training run

advancedwidespread

A training run is a single, end-to-end process of training an AI model, during which the model learns its task from the data.

In plain terms

An AI model is not created by programmers writing line by line what it should do. Instead, they show it countless examples, and the model derives the patterns itself. A complete run of this learning process, from start to finish, is called a training run.

It is like baking bread: you mix the ingredients, set the temperature and time, then put the tray in the oven. You monitor the process, and if something goes wrong, you might adjust the temperature slightly, but you cannot swap the dough. If the result is not good, you must start over with new dough and modified settings, which counts as a new run.

For the largest language models, a single run can last weeks or months, on tens of thousands of specialized chips, or even hundreds of thousands for the largest ones. The cost can be tens of millions, or even hundreds of millions of dollars for the largest. Therefore, when a company launches a new large run, it is considered an event in the industry: this forms the basis of the next model generation.

A training run is not the same as using the model. When you ask a chatbot a question, the model is no longer learning; it works from knowledge fixed during previous runs.

How it works

Preparation: developers compile the training data, select the model architecture, and set parameters, such as how fast it should learn and how long the run should last.
Learning cycle: the model receives data in small batches and attempts to guess the correct result for each batch (for example, a language model predicts the next word in a text). The system measures how far off the guess was and adjusts the model's internal numbers, the parameters, accordingly.
Iteration: this step takes place tens of thousands of times during a run, and even millions of times with large models, and the error gradually decreases. Developers follow curves to see if the model is truly improving.
Saving: the system periodically saves a snapshot of the model's current state, so that after a hardware failure you do not have to start over and can return to an earlier, better state.
Wrap-up and verification: after the main training run, large models are usually refined with additional, shorter training steps, then tested on tasks they did not see during training. Only then is a decision made on whether to release them.

An everyday example

You read in the news that an AI company's new model 'knows the world up to last spring'. This is because data were collected up to that date for the training run. What has happened since, the model does not know on its own; at most, it looks it up via internet search.

Who is affected

developerThey plan it, launch it, and monitor the run; a bad setting can waste days of work and significant compute time. → Code

entrepreneur, executiveIf your company trains or fine-tunes its own model, the number and duration of runs can be a significant cost item.

investorBehind the billions AI companies spend on data centers and chips lie training runs, and increasingly, the computing demands of serving ready-made models.

everyoneThe knowledge and blind spots of the chatbot you use stem from when its training run took place and what data it used.

A common misconception

MythAI continuously learns from what I write to it during conversation.

In realityA model's knowledge is fixed at the end of a training run and does not change during your conversations. Some chatbots have a 'memory' feature that can remember things about you, but these are only retrieved as notes; the model itself does not change. Companies can use conversations for a later run, but that constitutes creating a new model.

Limits and risks

The outcome of the run is not fully predictable: even with the same data and settings, a slightly different model may emerge because the process contains random elements.
A large training run consumes vast amounts of electricity and money, so only a few well-funded companies can train the largest models.
What is erroneous, biased, or missing in the training data gets baked into the model during training and is difficult to fix afterward.
Successfully completed training does not guarantee a good model: it may have merely memorized examples and perform poorly in new situations.

Similar, but not the same

inference – Inference is the use of the trained model when it answers your question, while the training run is the process in which the model is created.
fine-tuning – Fine-tuning is the targeted further training of an already trained model, that is, one type of training run, usually on a much smaller scale.
epoch (training cycle) – An epoch means the model has gone through the entire training dataset once. Smaller models may consist of many such rounds in one run, while large language models often see the data only roughly once.

Related concepts

Broader

machine learning, model training, deep learning

Narrower

pre-training, fine-tuning, epoch, checkpoint

On the edge

training data, Hyperparameter, computing capacity, GPU, overfitting

A brief history

The term originates from computer terminology, where "run" refers to a single execution of a program. The backpropagation method, which underlies today's training runs, has roots dating back to the 1970s (Linnainmaa, Werbos), but a 1986 study by Rumelhart, Hinton, and Williams made it widely known.

Where things stand

At the end of 2026, the training run is the basic operation in building any AI model, from small research experiments to industrial-scale systems. The size and cost of the largest runs continue to grow, so their energy demand and the data centers they require have become economic and political issues.

Status: 2026 November

Also known as

training run · training run · training run · training run · training run · model training run

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.