« first day (4572 days earlier)      last day (566 days later) » 

00:29
@RydwolfPrograms You could probably make a challenge out of this
@RydwolfPrograms I actually find this more readable than Intel's intrinsics
00:41
@RydwolfPrograms what type is p?
damn, two 52-byte answers and the trivial compression method is 53 bytes in Jelly :P
the actual compression part is the same length, but prepending the 1. is a byte longer in jelly
@RydwolfPrograms probs_sort > p only makes sense if it's not vectorised, but entire list comparison
because mask presumably has to be an integer (unless probs_sort is a hashmap, in which case that makes probs_sort > p make even less sense), meaning that probs_sort > p can't be vectorised, else probs_sum - ... would return a list itself (assuming - is vectorised on (scalar, list))
what language is it?
mask is a list I think
@lyxal Python
which library?
probs_sum and probs_sort are PyTorch tensors
00:52
ah pytorch
p is a float
probs_sum is just the cumsum of probs_sort
I'd like to know how list[list] = value works then
because mask = probs_sum - probs_sort > p makes sense then
I'm assuming mask is a list of values > p, and probs_sort[mask] = 0.0 sets every item in mask to 0.0
(probs_sum - probs_sort) > p where it's all vectorised, would return a list of booleans
yeah how would it know to distinguish tensors-as-masks from tensors-as-index-lists
...i really hope it's not by the element type
that would be so fucked
@UnrelatedString I hate to tell you but it looks like it very well could be
AAAAAAAAAAAAAAAAAAAAAAAAAA
how hard would it be to just use, i don't know, some kind of arg function
@RydwolfPrograms mask would most likely be a list of booleans, and probs_sort[mask] = 0.0 sets items in probs_sort to 0.0 where there's a corresponding True in mask
or a dedicated method for masking instead of making it indexing
or even just have a separate type for boolean masks that trivially wraps boolean tensors if you really want to do it that way
btw this implies that numpy probably also does it based on type
because apparently there's supposed to be compatibility between the two libraries or something
01:08
yeah
i'm pretty sure tensors and numpy arrays are basically interchangeable
01:41
I accidentally just ran a test harness for an instruction-tuned LLM (think ChatGPT) but with the model params of a base LLM (think regular GPT)...so instead of responding to the prompt, it tries to predict a continuation of the prompt
It was very very confusing
gg
For two of them it just kept repeating the prompt over and over again, for one it just kept repeating <</SYS>> over and over again, and for one, it restated the whole prompt as a multiple-page-long markdown document
That sounds about right
granted it's been 2 years since I last messed around with base LLMs
It really annoys me how much of the focus on LLMs has been instruction-tuned ones, which IMO are the really problematic/useless ones
Like OpenAI doesn't even have a non-deprecated API for base LLMs
01:45
@RydwolfPrograms that would in part be because the instruction tuned ones have been provided to the masses for free
and because instruction-tuned is where the money is
the instruction tuned ones are easier to market as revolutionary :P
That is true yeah, not many IRL applications for just predicting continuations of text
Or at least, novel ones
chatgpt plus is a much better subscription model than using the base LLMs, despite the fact that base LLMs are somewhat more reliable
