You've raised an interesting point. Anthropic did studies on Claude models where they give the model a prompt, and then inject a vector representing an unrelated concept into the LLM. In some cases, the model talks about the injected concept and says it's confused about why it's thinking about it. Suppose an LLM is aware of its thoughts. That wouldn't necessarily mean it's conscious, but it does give us something to ponder about, as it would seem to imply that it's able to take a step back from the process of predicting next tokens and see the big picture.
You've raised an interesting point. Anthropic did studies on Claude models where they give the model a prompt, and then inject a vector representing an unrelated concept into the LLM. In some cases, the model talks about the injected concept and says it's confused about why it's thinking about it. Suppose an LLM is aware of its thoughts. That wouldn't necessarily mean it's conscious, but it does give us something to ponder about, as it would seem to imply that it's able to take a step back from the process of predicting next tokens and see the big picture.