I’m wondering what would happen if we trained LLMs on LeetCode challenges—not using pre-written solutions, but allowing the model to come up with its own. The feedback signal would be whether the code passes the challenge or not, ignoring other aspects like formatting or variable naming, which might be terrible. Starting with a base model and fine-tuning it on LeetCode challenges, I’m curious what the resulting code would look like: an unreadable mess, clean and maintainable, or something in between.