A thread about the gullibility of language models. We know that language models are great at summarization. So if you build an agent and ask it to summarize and report on text from the internet, you would expect it to do reasonably well. Not so fast. 🧵