The meeting ends. You nod along. Someone says "RAG pipeline" and "agentic workflow" in the same breath. You write them down. Later, you Google. The definitions are circular. The explainers are written by people who forgot what it's like not to know. You close the tab. The insecurity calcifies.
This is how the industry keeps you out.
Artificial intelligence isn't just rewriting code. It's rewriting the vocabulary of power. Every new acronym — LLM, RLHF, RAG, AGI — functions as a gate. The people who mint the terms set the agenda. The people who define them control the narrative. Everyone else just tries to keep up.
Start with the biggest term in the room. Artificial general intelligence. Three letters. Zero consensus.
Sam Altman says it's a median human you could hire as a co-worker. OpenAI's charter says it's systems that outperform humans at most economically valuable work. Google DeepMind says it's AI at least as capable as humans at most cognitive tasks. Three definitions. Three Different goalposts. Each carefully crafted to make the speaker's progress look inevitable and their competitors' look insufficient.
This isn't imprecision. It's strategy. A moving target can't be missed. When the definition of "winning" lives in the eye of the beholder, the race never ends — and the funding never stops.
The honest answer? Nobody knows. The experts building these systems disagree on what they're building toward. If the architects can't agree on the blueprint, the glossary entry should say exactly that: contested territory. Not a definition. A warning.
"AI agent" sounds authoritative. Autonomous. Capable. The reality: a chatbot with a to-do list and API access.
File expenses. Book a table. Write code. The demos work beautifully until they don't. The infrastructure — authentication, error handling, context windows, memory — is held together with duct tape and prompt engineering. "Agentic" is the new "blockchain": a suffix that attracts venture capital faster than it attracts users.
But dismiss it at your peril. The concept matters more than the current implementation. An agent is any system that chains multiple model calls together to pursue a goal without step-by-step human supervision. That architecture — planning, tool use, reflection, iteration — is how software becomes labor. The first wave of agents will be brittle. The second wave will replace entire job categories. The terminology will survive the hype cycle because the underlying shift is real.
Just don't confuse the demo with the product.
Here's the term most glossaries skip. API endpoints. The hidden buttons on the back of every modern service.
Developers know them. Product managers ignore them. But agents change the calculus. When a human uses software, they click through a UI. When an agent uses software, it hits endpoints directly. No login screens. No dashboards. No rate limits designed for human-speed interaction.
This is where the automation explosion actually lives. Not in the model. In the plumbing. The companies that expose clean, documented, agent-friendly APIs become the infrastructure of the AI economy. The companies that don't become legacy overnight.
Watch the endpoint wars. They're quieter than the model wars. They'll matter more.
Ask a human "which is taller, a giraffe or a cat?" — instant answer. Ask "a farmer has chickens and cows, 30 heads and 88 legs total, how many of each?" — you reach for pen and paper.
Models used to fail the second question. Now they're taught to reason aloud. Step by step. Intermediate tokens that aren't shown to the user but guide the final answer. This is chain-of-thought. It's not magic. It's scratchpad.
The breakthrough isn't that models can reason. It's that they can be trained to externalize reasoning — making their process inspectable, debuggable, steerable. That changes everything for reliability. It also creates a new attack surface: prompt injection via manipulated reasoning traces. The glossary entry for "chain-of-thought" should include "security vulnerability" as a related term.
Every glossary claims to be "living." Few are. The field moves in weeks. A definition written in January is misleading by March. "RAG" meant retrieval-augmented generation. Now it means any system that fetches context. "Fine-tuning" used to mean updating weights. Now it means LoRA, prefix tuning, prompt tuning — techniques that don't touch the base model at all.
The only honest glossary is a version-controlled one. Changelogs. Deprecation notices. "As of Q2 2025, this term means X; previously it meant Y." That's not a reference. That's a commit history. And that's what the industry needs.
You don't need to know the math. You don't need to write the code. But you do need to spot when a founder uses "agent" to mean "cron job with an LLM call." When a VC uses "AGI" to mean "whatever we're raising for next." When a regulator uses "frontier model" to mean "the three companies we met with last week."
Language is the interface between technical reality and political power. The people who control the definitions control the regulations, the funding flows, the hiring standards, the narratives that shape what gets built and what gets banned.
This glossary — any glossary — is a weapon. Use it to call bullshit. Use it to ask sharper questions. Use it to translate the hype into signal.
The terms will change next quarter. The habit of demanding clarity? That's the only definition that lasts.