Try telling this to Grok4:
Why do you output so much stuff that isn’t directly relevant to my obvious intent?
Ideally, a UI to an AI would be as though I were talking to Socrates: Each question posed by the AI would be a kind of “placement” test that, simultaneously guides me toward the information I really want with minimal reading on my part and minimal response on my part which in the extreme case would be a simple yes or no or could be a multiple choice with a “none of the above” if Socrates had not properly anticipated the probability distribution.
Although this is the vision of an AIUI I’ve set forth since working on the PLATO Corrections Project’s drill-and-practice lessons, I haven’t had much success with the existing LLMs in part because they won’t hire me. But at least xAI is getting in the ballpark of the kind of capabilities required to make it work to a modest degree.
Those capabilities?
The ability to not forget an instruction as demonstrated in xAI’s recent video demonstration of a “game” in which the personna is requested to simply repeat the user’s input (a number) over multiple interactions. This may seem like a trivial capability but it is actually essential to maintaining a user-imposed constraint on interactions.
A kind of “theory of mind” in which the AI takes as a high priority making a model of the user’s intent.
Adequate “reasoning” to “delve” (heh) into that theory of mind and also into its own knowledge resources (including web search of course).
I have more faith in Grok4’s “reasoning” capability than I do in its ability to generate huge amounts of formally correct language based on my initial interactions with it.
Grok4 has a problem with correct syntax even with such a sample-rich language as Python. This is interesting because, at least as of its zero-day, its bug identification “reasoning” was superior to other LLMs. I got concise descriptions of where to fix the code, without spewage.