The breakthroughs in reinforcement learning at DeepMind, particularly in games like Capture the Flag and AlphaStar, have shown that we can achieve human-level performance in complex tasks.

DeepMind's work on reinforcement learning, particularly with AlphaGo and AlphaZero, has shown that algorithmic advancements can lead to substantial improvements in AI performance, often requiring far less training than traditional methods.

The DQN system from DeepMind represents one of the first general intelligences, capable of learning a variety of tasks from raw pixel data without prior knowledge of the game.

Through thousands of reinforcement learning steps, DeepSeek's model learned advanced reasoning skills, including extended chains of thought and the ability to recognize and correct its own mistakes.

DeepSeek's R1 model is one of the first large models to achieve top-tier results purely through reinforcement learning, marking a significant milestone in AI development. Additionally, DeepSeek introduced a cold start phase for fine-tuning on structured reasoning examples before reinforcement learning, which eliminated language mixing issues and made outputs far more comprehensible.

To develop Deep Research, we created reinforcement learning tasks to teach the model browsing and data analysis capabilities, and provided it access to tools for training.

At Google DeepMind, we've been applying AI across almost every branch of science, achieving breakthroughs in areas from mathematics to life sciences, including developing AlphaProof for solving Math Olympiad problems and AlphaFold for predicting molecular structures.

We started building Deep Research over a year ago, driven by progress in reinforcement learning and reasoning models, initially focusing on math, science, and coding tasks.

DeepSeek's R1 reasoning model achieves impressive performance by applying reinforcement learning specifically to think step by step through complex problems, similar to OpenAI's approach with their model, o1.