May 17, 2026 Why LLMs Don't Think in Embeddings (Yet) Coconut proved LLMs can reason without text. Why Anthropic, OpenAI and 40 researchers signed a paper hoping the idea doesn't spread. agents interpretability alignment research