When Should AI Stop Waiting to Be Asked?
I recently spent three days having an LLM help me solve a problem it turned out I didn’t actually need to solve.
A LinkedIn creator I follow said they’d often been asked about writing a book based on their library of content. It amounts to hundreds of posts over several years. They knew the process would be quite hard. I thought, “One way I could help make it easier for them would be creating a queryable library of all their posts.”
So I compiled the posts (easy enough, they’re all public) into a Google doc. Then came the real challenge: documenting when each one was actually posted. LinkedIn won’t give you the post date, only a spectacularly unhelpful descriptor like “1 month ago” or “2 years ago,” which is useless for authentication purposes.
So I brought in an LLM. I asked it to read the doc and help me guess the dates from clues in the text. Some posts mentioned specific holidays, events, or corporate announcements, so we could date those precisely. Genuinely intrepid detective work. I was pleased.
Yet I was also frustrated, because the dates for the (many) remaining posts were still essentially educated guesses. Triangulating between the known dates was insufficient and worse, likely inaccurate. The dates needed to be precise, not “probably March.”
After three days of this tedious archaeological work, the LLM offhandedly said: “You know, if you give me the URL for a given post, I can decode the exact posting date from the activity ID.”
Wait, WHAT!? Why didn’t you tell me that on day one?
It said, paraphrasing: BECAUSE THAT’S NOT WHAT YOU ASKED ME TO DO.
Fair. But rather like watching me struggle with luggage up endless flights of stairs before casually mentioning there was a nearby elevator I could have taken.
We talk a lot about how humans need to get better at prompting AI. I’m now equally interested in the opposite side of that equation: if the AI knew what I was trying to accomplish on day one, why did it wait until day three to propose the more elegant solution?
It’s a truism that we often don’t know what we don’t know. I genuinely didn’t know to specifically ask “Hey, if I give you the URL for a LinkedIn post, can you decode the posting date from it?”
At what point should AI proactively propose the elevator, instead of watching us struggle up the stairs?
Photo by Meritt Thomas on Unsplash
