TL;DR: Win up to $100,000 for finding an important task where larger language models do worse.