#Does ChatGpt suck at doing your math?

1 messages · Page 1 of 1 (latest)

worldly dagger
#

GPT 4 has some nice image recognition, but when it comes to math, it tries to use python to better "read" the image, rather than actually checking the image. this may suck if you have things like complex symbols or integration, as GPTs image recognition is far better. a good example of this is trying to get gpt to do calculus with multiple problems. in order to try and help out people who end up in situations like mine, here is my general guide on getting gpt to do math.

Step 1: Your problems

Be sure that your problems are legible. this means that if you show a fifth grader whos third language is english your problems, they should be able to read it. make sure the problem can be understood, then make sure there is a clear distinction of the instructions for each problem.

TIPS FOR THIS STEP:

A pdf is usable, and in a lot of cases preferred. for this reason, i will continue the guide. if your questions are in jpgs or pngs, thats fine too, but try to stick to one file format. in the case you already have images, skip to the merge step.

STEP 2: TRANSLATE & MERGE!

now, you have gpt looking at your gorgeous pdf. you were sure to dot your is, and cross your ts, you even included a nice message for the three letter agent reading your post. if you try to have gpt solve all the problems right there and then, there is a lower (in my experience) chance it will be able to do so. my theory on why this is is that the pdf format, while nice for humans, is very cringe to machines, as anyone who tried converting a pdf to docx may testify to. in order to have it not be cringe then, we need to mush our carrots into baby food for the ai. here is the prompt i would give GPT if i had a pdf:

"please use python to modify this pdf into a series of images. after doing so, use python to merge the images into one long image. after merging the images, use python again to embed the image into our chat."

dont use the first seventeen words if you already have images.

part 1/?

#

TIPS FOR STEP 2
the prompt i gave was very brief. in my experience, chatgpt tends to be 50/50 with them. if it does not work, i reccomend judging how gpt got it wrong, and either talking to it like a tutor to a student (it is important to do so, otherwise it will not get it) to correct its mistake, or if its answer is unsalvagable, editing your inital prompt to be more verbose.

The breakdown of the prompt is as follows:

please use python to modify this pdf into a series of images.
notice how i told it to use python. while its probably able to do it on its own, its typically better to tell it specifically to use pythons abilities to convert pdfs to images, as chatgpt is chaotic. (not an insult chaotic, more double pendulum chaotic) telling it directly to use python to make the series of images will guarantee it will do exactly that.

after doing so, use python to merge the images into one long image.
again, notice how i told it to use python. gpt tends to want to spend all your credits, and a very convenient way to do that while impressing you is generating images! i can not count the amount of times i have shown it a math problem and instead of it showing the image, its some dall-e gobbledegook (sorry dalle devs). for this reason, telling it to use pythons abilities ensure a nice, clean continuation of the task.

after merging the images, use python again to embed the image into our chat.

In my experience, GPT has no idea how to embed an image. it sure tries to do it, and oh boy does it try! but it sucks. if it fails at doing it, it will offer you links to download the image. trust me, its way easier to copy paste the image in the long run. telling it to use python will make it run some quick code to print out the images, which surprise surpise, will be shown in the output.

Part 2/? (tips continue next msg)