AI words you keep seeing, explained properly

Your flight home was cancelled and the airline has offered a voucher. Somewhere in your inbox there is a forty-page travel insurance policy, and somewhere in those forty pages is the answer to whether taking the voucher counts as compensation you have already received. You open a chat app to ask.

Between opening it and getting a usable answer you will walk past half a dozen pieces of vocabulary, most of it unexplained and some of it explained wrongly almost everywhere. What follows is that one conversation, with each word taken up at the moment you actually meet it, because a definition you read six months ago detached from any decision does not survive contact with a real question.

Start at the picker, where the numbers are

You have to choose a model before you can ask anything, and the list is full of names with numbers welded onto them: Qwen3 235B, Ministral 3 8B, Muse Glimmer 30B. Those numbers count parameters, the values a model learns during training and then uses to turn what you typed into what it writes back. The Open Source Initiative's definition of AI weights puts it about as plainly as anyone has: they are "the set of learned parameters that overlay the model architecture to produce an output from a given input".

The number is a measure of size, and size is a weak predictor of whether you will get a good reading of clause 7.3. Part of the reason is that the figure in the name overstates what actually runs on your message, because large models increasingly route each fragment of text through only a slice of themselves. Alibaba's own model card for the build behind our Qwen3 235B entry, Qwen3-235B-A22B-Instruct-2507, gives the shape directly: "235B in total and 22B activated". The rest of the reason is more basic: how well a model handles an insurance policy depends on what it was trained and tuned to do and on what it is allowed to read, and neither of those is recoverable from a parameter count.

You will also see open weights attached to some of those names. It means the trained parameters have been published for anyone to download and run, which is a narrower thing than open source: the same OSI definition asks that open weights arrive with "the data information and code used to derive those parameters", and published weights frequently do not.

The paperclip tells you what each model can see

You pick one and try to attach the policy. Whether that works depends entirely on which model is selected, and this is where multimodal stops being a property of AI in general and turns into a per-model fact. Google's Gemini documentation defines multimodal prompting as prompting "with text, image, audio, and video data", and the useful part is that list, because different models accept different items from it.

Across the 59 models active in our picker at the 19 October 2026 check, 46 accept an image, 24 accept a PDF file directly, and 6 accept video, all six of those being Gemini. So "can AI read my document" has no general answer, and the honest version is unglamorous: pick the wrong one and our app will take the file and then tell you to switch to a model that accepts PDFs. Check what the model can see before you assume the feature is there.

What a token is, and why the ratio moves

The policy goes in. What the model receives is tokens, chunks of text that do not line up neatly with words. OpenAI's own help article on the subject is careful about this in a way that most explainers are not: "A token count is not the same as a word count. The same text can produce different token counts depending on the model, its encoding, and the language." Their working estimate for English is that 100 tokens run to roughly 75 words.

That last clause about language is the part everyone drops, and it is the part that changes your afternoon. The same passage does not cost the same number of tokens in every language it is written in, so a document that sits comfortably inside one model's limits can crowd it once translated. If you have ever wondered why a chat in your second language runs out of room sooner, that is the mechanism.

Why the window fills faster than the number suggests

Those tokens land in the context window, the amount of material a model can hold at once for a single request. That much most glossaries get right. What they get wrong is presenting the advertised figure as a promise. It is a ceiling on what the model will accept, and the range across which it still answers well is a separate and smaller thing. We took that gap apart in our piece on context windows, including the well-documented habit of losing whatever is buried in the middle of a long input.

The version that matters while you are actually chatting is the meter beside the message box. It fills as your conversation grows, and once it leaves its neutral band it tells you the point past which answer quality tends to drop for the model you have chosen. Forty pages of policy plus twenty messages of back and forth is precisely the shape of conversation where that stops being an abstraction, and the right response is usually to open a clean chat and paste in a short summary.

The model says it is thinking. That word is doing a lot of work.

You ask it to compare two clauses. Instead of answering, it shows a shimmering "Thinking..." line, and when the answer arrives there is a collapsible block above it labelled Reasoning that you can open and read.

Nearly everyone reads that block as a window into the model's mind. The model makers say otherwise in their own documentation, which is a strange thing to have to point out. OpenAI's reasoning guide states that "we don't expose the raw reasoning tokens emitted by the model" and that what you can view is a summary of the model's reasoning. Anthropic's thinking documentation goes further and names who writes the summary:

Summarization is processed by a different model from the one you target in your requests. The thinking model does not see the summarized output.

So the trace you are reading is an account of a process you never see, assembled afterwards by other machinery. It is genuinely useful for one thing, which is catching that the model has misread which clause you meant, and you should treat it with suspicion for everything else. Reading it as the order in which the model actually worked will mislead you, and reading it to find out why an answer came out wrong will mislead you twice. Forty-one of the 59 models in our picker can run in this mode, and on 27 of those the thinking cannot be switched off at all, so a slow first response usually means the model is working.

You cannot hear a hallucination coming

The answer comes back sure of itself. The voucher counts as compensation received under clause 7.3, it says, with a clean paraphrase of the clause. Confident phrasing is simply the house style of these systems, and it arrives in exactly the same register whether the model has read your document correctly or has smoothed over a gap with something plausible. That second behaviour is what hallucination names.

The thing to take from the word is narrow and it is this: the writing gives you nothing to go on. There is no hedge that reliably shows up when the model is unsure and no flourish that reliably shows up when it is on solid ground. The same fluency comes out either way, which is why "but it sounded so certain" is the most common sentence in every story about someone getting burned by one of these tools. Our longer piece on hallucinations covers the failure patterns and what genuinely reduces them. For your policy, the move is dull and it works. Ask it to quote clause 7.3 verbatim, then open the PDF and read clause 7.3.

Which two of these are worth keeping

If we had to keep two of these words and let the rest go, we would keep the two that change behaviour. The first is that the visible thinking is a summary written by something other than the model that did the thinking, because knowing that stops you granting authority to a trace that has not earned any. The second is that a confident tone tells you nothing whatsoever about whether the answer is right, because knowing that sends you back to the source document while you still have time to act on it.

Parameters, open weights and multimodal are worth recognising when they scroll past, and that is the whole of their value. Tokens and context windows matter for one decision, which is when to abandon a long chat and start again. A glossary that leaves you knowing eight definitions and behaving exactly as you did before has not earned the time you spent on it, and that is a fair test to apply to the next one you are handed, including this one.