# Who founded DeepMind in 2010, and what were they trying to build?

Date: 2021-01-05 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/when-was-deepmind-founded/

**Answer:** DeepMind is an artificial intelligence research company founded in London in 2010 by Demis Hassabis, Shane Legg and Mustafa Suleyman, with the goal of building general-purpose learning systems rather than programs that do one narrow task. Its approach, combining ideas from neuroscience with machine learning and testing them on games, produced the Atari-playing DQN, AlphaGo and, in late 2020, AlphaFold's leap in protein structure prediction.

In 2010, very few companies said openly that they wanted to build general artificial intelligence, a system that could learn to do almost anything. AI had been through cycles of grand promises and disappointment, and most successful products focused on narrow jobs such as filtering spam or ranking search results. Three young researchers in London decided to aim at the general problem anyway. Their company was called DeepMind.

The founders were Demis Hassabis, Shane Legg and Mustafa Suleyman. Hassabis and Legg first crossed paths at University College London, in its Gatsby unit for computational neuroscience, a research group that studies how the brain computes. DeepMind's own history describes its starting point as an interdisciplinary attempt to build general AI systems, drawing on machine learning, neuroscience, engineering, mathematics and simulation.

## When and where it started

The UK's official company register, Companies House, records that the company now called DeepMind Technologies Limited was incorporated on 23 September 2010. It carried a temporary placeholder name until 15 November 2010, the date Wikipedia gives for its official launch. So depending on which milestone you count, DeepMind began in September or November 2010. It was based in London and remains headquartered there.

Investors included the venture firms Horizons Ventures and Founders Fund, along with entrepreneurs such as Peter Thiel and Elon Musk. Jaan Tallinn was an early investor and adviser.

## The big idea: learn, don't hand-code

Most AI programs in 2010 were built for one task, with a lot of knowledge supplied by their programmers. DeepMind wanted systems that start with very little and learn from experience, the way an animal or a child does.

Games were the chosen testing ground. They are complex enough to be interesting, they have a clear score, and a computer can play millions of rounds without anyone getting hurt. According to Wikipedia, the company began by teaching its software to play Atari video games from the 1970s and 1980s, such as Breakout, Pong and Space Invaders, without telling it the rules.

Here is roughly how such a learning loop works in Breakout, where you move a paddle to bounce a ball into a wall of bricks:

| Round | What the program sees | What it tries | What happens | What it learns |
|---|---|---|---|---|
| 1 | Raw screen pixels | Moves randomly | Ball falls past the paddle | Nothing much yet |
| 50 | Ball heading left | Moves left, partly by chance | Ball bounces, a brick breaks, score rises | Moving toward the ball pays off |
| 5,000 | Ball heading left, fast | Moves left early | More bounces, higher score | Anticipate where the ball will land |
| Much later | A gap in the brick wall | Aims the ball at the gap | Ball gets behind the wall and clears many bricks | A long-term strategy |

The program is never told what a paddle or a ball is. It sees pixels and a score, and learns which actions lead to higher scores over time. DeepMind's history says its DQN system learned 49 different Atari games this way.

> **Try it:** Play any simple online game you have never seen before without reading the instructions. Write down how many tries it takes before you understand what earns points. That trial-and-error process is what DeepMind's early systems automated.

## What DeepMind has done since

The Atari research was published in 2013, and according to Wikipedia it reportedly helped prompt Google to buy the company. Google confirmed the acquisition on 26 January 2014; the price was reported at between $400 million and $650 million. DeepMind kept its London base and its research focus.

The milestones since then, up to the start of 2021:

- **AlphaGo** beat European Go champion Fan Hui 5–0 in October 2015, the first time a program had defeated a professional Go player, and beat Lee Sedol 4–1 in March 2016.
- **AlphaZero** taught itself chess, shogi and Go by playing against itself, and **MuZero** (2019) mastered games without even being told their rules.
- **WaveNet** (2016) generated realistic synthetic speech and was later used in Google products.
- **AlphaStar** defeated a top professional StarCraft II player in 2019.
- **AlphaFold** won the CASP13 protein-structure prediction assessment in December 2018. At CASP14, which reported results at the end of November 2020, its predictions reached an accuracy that the organisers considered comparable with laboratory methods.

There have been controversies too. In 2017 the UK's data protection regulator found that a London hospital trust had not complied with the law when it shared patient records with DeepMind for testing a health app. DeepMind responded that it needed to do better, and in October 2017 it launched a research unit, DeepMind Ethics & Society, to study the social questions AI raises. In December 2019, co-founder Mustafa Suleyman announced he was leaving for a policy role at Google.

## Why DeepMind's founding matters

DeepMind showed that a research lab could be organised like a company, with large teams, big computing budgets and long-term goals, and still publish fundamental science. Its success with games turned reinforcement learning, a fairly specialised corner of AI, into one of the most watched areas of research.

AlphaFold's CASP14 result, announced only weeks ago, is the clearest sign of where the founders hoped to go: from games to real scientific problems. Predicting the shape a protein folds into is a long-standing problem in molecular biology, one DeepMind began working on in 2016.

## What is reinforcement learning, the method behind DeepMind's best-known results?

Reinforcement learning is a way of training a computer by reward rather than by example. Instead of showing it thousands of labelled answers, you let it act, give it a score for the outcome and let it work out which actions lead to higher scores over time. The Breakout table above is a simple reinforcement learning story.

DeepMind's key step was to combine this with deep neural networks, which can turn raw inputs such as screen pixels into useful features. The combination, called deep reinforcement learning, let one method handle many different games.

It is not the whole story. AlphaGo, for example, first studied records of human games before improving by playing against itself. But the habit of letting machines learn by trial, error and reward is DeepMind's signature, and it traces back to those first experiments with old Atari games.

## Sources

- About Google DeepMind: https://deepmind.google/about/
- DEEPMIND TECHNOLOGIES LIMITED overview: https://find-and-update.company-information.service.gov.uk/company/07386350
- Google DeepMind: https://en.wikipedia.org/w/index.php?title=Google_DeepMind&oldid=997082003