Glossary

Plain-language definitions for the technical terms we use across AI, emerging tech, hardware, FitTech, HealthTech, SportsTech, and wellness coverage. Linkable per term.

A

Agentic AI: A loose term for AI systems that act with relative autonomy — choosing actions, calling tools, and iterating on results — rather than only producing a single response. Often used interchangeably with "AI agent" but emphasizes the multi-step behavior.
AI Agent: A program that wraps a large language model in an action loop, letting it call tools, read and write files, browse the web, and execute code based on its own reasoning rather than a fixed script.
Ambient AI: An AI system that runs passively in the background and intervenes only when needed — most commonly in healthcare, where ambient scribes (Abridge, Suki, Nuance DAX) transcribe and structure clinician-patient conversations into notes.

B

Brain-Computer Interface (BCI): Hardware and software that translates neural activity into computer commands, either through invasive electrodes implanted in the brain or non-invasive sensors placed on the scalp.

C

CGM (Continuous Glucose Monitor): A small wearable sensor (Dexcom, FreeStyle Libre, Stelo, Levels) that measures interstitial glucose every few minutes, plotting trends and spikes in real time. Originally prescribed for diabetes management; increasingly used by metabolic-health consumers.
Chain of Thought (CoT): A prompting technique that asks a language model to show its reasoning step-by-step before producing the final answer. Improves performance on math, logic, and multi-step problems. Reasoning models (o1, o3, Claude with extended thinking) extend this with longer internal deliberation.
Chiplet: A small, specialized die that is packaged together with other dies to act as one processor, instead of building everything on a single large chip. Chiplets improve manufacturing yield and let vendors mix process nodes, for example a leading-edge compute die with an older, cheaper I/O die.
Circadian Rhythm: The roughly 24-hour internal biological cycle that regulates sleep, body temperature, hormone release, and metabolism. Disrupted by inconsistent sleep timing, late-evening blue light, and shift work; modern wellness tech (Oura, Whoop, dawn-simulation lights) tries to optimize alignment.
Cold Plunge: Deliberate immersion in water at temperatures of roughly 10–15°C (50–59°F) for two to five minutes. Used as a recovery and stress-adaptation protocol; evidence is mixed on long-term performance benefits.
Constitutional AI: A training approach developed by Anthropic where the model is given a written "constitution" — a set of principles — and trained to self-critique and revise its outputs against those principles. Reduces reliance on human-rated reward modeling for safety alignment.
Context Window: The maximum number of tokens (≈ ¾ of a word) a language model can consider in a single request. Modern frontier models support 100K–2M-token windows; longer windows let you load entire codebases or documents but cost more per request and may degrade in accuracy at the back of the window.

E

Edge Computing: Running compute close to where data is generated — on devices, gateways, or regional points of presence — rather than in a centralized cloud, to reduce latency and bandwidth costs.
Embedding: A dense vector (typically 768–4096 numbers) representing the semantic meaning of text, an image, or other input. Embeddings power semantic search, retrieval-augmented generation, recommendations, and clustering. Generated by purpose-built embedding models (text-embedding-3-large, voyage-3, etc.).
Esports: Competitive, organized video gaming played at a professional level, including leagues, tournaments, sponsorship structures, broadcast platforms, and the supporting performance-and-analytics tech stack.

F

Federated Learning: A training paradigm where the model travels to the data rather than the reverse — each participating device or institution trains locally and shares only weight updates. Used in healthcare (HIPAA-bound data) and mobile (on-device personalization without uploading raw data).
Fine-tuning: Additional training applied to a pre-trained foundation model on a smaller domain-specific dataset. Common variants: full fine-tuning (updates all weights, expensive), LoRA (low-rank adapter, cheap and reversible), and instruction tuning (alignment with task formats).
Foldable Phone: A smartphone with a flexible OLED display that folds along a hinge, either book-style (opening into a small tablet) or clamshell-style (folding a normal phone in half). The main engineering trade-offs are hinge durability, the visible crease, thickness and price.
Foundation Model: A large model trained on broad data (text, code, images) at scale, intended to be adapted to many downstream tasks via fine-tuning or in-context learning. Examples: GPT-5, Claude 4, Llama 4, Gemini 2.5. The "foundation" framing was popularized by Stanford in 2021.
Function Calling: A model capability where the LLM, instead of returning prose, returns a structured JSON object describing which tool to invoke and with what arguments. Enables agentic workflows: weather APIs, database lookups, file operations, browser automation. Also called "tool use."

