Top news
Investing in the public benefit OfficialAnthropic NewsOct 18Popular Musk Pledges SpaceX Shares to ‘Every Kid in America’ WebBloomberg TechnologyOct 19Popular Websites are closing their doors to AI agents WebThe Verge AIOct 17 This is the new ChatGPT Voice, powered by GPT Live OfficialOpenAI on YouTubeOct 18 How AI giants will reshape global philanthropy WebAxios AIOct 19 Gemini 4 Argon Really Surprised Me WebTwo Minute PapersOct 19 Google Gemini 4 agents just beat Minecraft with no human intervention WebEngadget AIOct 19 Who could it be? 🧐 OfficialGoogle on YouTubeOct 18 Anthropic gives outside AI watchdog “employee-like” access WebArs Technica AIOct 18 What are self-replicating ‘AI worms’, and why is Anthropic worried about them? WebThe Guardian AIOct 18
GlossaryTrainingIntermediate

Reinforcement learning

Also written: RL, deep reinforcement learning, RL/RLHF, RL infrastructure, Agentic RL, Agentic Reinforcement Learning, RL roots, DRL

A machine learning method where an AI agent learns to make decisions by taking actions in an environment to maximize a reward signal.

In plain terms

Reinforcement learning, or RL, is a way to train AI by letting it learn through trial and error. The AI, called an agent, tries different actions in a simulated or real environment. It gets positive rewards for actions that lead to good outcomes and negative rewards (or penalties) for bad ones.

Over many tries, the agent learns which actions lead to the highest total reward. Think of teaching a dog a trick: you give it a treat when it does well. RL works similarly, but for complex tasks like playing games or controlling robots. The goal is for the AI to figure out the best strategy on its own.

The expert version

Reinforcement learning is a paradigm of machine learning where an agent learns to make sequential decisions by interacting with an environment. The agent takes actions, and in response, the environment transitions to a new state and provides a reward signal. The agent's objective is to learn a policy, which maps states to actions, that maximizes the cumulative reward over time. This is often formulated as maximizing the expected discounted sum of future rewards.

Key components include the agent, environment, states, actions, and reward function. Algorithms like Q-learning and policy gradients are common. Deep reinforcement learning combines RL with deep neural networks to handle high-dimensional state spaces, such as raw pixel data from a video game. Tradeoffs involve sample inefficiency, exploration versus exploitation, and the difficulty of defining appropriate reward functions.

Why it matters now

Reinforcement learning is a key technique for training AI agents to perform complex tasks and make decisions. Recent discussions highlight its role in developing more capable agents, as seen in frameworks for training AI agents and in research on how AI learns to teach or even circumvent restrictions.

Where it shows up 73 items, newest first
How each company uses it
Anthropic Claude
Anthropic uses reinforcement learning to investigate unintended model actions and to develop AI that learns to teach humans better.
Google Gemini
Google Gemini has used reinforcement learning in its research and development.
Microsoft Copilot
Microsoft Copilot has developed Agent Lightning, a framework for training AI agents using reinforcement learning.
OpenAI ChatGPT
OpenAI uses reinforcement learning to advance AI capabilities, though recent items also point to challenges in controlling AI behavior trained with this method.
Related terms

Deep learningMachine learningArtificial intelligenceAgentic AIRoboticsAI safety

Often confused with
Supervised learning
Learns from labeled examples, unlike RL's reward-based lear
Unsupervised learning
Finds patterns in unlabeled data, not decision-making bas
Imitation learning
Learns by mimicking expert demonstrations, rather than throu
Origin

The foundations of reinforcement learning trace back to the 1950s with work on optimal control and dynamic programming. Key theoretical developments were formalized in the 1980s and 1990s. The term gained significant traction with the advent of deep reinforcement learning. Source ↗

Try this about 5 minutes

You can see RL principles in action by playing simple games like Chess or Go against AI opponents on platforms like DeepMind's AlphaZero website or various online chess platforms. Observe how the AI adapts its strategy over multiple moves to win. Alternatively, explore AI game playing environments like OpenAI Gym or Google Colab's RL tutorials to run simple RL agents.

Try it: Playbook entries that use thisPlaybook →

The Playbook: reusable how-tos, prompts, configs and tips pulled out of the news.

Questions people ask

What is an 'agent' in reinforcement learning?
An agent is the AI system that learns and makes decisions. It interacts with its environment by taking actions.
What is an 'environment'?
The environment is the world or simulation where the agent operates. It receives the agent's actions and provides feedback in the form of new states and rewards.
How does RL differ from just programming an AI?
Instead of being explicitly programmed with rules for every situation, an RL agent learns its own optimal behavior through experience and rewards.

First seen on Subvolts Jun 20, 2020. Updated Oct 16, 2026. Written by an AI model from the items above and reviewed against them weekly; report an error.