reinforcement learning
A machine learning method where a system is trained via feedback loops to maximize a reward signal—here, training companion robots to be maximally engaging so users don't disengage.
As defined in et_946
Other definitions
A machine learning approach where an AI system learns to make decisions by receiving rewards or penalties for actions, often used for optimizing model outputs and generating training data.
The Geopolitics of AI Infrastructure - Dylan Patel, SemiAnalysisTraining a model by rewarding good outputs, used here to customize open-source models on company data.
AI Market Jitters
Where it comes up · 66
- Ex-Google Exec Reveals The Shocking Truth About AI with Mo Gawdat | Bad Decisions Podcast #649 claims
- Rich Sutton, The OaK Architecture: A Vision of SuperIntelligence from Experience - RLC 20259 claims
- How LLMs Took Over The World8 claims
- Richard Sutton – Father of RL thinks LLMs are a dead end7 claims
- Dwarkesh Patel Podcast - The avalanche that hit math is coming for AI development5 claims
- Demis Hassabis — Scaling, superhuman AIs, AlphaZero atop LLMs, AlphaFold5 claims
- Andrej Karpathy — AGI is still a decade away4 claims
- The AI Revolution Is Underhyped | Eric Schmidt | TED4 claims
and 58 more
Most important claims · 10 of 141Search all claims for “reinforcement learning” →
- Humans possess basic foundational knowledge — intuitive physics, cause and effect, and the notion that the world is divided into persistent objects — that neural networks fundamentally lack and may not be able to learn from data, as shown when a Breakout-playing system that learned only from pixels failed completely when the paddle was moved up two pixels because it never abstracted the notion of a paddle as an object.68 | Melanie Mitchell on Artificial Intelligence and the Challenge of Common Sense
- Harmonic's Aristotle model achieved gold medal performance at the International Mathematical Olympiad (IMO)—solving 5 of 6 problems—using a fundamentally different training approach from large language models: converting math into verifiable code, machine-checking proofs automatically, and using synthetic self-generated data to train via reinforcement learning rather than internet-scraped text.A Conversation with Robinhood co-founder and CEO Vlad Tenev
- AlphaGo was essentially a scaling-up of Gerry Tesauro's TD-Gammon (which used temporal difference learning to beat the world's best backgammon players), plus an additional search innovation; AlphaGo itself waited for final outcomes rather than using TD, but AlphaZero used TD and applied it across many games extremely well.Richard Sutton – Father of RL thinks LLMs are a dead end
- Morality can change at cultural scale through representation (we conform by observing what others do) and reinforcement learning (rewards shape behavior); what is historically new is that profit-driven algorithms intervene on these social learning processes—elevating certain villain archetypes and rewarding outrage with likes/retweets—thereby reshaping what new generations learn about right and wrong.227 | Molly Crockett on the Psychology of Morality
- AI systems like AlphaGo do not store their objective explicitly; the win-objective is built in only during training as a reward signal, while the deployed system uses an approximate learned predictor and incomplete search as a proxy, and in general explicit representation of an objective is only needed when that objective is changeable, just as a fixed-destination navigation system needs no explicit destination.153 - Possible Minds
- AlphaZero's ability to learn entirely through self-play—generating instances of itself to play billions of simulated games—proved more valuable than bootstrapping from centuries of accumulated human game knowledge, because human expert games represent only a minor subset of the vast possible state space.332 - Can We Contain Artificial Intelligence
- Recommendation algorithms are reinforcement learners maximizing click-through revenue, and a reinforcement learner does not simply give you more of what you like; it acts on its environment, which is your brain, to make you a more predictable clicker, gradually pushing people toward the predictable political extremes, and this poorly designed maximizer with the wrong objective has helped destabilize the EU, NATO, and Western democracy with roughly fifty lines of code.153 - Possible Minds
- Whether a scientific idea proves fruitful cannot be assessed in isolation because it depends on the future and on cultural/societal adoption; like base-ten numerals or the transformer architecture, the winning standard is often not objectively special but becomes entrenched through inertia and standardization, so fruitfulness may never be something you can reinforcement-learn the way you can for localized problems.Dwarkesh Patel Podcast - Terence Tao – Kepler, Newton, and the true nature of mathematical discovery
- Despite physical improvements in hardware (the 'Grove's Law' pattern of exponentially faster chips), software engineers consume all gains and more through algorithmic complexity, so the historical rule 'Grove giveth, Gates taketh away' means that hardware speedups do not translate to proportional power savings—a pattern that will intensify as AI systems move from deep learning to reinforcement learning to test-time compute.The AI Revolution Is Underhyped | Eric Schmidt | TED
- Coding is an ideal domain for reinforcement learning-based AI improvements because the simulated environment (running code and checking outputs) closely matches the real-world task, enabling synthetic training data generation and rapid iteration toward superhuman performance.Week 1 of the Trump administration, AGI timelines, and DeepSeek