A* Search
Heuristic search algorithm that finds the shortest path through a graph efficiently - the basis for pathfinding in games, robotics, and logistics.
Concepts
Plain-language explanations of the ideas behind modern AI.
355 entries, all primary-sourced
ACT-R
A cognitive architecture by John Anderson that models human thought as production rules over declarative memory, with modules tied to brain regions.
Active Learning
A training approach where the model itself chooses which unlabeled examples a human should label next, to learn more from fewer labels.
Adversarial Examples
Inputs altered with tiny, often imperceptible changes that reliably trick a machine learning model into a confident wrong answer.
Agent Memory
Techniques that let an AI agent store and recall information beyond its context window so it can remember across time.
Agentic Coding
The practice of letting an AI agent autonomously read, write, run, and fix code rather than just suggesting completions.
AI Agents
AI systems that use a language model to plan, take actions through tools, observe results, and iterate toward a goal.
AI Alignment
The effort to ensure AI systems pursue the goals and values their designers and users actually intend, and avoid harmful behavior.
AI Compute
Compute - the calculation spent training a model - is AI's key input, growing far faster than Moore's Law and forecastable enough to plan capability around.
AI Control
A safety agenda that assumes a model may be misaligned and designs protocols to stay safe even if it actively tries to subvert them.
AI Data Center
An AI data center is a power-dense facility of networked GPUs for training and serving models; new campuses are now measured in gigawatts, not megawatts.
AI Demand Forecasting in Retail
Retailers use machine learning to predict product demand from sales history, search, weather, and events, so the right inventory is in the right place.
AI Energy
Data centers used 415 TWh - 1.5 percent of world electricity in 2024 per the IEA - set to more than double to 945 TWh by 2030, with AI the top driver.
AI in Nuclear Command and Control
The debate over whether and how to use AI in the systems that warn of, decide on, and execute nuclear weapons use, and the push to keep humans in control.
AI Regulation
The emerging body of laws, executive actions, and international declarations that governments use to govern the development and use of artificial intelligence.
AI Safety Case
A structured, evidence-based argument that an AI system is unlikely to cause a catastrophe, borrowed from safety-critical engineering.
AI surrogate model
A fast machine-learned stand-in for a slow, exact simulation, trained to reproduce its outputs at a tiny fraction of the cost.
AI therapy and the chatbot-as-therapist debate
The contested practice of using AI chatbots for mental-health support, and the safety and deception concerns regulators and psychologists have raised.
AI winter
A period when disappointment in artificial intelligence causes a sharp drop in funding, interest, and progress, following a stretch of inflated expectations.
Algorithmic bias
The tendency of machine-learning systems to encode and amplify biases present in their training data, producing systematically unequal outcomes across groups.
Algorithmic Credit Scoring
Algorithmic credit scoring uses statistical and machine-learning models to predict whether a borrower will repay, increasingly with non-traditional data.
Algorithmic impact assessment
A structured evaluation of an AI system's potential harms, such as discrimination, conducted before and during deployment to guide mitigation.
Analog In-Memory Computing
Doing neural-network math inside the memory itself, using analog physics to multiply and add - cutting the energy lost shuttling data to and from a processor.
Anomaly Detection
The task of finding rare data points that differ from the norm, used to catch fraud, faults, and intrusions, often without labeled examples.
Artificial General Intelligence (AGI)
The idea of an AI system with broad, human-level competence across most tasks - a goal that is real but deliberately ill-defined and contested.
Attention Mechanism
A technique that lets a model focus on the most relevant parts of its input when producing each piece of output.
Augmentation vs Automation
The framing of whether AI should replace human work or amplify it, a choice that shapes who captures the gains from the technology.
Automated Planning
The AI subfield of computing a sequence of actions to reach a goal, from the 1971 STRIPS planner to modern logistics and agent systems.
Automated Program Repair
Techniques that automatically fix bugs in software, now dominated by large language models that generate patches.
AutoML (Automated Machine Learning)
Automating the steps of building a model - preprocessing, model choice, and tuning - so non-experts can produce strong models.
Autoregressive Model
A model that generates a sequence one token at a time, each prediction conditioned on everything generated so far - how GPT-style models produce text.
AWS Trainium and Inferentia
Amazon's custom AI chips for training (Trainium) and inference (Inferentia), its in-house alternative to buying GPUs.
Backpropagation
The learning algorithm that efficiently trains multi-layer neural networks by propagating error signals backward through the network.
Bayesian Networks
A graph-based way to represent uncertain knowledge and reason about probabilities, invented by Judea Pearl and foundational to AI before deep learning.
Bayesian Optimization
A sample-efficient strategy for optimizing expensive functions by modeling them probabilistically and choosing each next trial intelligently.
Beam Search
A decoding strategy that keeps several of the most promising partial sequences at each step instead of committing to one - long the default for translation.
Behavior Cloning
The simplest imitation learning method: train a policy with supervised learning to copy the action a demonstrator took in each state.
Behavior Trees
Behavior trees model game-character decisions as a tree of tasks, popularized by Bungie's Halo 2 AI in 2005.
Benchmark Contamination
When test questions leak into a model's training data, inflating its benchmark scores and making them an unreliable measure of true ability.
Benchmark data contamination
When a model has seen benchmark questions during training, its scores reflect memorization, not ability, inflating results and breaking fair comparison.
bfloat16
A 16-bit number format from Google Brain that keeps full 32-bit dynamic range, now standard for training neural networks.
Bias-Variance Tradeoff
The core tension in machine learning between models too simple to capture the pattern and models too flexible that fit the noise.
Bootstrap Resampling
Estimating the uncertainty of any statistic by repeatedly resampling the observed data with replacement.
Brain-Computer Interface
A system that reads brain activity and turns it, using machine-learning decoders, into commands - moving a cursor, typing, or speaking.
Capability Elicitation
The problem of drawing out a model's true abilities, since a model can underperform what it can actually do.
Catastrophic Forgetting
The tendency of a neural network to lose what it learned on earlier tasks when trained on a new one - a core obstacle to continual learning.
Cellular automata
Cellular automata are grids of cells that update by simple local rules, showing how complex global behavior can emerge from simple parts.
Chain-of-Thought Faithfulness
Whether a model's written reasoning truly reflects how it reached its answer, and whether that reasoning can be safely monitored.