AI Star Charts? Make AI do some work..?

[gritty rock]

@umbral summit

¼#'r Successfully Uploaded Some Code

I asked AI to do the following:

Look up the constellation map and get all the relative positions of the most popular/prominent stars.
Write a javascript function that puts spans with class="star" set to position: absolute with top/left relative offsets so the star classes are accurately arranged to be in the correct positions to fit the current viewport size.
For each span set the ID value to match the actual star name and add the constellation names as classes to the matching star spans.

AI gave the the attached output which almost seemed handy until I viewed it.
What syntax of prompting would get the AI to generate relative positions (% vs. px)?
I suppose any "good" result would need star size/intensity to look normal? 🤔

Uploaded these files to a Gist

starchart.html

View The Gist

For safety reasons we upload approved file extension files to a Gist.

[torpid umbra]

Why does the "AI" need to generate perfect code / why are you asking for prompting advice instead of just changing 2 lines in the code?

💀

Think you might want to use a canvas here instead of just positioning the elements btw

[umbral summit]

Why always fix the output from AI vs. learn to get AI to do the work properly?

[torpid umbra]

May I ask if you have done any ML research? 🤨

[umbral summit]

I might be owning too much of the problem, but that's allowable.

I'm only human. LOL

[idle ingot]

Why do people think LLMs can replace a lack of skill

[torpid umbra]

What's your definition of relative positions here? Relative to what?

[umbral summit]

You're spelling 'grunt work' oddly.. Skill would be wording this right so I'm not doing the leg work.

How far in % from the top left corner.

[torpid umbra]

Unpopular opinion: The grunt work is mostly stuff you can just grind out with fast typing

[stable pike]

seems like this would be solved with a short awk script

[torpid umbra]

vim macro moment

[torpid umbra]

That's what it is doing though? It's dividing the star's coordinates by the viewport height and width (probably not actually what you want, the coordinates don't really make sense)

(What would make sense is having coordinates and a camera position and then using linear algebra to figure out what you can see and where the stars should be placed)

[umbral summit]

Correct, if I presume the viewport size that the AI aimed for is 4k I should be able to convert the px values to percentages.
Though in reality I'd rather word my query so that the AI doesn't poop the bed, if that's even an option.

[torpid umbra]

Small hint: That LLM doesn't actually think, it tries to predict something that looks like a reasonable response to your prompt

Who came up with the coordinates? To me they look like some old 4:3 screen instead of a 4k resolution lol

X values are all below 1300 pixels, y values are all below like 900 or something

[umbral summit]

I'd make 4 AIs that work as a team.
1 to try and find the answer.
3 to judge the output and send it back if it's not close enough.
The 2 judges with the same decision are right.

[torpid umbra]

That has been a relatively common model architecture for a couple of years at this point, right?

[umbral summit]

Not sure.. Just a funny way to poop stack IMO.

[torpid umbra]

Or at least the router + expert section thing seemed to be common in larger models

[idle ingot]

Sounds like a GAN lol

[torpid umbra]

See that's the difference between me and wang

He actually knows the theory and stuff and I just throw random shit at the GPU cluster until it works