Explaining The Doom and Gloom

First, I want to explain why many people seem overly worried about people who are using LLMs regularly. Look up the term “cognitive surrender”, a LOT of folks who have been in tech for a while know at least one person who has fallen into this trap, where they stop thinking for themselves, and just have the LLM feed them everything. I’m not even talking about idiots off the street, I’m talking about folks who have been programming for 10+ years, basically forgetting how to organize common programming patterns without AI holding their hands through it, even though only 3 years ago, that same person was explaining it to new hires. (( Okay, maybe I AM talking about someone I know, but I know many of us have known someone to fall into this )).

We Don’t Memorize Phone Numbers Anymore

There’s a literal TED talk from a long time ago by Amber Case, called We Are All Cyborgs Now. This talk explains this phenomenon where, globally, individuals have come to trust in the technology we carry around with us enough that we don’t feel the need to remember phone numbers. It talks about a lot of other stuff, and I highly recommend watching it.

For context, understand that thinking and remembering things is hard work. There’s a reason why well-off countries force children to spend 12 years learning how to learn and study.

Okay But, I’m Different

No, you are not different, but you can also force yourself to not fall into the cognitive trap. Trust is the key that allows someone to fall into “cognitive surrender”. So here is where I’ll tell you to go ahead and use LLMs if you want to. Here’s the thing, LLMs can often explain things in ways that might be easy to understand. BUT you, as a user of LLMs, need to understand that the absolute best “self-reported accuracy score” from any LLM right now is only 91%. That is a number that pushes best case scenario, and often against tests that all the LLM systems specifically try to optimize for.

Here is my take on this. A self-reported 91% means that in practice, MORE THAN 10% of the time, what the LLM feeds you is wrong in some way.

Keep this firmly in mind whenever asking an LLM a question.

Extra Caution While Asking Questions

So, when it comes to trying to learn and understand some new subject matter, believing something subtly wrong in your head, too early, could stick with you way longer than you would ever expect, skewing the way you read everything else. Take the extra time to fact check EVERY LLM ANSWER. Literally, search the points of the answer, and see if you can find some human written example that shows it to be true.

I honestly think there are cases where a wrong LLM answer, can still lead to a better outcome. Here’s an example: Say I’m looking to understand how to replace a cracked drain under my sink. It cracked, so I don’t know the orientation it was originally in, so I ask an LLM. That LLM tells me about an “S-Trap”, and how to set that up. So, I have the keyword, “S-Trap”. Here, through an actual web-search, I learn that “S-Traps” are actually illegal in most places (excepting literal toilets). But, that article links to another article about why it is illegal, and that article shows me a “P-Trap”, which is actually what I need.

I may have never known how to find the term “P-Trap” without being told about the wrong thing. At least the LLM knew to feed me a domain specific keyword that I may not have otherwise known.

LLM or No

If having to do a search to verify every single LLM answer sounds like more of a pain than just finding and reading a book that covers the subject, then you can start to understand why so many people try to warn others away from LLM use.

If it still sounds like using an LLM would still be faster for what you want to do, then that’s your answer.

I mean, I’d still judge you for using a hype technology that really is being marketed way before it’s actually ready, but I honestly don’t have any tracking on this blog, so don’t worry about what I’d think. Go help burn down the world, I literally can’t stop you.