A model's weights freeze the moment its training run ends

A December 2022 fact, asked of a model finished reading in 2021

Ask GPT-4 who won the 2022 World Cup and it has no way to get the answer right, not because Argentina's win over France is obscure, but because of when the model stopped reading. OpenAI's own account of the model states the boundary directly: "GPT-4 generally lacks knowledge of events that occurred after the vast majority of its pre-training data cuts off in September 2021, and does not learn from its experience."¹ The World Cup final was played fifteen months after that cutoff. Every sentence GPT-4 can produce about it was assembled from whatever came before September 2021, and nothing about it came from anywhere else.

The whole exchange, prompt in, answer out, is produced by one thing: a fixed set of numbers called weights, set once during training and arranged into a network that reads your words and produces the next ones. Nothing else is in the box: no live connection to the internet, no memory of yesterday's conversation, no clock. Ask the same question twice, a year apart, and the model runs the identical fixed function on your words both times.

The natural guess is that a system used by hundreds of millions of people every day keeps learning as it goes, the way a person does. It doesn't.

Why the weights don't just keep updating

Training is a distinct phase that ends. After it ends, the weights are constants a program reads, not a process still running. GPT-4's own report says the model "does not learn from its experience" for exactly this reason: once pretraining is over, no further gradient step touches those numbers. Meta describes the same shape of limit for a different model family. Llama 3 was pretrained on "a variety of data sources containing knowledge until the end of 2023,"² a single closed window. Two labs, two model families, the same shape of claim: a training run collects data up to some point, stops, and the weights that come out of it are what the deployed model has, permanently.

Why not just keep training as new data arrives? Because a network that keeps learning overwrites what it already learned. DeepMind's work on catastrophic forgetting describes exactly this failure. Train the same network further on a new task, and "knowledge of previously learnt task(s)... [is] abruptly lost as information relevant to the current task... is incorporated."³ The reason is mechanical: "weights in the network that are important for task A are changed to meet the objectives of task B." A system that has to answer the same question the same way next year cannot afford to keep quietly rewriting the numbers that produced this year's correct answer. Freezing the weights once training stops is the design that makes a deployed model a fixed function at all.

Because of that stop point, no gradient step ever touched a match played fifteen months after data collection stopped. The words "Argentina" and "World Cup" may well sit together in GPT-4's weights, from years of football writing before 2021, but nothing ties either word to the score of a match played after the cutoff. Asked directly, the model can do one of two things. It can say plainly that it doesn't know, or it can produce a fluent, wrong answer built from whatever pattern in its weights looks closest, the construction this desk's earlier lesson on hallucination already describes. A frozen model with no real record of an event is exactly the setup that produces a confident, invented one.

What a search tool adds, and what it doesn't

The 2020 paper that introduced retrieval-augmented generation gave this split its working vocabulary. A language model's weights hold "parametric" knowledge, learned once during training. A separate system can supply "non-parametric" knowledge instead, drawn from "a dense vector index... accessed with a... neural retriever," fetched fresh at the moment of the question.⁴ The distinction matters because from the outside, the two look identical: the same fluent answer, in the same chat window. What differs is where the fact came from. Parametric knowledge was baked in during training. Retrieved knowledge arrived seconds ago, in the prompt, the same channel this desk's lesson on in-context learning already showed can carry a pattern the model was never trained on.

Give a search tool to a model asked about the World Cup, and it can now answer correctly, fetching a page that names Argentina's win and reading it back. Google says exactly this about its current model: "Gemini 3 models have a knowledge cutoff of January 2025. For more recent information, use the Search Grounding tool."⁵ It does not say the tool moves the cutoff forward, or teaches the model the new fact. It hands the model a page to read, once, for this one exchange. Close the chat, and the weights are exactly what they were before the World Cup came up: a fixed set of numbers, untouched.

This is the gap between a product getting better and a model learning something. When a chatbot starts answering questions about last week's news, what typically changed is the product wrapped around the model, a search tool bolted on, not the weights underneath. Gary Marcus, writing about a chatbot caught denying an already-reported story, calls exactly this kind of fix a "band-aid".⁶ It patches the one conversation in front of you. The next model a lab releases still needs its own training run and its own new set of frozen numbers to know anything by default.

Even the labs don't write down one clean date

The frozen-weights part of this story is settled: every lab that publishes a model card says some version of it. What is genuinely unsettled is whether "the cutoff" names one clean date at all. Anthropic's own documentation lists two different cutoff dates for the same model, not one, in tables headed "reliable knowledge cutoff" and "training data cutoff":⁷

Anthropic explains the split plainly: "Reliable knowledge cutoff indicates the date through which a model's knowledge is most extensive and reliable. Training data cutoff is the broader date range of training data used."⁷ The company that trained the model is telling its own customers that data collected right up to the training cutoff is present in the weights but thin, patchy enough that the label warns against leaning on it. If the company that owns the weights won't stand behind their own tail end, the cutoff was never one instant. It is a range with a soft edge.

The soft edge gets worse once anyone measures it instead of reading the vendor's label. A 2024 study set out to test whether a training corpus's claimed cutoff matches what a model trained on it can actually demonstrate, and named the measured version the "effective cutoff," distinct from "the LLM designer reported cutoff."⁸ The gap it found in one widely used open corpus, RedPajama, was not months. The corpus's own file listing cites a 2023 Wikipedia dump. But the study traced over 80% of the Wikipedia text actually included back to older snapshots crawled years earlier.⁸ Its effective, demonstrable cutoff landed around 2019, four years before the date printed on its label. A reported cutoff describes what a data-collection script was told to fetch. An effective cutoff describes what the resulting weights actually carry. Nothing guarantees the two match.

A cutoff can drift the other way too: forward, after the fact, without the model ever learning anything at inference. A 2026 study trained models on a pretraining set that included contaminated information, then continued training them on clean data until the contamination's measurable effect nearly vanished. The information was not erased. Running the same models through a later training stage, supervised fine-tuning or reinforcement learning, brought it back: "the leaked information is merely submerged not erased and is readily rediscovered during post-training."⁹ That resurfacing still happens through training. No live conversation pulled the fact back in; a later training stage did, one that ran after the pretraining cutoff most people mean when they cite "the cutoff" at all. The frozen-function fact from earlier in this lesson still holds at every instant. It is only that training turns out to be more than one stage, and a date attached to the first of them does not describe the last.

Anthropic and Google publish their models' cutoffs as facts about the model, stated by the company. Neither claims the model reliably reports its own cutoff, and the entire premise of measuring an "effective cutoff" is that asking directly does not settle the question.⁸ Ask GPT-4 both what its own cutoff is and who won the 2022 World Cup, and both answers come from the identical process: whatever the weights happen to encode, delivered with the same confidence either way.