ethics and Society Work

In October 2017, DeepMind started a new research group. It is called DeepMind Ethics & Society. Their goal is to understand the ethical issues of AI. They fund outside research on topics like privacy, fairness, and AI risks. They want to make sure AI benefits society.

This group is separate from other partnerships. DeepMind is also part of the Partnership on Artificial Intelligence to Benefit People and Society. This group includes many companies and organizations.

DeepMind Professors

DeepMind supports three university professors who study machine learning:

At the University of Cambridge, held by Neil Lawrence.
At the University of Oxford, held by Michael Bronstein.
At the University College London, held by Marc Deisenroth.

Images for kids

Logo from 2015–2016
Logo from 2016–2019