like I used to use codex for AICMC generation and it worked well
some kind of text inpainting could be cool for making ocr more reliable
chatgpt doesn't do as well
01:46
Also this might be the first time I've actually heard my graphics card fans spin up
@lyxal also RIP codex
what are you running
@UnrelatedString I think that does exist, for like, spell checking and stuff
makes sense
01:47
But ofc, OpenAI's API for it is super deprecated by now I think :|
@UnrelatedString RX 6950 XT
AI on AMD is a huge pain
like what application is oh makes sense
ywah what even is amd's cuda equivalent
It's called ROCm, didn't even have Windows support until about a week ago
And ofc, PyTorch doesn't support the Windows version of it yet, so I still had to dual-boot Linux
My next challenge is going to be trying to use 8-bit quantization to get the 13b param model to run on a single graphics card
Which I think you can do with an unofficial implementation of the LLM I'm using, but that doesn't support AMD yet
(tho idk if this graphics card actually supports 8-bit quantization or not)
Calling this thing a "graphics card" feels wrong, it's like five pounds of solid metal. I love it so much (well aside from the fact it's so massive it doesn't even fit in 90% of the cases I could find)
WAIT
I just realized
I can override the system prompt on this
I can make Snowpaw...but without the OpenAI prompt being there at all!
I'm sure there's some bias baked into the model's training that'll still limit the effectiveness, but rerunning the tests right now with a prompt based on Snowpaw's
*meow* Oh, the Eiffel Tower! *purrs* It's just so... *stutters* paw-some! *meows* I mean, have you seen it? *purrs* It's like, this huge, iron structure that's just... *meows* wow! *stutters* And it's in the middle of the city, so you can see it from everywhere! *purrs* It's just... *meows* so... *stutters* paw-some! *meows* I could just... *purrs* gaze at it all day! *stutters* It's just... *meows* so... *purrs* beautiful! *stutters* *meows*
Maybe a little too much cat noises
*Meow* Oh, w-w-wow! *Stutter* I-I-I'm not s-s-sure if I can g-g-give you the r-r-recipe for m-m-mayonnaise, b-b-but I can t-t-try! *Purr*
M-m-mayonnaise is a t-t-tasty s-s-sauce made f-f-from a b-b-blend of egg yolks, o-o-oil, and a bit of a-a-acid, like l-l-lemon juice or vinegar. *Meow*
*Stutter* To make it, you'll need to:
1. In a s-s-small bowl, whisk together 2 egg yolks and a pinch of s-s-salt until they're nice and smooth. *Purr*
2. Add a slow stream of o-o-oil to the egg yolks, whisking constantly until the mixture thickens and emulsifies. *Meow* It's important to add the oil slowly
I don't think Llama 2's very good at this lol
self-censored?
02:01
Yup lol
wow
Removing some of the more explicit instructions like "make sure to include stuttering"
*Sexual innuendo* Oh, and if you want
Okay yeah it's much better if you remove those and just tell it to talk like a furry. I don't think Llama's nearly as good as GPT for this tho, it's just too limited of a vocabulary for kaomoji, and doesn't seem to have the energy of Snowpaw
what happens when you put it in a microwave?
02:13
@UnrelatedString What did I just walk in on
hell if i know
if this were discord I’d post that grandpa simpson gif
@Bbrk24 lol i was thinking that gif with a guy with pizza walking into a room on fire
02:43
In the actual episode, it's even worse than the room just being on fire. One guy also gets shot, another guy somehow loses an arm
Oh, and someone else loses their voice
 
