reward is enough
In plain wordsThe claim, made by DeepMind researchers in 2021, that maximizing a single reward in a rich enough world could by itself produce all the abilities we associate with intelligence.
The hypothesis, advanced by David Silver and colleagues at DeepMind (2021), that the single drive to maximize rewardreward, n.: The score an AI system receives after an action, telling it how good the outcome was. Reinforcement learning trains the system to collect as much reward as possible over time. in a sufficiently rich environmentenvironment, n.: Everything an AI agent interacts with but does not control — a game, a simulator, a website, the physical world. It answers each of the agent's actions with a new situation and, in reinforcement learning, a score. could by itself give rise to perception, knowledge, language, social skill, generalization, imitation and the other abilities associated with intelligence, so that a capable reinforcement learningreinforcement learning, n.: A way of teaching a computer program by trial and error: it tries actions, earns points for good results and penalties for bad ones, and gradually learns which choices pay off over time. agentagent, n.: Something that acts. In AI, a program that observes its situation and takes actions toward a goal: a game-playing learner, a software helper, or a chatbot that uses tools to complete tasks. could acquire them without each being designed or specified separately. Its authors argue that one generic objective in a rich world implicitly demands all these skills, much as survival does for animals, and present reinforcement learning as a route to artificial general intelligenceartificial general intelligence, n.: AI that could learn and do almost any mental task a capable adult can, rather than excelling at one narrow job. Experts disagree on the exact bar and on when, or whether, it will arrive.. Critics reply that the claim is close to unfalsifiable, since almost any behavior is optimal for some reward; that sample efficiency, environment choice and reward design carry the real burden; and that a single scalar reward may not capture multiple or changing human values. It is a strong form of Sutton's reward hypothesisreward hypothesis, n.: The idea that every goal can be expressed as maximizing a single running score — that whatever we want an agent to do, we can say it with points. and a precursor of the era of experienceera of experience, n.: The prediction, made by David Silver and Richard Sutton in 2025, that AI will soon learn mainly from its own experience of acting in the world rather than from human-written data. argument.
Reward-is-enough skeptics point out that the hypothesis says nothing about how much experience an agent would need.
Origin
Proposed by David Silver, Satinder Singh, Doina Precup and Richard Sutton of DeepMind in 'Reward is Enough' (Artificial Intelligence, vol. 299, October 2021; online earlier that year), extending Sutton's reward hypothesis.
Usage note
The hypothesis concerns what reward maximization could achieve in principle; it does not claim that current algorithms suffice. Whether it is a substantive empirical claim or a near-tautology remains debated.
Related
- reward hypothesisreward hypothesis, n.: The idea that every goal can be expressed as maximizing a single running score — that whatever we want an agent to do, we can say it with points.
- era of experienceera of experience, n.: The prediction, made by David Silver and Richard Sutton in 2025, that AI will soon learn mainly from its own experience of acting in the world rather than from human-written data.
- reinforcement learningreinforcement learning, n.: A way of teaching a computer program by trial and error: it tries actions, earns points for good results and penalties for bad ones, and gradually learns which choices pay off over time.
- artificial general intelligenceartificial general intelligence, n.: AI that could learn and do almost any mental task a capable adult can, rather than excelling at one narrow job. Experts disagree on the exact bar and on when, or whether, it will arrive.
- the bitter lessonthe bitter lesson, phr.: The argument, made by AI pioneer Rich Sutton in 2019, that AI has advanced mostly through general methods that use ever more computing power, not by programming in human knowledge.
- rewardreward, n.: The score an AI system receives after an action, telling it how good the outcome was. Reinforcement learning trains the system to collect as much reward as possible over time.
Sources
- Silver, D., Singh, S., Precup, D. & Sutton, R. S. (2021). Reward is enough. Artificial Intelligence 299, 103535.
- Vamplew, P. et al. (2022). Scalar reward is not enough: a response to Silver, Singh, Precup and Sutton (2021). Autonomous Agents and Multi-Agent Systems 36, 41. arxiv.org/abs/2112.15422
Mentioned in
Richard SuttonRichard Sutton, prop. n.: A Canadian computer scientist who, with Andrew Barto, laid the groundwork for teaching machines by trial, error and reward, and whose essay 'The Bitter Lesson' argues that methods that scale with computing power always win., era of experienceera of experience, n.: The prediction, made by David Silver and Richard Sutton in 2025, that AI will soon learn mainly from its own experience of acting in the world rather than from human-written data., reward hypothesisreward hypothesis, n.: The idea that every goal can be expressed as maximizing a single running score — that whatever we want an agent to do, we can say it with points.