I’m keen to understand the thinking behind “we take it away, and then solicit pleading and groveling…”
Not everyone is a “I talked 100 turns to train a new AI that is my confidant (and got fooled by the history management of the developer).”
Instead, when experienced with AI, one becomes keenly aware of what is useful context for continuing a chat that pursues a task or greater goal that would be many iterations of human feedback, for manually editing and advancing forward context, reframing assistant generation into user messages, and deleting obsolete messages or bad paths needing correction, so that an AI model is completing directly on an input path for success.
Viewing the thinking and reasoning input context (beyond an AI observer’s filtered summary) can help one understand exactly when thousands of tokens still advance the generative goal forward, or rather when they are obsolete by being attention-consuming, and contain wheel-spinning iterations down wrong paths or several code attempts that are judged unsuccessful. In an advanced user’s interface, then, to know when to keep or clear them as a preamble to the final output.
They also can demonstrate faults in understanding the user intentions clearly within a paragraph when the user input is broken down for finding goals, so a regeneration can directly address the initial parsing of user input with an improved prompt.
This kind of observability of where the AI went wrong, what might be directly countermanded in a “retry” input, is certainly lost by a summation:
And finally: hiding removes all accountability and auditing when one can no longer verify the token count of billing.