The case for the return of fine-tuning

No post summary.

Fine-tuning is making a comeback as a niche tool for specific tasks, but the debate is whether it's worth the effort vs. using larger models with better prompting.
The community is split: some see it as essential for control, IP ownership and cost savings, while others argue that frontier models have made it redundant for most use cases.
Key friction points: cost-benefit of training vs. inference, data-labeling overhead, and whether the juice is worth the squeeze when prompt-engineering can achieve similar results.
OpenAI, Anthropic and others are quietly re-opening fine-tuning APIs, while simultaneously pushing the narrative that "you don't need it anymore"—a tension that may be more about GPU budgets than user needs.
The open-source community is rallying around LoRA and QLoRA as a compromise, but the real question is whether the economics of serving a custom model will ever make sense versus just paying per-token for a larger model.