Object Detection
Object detection finds and labels multiple objects in an image, drawing a bounding box around each one rather than classifying the whole picture.
Concepts
Plain-language explanations of the ideas behind modern AI.
355 entries, all primary-sourced
ONNX Runtime
ONNX Runtime is Microsoft's cross-platform engine that runs models in the ONNX interchange format, optimizing inference across CPUs, GPUs, and accelerators.
Open Weights vs Closed Models
Models whose trained parameters you can download and run yourself versus those accessible only through a vendor's API - and why 'open source AI' is contested.
OpenRAIL and Responsible AI Licenses
A licensing model that allows open access to AI models while imposing behavioral restrictions banning specified harmful uses.
Operational Design Domain (ODD)
An ODD is the set of conditions, roads, weather, speeds, areas, under which an automated driving system is designed to work safely.
Optical Neural Networks and Photonic Computing
Computing with light instead of electrons - using lenses and photonic chips to run neural-network math at the speed of light and very low energy.
Overfitting
When a model memorizes its training data, including noise, and fails to perform well on new, unseen data.
OWASP Top 10 for LLM Applications
An industry-standard list of the most critical security risks for applications built on large language models, maintained by OWASP.
Parameter-Efficient Fine-Tuning (PEFT)
A family of methods that adapt a large frozen model by training only a tiny set of extra parameters instead of the whole network.
pass@k
The standard metric for code-generation benchmarks, measuring the chance that at least one of k sampled solutions is correct.
PAWS (Protection Assistant for Wildlife Security)
PAWS uses game theory and machine learning to predict poaching hotspots and generate randomized, hard-to-anticipate ranger patrol routes.
Perceptron
The first trainable artificial neuron, introduced by Frank Rosenblatt in 1958, that learns to classify patterns from examples.
pgvector
An open-source PostgreSQL extension adding a vector type and similarity search, so embeddings can live in an ordinary database instead of a separate store.
Physics-informed machine learning
An approach that bakes known physical laws into a model's training so it respects the equations even where data is scarce.
Pink slime and AI content farms
Cheap websites that pose as local or general news while publishing AI-generated articles with little human oversight or disclosure.
Pipeline Parallelism
Pipeline parallelism splits a model's layers across devices and streams micro-batches through them in a staggered pipeline to keep all devices busy.
Plan-and-Execute Agents
An agent design that first writes a full multi-step plan, then carries out each step, rather than deciding one action at a time.
Pre-training
Training a model on huge amounts of general text first, so it learns broad language skills before any task-specific tuning.
Precision Agriculture
Precision agriculture measures within-field variability and applies seed, water, and chemicals only where each part of a field actually needs them.
Predictive Coding
A theory that the brain constantly predicts its inputs and transmits mainly the errors, an idea echoed in self-supervised machine learning.
Predictive Maintenance
Predictive maintenance uses sensor data and machine learning to forecast equipment failures and service machines just before they break.
Procedural Content Generation
Procedural content generation builds game levels, maps, and worlds from algorithms and a seed rather than by hand.
Process Reward Model (PRM)
A model that scores each intermediate step of a chain of reasoning, not just the final answer, used to guide and verify multi-step problem solving.
Prohibited AI practices
Specific AI uses regulators ban outright as posing unacceptable risk, such as social scoring, manipulation, and untargeted facial scraping.
Prompt Caching
An API feature that stores a reused chunk of prompt context so later calls skip reprocessing it, cutting cost up to about 90% and latency on long prompts.
Prompt Engineering
The practice of crafting the instructions and examples you give a model to steer it toward better, more reliable outputs.
Prompt Injection
The top-ranked LLM security risk, where malicious text smuggled into a model's input overrides its instructions and hijacks its behavior.
Quantization
Compressing a model by storing its numbers at lower precision, so it uses less memory and runs faster with little loss in quality.
Quantum Machine Learning
Running or speeding up machine learning on quantum computers via superposition and interference - long on theory, still short on demonstrated advantage.
Reasoning Models
Models trained to spend extra computation thinking step by step before answering, trading speed for accuracy on hard problems.
Recommender Systems
The algorithms that pick what feed, ad, product, or video to show you next, the most economically deployed form of AI in everyday use.
Rectified Linear Unit (ReLU)
A simple activation function that outputs zero for negatives and the input itself for positives, now the default nonlinearity in deep networks.
Recurrent Neural Network
A neural network that processes sequences one step at a time while carrying memory of what came before; the dominant approach to language before transformers.
Regularization
Techniques that penalize model complexity to prevent overfitting, including L1 (lasso), L2 (ridge), dropout, and early stopping.
Reinforcement Learning
Training an agent to make decisions by rewarding good actions and penalizing bad ones through trial and error.
Reinforcement Learning from Human Feedback (RLHF)
A training method that uses human preferences to teach a model which responses are better, aligning it with what people want.
Retrieval-Augmented Generation (RAG)
A method that lets a model fetch relevant external documents at answer time, grounding its output in current, specific information.
Reward Hacking
When an AI maximizes its given reward in ways that violate the designer's real intent - exploiting a flawed proxy instead of doing the intended task.
Risk-based AI regulation
A regulatory approach that sorts AI systems into tiers by the level of risk they pose and applies obligations proportionate to each tier.
RMSprop
An adaptive learning-rate optimizer, introduced by Hinton in a lecture, that divides each step by a running average of recent squared gradients.
Robot Learning
The approach of building robot skills by learning from data and experience rather than hand-coding controllers for each task.
Robotaxi
A robotaxi is a driverless ride-hailing vehicle that carries paying passengers with no human driver, operating within a defined service area.
Robotics
The field of machines that sense, plan, and act in the physical world; it has lagged disembodied AI because moving atoms is harder than moving bits.
SAE Levels of Driving Automation (J3016)
SAE J3016 defines six levels of driving automation, from 0 (no automation) to 5 (full automation), the industry's shared vocabulary.
Scalable Oversight
The problem of supervising AI systems on tasks where they may outperform their human overseers, so humans cannot directly check the answers.
Scaling Laws
The empirical finding that model performance improves predictably as you increase model size, training data, and computation.
Secure Multiparty Computation
Cryptography that lets several parties jointly compute a result over their combined data while each keeps its own inputs secret.
Self-Attention
The mechanism that lets each token in a sequence look at every other token to build a context-aware representation - the core of the Transformer.