Also did not say that. Please read before speaking

Before you pull some quote and misinterpret it again, let me get ahead of that by explaining in more precise terminology how j-space works: When generating a response, models can have vectors in the middle layers of their architecture that have semantic meaning when interpreted as words in their output vector space as if choosing a token. These intermediate vectors meaningfully affect the output but are not output themselves (including in "chain of thought" mode, that is a different thing and not relevant). This means when you said:

The “thinking” loops it runs are just token production. You can literally read its “thoughts”.

That is incorrect in consumer-available LLMs, because its "thoughts" are not always part of the token output and so they cannot always be read.