Open weights, RAG, agent, frontier model — ever found yourself in the middle of an AI conversation nodding along like you totally get it, when in reality you understood almost nothing?
That scenario is way more common than people think, and it is definitely not your fault. The AI vocabulary has grown at the same absurd speed as the technology itself, and today it is practically impossible to follow any industry news without running into at least five terms that sound like they came out of a manual written by robots for other robots.
The real problem is not the complexity of the concepts — most of them are actually much simpler than they seem. The problem is that almost nobody explains them honestly, without piling jargon on top of more jargon. That is exactly what Oren Etzioni, researcher and one of the most respected voices in AI, had in mind when he originally published an opinionated — and slightly snarky — glossary on GeekWire covering the terms that confuse the most people, including those who work with this stuff every single day.
As he puts it: jargon is boring. The goal here is to break these terms down into plain language, with practical context, real-world examples, and the kind of honesty that is usually missing from textbook definitions 👇
Model, LLM, and frontier model
Let us start by clearing up a basic mix-up. ChatGPT is the app you open. An LLM, or large language model, is the artificial intelligence running behind it. Everyone says LLM, but very few people could define it on the spot: it is a system trained to predict the next word in a sequence of text, and it is from that seemingly simple mechanism that everything else follows.
The term frontier model is not really a technical category. It just refers to the handful of most powerful and capable models at any given moment. The funny part is that this trophy changes hands all the time, and the ranking is usually announced by the very companies being ranked. In other words: it is marketing dressed up as a technical classification.
Prompts, tokens, and parameters
A prompt is the thought, question, or instructions you give the model — plus everything the app added before you even noticed. The model takes that prompt and generates words, both in its internal reasoning process and in the response that shows up on your screen.
Tokens are, roughly speaking, the words going in and coming out. The model slices your prompt into tokens, produces more tokens when replying, and that is exactly what the industry charges you for. Here is an important detail: the number of parameters gets announced like engine horsepower, but the figure nobody advertises is how many tokens are needed to answer your question. And guess what? That is precisely the number that shows up on the bill at the end of the month.
Receive the best innovation content in your email.
All the news, tips, trends, and resources you're looking for, delivered to your inbox.
By subscribing to the newsletter, you agree to receive communications from Método Viral. We are committed to always protecting and respecting your privacy.
Parameters, also called weights, are the internal numerical values of the model. Think of them as the fine-tuning knobs on an instrument: each parameter is a small setting that, together with billions of others, defines how the model interprets and responds. A frontier model has hundreds of billions of them, and the largest ones are already reaching into the trillions. The curious thing is that nobody can explain what any individual parameter does — they encode statistical patterns, not facts stored like a database.
Pre-training, post-training, and fine-tuning
Pre-training is feeding the model a huge chunk of the internet so it learns to predict the next word in a sentence. This is the expensive part of the process, costing hundreds of millions of dollars and producing something that knows a lot but cannot follow an instruction properly.
Post-training is when people rate the model’s responses, and it learns to deliver more of what scored well. Fine-tuning is post-training done by you, on someone else’s model, using your own data. To put it bluntly: pre-training gives you a model that is bad at answering questions; it is post-training that turns everything into a usable product.
This refinement process usually involves a technique called RLHF — Reinforcement Learning from Human Feedback. It is what teaches the model to be more helpful, safer, and more aligned with what people actually expect from an assistant.
Training from scratch versus distillation
Training from scratch means buying (or renting) the computers and doing all the work of building and training a model. Distillation trains a cheap model using the outputs of an expensive model, inheriting the behavior without paying the full price. The catch is that distillation violates the terms of service of most AI companies.
This issue made headlines when OpenAI accused DeepSeek of distilling its models — a position that is at least a little ironic for a company that trained its own systems on the entire internet without asking anyone for permission. Learning from other people’s work seems great, right up until the other people are you.
