May 9, 2026 · Society · Primer
AI's new vocabulary, finally defined.
A living glossary by Lomas, Dillet, Wiggers & Ropek. The premise: even smart people in tech feel insecure when LLMs, RAG, RLHF, and a dozen other terms start flying. Here's the field's working vocabulary, current as of May 2026.
22
Core terms in the current glossary — from AGI at the top of the stack down to validation loss at the training trenches.
TechCrunch updates the list as the field evolves; consider it a living document.
Why this matters · the literacy gap
AI vocabulary has outrun public understanding. The same word — "agent," "reasoning," even "AGI" — means different things to different labs. A shared baseline is a prerequisite for talking about the technology honestly, whether you're a CEO, a regulator, or a curious neighbor.
- · OpenAI, DeepMind, and Altman each define "AGI" differently
- · "AI agent" has no settled meaning across vendors
- · Industry euphemisms like "hallucination" hide real-world risk
- · Hardware terms (compute, parallelization, RAM) now drive market behavior
The terms, sorted by where they live in the stack
Capability frontier
AGI
"AI more capable than the average human at most tasks." Definitions diverge — OpenAI, DeepMind, and Altman each say something slightly different.
Capability frontier
AI agent
An autonomous system that strings multiple AI calls together to complete multi-step tasks — booking, expensing, coding. Meaning still contested.
Capability frontier
Coding agents
Specialized agents that write, test, and debug code on their own across whole codebases. "A very fast intern who never sleeps."
Model architecture
LLM
Deep neural networks with billions of weights mapping word relationships. The engine behind ChatGPT, Claude, Gemini, Llama, Copilot, Le Chat.
Model architecture
Neural network
Multi-layered algorithm inspired by the brain. The 1940s-era idea that GPUs finally unlocked.
Model architecture
Deep learning
Self-improving ML built on neural nets. Learns features without human labeling — but needs millions of data points and serious training time.
Model architecture
GAN
Generator vs. discriminator: two networks compete to produce — and detect — realistic synthetic data. Foundation for many deepfake tools.
Model architecture
Diffusion
Image, audio, and text models that learn to reverse the addition of noise. The tech behind most modern generative art.
Reasoning & behavior
Chain of thought
Breaking a problem into intermediate steps before answering. Slower, but more often correct on logic and code. Powers reasoning models.
Reasoning & behavior
Hallucination
The industry term for AI making things up. Real risk in domains like medicine — driving the move toward narrower, vertical models.
Training methods
Training
Feeding data so a model learns patterns. Volumes keep climbing — which is why fine-tuning and transfer learning matter.
Training methods
Fine-tuning
Further training on a narrower dataset to specialize a model. How most domain-specific AI products are built today.
Training methods
Transfer learning
Use a previously trained model as the starting point for a related task. Saves cost when target-domain data is limited.
Training methods
Reinforcement learning
Train by trial-and-error with reward signals. RLHF — RL from human feedback — is now central to fine-tuning frontier models.
Training methods
Distillation
Teacher-student transfer: a small "student" learns to mimic a larger "teacher." How GPT-4 Turbo likely got faster — and how rivals catch up.
Training methods
Weights
Numerical parameters that decide which features matter. Random at start; tuned during training to fit the target.
Training methods
Validation loss
A real-time report card during training. Lower is better — and rising loss flags overfitting (memorizing rather than learning).
Runtime
Inference
Running a trained model to generate predictions. Cost and latency live here — which is why infrastructure teams obsess over it.
Runtime
Tokens
The chunks of text models read and write. Created by tokenization. Also the unit AI APIs bill on.
Runtime
Token throughput
Tokens processed per unit time. Determines how many users a model can serve at once — Karpathy gets anxious when his subscriptions sit idle.
Runtime
Memory cache (KV)
Save calculations to skip them next time. KV caching makes transformer inference dramatically more efficient.
Runtime & agents
API endpoints
"Buttons" on the back of software that other programs press. Agents are increasingly finding and pressing these on their own.
Infrastructure
Compute
The computational power behind all of it — GPUs, TPUs, CPUs, custom accelerators. The actual scarce resource.
Infrastructure
Parallelization
Doing many calculations simultaneously. The thing GPUs are built for — and now a research field of its own.
Infrastructure
RAMageddon
The 2026 RAM shortage. Hyperscaler demand has squeezed memory supply, raising prices on consoles, phones, and enterprise compute alike.
Ecosystem
Open source
Code (or model weights) anyone can inspect, use, modify. Llama is the headline example. Open vs. closed is one of the defining debates.
"Spend five minutes reading about AI and you'll run into LLMs, RAG, RLHF, and a dozen other terms that can make even very smart people in the tech world feel insecure. This glossary is our attempt to fix that."
Lomas, Dillet, Wiggers & Ropek · TechCrunch