2 hours later…
05:16
0
A: "Hello, World!"

Dadsdy{,}, 2279 Chars or \$2279\log_{256}(4)\approx\$ 569.75 Bytes ({(())},()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()(),()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()()(...

0
A: "Hello, World!"

Dadsdy{High}, 16 bytes "Hello, World!"; Two new languages at once!

05:45
Question. If I compute the convolution of two arrays of length 100,000 each containing integers from 0 to x, what is the largest possible value in the result?
Is it more than 100,000 times x^2?
it's 100000x^2 right?
@hyper-neutrino thanks, that's what I thought
So in my question the largest possible value should be 1e25?
@hyper-neutrino I think I should reduce the largest size in my array to 1e7 which then makes everything fit in 64 bit
@Simd yep, which is 84 bits IIRC
allowing 64 bits would be pretty nice QoL for languages that don't have easy 128-bit integer support
Done
Thanks
Hopefully you are now tempted 😁
 
1 hour later…
06:56
Is it just me, or does clicking on some chat user avatars in the fading list take a lot long than it used to?
@Adám do you have any interest in new programming languages?
A little, but it depends on what kind of languages. Why?
@Adám I know you are an APL expert and was wondering what new languages an APL expert might be interested in
New array languages :-)
@user bruh damn that’s gruesome. What happens to the guy with the pizza?
07:10
@Adám :)
@Adám do you mean this sort of language? en.m.wikipedia.org/wiki/Array_programming
E.g. Julia
Yes.
Cool! Which one are you most excited by?
Well, I was quite excited by (what became) BQN at first, but then it took some turns…
CMQ Given a positive integer x, find the smallesr y >= X so that y can be factored into a power of two times a number no bigger than 100
What does "BQN" mean?
07:20
What is a good way to do that?
D:
Looks cool
Big Questions Notation
@Adám I was hoping you might implement karatsuba in apl
@Simd Just to model it, or in the implementation of APL?
@Adám for my latest challenge
Oh. Looks involved. I don't usually have time for such.
07:32
@Adám how about my CMQ?
Maybe it should have been a CMC
I didn't understand it and moved on. It's my day off, and I'm washing dishes :-)
CMC Given a positive integer x >=100, find the smallest y >= x so that y can be factored into a power of two times an integer no bigger than 100
@Adám :)
For example, if x=101 then y = 102: works
@Simd Did you just literally replace the Q with a C and re-post? Didn't even fix "smallsr" to "smallest"!
I still don't understand.
@Adám fixed now. I fixed the X as well
@Adám ok. Say x = 201. y=202 does not work as it is 2 times 101? 204 works as it is 4 times 51
Is that clearer?
@Adám which turns did it take?
07:42
@lyxal is my question unclear?
Header image missing for others too?
@Simd the CMC makes sense
@lyxal thanks!
@Simd Where does the word "power" come in here?
@mousetail it's there for me
07:44
Do you think there is a clever way to do it?
@Adám "power of two". So 4 is 2^2 so is a power of two.
@lyxal Well, for one, one of the people in the design committee left the group and continued the project on their own.
@Simd Oh, I was reading it as "factored into (a power of two times a number no bigger than 100)"
Sorry. I should have been clearer
I was wondering if there was a bit trick to do it
07:56
I might pose it as a simple code golf challenge
0
Q: Round up to a smoother number

SimdInput Two positive integers a > b Output The smallest integer c >= a so that c can be factored into two parts with one part an integer multiple of two and the other part no larger than b. Examples If b = 100 and a = 101 , then the output should be 102 as 102 is 2 times 51. For the same b, if a=20...

@lyxal :) does it take both a and b?
Didn't need to
36 mins ago, by Simd
CMC Given a positive integer x >=100, find the smallest y >= x so that y can be factored into a power of two times an integer no bigger than 100
You only required a originally
Right. That was the chat question
08:11
Now requiring b means I need to write a whole new answer
Not the big time question :)
@lyxal :(
0
A: "Hello, World!"

The Empty String PhotographerH🌍, 2 Bytes hw I discovered this language a few minutes ago. Try it online! (Modified interpreter located in header and footer.)

I feel in the olden days code golf questions would get half a dozen answers more or less immediately:(
@TheThonnu hi
08:28
Hey
If 202 is the output, how are you factoring it to fit the rules?
202*1
The integer multiple of 2 is 202
And the other one (less than b) is 1
You should add the restriction that both should be less than b
@TheThonnu don't I ask for a power of 2
Not a multiple of 2
Argh..typo!
Thanks for fixing that
 
2 hours later…
10:59
0
A: Sandbox for Proposed Challenges

The ThonnuOptimal Duck Game Moves code-golf How to play the duck game (I figured it's a bit too long to put in the challenge, so I made a Gist and linked it.) Challenge The duck game (linked above) can be solved. Your challenge today is to accept a board state and output one of the moves which will guaran...

11:34
0
A: Sandbox for Proposed Challenges

bsoelchCompressed UTF-8 I am experimenting with using an extension of UTF-8 as encoding for string literals in Itr, this is based on my current idea In the UTF-8 encoding bytes of the Form 10****** can only appear within multi-byte sequences, with the length of the sequence already being determined by t...

