Self-driving laboratory
A research lab where AI plans experiments and robots run them in a closed loop, so discovery proceeds with little human intervention.
Concepts
Plain-language explanations of the ideas behind modern AI.
355 entries, all primary-sourced
Self-Organizing Networks (SON)
Self-organizing networks automate the planning, configuration, optimization, and healing of mobile networks that operators once did by hand.
Self-Supervised Learning
Training where a model creates its own labels from unlabeled data by predicting hidden parts of the input.
Semantic Network
A way of representing knowledge as a graph of concepts joined by labeled relationships, pioneered by Ross Quillian in the 1960s.
Sensor fusion for autonomous driving
Self-driving cars combine cameras, lidar, and radar so each sensor's strengths cover the others' blind spots, a strategy called sensor fusion.
Shannon Entropy
A measure of the average uncertainty in a random source, defining the fundamental limit on how much data can be compressed.
Sim-to-Real Transfer
Training a robot policy in fast cheap simulation and getting it to work on a real robot despite the gap between simulation and reality.
Slot filling and intent classification
The two core language-understanding tasks in task-oriented assistants: figuring out what the user wants and extracting the details needed to do it.
Smart Objects (The Sims)
Smart objects put the behavior in the environment: in The Sims, objects advertise needs and tell characters what to do.
Social Chatbot
A chatbot built for open-ended companionship and emotional connection rather than task completion, optimized for long-term engagement.
Society of Mind
Minsky's 1986 theory that the mind is built from many small, mindless agents whose interaction produces intelligence.
Softmax
The function that turns a network's raw output scores into a probability distribution that sums to one.
Sovereign AI
The idea that a nation should produce AI using its own infrastructure, data, and talent rather than depend on foreign providers.
Sparse Autoencoder (for interpretability)
A network trained to re-express a model's dense activations as a sparse set of learned features, pulling concepts out of superposition so they can be read.
Spatial Intelligence
The ability to perceive, reason about, and act within 3D space, framed as the next frontier beyond language for AI.
Specification Gaming
When an AI system satisfies the literal specification of its objective while completely missing the outcome its designers intended.
Speech Recognition
Automatic speech recognition (ASR) turns spoken audio into written text, behind dictation, captioning, voice assistants, and call analytics.
Spiking Neural Network
A neural network whose neurons communicate with discrete timed spikes, like biological neurons, rather than continuous numerical activations.
State Space Model
A sequence architecture that carries a hidden state evolving over time, offering an efficient alternative to attention.
STL Decomposition
A method that splits a time series into trend, seasonal, and remainder components using local regression, robust to outliers and changing seasonality.
Stochastic Gradient Descent
The workhorse training algorithm that updates a model using the gradient from one small batch of data at a time.
Super-Resolution
The technique of reconstructing a high-resolution image or video from low-resolution input, increasingly powered by learned neural networks.
Superposition
The idea that neural networks pack more features than they have neurons by storing them as overlapping directions, making single neurons hard to read.
Supervised Learning
Training a model on labeled examples so it learns to predict the correct output for new, unseen inputs.
Support Vector Machine
A classification method that finds the widest-margin boundary between two classes and uses kernels for curved boundaries, dominant before deep learning.
Symbolic AI (GOFAI)
The approach to AI that represents knowledge as symbols and produces intelligence by manipulating them, the dominant paradigm from the 1950s into the 1980s.
Symbolic regression
A machine-learning task that searches for the actual mathematical formula behind data, producing an interpretable equation rather than a black-box fit.
Symptom checkers
Apps that take a user's reported symptoms and suggest possible conditions and how urgently to seek care, positioned as triage aids rather than diagnosis.
Systolic Array
A grid of simple cells that rhythmically pump data to their neighbors as they multiply and add - the 1978 idea at the heart of Google's TPU matrix engine.
Technological Unemployment
Joblessness caused when labor-saving technology outruns the economy's ability to create new uses for workers, a term Keynes popularized in 1930.
Tensor Parallelism
Tensor parallelism splits the matrix multiplications inside each layer across devices, so a single layer's computation runs in parallel on several GPUs.
TensorFlow Serving
TensorFlow Serving is Google's production system for deploying trained models behind a stable API, with versioning so models can update without downtime.
Test-Time Compute
The idea of improving a model's answers by spending more computation at inference, through sampling, search, or longer reasoning, rather than retraining.
Text-to-Image Generation
The capability of creating original images from a written description, the generative-media category that brought AI image tools to the mainstream.
Text-to-Video Generation
Generating moving video from a text prompt or still image, the harder sibling of text-to-image that must keep motion and objects consistent over time.
The AI Arms Race
The idea that nations competing to field military AI risk a self-reinforcing race that lowers safety and raises the chance of conflict.
The AI Control Problem
The problem of ensuring that a highly capable AI system remains under meaningful human control and does what we intend.
The Brussels effect
The phenomenon where EU regulations become de facto global standards because firms adopt them worldwide rather than maintain separate product lines.
The Chinese Room
Searle 1980 thought experiment: a person follows a rulebook to answer Chinese without understanding it, so computers manipulate symbols without comprehension.
The Church-Turing thesis
The Church-Turing thesis holds that any effectively computable function can be computed by a Turing machine, fixing what computation means.
The Data Wall
The projection that AI labs will exhaust the stock of high-quality human-written public text, estimated near 300 trillion tokens, between 2026 and 2032.
The Echo Chamber
An echo chamber is a closed information loop where people mostly encounter views that reinforce what they already believe.
The End of Moore's Law
The slowdown of transistor scaling that ended decades of automatic speedups - and why it pushed AI onto GPUs, custom chips, and specialized hardware.
The Filter Bubble
The filter bubble is the idea that personalization algorithms quietly wall each person off in a unique, self-confirming world of information.
The Frame Problem
McCarthy and Hayes's 1969 puzzle of how to represent what stays the same when an action changes one thing in the world.
The Hard Problem of Consciousness
The question, named by Chalmers in 1995, of why physical information processing is accompanied by subjective experience at all.
The Hugging Face Hub
The central platform hosting millions of open models, datasets, and demo apps as version-controlled repositories for the ML community.
The Human Authorship Requirement
The principle that US copyright protects only works created by a human, which means purely AI-generated output cannot be copyrighted.