Part of what makes LoRA so effective is that - like other forms of fine-tuning - it’s stackable. Improvements like instruction tuning can be applied and then leveraged as other contributors add on dialogue, or reasoning, or tool use. While the individual fine tunings are low rank, their sum need not be, allowing full-rank updates to the model to accumulate over time.
This means that as new and better datasets and tasks become available, the model can be cheaply kept up to date, without ever having to pay the cost of a full run.
By contrast, training giant models from scratch not only throws away the pretraining, but also any iterative improvements that have been made on top. In the open source world, it doesn’t take long before these improvements dominate, making a full retrain extremely costly.
We should be thoughtful about whether each new application or idea really needs a whole new model. If we really do have major architectural improvements that preclude directly reusing model weights, then we should invest in more aggressive forms of distillation that allow us to retain as much of the previous generation’s capabilities as possible.
