Reinforcement Learning Fine-Tuning Is the New MLOps Primitive

The Ramp story is specifically about RL fine-tuning, not supervised fine-tuning. They used reinforcement learning to train a 9B model on catalog review episodes, graded by a scorer (likely a frontier model itself), and the result outperformed those same frontier models on the task. This is a meaningful technical detail: RL fine-tuning lets you optimize for actual task outcomes rather than imitation of existing outputs.

The Kimi Linear architecture paper appearing on HN the same day, focused on efficient attention, fits the same broader theme: the research community is actively working on making capable models cheaper to run and train. PyTorch being described as a 'reference language' rather than just a framework also fits here: the tooling is stabilizing around a set of primitives that make this kind of work more accessible.

The barrier to running this kind of experiment is dropping. A year ago, RL fine-tuning at this scale was a research project. Now it is a $500 weekend experiment that a startup can run.