Building reward functions for reinforcement learning models is hard. Imagine trying to code exactly *what* you want a robot dog to do – it's a coding nightmare. Reinforcement Learning from Human Feedback (RLHF) aims to simplify this process by learning reward functions directly from human preferences. Show a person two videos of a robot doing different things, ask which they prefer, and use that preference to shape the reward. The problem? It takes *forever*. Thousands of comparisons might be needed just for a simple task. Researchers are exploring whether Large Language Models (LLMs) can accelerate this process. A new technique called In-Context Preference Learning (ICPL) uses LLMs to generate reward function code directly. It then shows videos of the resulting robot behavior to a human, who ranks their preferences. The best and worst examples are fed back to the LLM, which refines the code in the next iteration. Experiments show ICPL needs drastically fewer human comparisons than traditional RLHF, even rivaling methods that use hand-coded reward functions. Impressively, ICPL even tackled a subjective task: making a simulated humanoid jump like a real human. By iteratively incorporating human feedback, the LLM learned to generate reward functions that produced surprisingly realistic jumping behavior, far beyond simply maximizing airtime. While limitations remain—some tasks are too complex for humans to judge from video alone, and the initial diversity of generated reward functions impacts performance—ICPL opens exciting new avenues for aligning complex AI behaviors with human intentions.