The Nineteenth Byte

« first day (5126 days earlier) ← previous day next day → last day (12 days later) »

rydwolf

00:05

2 hours later…

l4m2

01:54

CMC Output 1 for 1/G(64) possibility and 0 otherwise. Should usually halt in reasonable time.

5 hours later…

New Posts

07:22

Q: Max Island Area

You are given a two-dimensional array consisting of 0s and 1s. The number of rows and columns can vary. For example, 10 by 20. 0 represents water. 1 represents land. Your task is to find the area of the largest island. Example: [[0,0,0,0],[0,1,1,0],[0,1,1,0],[0,0,0,0]] Constraints and additional...

3 hours later…

Themoonisacheese

10:09

@rydwolf 👀

2 hours later…

mousetail

12:38

I've been tasked with a feature that allows converting output from chatGPT into structured multiple choice questions. The prompt has some variant of "don't use emojis as bullet points" in some form written 3 times but still it keeps using emojis as bullet points. It would be fine if it always did it but the inconsistency makes it very hard to parse

Themoonisacheese

@mousetail have you tried threatening it?

i find "if you do x, a puppy will be killed. doing x multiple times will result in multiple puppy deaths" to be particularly effective

alternatively: replace all characters in the emoji codepages with -

you could also try to ask it for a json representation of the list, it's much less likely to use emojis if it's writing code

mousetail

It's indicating the correct answer with emoji choices for bullets instead of writtiing "(c is the correct answer)" at the end like it's instructed

Themoonisacheese

then the json thing is probably a better idea

basically no examples in the training set will have labelled the correct answer in the answer when doinng this as json, as opposed to natural text

mousetail

The issue is a non-technical person will read the raw output and select which of the questions they want to include, perhaps after making some modifications.

Themoonisacheese

do you not get to parse the input at all before this?

if you do: you already know what to do
if you don't: surely that person can also be tasked with removing the emojis in some way?

mousetail

12:46

The basic idea is they will just ask the question inside the chatGPT UI directly and then copy paste the bits they like into my tool. I can edit the prompt template

They can remove the emojis but at that point they might as well copy paste the individual answers thus not actually saving them time

Themoonisacheese

bruh moment

mousetail

This project is basically intended to keep by busy for a week or two until my next actually important project starts, it has 0 priority but I do want to do a good job if I can

Themoonisacheese

what if they choose answers from the output then ask chatgpt to convert those into json, with it having been instructed to follow a specific spec for tha tjson?

lyxal

tjson makes me think json with static typing

which is either a really cool thing or a really horrible idea

Themoonisacheese

@lyxal tjson.org

lyxal

12:48

Themoonisacheese

good job guys :D

lyxal

smh 2000 notifications from my bank

mousetail

Identity stolen?

Themoonisacheese

he tried loading a site with a bad ssl cert, happens

mousetail

For some reason I thought that was meant to be to the tune of a song

Themoonisacheese

12:57

no i get it

it wasn't but i get it

lyxal

this is worse than that time jason typed my credit card details into ebay

and bought 5 iphones with tiktok preinstalled

Themoonisacheese

singapore, i'm senatorean

3 hours later…

Simd

16:13

I want to ask a fastest-code question about computing the convolution of arrays of integer valued long doubles. It's not at all clear how to do it!

mousetail

Why the "!"

Simd

I feel it ought to be obvious. I mean for float64s the answer is use the FFT if the arrays are at all long

but for long doubles, which isn't that many more bits, I have so far failed to find any method better than naive schoolbook multiplication

Mukundan314

does long double have intrinsics in avx2?

Simd

@Mukundan314 I think not but I am no expert

I mean 80 bit long doubles, in case that isn't clear

I did wonder if a chinese remainder theorem idea might work

Mukundan314

@Simd how large are the integers and how accurate does the result need to be?

Simd

16:37

@Mukundan314 they are up to 1e500 and the answer for each number in the output should be correct up to 1%

I think my CRT idea is doomed

2 hours later…

New Posts

18:28

Q: Does this standard loophole contain a loophole?

This standard loophole, disallows "Using the program name to store data without counting those bytes". However there is something curious about the "without counting those bytes" portion. Quoting from the answer: When using the filename to store data, add them as extra bytes to your solution sim...

discussion

5 hours later…

lyxal

23:52

@NewPosts two ideas: 1) require solutions to be executable with any file name (a more sensible solution, but would require a wrapper language for pxem)

2) (this is a hot take, but a serious one) requiring an explicit filename makes that answer its own language

Unrelated String

I like the idea of requiring pxem to be wrapped

explicit filename = own language is absolutely sensible though

ofc

lyxal

both ideas would effectively require a pxem wrapper

Unrelated String

That would also still require pxem to be wrapped to not be metagolfscript :P

exactly

lyxal

@UnrelatedString that's why it's a hot take :p

it would potentially lead to absurd amounts of chicanery, but I imagine downvotes would handle that :p

Unrelated String

It would just be how we handle flags but requiring more effort to abuse in exactly the same ways, which we already forbid lmao

you CAN flood your C solution with -Ds. Will you? fuck no

or ig maybe we don't strictly forbid that but again, downvotes

lyxal

23:56

It's not exactly MGS though

and the hot take comes from the fact that you're submitting a solution as a language

Unrelated String

submitting a solution as a language. is the whole point of MGS

lyxal

I mean, it's MGS in the sense "it's whatever I want the language to do", but not MGS in the sense "there's still a logical structure to the language name, and it's always going to be MGS"

Unrelated String

Also we can grandfather existing Pxem solutions in or even let them use ad hoc scoring a la Minecraft command blocks

lyxal

@UnrelatedString my point being that this idea will effectively make all pxem submissions MGS without a wrapper

so explicit filename = language would effectively ban unwrapped pxem

Hence why it's MGS but kind of not MGS in a very loose sense of not MGS

« first day (5126 days earlier) ← previous day next day → last day (12 days later) »

all rooms