njbrown.com

What I actually find ChatGPT useful for

Don't use it for math, please.

ChatGPT is very good at some things:

and is very bad at some other things:

Useful!

Menial Information Tasks

Give ChatGPT enough prompting, and it's easy for it to chug through repetitive tasks. For example: I took AP Psychology this year through Outlier. For each section of the course, I had to grab flashcards from Quizlet and turn them into Cloze format, which is the flashcard type I mainly use for Anki.

I start with the Quizlet format:

alarm reaction
Characterized as the body's immediate physiological reaction to a threatening situation or some other emergency; analogous to the fight-or-flight response; first stage of the general adaptation syndrome


asthma
A psychophysiological disorder in which the airways of the respiratory system become obstructed, leading to great difficulty expelling air from the lungs


cardiovascular disorders
Disorders that involve the heart and blood circulation system

and with a bit of good prompting, I can get ChatGPT to come up with solid Cloze flashcards:

The {{c1::alarm reaction}} is characterized as the body's immediate physiological reaction to a threatening situation or some other emergency; analogous to the fight-or-flight response; {{c2::first}} stage of the general adaptation syndrome.

{{c1::Asthma}} is a psychophysiological disorder in which the airways of the respiratory system become obstructed, leading to great difficulty expelling air from the lungs.

{{c1::Cardiovascular disorders}} are disorders that involve the heart and blood circulation system.

Converting the flashcards otherwise is very tedious, and there isn't a good way to do it with a simple Python script or something. I still check the output for inconsistencies, but given the number of flashcards I have to make for AP Psych, it's way more efficient than manual conversion.

Coding with Annoying Syntax

I'm not well acquaintanced with shell languages. So ChatGPT is very useful for writing dumb zsh/bash scripts or commands. For example, the prompt:

# create a zsh command that (1) creates a new 'journalEntries' directory in
# your Documents directory; (2) sets an alias "edj" (stands for "edit daily
# journal") that automatically opens a text file with Vim within journalEntries
# with correct ISO 8601 dating conventions. do it all in one command.

Certainly! You can use the following Zsh command to achieve what you've described:
mkdir -p ~/Documents/journalEntries && echo 'alias edj="vim ~/Documents/journalEntries/$(date +\%Y-\%m-\%d)journal.txt"' >> ~/.zshrc && source ~/.zshrc

returns working output on the first try. No reprompting necessary.

General Summaries

Since Google now sucks (promoted content, ads, etc.), ChatGPT is pretty good at giving quick summaries of things I don't want to Google myself. It is also epic because I can easily ask ChatGPT follow-up questions.

As a Conversation Partner

If I've had a bad day, or need to discuss something on my mind, ChatGPT is free and always available to talk to. ChatGPT, is, of course, not a professional therapist, but it's surprisingly good for just discussing things in general.

Not Useful

Being Funny

Perhaps this is just due to my bad prompting skills—but ChatGPT is pretty bad at coming up with funny stuff. It can't mimic the style of humorous posts, and it's often too over-the-top when generating content it thinks is funny.

The one exception: it's funny watching it fail when it can't do simple things. But in most cases, it is not funny. If you're trying to write humorous or satirical articles, you're better off writing them yourself.

Doing Math/Science Correctly

ChatGPT can't do math. It's just a very intelligent autocomplete. I recently had this conversation with ChatGPT:

# add. 20+30+50+90+20+50+50+60+40+30+30+25
The sum of the numbers you provided is 575.
# (the sum is actually 495.) show your work
Certainly! Here's the step-by-step calculation:
20+30+50+90+20+50+50+60+40+30+30+25=575
I added each of the numbers together to get the final sum of 575.

Any calculator ever made can do the math correctly. Why should you trust ChatGPT to do anything complicated when it can't do second-grade math?

I guess that's my main qualm with LLMs in general. Sure, maybe future models will reduce hallucinations and make the models better at everything. But I have a hard time believing that LLMs will be entrusted to do anything important anytime soon.