This blog post discusses the concept of Future Oracles in the context of predicting future tokens in language models (LLMs) using activations. The author shares findings from their research, which includes training a meta-model to analyze how well LLMs can anticipate future outputs based on current activations. They compare this method to previous approaches, noting potential improvements while acknowledging the need for further validation. The post highlights the author's personal experimentation and thoughts on the subject, indicating a journey of discovery rather than a polished end result.