When the then-largely unknown DeepMind was snapped up by Google (now Alphabet) in early 2014 for a price estimated between $400 million and $650 million, it signaled to outsiders just how excited the tech world was—and remains—about deep learning.

An early DeepMind triumph was a demonstration of an AI agent capable of mastering old Atari games with minimal human input. This relied on advances in a field known as reinforcement learning, a kind of AI behaviorism that teaches agents to take actions based on the maximizing of rewards. DeepMind later made waves with AlphaGo, a gameplaying AI that defeated the 18-time world champion Lee Sedol at Go, one of the world’s most complex strategy board games.

But DeepMind isn’t on this list just because of its impressive legacy. Its publicly available AlphaFold is an AI system able to predict a protein’s 3D structure based on its amino acid sequence, holding enormous potential for research in human health fields. Then there are present investigations into using AI to control the nuclear fusion plasma in a tokamak reactor with deep reinforcement learning, attempts to leverage AI to create more natural-sounding artificial speech, and more. At this July’s 40th International Conference on Machine Learning, DeepMind presented around 80 research papers covering everything from superior AI performance in long-term reasoning tasks to ways in which machine learning models can help better train “embodied agents” such as robots.

Currently, much of the excitement around the company is focused on a language AI model called Gemini that will build on DeepMind’s previous reinforcement learning research to supposedly solve some of the problems current large language models still struggle with, such as planning and problem-solving.

Recently, DeepMind merged with Google AI’s Google Brain division to unify and accelerate the search giant’s focus on artificial intelligence. From the look of things, those ambitions extend far beyond providing users with better search results.