H

Hallucination: When a language model fabricates information that sounds plausible but is incorrect — invented citations, made-up function signatures, false facts. Mitigated (not eliminated) by retrieval augmentation (RAG), tool use, constitutional training, and explicit "I don't know" prompting.
HBM (High-Bandwidth Memory): A type of DRAM built from stacked memory dies placed right next to the processor, giving far more bandwidth than standard memory. AI accelerators depend on it, so when AI demand absorbs memory-factory capacity, prices of ordinary DRAM for phones and PCs can rise too.
Heart Rate Variability (HRV): The variation in time between consecutive heartbeats, measured in milliseconds. Higher HRV correlates with parasympathetic dominance and is widely used as a recovery and stress proxy in fitness wearables.

I

Inference: Running a trained model to generate predictions or outputs from new input. Distinct from training (which updates model weights). Inference cost — compute, latency, energy — is the dominant ongoing cost of operating an LLM-powered product.

L

Large Language Model (LLM): A neural network trained on large text corpora to predict the next token in a sequence. Modern LLMs handle multi-turn conversation, structured tool calls, and multimodal input (text, image, audio, video).
LoRA (Low-Rank Adaptation): A parameter-efficient fine-tuning technique that freezes the base model weights and learns small low-rank update matrices instead. Reduces fine-tuning cost by 99%+, makes adapters portable and combinable, and is the dominant approach for personalizing open-source LLMs.

M

MCP (Model Context Protocol): An open protocol developed by Anthropic for connecting AI assistants to external tools, data sources, and services through a standardized interface. Native in Claude, supported by Cursor and ChatGPT. Server implementations exist for filesystems, GitHub, Slack, Postgres, and more.
Mixture of Experts (MoE): A model architecture where only a subset of "expert" sub-networks is activated per token, rather than the whole model. Enables much larger total parameter counts (e.g. DeepSeek-V3, Llama 4) at inference costs comparable to a smaller dense model.

N

Neuromorphic Computing: Chip architectures that model brain-like computation: spiking neurons, event-driven processing, and analog or mixed-signal circuits. Examples include Intel Loihi and IBM TrueNorth.
NPU (Neural Processing Unit): A processor block designed specifically for the matrix math behind neural networks. Phone and laptop chips include one so on-device AI features (transcription, photo editing, local assistants) run faster and use less battery than on the CPU or GPU.

P

Post-Quantum Cryptography: Cryptographic algorithms designed to resist attacks from large quantum computers. NIST standardized the first generation in 2024 (ML-KEM, ML-DSA, SLH-DSA). Major cloud platforms began rolling out hybrid post-quantum TLS through 2025–26.
Process Node: The label for a chip-manufacturing generation (e.g. 3nm, 2nm). Newer nodes generally pack more transistors into the same area for better performance per watt. The number is a marketing name rather than a literal measurement of any transistor feature.
Programmable Matter: A speculative class of material whose physical properties (shape, density, conductivity) can be changed under software control. Active research areas include modular robotics, claytronics, and 4D printing.
Prompt Injection: A class of security attack against LLM-powered systems where malicious instructions hidden in user input or retrieved content override the system's intended behavior. Especially dangerous for agents with tool access (file write, email send, code execution). No fully reliable defense exists.

Q

Quantum Annealing: A quantum-computing technique that finds the minimum of an objective function by gradually evolving a quantum system. Used for combinatorial optimization on D-Wave hardware.

R

RAG (Retrieval-Augmented Generation): A pattern that augments LLM generation with retrieved context from an external knowledge store — typically embeddings + a vector database. Reduces hallucination, supports private/proprietary data, and keeps answers fresh without retraining. Used by every serious enterprise AI deployment.
Reasoning Model: A class of LLM (OpenAI o-series, DeepSeek-R1, Claude with extended thinking, Gemini Thinking) that uses extended chain-of-thought during inference to solve complex problems. Slower and more expensive per query but markedly stronger on math, code, and multi-step reasoning.
Recovery Score: A composite metric (typically 0–100) computed by wearables from inputs like HRV, resting heart rate, sleep duration, and sleep quality. Used to prescribe training load for the day.
Reinforcement Learning from Human Feedback (RLHF): A training technique that aligns LLM outputs with human preferences. Humans rank model responses; a reward model is trained on those rankings; the base model is fine-tuned to maximize that reward.
Resting Heart Rate (RHR): The number of times the heart beats per minute at full rest, typically measured first thing in the morning before getting out of bed. Lower RHR generally indicates better cardiovascular fitness. Sudden elevation often precedes illness, overtraining, or poor sleep.
RPE (Rate of Perceived Exertion): A subjective 1–10 scale (often the Borg CR10) by which an athlete reports how hard a set or session felt. Pairs with objective load metrics (HRV, heart rate, watts) to calibrate training.

