Training

Training is the process of iteratively updating a neural network's weights to learn patterns in data — mathematically, it is gradient-based minimisation of a loss function. Training a modern LLM is a massive engineering effort: trillions of Tokens, thousands of GPUs running for weeks, distributed parallelism strategies, MFU tracking and constant failure management. It usually unfolds in several phases — Pre-training, Post-training and Fine-tuning — each optimising for a different objective. Don't confuse training with Inference: one creates the model, the other puts it to work.