AI terminology has become so muddled that even industry insiders struggle to define basic terms like "LLM" and "frontier model" on the spot. Oren Etzioni, a prominent AI researcher, has published a comprehensive glossary that cuts through the jargon and explains what these terms actually mean, exposing how loose language often hides important distinctions about how AI systems work and what they cost.

Why Does AI Terminology Matter So Much?

The problem isn't just academic. When companies use vague language, it becomes harder for customers, investors, and policymakers to understand what they're actually buying or funding. Etzioni notes that terms like "frontier" aren't even technical categories; they're rankings announced by the companies being ranked, making them inherently self-serving. Similarly, "open weights" has become a way for companies to claim the open-source mantle without actually sharing the data or code that made their models work.

The stakes are highest when it comes to costs. Training a model costs hundreds of millions of dollars and happens once. Inference, the process of running the model to answer a question, costs a fraction of a cent per use but happens billions of times daily. Yet companies announce training costs in press releases while inference costs remain hidden until the bill arrives. As Etzioni points out, "Training costs get announced. Inference costs get discovered. Only one of them shows up in a press release".

As Etzioni

How to Decode the Most Confusing AI Terms?

• LLM vs. ChatGPT: ChatGPT is the app you open on your phone or computer; an LLM (large language model) is the AI engine running inside it. Many people use the terms interchangeably, but they describe different things: one is the interface, the other is the underlying technology.
• Pre-training vs. Post-training: Pre-training feeds the model most of the internet so it learns to predict the next word in a sentence; this is the expensive part that costs hundreds of millions. Post-training is where people rank its answers and it learns to give more of what ranked well, turning a knowledgeable but instruction-resistant model into an actual product.
• Tokens vs. Parameters: Tokens are roughly the words going in and coming out of a model; they're what the industry charges by. Parameters, also called weights, are the numbers inside the model that determine how it behaves. A frontier model has hundreds of billions of parameters, with the largest running to trillions, and nobody can tell you what any single one does.
• Chatbot vs. Workflow vs. Agent: A chatbot answers your question and stops. A workflow runs the steps you defined in your order. An agent receives a goal instead of steps and works out for itself what to do, calling other software and checking results until it's done or stuck. The useful test: if it decides its own next step, it's an agent.
• Hallucination vs. AI Slop: A hallucination is a confident falsehood, like a citation to a paper that doesn't exist. The model isn't lying; it has no notion of truth to violate. AI slop is different: accurate, fluent, and worthless, like a LinkedIn post that says nothing in 300 fluent words.
• RAG and Memory: RAG (retrieval-augmented generation) searches a document collection and drops relevant passages into the context window before the model answers. Memory is a feature that saves facts about you and slips them back into the context window later. Nothing in the model actually remembers you; the app keeps a file on you and pastes it in before every conversation.

What's the Real Issue With "Open Weights"?

One of Etzioni's sharpest observations concerns the term "open weights," which has become a marketing tool rather than a genuine commitment to openness. When a company releases a model with open weights, an AI expert can download the model and run it on their own server. But you don't get the data or the code that made it, which means you can't truly understand or modify how the model was built.

True open source means data and software that experts can use and modify. Almost no major model offers this; AI2's Olmo is a rare exception. Etzioni calls the current practice "open washing," a term that captures how companies claim the open-source mantle without giving much away. This distinction matters because it affects whether researchers can actually audit, improve, or build upon the work.

How Do Companies Hide Costs in Their Messaging?

Etzioni highlights a critical gap between what companies advertise and what actually determines your bill. When a company announces a model with 1 trillion parameters, that number gets quoted like horsepower in a car commercial. But the number that shows up on your invoice is how many tokens it takes to answer your question, and that's the number nobody advertises.

Inference has become the dominant cost in AI systems. In 2023, inference accounted for roughly one-third of all AI compute spending. By 2026, that share had grown to about two-thirds, meaning most of the computational resources in the AI industry are now spent answering questions rather than building models. Yet this shift remains largely invisible in how companies market their products.

Why Does the Distinction Between Alignment and Censorship Matter?

Etzioni's glossary also tackles the politically charged distinction between alignment, guardrails, and censorship. Alignment is the research problem of getting a model to do what people want when nobody's watching. Guardrails are the rules behind its refusals, like "no, I won't tell you how to make a bioweapon." Censorship is a guardrail that blocked something you wanted.

The same refusal can be called "safety" in a press release, "guardrails" in the documentation, and "censorship" by critics. The terminology you choose shapes how people perceive the decision, which is precisely why Etzioni's effort to define these terms clearly matters. Clear language forces clarity of thought.

By publishing this glossary, Etzioni has provided a reference guide that cuts through the marketing speak and helps readers understand what AI companies are actually claiming. In an industry where terminology often obscures more than it reveals, plain English definitions become a form of accountability.