if you're already starting to fine-tune open source models for smaller tasks, using something like the @runanywhereai sdk can let you offload some of the inference to user devices, lowering cloud inference costs and latency but there's probably more fun/unique applications i'm not thinking of