S

Sensor Fusion: Combining inputs from multiple sensors (e.g. accelerometer + gyroscope + GPS + heart rate) so the system can infer state (e.g. activity type, form, fatigue) that no single sensor could detect alone.
Sleep Debt: The cumulative deficit between sleep needed and sleep actually obtained over a multi-day window. Even modest chronic debt (1–2 hours/night) measurably degrades reaction time, decision-making, and immune function. Cannot be fully "paid off" on weekends.
Smart Glasses: Eyewear with built-in cameras, microphones, speakers and sometimes a small display, usually paired with a phone. Camera-and-audio models focus on capture and voice AI, while display models overlay notifications, navigation or translations in your field of view.
SoC (System on a Chip): A single chip that combines the CPU, GPU, NPU, memory controllers, modem or connectivity and other components that would otherwise be separate parts. Nearly every phone, tablet and many laptops (Apple A- and M-series, Google Tensor, Qualcomm Snapdragon) are built around one.
Spatial Computing: A computing paradigm where the interface lives in 3D physical space — through AR/VR/XR headsets (Apple Vision Pro, Meta Quest, Magic Leap) and depth sensors that map the user's environment. Apple's framing of the category in 2024 popularized the term over "AR/VR."
Sports Analytics: The application of statistical models, machine learning, and tracking data to athlete and team performance. Spans on-field metrics (expected goals, advanced WAR), workload and injury prediction, opponent scouting, and fan-engagement personalization.

T

Tokenization: The process of converting raw text into discrete tokens (sub-word units) that an LLM can process. Modern English tokens average ≈ 0.75 words. Tokenization choices affect cost (per-token billing), context window utilization, and model performance on non-English text and code.
Tool Use: A model capability where the LLM can call external functions — search engines, code interpreters, databases, APIs — and incorporate results into its response. Synonymous with "function calling" in practice. The foundation of agentic AI.
TOPS (Trillions of Operations per Second): A headline measure of how many simple math operations an NPU or AI accelerator can perform each second, usually quoted at low precision (INT8 or lower). It is useful for rough comparisons, but real-world AI speed also depends on memory bandwidth, software support and precision.
Total Daily Energy Expenditure (TDEE): The total calories an individual burns in a day, calculated as basal metabolic rate (BMR) multiplied by an activity factor. Used as the starting point for fat-loss, recomposition, or muscle-gain calorie targets.

V

Vector Database: A database optimized for storing and querying high-dimensional embedding vectors. Supports approximate-nearest-neighbor search at scale for semantic similarity. Examples: Pinecone, Weaviate, Qdrant, pgvector (PostgreSQL extension), Chroma. Pairs with embedding models in RAG pipelines.
VO2 Max: The maximum rate at which a person can consume oxygen during exercise, expressed in ml/kg/min. It is the strongest single predictor of endurance performance and correlates with cardiovascular health.

W

Wearable: A consumer-grade device worn on the body (wrist, finger, chest, ear) that continuously measures physiological signals — most commonly heart rate, motion, sleep, and skin temperature.

X

xG (Expected Goals): A soccer (and increasingly hockey) analytics metric estimating the probability that a given shot becomes a goal, given its location, angle, body part, and game context. Aggregated over matches, xG normalizes for finishing luck and is a stronger predictor of future performance than raw goals.

Z

Zero-Knowledge Proof (ZK Proof): A cryptographic technique that lets one party prove a statement is true without revealing any information beyond the truth of the statement itself. Used in privacy-preserving identity (Polygon ID, zkLogin), L2 blockchain rollups (zkSync, StarkNet), and confidential compute.
Zone 2: A low-intensity aerobic training zone, roughly 60–70% of maximum heart rate, where the body burns fat primarily and stays below the lactate threshold. Used to build mitochondrial density and aerobic base.