AI as a mental crutch

(Feature image created with DALL-E, providing feedback on my image proompting skills)

Every couple of years I find myself building a new Linux virtual machine baseline for some project. Even though I’ve documented the process thoroughly the last few times there is always some quirk that has me starting mostly from scratch each time. This time I started off with setting the home page in Firefox to perplexity.ai and using it to find all of those mundane commands I forget three weeks after going back to Windows for my day to day work.

This time I hit a snag pretty early in that I was getting an error that made no sense to me (the specifics of which aren’t relevant to this train of thought). Perplexed, I asked Perplexity “wtf?” in my best prompt formatting (which, admittedly, is a WIP) and it gave me a few things to try. Some (not all) of them made perfect sense and I gave them a try. They failed.

I compared everything I was looking at against a similar appliance and didn’t see any obvious differences. I tried variations of prompts with Perplexity to get a more directly relevant response, which either resulted in what had already been suggested or even less relevant responses (I did mention my prompting skills need, work, right?).

I then tried ChatGPT, which gave me the same answers that differed only in their verbosity and longer pauses between response blocks.

Finally, I ran the same search I started with in Google, which returned the usual multiple links from our old friend Stack Overflow. I did like I did before using GPT backed by LLMs and narrowed the time frame down to the last year to eliminate answers 10 years out of date (and sometimes links to my own past explanations that are equally out of date) and found a summary that looked closer to my actual problem than the bulk of the answers (which were clearly the source of the responses from both GPT sources I had tried earlier).

And there was my answer. Not just to this one problem, but to the kind of sloppy approach I had fallen into using AI. The thread started with an exact description of the same problem, with lots of the same answers that had been of no help. And then the original poster replied to his own thread with the solution (a habit of frequent Stack Overflow contributors I have always admired and sometimes remember to emulate), along with how he wound up in the situation. Again, the specific error isn’t relevant to this tale, but the source is using the the first search result that seems to answer the question rather than reading it all the way through and seeing the subtle difference between what was needed and what was provided.

No AI response will tell you about the screw ups that caused the problem (are they embarrassed for their human creators or just don’t think it’s relevant?) and the path to realizing the mistake and then recovering (and learning). But real people will and that is how we learn from each other.

So having copilot proof your work is great and using promoting to get a start on something you’re stuck on is a great productivity boost. But relying solely on the technology to do all the work is how we wind up forgetting how to think and learn and build better technology to give us time to think and learn. In short, don’t trade the mental crutch for a creative wheelchair.

Facebooktwitterredditlinkedinmail
© Scott S. Nelson

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.