“The effective context window of a large language model is significantly smaller than its advertised size, with intelligence and quality depreciating once it is 20-40% full.”