Training vs inference
Both run numerical computations through a model. The critical difference is whether the process uses an objective and optimizer to update persistent parameters.
| Question | Training | Inference |
|---|---|---|
| Purpose | Improve parameters for an objective | Produce outputs using existing parameters |
| Input | Examples and a learning objective | A supplied sequence, possibly with tool results |
| Main work | Forward pass, loss, backward pass, update | Forward computation and decoding |
| Persistent change | Optimizer updates model parameters | Ordinary inference keeps parameters fixed |
| Useful question | Did held-out performance improve? | Is this response useful and supported? |
Reading a worked example in a prompt can change the next response without creating a training update. Fine-tuning on examples changes parameters across updates.
Follow the evidence
Read the connected lessons for detailed mechanisms, assumptions, and primary references.
Pretraining: learning from continuations
Large collections of examples teach a model reusable patterns before any particular assistant conversation.
Gradient descent: changing the weights
A gradient points uphill in loss. An optimizer uses it to take a carefully sized step in the other direction.
Autoregression: one choice becomes the next input
Generation repeatedly predicts, selects, appends, and predicts again.
In-context learning: adapting without weight updates
Instructions and examples in the input can change a model’s behavior while its learned parameters stay fixed.