Training versus inference
Training is how you build a model. Inference is what happens every time it responds: the model runs and produces a result. Training is a one-time cost that lasts months and burns hundreds of millions of dollars. Inference is a cost you pay forever — each response costs a fraction of a penny, but that adds up when it happens billions of times a day.
Here is a truth that does not get enough attention: training costs are announced at press conferences, but inference costs only become clear over time. According to Deloitte projections, inference is expected to consume about two-thirds of all AI computing power by 2026, compared to just one-third in 2023. In other words, the real spending is shifting to the part nobody puts in the official press release.
Open weights, open source, and API-only
We normally use LLMs through apps like ChatGPT, Claude, or Gemini. But specialists often want the model itself, not just an app wrapped around it.
Open weights means a specialist can download the model and run it on their own server — but without access to the data or the code that created it. True open source means data and software that anyone can use and modify, something almost no major model actually offers. Olmo, from AI2, is one of the rare exceptions. An API-only model gives you nothing: you send your text to the company’s computers, the answer comes back, and you pay for each use.
An honest observation worth making: open weights is usually a way to claim the open-source halo without actually giving away much. A Stanford researcher even called it open distribution — openly distributed, but far from truly open source.
Context window, memory, and RAG
The context window is the amount of text the model can hold in its head at once, including your question and everything pasted into the conversation. Memory is a feature that saves facts about you and injects them back into the context window later.
RAG, which stands for retrieval-augmented generation, searches a collection of documents and inserts the relevant excerpts into the context window before the model responds. This reduces hallucinations and makes the model much more reliable for tasks that require up-to-date or company-specific data.
An uncomfortable truth: nothing inside the model actually remembers you. The app keeps a file with your information and pastes it at the beginning of every conversation — a much less glamorous way of describing the exact same feature sold as memory.
Chatbot, workflow, and agent
A chatbot answers and stops there. A workflow executes the steps you defined, in the order you set. An agent receives a goal instead of steps and figures out on its own what to do, triggering other software and checking results until the task is done or it gets stuck.
An example makes everything clearer: ask about a delayed flight and a chatbot recites the airline’s policy; a workflow submits the refund form you built; an agent just books a new flight for you. The test is simple and useful: if it decides its own next step, it is an agent. If you decided the steps, it is a workflow.
Hallucination, AI slop, and AI cream
A hallucination is a falsehood stated with total confidence, like citing an article that does not exist. The model is not lying — it does not even have a concept of truth to violate. It simply produces text that looks like the right kind of answer.
AI slop is a different kind of failure: content that is accurate, fluent, and completely useless. Think of that LinkedIn post that says absolutely nothing in 300 flawless words. It is no coincidence that slop was named the 2025 word of the year by the Merriam-Webster dictionary. AI cream is the rare and desirable case: excellent writing produced with the help of AI. The irony is that nobody sets out to produce slop — everyone believes they are making cream.
Alignment, guardrails, and censorship
Alignment is the research problem of making a model act the way people want, even when nobody is watching. Guardrails are the rules behind the model’s refusals — that classic no, I am not going to explain how to make a bioweapon. And censorship is, essentially, a guardrail that blocked something you wanted.
Notice the wordplay: the same refusal becomes safety in the press release, a guardrail in the technical documentation, and censorship when people complain on social media. It all depends on who is telling the story.
Why understanding this vocabulary actually matters
Mastering these terms is not just about sounding smart in a conversation. It is the difference between making informed decisions about which tools to adopt, being able to evaluate AI vendors with a critical eye, and seeing what is behind the industry’s grand announcements. When you understand that a frontier model is a ranking announced by the ones being ranked, or that inference is the cost that shows up after the party, you start reading tech news with a whole different perspective.
The field of artificial intelligence is advancing at a breathtaking pace, and new terms pop up almost every week. But the fundamental concepts we covered here remain solid and continue to be the foundation of everything. Those who truly understand them — without being intimidated by jargon or impressed by marketing — come out ahead in any discussion about the present and future of this technology 🚀
