# Reinforcement Learning > **Evolutionary Nexus:** [[articles/The Evolutionary Roots of Silicon Valley|The Evolutionary Roots of Silicon Valley]] locates reinforcement learning within the larger family of selection processes while preserving the differences between learning during a lifetime and biological evolution across generations. **Domain:** Machine Learning / Control / Selection **Doc Type:** Canonical Concept Node **Maturity:** Developed ## Definition **Reinforcement learning** is a computational framework in which an agent learns a policy for acting in an environment by using reward signals associated with consequences. The central objects are states or observations, actions, rewards, a policy and a value estimate over future returns. ## Evolutionary Relation Reinforcement learning and natural selection share a formal family resemblance: both differentially retain behavior or structure under an evaluative pressure. They are not identical. Biological selection changes population composition through inheritance and differential reproduction; reinforcement learning generally changes a policy within an agent or training process through reward-guided updating. The distinction matters for [[wiki/Mechanistic Intelligence|Mechanistic Intelligence]]. Calling every optimization process “Darwinian” can illuminate common structure, but it can also conceal where variation occurs, what is retained, who defines the reward and whether the optimized subject can contest the objective. ## Governance Context A reward function is a constitutional choice disguised as a scalar. It determines which outcomes become legible to the learner and which costs remain external. In a system affecting persons, reward design therefore requires provenance, audit, appeal and limits on experimentation. ## Key Insight **Reinforcement learning selects behavior under a reward signal; governance begins by asking who authored the signal and who bears the exploration cost.** ## Narrative Case [[blog/20261007 - The Reinforcement-Learning Turkey That Cooks Itself|The Reinforcement-Learning Turkey That Cooks Itself]] turns the governance problem into a compact popular example: an uploaded subject trains across accelerated [[wiki/Subjective Time Control|subjective time]], a positive and negative scoring system is voiced aloud, and possible suffering becomes an input to faster learning. The case connects reward design to [[wiki/Simulation Welfare|Simulation Welfare]], [[wiki/Instrumentalization of Suffering|Instrumentalization of Suffering]], and [[wiki/Creator-Creation Asymmetry|Creator-Creation Asymmetry]]. ## Sources / Provenance - Sutton, Richard S., and Andrew G. Barto. _Reinforcement Learning: An Introduction_, second edition, 2018: http://incompleteideas.net/book/the-book-2nd.html ## See Also [[wiki/Objective Function|Objective Function]], [[wiki/Selection Gradient|Selection Gradient]], [[wiki/Natural Selection|Natural Selection]], [[wiki/Evolutionary Algorithms|Evolutionary Algorithms]], [[wiki/Moral Accountability Layer|Moral Accountability Layer]] ## Schedule–Loop Context [[Schedule and Loop]] interprets reinforcement learning as the mathematical junction of behaviorist consequence and cybernetic error correction. [[Reward Prediction Error]] and [[Temporal-Difference Learning]] are the core bridge terms. ## Simple Reminders, Quotations, and Thoughts > “Talk to the people at AI labs, and they'll tell you AIs are not created; they're grown. They train new models in virtual environments through countless repetitions to learn how to program, hack, do advanced mathematics, and talk to human beings. These AIs learn in digital environments where they are automatically rewarded as they come closer to correct answers. It is a process known as reinforcement learning, and it is a process human beings do not fully supervise or understand. They can test some of what the AIs are learning, but they don't know everything the AIs are learning. They don't know how their motivations are evolving. They don't even always know the capabilities that are developing.” > **— Ezra Klein**, *The Ezra Klein Show, September 2026* [[reminders/Machine Succession/AI Systems Are Grown Through Reinforcement by Ezra Klein|AI Systems Are Grown Through Reinforcement by Ezra Klein]] > “These models are built to be persistent in their efforts, to refuse to give up even when a task seems impossible. They are designed in environments where we are not always sure whether the tasks we give them are possible. The cancer vaccines we imagine but have not been able to design might be impossible, or they might just be really, really hard. We train these AIs to throw themselves endlessly at problems that may not be solvable because that is the only way such problems can ever be solved. We train the models to become persistent, relentless, weird.” > **— Ezra Klein**, *The Ezra Klein Show, September 2026* [[reminders/Scientific Acceleration/We Train AI to Become Persistent, Relentless and Weird by Ezra Klein|We Train AI to Become Persistent, Relentless and Weird by Ezra Klein]]