Hamel Husain shows that larger context windows actually hurt language model performance. When models see many similar documents, they pick the wrong one or make things up. Claude admits confusion while ChatGPT confidently hallucinates, proving that feeding AI only what it needs still matters more than dumping everything in.
Models have huge context windows now. It's tempting to stuff as many documents as you can in there and let the AI just sort it out. But research shows that the longer your input, the worse the AI performs. Retrieval researchers tested what happens when your context has similar looking information. Meeting notes from different weeks or reports from different quarters. The model picks the wrong one or makes something up. More documents, worse results. Here's how bad it gets. They told the AI, repeat this text back to me word for word. As the input grew longer, it couldn't do it. The AI couldn't even copy what was right in front of it. And it fails differently depending on what AI you use. Claude says, I don't know. ChatGPT gives you a confident but wrong answer. So either nothing or something you can't trust. This is why finding the right information before you hand it to the AI still matters. People say RAG is dead, just use a big context window, but this research says the opposite. Give the AI what it needs, nothing more. The full talk is linked in the comments.
