The model is a snapshot: training cutoffs

Note 01 described training as a game played trillions of times over an enormous pile of text: cover the next word, score it, nudge the numbers. What it did not say is that the game ends. On some date, the pile of text stops being added to, the nudging stops, and the numbers are frozen. From then on, that is the model. A model is a snapshot of the written world up to one date. Everything after that date, it has never read.

That date is called the training cutoff, and it is the plainest fact about a model that people forget, because nothing on the screen reminds you of it. The chat tool answers today's question in today's tense, fluently, from a world that stopped some months or years ago.

Why it answers anyway

Nothing in the loop from Note 01 checks a calendar. Ask "who is the CEO of Akamai?" and the model scores next words from the patterns it learned; the name that followed that question most often in its text wins. If the person changed after the cutoff, the model has no way to feel that. The old name is still the highest-scoring continuation, and it comes out with full confidence. The same happens with prices, versions of software, laws, the newest phone, and, most reliably of all, the newest AI model: a model will tell you with a straight face that it is the latest, because when its text was collected, it was.

Better-trained models have learned to add a hedge: "as of my last update, this was the case, but it may have changed". That hedge is worth reading properly. It is not the model looking at a clock. It is a phrase it was trained to produce around dates and people and prices, because those are the kinds of things that change. It is a good habit, and it is still a prediction.

How a snapshot learns something new

It does not, and this is the second half of the note. The frozen numbers do not update when you tell the model something, and they do not update when it searches the web. There is exactly one way for a model to use information from after its cutoff: someone puts the text in front of it, in the same place your question goes, so that it becomes part of "the text so far" for this one answer.

So "search on" in a chat tool does not mean the model has been updated. It means the product may run a separate step: send your question to a search engine, take a few results, and paste their text in front of the model along with your question. The model then reads that text the way it reads anything you paste, and writes an answer from it. That is a bolt-on, not a change to the model, and two things follow. The bolt-on decides when to run, not you; in my practical, with search switched on, the model declined to search and said so. And whatever it fetched is used for that one answer and then gone.

The same door explains a small thing that surprised me. With search off, the model told me it was answering "from the current date of September 8, 2026". It cannot know that; its text stopped in January 2025. It knew because the product had written today's date into instructions the model reads before my question, instructions I never see. That is Note 11, and this is the first time the notebook has bumped into it.

What this means at the keyboard

Before asking anything that could have changed, ask yourself when it could have changed. If it is after the model's cutoff, the model cannot know, and a fluent answer is a guess dressed as a fact. Three habits follow. Know the cutoff of the tool you use; every provider publishes it, and it is usually a year or more behind today. For anything dated, either turn search on and check that it actually ran, or paste the current text in yourself. And read the hedge "as of my last update" as an instruction to go and check, not as a footnote.

The practical, and what it showed

Practical 04 is five questions in a chat tool: ask the model its cutoff; ask for last month's news with search off, then on; ask a fact that could have changed, with search off, then on. Ten minutes. The interesting results are the two failures the practical is fishing for: a model that invents a news item rather than saying it cannot know, and a model that is right only because search quietly ran.

My run caught neither, which is itself the result. Asked for last month's news without search, the model said plainly that it could not know, named its cutoff, and said that anything it offered would be "guessing or potentially hallucinating". With search on, it still did not search; it said it would need to. On the CEO question it gave the right name both ways, hedged without search and unhedged with. A well-behaved run, and the note's job is to say why that behaviour is trained, not guaranteed: an older or smaller model asked the same questions will often produce a confident August headline that never happened. Run it in a second tool and see.

it knew today's date. it had never read a word from this year. someone told it.

One layer deeper *

The model's own cutoff is a prediction too. When you ask a model its training cutoff, it does not look one up; it produces the most likely answer, which is usually a date it was told in its hidden instructions or one that appeared often in its training text. Models misreport their cutoff fairly often, sometimes by months. The reliable number is in the provider's documentation, not in the chat.

The cutoff is not a wall. Text about a period thins out before the cutoff rather than stopping dead: the internet keeps writing about an event for years after it happens, so the last few months before the cutoff are underrepresented compared with how they will eventually be covered. In practice a model is reliable up to a little before its stated cutoff and patchy in the final stretch.

Search is one case of a general idea. Fetching text and placing it in front of the model at answer time is called retrieval, and "chat with your documents" products, search-enabled assistants, and most enterprise AI tools are the same pattern with different sources. It is the subject of Section 4. The thing to hold on to from this note is where the door is: the text so far, which has a fixed size (the context window, Section 3).

Updating the numbers is possible but rare. A provider can continue training a model on newer text, or train a new one; that is why models have version dates. Nothing a user does through a chat tool changes the numbers. Features called "memory" store text about you and put it back in front of the model each time; they do not train it. That distinction is Note 09.

Why models misjudge their own generation. The most common cutoff mistake in daily use is a model's belief about which models exist. Ask which AI model is newest and it names one from before its cutoff, because that is the newest one it read about. It is a harmless example of the mechanism and a useful test: if a tool gets this wrong, it is answering from the snapshot, and everything dated deserves the same suspicion.

Sources

Provider documentation for published training cutoffs (OpenAI and Anthropic model pages, checked 8 September 2026). Dhingra et al., "Time-Aware Language Models as Temporal Knowledge Bases" (TACL 2022) for how knowledge of a period thins before the cutoff. Lewis et al., "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks" (2020) for the search-as-bolt-on pattern. The practical results are my own, Claude Haiku 4.5 in a consumer chat tool, 8 September 2026. Terms used: model, context window, training cutoff.