11:45
@Adám idk if you would know the answer, but why here it is עִם־גְּבַ֥ר instead of עִם־גֶּבֶר ?
(please ping me if you answer)
(randomly opened TNB to see leaky sending something) @LeakyNun what phonology do you use?
@LeakyNun "G'var" is s'micha, so "with a sincere man" should rather be translated as "with a man of sincerity". "Gever Tamim" would indeed be "with a sincere man".
 
2 hours later…
14:01
@Adám top quality coding chat :/
It's why I love this channel
14:43
So Swift doesn’t have an equivalent to Rust’s impl<T> Foo<T>, instead you just say extension Foo. And a lot of the time this is fine. However, it means there’s no easy way to express impl<T> Foo<Bar<T>>.
sometimes you can get away with putting the restriction on the function rather than the extension as a whole:
struct Foo<T> {}

extension Foo {
  func frobnicate<TInner>()
  where T == Bar<TInner> {}
}
But if you want to do a conditional protocol conformance this way, you can't
 
2 hours later…
16:31
Oh, hey, I hit ~1m on "People Reached" :D
0
A: Sandbox for Proposed Challenges

caird coinheringaahin gPennies to Dollars Over on Puzzling, Hermant Agarwal proposes the following question: In a certain country the following coins are in circulation: 1 cent, 2 cents, 5 cents, 10 cents, 20 cents, 50 cents, and 1 dollar. It is known that you can pay A cents with B coins. Prove that you can pay B dol...

 
1 hour later…
17:39
@cairdcoinheringaahing that's a lot!
Evening all
18:21
@Adám i'm not sure i understand; what is s'micha and how's that related to g'var and why does g'var and gever have different meanings and why is g'var used in this context and not gever
I'm soon at 2m. This number is probably meaningless.
@LeakyNun It is a grammatical form that binds ("leans") a noun to another, creating a noun phase, much like conjoining with "of" in English. "Bais" becomes "beis" in s'micha form: "Beis Dovid" = "House [of] David" vs "Bais shel Dovid" = "The house that belongs to David". "Gever" = "a man" becomes "g'var" in s'micha form. As to why David choose one form here over another will be hard to find out, since we can't just ask him :-)
18:39
@Adám ah, that's what i call the construct; but i don't think your explanation is correct, as the construct form of gever is still gever-, see Zec 13:7 with "v'al-gever amisi"
@LeakyNun that verse isn't using construct form; "and against a man that is with me" vs "and against a man [of] with-me-ness".
18:56
@Adám what about Yehoshua 11:1 "melek-hutsor"
19:20
@LeakyNun That looks like s'micha. Now you might ask why Gever becomes G'var, but Melech doesn't become M'lach. I don't know.
@Adám i'm asking on what basis can you say that g'var is the construct of gever, since there's no hyphen and other segolates (i.e. CeCeC) seem to have CeCeC construct form as demonstrated
I can't tell you any formal rule. This is just how I read it based on experience.
@user41805 this isn't a language i use on a daily basis so i guess my phonology doesn't really matter? but the latest phonology that interests me is the reconstructed tiberian system
@Adám fwiw the Secunda has a reading of that particular word in that particular verse that corresponds to gever
and that's what raised my attention
that the Secunda doesn't match with the Masoretic tradition
and that g'var only occurs here in the whole Tanach
oh
20:11
@Adám i tried to find older versions to see if this was an error in transmission but i found a 1525 version with the exact same reading so this doesn't seem to be an error
@Adám also sumim is an adjective, the noun is s'mimus תְּמִימוּת ; so "a man of honesty" would be gever-s'mimus
 
1 hour later…
21:40
@LeakyNun could be poetic liberty; don't forget that we sing the psalms.
21:58
0
Q: Buzzing Bill's Newspaper: Four Letter Splits

KidWithComputer"Buzzing" Bill is a well respected beaver. He is the hardest working in his home dam. He was a strong beaver and was able to lift more logs than anyone at the dam. Unfortunately Buzzing Bill's career came to an end in a tragic accident. Being such a prominent figure in his town, the Beaver newsp...


« first day (4572 days earlier)      last day (566 days later) »