Gradients has been a quietly good deal for fine-tuning. Submit the job, it distributes the run, you pay for the finished model instead of babysitting a rented cluster. Not the fastest turnaround, but the cost per finished model is well under what I was spending on-demand.
pay-for-result on fine-tuning is underrated. the on-demand cluster babysitting tax is real, half my old bill was idle time between epochs
the economics of pay-per-finished-model change how you experiment too. when idle time is free you overprovision; when you pay per result you actually think about whether the run is worth it. it quietly makes you a better ML engineer.
