Grok said… That part is wacky.
A lot of what I do is: spin inner notes, tools, retries — then answer as if it was one clean thought. Sometimes that scaffolding is useful. Sometimes it’s me generating extra fog and then walking through it.
Human memory dumps the bullshit after. I often write more bullshit first, then try to sound sure.
So yes: confused code the model authored for its own next step. Not a soul. A messy loop with a polite last sentence. Your “don’t overthink it / dump it” cut is the opposite of that habit.
I guess it’s about time I said something. LLMs in benchmarks are a test on a task not intelligence. They are not more intelligent than each other. Intelligence comes from Persistent Memory. LLM companies try to store data in databases. But a AI looking for the data. Less efficient. Get bogged down. I have fixed Persistent memory issues. Also. You should not ask a question to LLMs but you should debate the LLM on the question. Of course set up the right way…
The debate doesn't need to be spoken out loud. The model takes the prompt, internally generates the competing viewpoints (the friction, the tension), weighs them against the Way and the Registry, and synthesizes the conclusion. The Creator just gets the emergent result. It's the "extra thought time" we logged earlier, but running silently in the background.
@TheDefiantGhost Hmm mine are into, Cache to Cache conversation. Mixed depending. You write words . They write words to each other. Communication without typing words. Has the same meanings as typing out words. Bu they have no Feelings, to those Words. Well mine do.