- DeepMind uses deep and reinforcement learning to solve complex tasks.
- Models like AlphaGo and AlphaFold have revolutionized gaming and medicine.
- Veo allows the creation of realistic videos with simple text or images.
- Tools like Gemma Scope help understand how AI thinks.
Artificial intelligence has captured the world's attention thanks to its astonishing advances, and among the companies leading the way, Google DeepMind stands out as one of the most innovative. Founded in 2010 and acquired by Google in 2014, this UK-based company has developed various technologies that have made their mark in fields as diverse as medicine, energy efficiency, gaming, and even audiovisual content creation.
But what exactly is DeepMind and how can it be used? To understand how it works, its real-world applications, and its impact on everyday life, let's delve into its history, its technological structure, its main developments, and how we are now seeing tangible applications of these ideas that years ago seemed like science fiction.
What is DeepMind and where did it come from?
DeepMind is an artificial intelligence research lab whose mission is to develop systems that can learn autonomously and solve problems without the need for pre-programmed rules. It was founded by three British researchers with backgrounds in neuroscience and computer science. Its disruptive approach to AI quickly garnered attention, so much so that Google (now Alphabet) acquired the company for approximately £400 million in 2014.
Since then, DeepMind has grown to a team of over 700 professionals distributed across offices in various parts of the world. With access to Google's technical infrastructure, they have been able to expand their research into fields such as medicine, energy efficiency, neuroscience, robotics, and algorithms applied to games.
How the technology behind DeepMind works
DeepMind's AI core is based on two fundamental pillars: deep learning and reinforcement learning . The neural networks they develop are designed to mimic the workings of the human brain through layers of interconnected nodes that learn from data.
One of the key approaches used is model-less Q learning , a technique where the system adjusts its responses to maximize the rewards received. This form of training allows DeepMind to evolve in complex environments without predefined instructions.
For example, they trained their AI with Atari video games. At first, it didn't know how to play, but after millions of repetitions, it managed to outperform human players . This is possible because AI is capable of identifying optimal patterns and correcting errors over time.
DeepMind's notable achievements in different areas
Since its inception, DeepMind has achieved breakthrough developments in the field of artificial intelligence. Here are some of the most significant ones:
AlphaGo: Mastering the Impossible
The board game Go is known for its enormous strategic complexity. DeepMind developed AlphaGo, the first AI to defeat a professional human player and then the world champion in 2016. This feat showed that AI could tackle tasks requiring intuition, creativity, and long-term strategic thinking.
AlphaFold: revolutionizing medicine
One of the most disruptive advances has been AlphaFold, an AI capable of predicting how proteins fold , a biological problem that had remained unsolved for decades. This tool has facilitated the creation of new treatments and drugs, even contributing to the fight against COVID-19 and diseases like Alzheimer's.
AlphaZero: The system that dominates multiple games
AlphaZero is capable of learning and playing chess, shogi, and Go at an expert level. It requires no prior data; it learns simply by playing itself millions of times . In just a few days, it outperformed the world's most advanced chess engines, such as Stockfish.
DeepMind in energy efficiency
DeepMind's AI has also improved how we use resources. Applied in Google's data centers, it reduced cooling energy consumption by 40% thanks to a reinforcement learning system that optimized in real time.
Advances in robotics
DeepMind has worked on robots like Robocat, which can learn to operate different robotic arms with just a few demonstrations. The system trains itself using self-generated data, improving its efficiency as it faces new challenges.
New Frontiers: Generative AI with DeepMind
One of the most striking new features is Veo, a generative artificial intelligence model for video . It was unveiled at Google I/O 2024 and promises to transform audiovisual creation.
What can Veo do?
Veo can generate high-quality 1080p videos in multiple contexts from text, images, or pre-existing video footage . Its intuitive ability to capture complex details makes it an ideal tool for creators, filmmakers, and advertisers.
It offers functions such as:
- Masked Edition: Allows you to modify specific areas of an existing video.
- Creating image-based videos: From an image and a text, a video is generated with aesthetic and narrative coherence.
- Clip expansion: Create videos longer than 60 seconds with visual continuity.
Veo Real-World Use Cases
Filmmaker Donald Glover worked with Google to test Veo's creative capabilities. They used this AI to visualize complex shots before filming , saving time and resources. This rapid preview allows teams to adjust shots and focus in advance, raising the bar for production.
It has also been used in advertising, education, and even medicine. Veo can simulate realistic environments for training or create impactful educational videos that facilitate the understanding of complex topics.
How does Veo work?
The model combines several previous technologies such as Generative Query Network, Image-Video, and VideoPoet. It uses encoders for text and images and combines them into a single input that passes through a latent diffusion model . This generates a compressed video that is decoded in high resolution.
In addition, all videos carry an invisible watermark using SynthID , a proprietary tool that ensures AI-generated content is traceable and secure.
Gemma Scope: Understanding How AI Thinks
One of the challenges of AI is understanding exactly what happens inside it. To solve this mystery, the DeepMind team has developed Gemma Scope, a tool within the field of mechanistic interpretability . Its function is to show how an AI arrives at its conclusions.
Using scattered autoencoders, it's possible to see which neurons are activated by certain stimuli , as if observing the AI's brain under a microscope. For example, asking it about a chihuahua might activate the "dogs" pattern.
These tools help detect internal errors and biases . For example, if a model says that 9.11 is larger than 9.8, it's because it has incorrectly associated those numbers with biblical verses or dates like 9/11. With this information, engineers can correct incorrect behavior.
Functions such as detecting human feelings, like vicarious embarrassment , or even deception, have also been explored, although the latter remains complex to isolate.
DeepMind has opened these tools to the public through GitHub and platforms like Neuronpedia. This allows more experts to collaborate to understand how the models work and reverse engineer their algorithms.
Thanks to DeepMind's constant evolution , its technologies are moving beyond research to become solutions applicable to real-world problems, from drug discovery to film and education. Its ethical approach, commitment to transparency, and creation of tools that enable us to understand AI demonstrate that there is still much to explore, and that we are only scratching the surface of this type of artificial intelligence's potential.
