Few artificial intelligence (AI) labs have made as many breakthrough discoveries as the UK-based startup DeepMind. The lab, founded by Demis Hassabis, Shane Legg, and Mustafa Suleyman in 2010, has consistently stood at the forefront of AI development in the realm of deep learning, unsupervised learning, reinforcement learning, and neuroscience.
After being acquired by Google for $500 million in 2014, the company has also gone on to play a key role in the development of services, including Google Search, YouTube, and Gmail. In 2023, DeepMind merged with Google Research to form Google Deepmind.
While it had taken until 2021 for DeepMind to turn a profit for the first time, for years, it has played a critical role in bringing to life high-profile projects, including AlphaFold, AlphaGo, and WaveNet, and discovering new ways to use AI to improve human decision-making.
DeepMind’s Top 5 AI Breakthroughs
Since its formation over a decade ago, DeepMind has contributed to a substantial number of innovations in AI research. Some of the most significant are broken down below.
One of DeepMind’s biggest breakthroughs in the last decade was an AI program called AlphaFold, which launched on 22 July 2021. AlphaFold uses AI to process the amino acid sequence of proteins and predicts the shape of proteins by generating a 3D model.
Before the release of AlphaFold, scientists only knew the 3D structures for just 17% of proteins in the body. Now after the launch of AlphaFold, scientists have access to over 200 million protein predictions, and 98.5% of 3D structures for human proteins can be predicted.
Fundamentally, AlphaFold helped to increase visibility over the structure of human proteins, which had traditionally remained opaque as one of the biggest problems in biology. Before AlphaFold, scientists would have to experiment independently to discover the underlying structure of a protein.
Today, over 1.2 million researchers across 190 countries use the AlphaFold Protein Structure database and have used the solution to help develop a vaccine for Malaria and discover new treatments for liver cancer.
By using AI to automate protein mapping, DeepMind has not only helped scientists to view protein structures but has actively helped to accelerate biological research and future drug discovery.
Another research project that gathered significant international interest is AlphaGo. AlphaGo is an AI-driven program that uses machine learning and deep neural networks to play the board game Go. AlphaGo analyzes past games and board configurations to predict the next move to take when playing Go.
Unlike other AI-playing Go programs, which used search trees to test all possible moves and positions, AlphaGo was given a description of the Go board as input and then trained to play against itself thousands of times to improve its decision-making capabilities.
This approach was such a success that not only did the program manage to defeat three-time Go European Champion Fan Hui in October 2015, but in March 2016, it went on to beat Lee Sedol, an 18-time Go world champion. AlphaGo also achieved a 9-dan professional ranking, the highest certification of the game.
More broadly, AlphaGo demonstrated a framework for how AI could be applied to dissect a complex decision-making process, learning common patterns and predicting the most effective response or outcome.
WaveNet, released in 2016, is another one of DeepMind’s core creations. It is a generative model for raw audio that is trained on a large volume of speech samples and has the ability to generate natural-sounding speech based on text or audio input.
Instead of cutting and recompiling voice recordings like other text-to-speech systems, WaveNet instead used a convolutional neural network trained on images, videos, and sounds to learn and emulate the structure of human language.
This meant that it could compose waveforms from scratch and generate speech that mimics the sound of a human voice. Now, WaveNext is used in a range of popular applications, including Google Assistant, Google Search, and Google Translate.
The main achievement of DeepMind with WaveNext was that the organization managed to build a solution that made computer-generated speech sound more natural and less robotic than traditional text-to-speech solutions.
4. Google Bard
Bard is built on the Pathways Language Model 2 (PaLM), a language model trained on publicly-available data, including web pages, source code, and other data, and enables users to process users’ natural language queries and responses in natural language.
The release of Bard came just months after OpenAI had established itself as a Large Language Model (LLM) industry leader with the release of ChatGPT in November 2022 and was critical for remaining competitive in the AI arms race.
At this stage, the key differentiator between Bard and ChatGPT is that the former can search the Internet on demand and generate responses, whereas ChatGPT relies on training data taken from before 2021.
It’s worth noting that while Bard got off to a rough start, losing Google $100 billion in market value after incorrectly claiming that the James Webb Space telescope took the first pictures of a planet outside the solar system during a demo. Since then, however, it has become a key competitor against OpenAI, reaching 142.6 million visitors in May.
Just months after working with Google AI to release Bard, DeepMind proceeded to release RT-2 in July 2023, the first vision-language action (VLA) robotics transformer model. RT-2 processes text and images taken from across the web and uses them to output robotic actions.
RT-2 can be used to control robotics equipment, teaching robots how to do basic tasks, such as identifying a piece of trash and throwing it away. It also has the ability to respond to user commands with reasoning in the form of object categories or high-level descriptions.
It also has the ability to perform semantic reasoning, such as identifying whether an object could be used as an improvised hammer (e.g., a rock) and what type of drink would be best for a tired person (an energy drink).
As a VLA model, RT-2 is innovative in the sense that it provides users with a new way to interact with robots, teaching them how to perform certain tasks and execute particular output actions. This opens up a whole new range of use cases for AI in robotics.
For years, DeepMind has maintained a reputation as one of the top AI labs in the world. With its track record in bringing to life high-profile projects AlphaFold, AlphaGo, WaveNet, Google Bard, and RT-2, DeepMind will remain a key player in AI research for years to come.
If there is anything to be learned from DeepMind’s research, it is that AI can be used to solve almost any problem you can think of, no matter how complex the scenario or requirement.