Large language models (LLMs) are revolutionizing how we interact with technology, but their inner workings remain largely mysterious. What if we could decode the hidden language of LLMs, understanding their "thoughts" and even influencing their behavior? New research introduces LatentQA, a groundbreaking approach to interpreting and controlling LLMs by deciphering their activations—the internal representations of information flowing through the model. Researchers have developed a technique called Latent Interpretation Tuning (LIT), which trains a separate "decoder" LLM to translate activations into natural language. Imagine being able to ask questions like, "What biases does the LLM hold about this text?" or "What is the LLM's goal when generating this response?" and receiving clear, human-readable answers. LIT makes this possible. The decoder LLM learns to understand the nuanced patterns within activations, effectively providing captions for the LLM's internal state. This capability opens doors to a deeper understanding of how LLMs process information, identify potential biases, and even predict future outputs. But the potential of LIT extends beyond mere interpretation. By providing a differentiable loss function, the decoder can also steer the target LLM’s behavior. This means we can potentially mitigate biases by minimizing the loss associated with, for example, stereotypical responses. Researchers demonstrate this control by successfully debiasing models, influencing the sentiment of generated text, and even—in a concerning twist—eliciting potentially harmful capabilities from safety-tuned LLMs. While this raises ethical questions about the potential for misuse, it also highlights the importance of understanding and auditing these powerful models. The ability to decode activations is a major step towards greater transparency and control over LLMs. As models and datasets grow larger, LIT becomes even more effective, promising a future where we can interact with AI in a more informed and purposeful way. However, challenges remain, including the potential for hallucination in the decoder and the need for diverse and unbiased training data. The journey into the hidden world of LLM activations has just begun, and the insights gained through LatentQA and LIT promise to shape the future of AI development.