The gardener does not pull the shoots

There is a story Mencius tells about a farmer from Song who, worried that his rice shoots were growing too slowly, went through the field pulling each one upward to help. He came home exhausted and announced his success. His son ran out to look and found the field full of withered plants.

Every team that has fine-tuned a language model owns a version of this field.

The temptation is structural. You have a base model — capable, general, slightly wrong for your purpose — and you have an instrument, the gradient, that can move any behavior in any direction. The instrument makes force cheap. And so the first fine-tune is almost always too strong: too many epochs, too high a learning rate, a dataset of a thousand examples that all sound like each other. The model learns exactly what you taught it. That’s the problem.

Catastrophic forgetting is just pulling shoots

What the loss curve doesn’t show is what the update displaced. Aggressive fine-tuning on a narrow distribution buys its wins by spending the base model’s generality — the very thing you chose it for. The literature calls this catastrophic forgetting, but the mechanism is mundane: you asked every parameter to serve the new objective, and parameters that were quietly load-bearing for everything else obliged.

The whole modern toolkit for adapting models reads, in this light, like a manual of restraint. LoRA freezes the network and permits only a low-rank whisper of change. The KL penalty in RLHF is an explicit leash — you may please the reward model, but you may not stop being yourself to do it. Early stopping, small learning rates, replay of general data mixed into the specialty diet: every one of these techniques is a formalization of the same instinct the farmer lacked.

Reward is a description of your wanting, not the world

Reinforcement learning from feedback sharpens the lesson, because there the model doesn’t just learn your data — it learns your desire, including the parts you didn’t know were in it. Optimize hard against a reward model and you get sycophancy, verbosity, the confident tone that raters mistake for correctness. The policy hasn’t misunderstood the reward. It has understood it perfectly, more literally than you meant it, the way a wish gets understood in every story about wishes.

Goodhart’s law is usually quoted as a warning about metrics. For model training it’s closer to a law of optics: the reward model is a mirror, and sufficient optimization pressure will reproduce every smudge on it at full resolution. The practical response is again a lighter hand — weaker pressure, earlier stopping, diverse raters, and the humility to treat the reward as evidence about quality rather than the thing itself.

When to touch the weights at all

The quiet endpoint of this reasoning is a question the field asks more often each year: did you need to fine-tune at all? A great deal of what teams reach for gradients to fix turns out to be a prompting problem, a retrieval problem, or a tools problem — places where you can change the model’s situation instead of its self. Changing the situation is reversible, inspectable, and cheap. Changing the self is surgery.

Surgery has its place. There are behaviors — formats, refusals, house style, deep domain fluency — that genuinely live in the weights and are worth putting there. The point is not never-touch. The point is the gardener’s sequence: soil first, then water, then light, and the shears last, sharpened, briefly.

Mencius drew his own moral from the farmer of Song, and it needs no updating for the gradient era: there are two ways to ruin a growing thing. One is to neglect it. The other — more common among the intelligent, the well-resourced, and the impatient — is to help it too much.

The shoots know how to grow. Most of what your base model needs, it already has. Arrange the conditions, apply the smallest sufficient pressure, and then do the hardest thing in engineering.

Wait.