The Nineteenth Byte

Weird Glyphs

00:23

@DLosc I think you can find the optimal solution for an arithmetic sequence of n prime numbers by looking at oeis.org/A005115.

DLosc

Cool, TIL

Neil

00:51

I think you can find the solution for a sequence by looking at oeis

hyperneutrino

07:08

@Adám hmm. that is deceptively simple lol

Themoonisacheese

08:16

→ 1 brittle linked list moved to The Nineteenth Bakery

Simd

10:13

hi all

Themoonisacheese

heyo

Simd

deepseek times out every time. Do you think the multi billion dollars western companies are killing it?

Themoonisacheese

probably no cores to run that many requests

since they can't buy the newest chips they're stuck with old gen and consumer GPUs

download it and run it locally? it will be better for your data too, the deepseek app is a mess of phoning home to china

Simd

To run it locally I would need to buy a million dollars worth of kit wouldn't I?

Themoonisacheese

well no that's the entire thing

Simd

10:17

the russians are really good at getting round all sanctions. I am surprised the chinese can't do that same

Themoonisacheese

you probably won't be able to run the largest models for a lack of VRAM but the 1.5B and the one just above that can run on 8Gb of VRAM

Simd

yes but that's sort of pointless isn't it? It's the full version that is amazing

qwen 2.5 coder is perfectly good if I don't need amazing

Themoonisacheese

the small version are actually pretty great as well

which is the reason sam & co are freaking out

Simd

do you find it better than qwen 2.5 coder?

Themoonisacheese

i find all AI assistants to be thoughroughly unhelpful

the thing about this is that it's being marginally less unhelpful for a fraction of the price, not that it's so great

Simd

10:19

@Themoonisacheese tell me more

I never pay a penny so price isn't a factor for me

but it seems to be unlimited in the number of queries and free

Themoonisacheese

@Simd yes but the money in question is spent on compute resources for running it, so cheaper = you can run it on less resources

Simd

understood

Themoonisacheese

including consumer hardware

Simd

trying... ollama pull deepseek-r1:14b

Themoonisacheese

@Simd any problem covered enough on the internet such that an AI would give you an accurate and useful answer is also covered enough on the internet such that you can find the same information at its source easily. you can also find harder info, in a way i find AI assistants aren't able to

thus: if my problem is easy enough for AI, i didn't need AI, and if it's too hard for AI i can probably still figure it out myself

Simd

10:24

@Themoonisacheese I mostly use it as a more targetting google search. It's pretty cool for that.

Themoonisacheese

@Simd most likely you don't have enough VRAM as it stands. give it like a week and someone will have resampled it down to a level you can run on a consumer GPU, much like they did with llama

Simd

I don't even have a gpu :)

cpu only here my friend

Themoonisacheese

oh lol you stand no chance

Simd

I do run some models ..It's fine if you are happy seeing the reply as the same speed I type

Themoonisacheese

@Simd you do you but cool != helpful != accurate, in the cases where it matters

lol

Simd

10:25

@Themoonisacheese can you give a specific example?

Themoonisacheese

sure

here, i asked chatGPT for a python program that spawns and talks to another program using stdin/stdout chatgpt.com/share/679a0261-dca8-8011-8099-53748b65ab0e

it's a simple problem, just one i hadn't encountered before

it's wrong about pretty much all of it, its code doesn't work and debugging its mess took me longer than it would have taken me to read the ypthon docs on the subject

lyxal

I also have an example proving cool doesn't imply helpful and accurate

lyxal.github.io/Joeasus

Themoonisacheese

and this is scraping the barrel of complexity. anything i do in my day job (which actually isn't that complicated either) it frequently would simply be just wrong

lyxal

It kind of works when you don't give it inputs but very quickly falls apart when you do :p

Simd

@Themoonisacheese bpa.st/S3SA is that any good?

Themoonisacheese

10:32

@Simd well that's the thing how would i know until i run it

Simd

I started to try to do some coding for pay and the python problems I have to solve are too weird for AI :)

Themoonisacheese

it looks good, but that's the entire problem

Simd

I have to use multiprocessing to run a method in a class on 8 different cores....sounds simple, right?

lyxal

@lyxal AI's lack of mathematics will always be amusing lol

Simd

problem a) multiprocessing can't cope with "bound instances" of a class. b) it can't cope with lambda functions c) it can't cope if any of the parameters you are calling the method with are not picklable

it's almost like multiprocessing is designed not to work in the real world

Themoonisacheese

10:37

almost as if python isn't a language that was designed for performance

and then they tacked on multiprocessing to try to make up the 70:1 disadvantage they have against compiled languages

mousetail 'he-him'

Multiprocessing is nothing but an ugly hack to get around the GIL

Simd

it can be much more than 70! codegolf.stackexchange.com/questions/26323/…

@mousetail'he-him' yes but.. there is multiprocess and I can't understand why multiprocessing doesn't at least do what multiprocess doesn

mousetail 'he-him'

It can't, there is no shared runtime

Simd

I mean the multiprocess 3rd party module

mousetail 'he-him'

It doesn't use pickle right?

Simd

10:41

it uses dill I think

mousetail 'he-him'

It might be able so serialize some things normal pickle can't while not being able to serialize some things that normal pickle can

Simd

that is possible

@Themoonisacheese I am now running deepseek-r1:32b locally on a cpu :)

Themoonisacheese

@Simd lol see you in an hour for the next token

mousetail 'he-him'

You should have an intergrated GPU aat least?

Themoonisacheese

like that's gonna help

mousetail 'he-him'

10:46

Faster than CPU at least

Themoonisacheese

[citation needed]

Simd

@Themoonisacheese I don't type that slowly! :)

Themoonisacheese

in keeping with my tradition of only researching products after i buy them, anyone have opinions about framework laptops?

Simd

@Themoonisacheese the 32b version is clearly much weaker than the full version in my preliminary experiements, sadly

Themoonisacheese

wrong reply i think

but alright

Simd

11:08

this is quite hard for AI. "In bash, how can I iterate over 1 + x for x in (0.1, 0.2, 0.3... up to 1) ?"

Themoonisacheese

it's quite hard for bash tbh

since it doesn't have float support

Simd

yes... but it is doable with seq and bc

for x in $(seq -s ' ' 0.1 0.1 1); do result=$(echo "$x + 1" | bc); echo "Result: $result"; done

if the AI you are using can't do that, throw it away!

that is '[space] '

att

@Themoonisacheese uhh thank you for being my beta user

i think i'd like one some day but i want to wait for it to mature some more first

Themoonisacheese

i'll admit it was a 1400 euro impulse purchase

i'll tell you how it goes if they would ever start shipping it to me

(i am aware it is chinese new year)

oh my god it finally happened

someone sent me a screenshot pasted in a word document

att

what

Themoonisacheese

11:17

it's an old IT tech legend

att

also happy chinese new year

4

Themoonisacheese

the user that is unaware of any file format that is not docx

Simd

@att and to you!

att

~~Back In My Day we only had doc files~~

it wasn't a screenshot of a doc at least? (:

Themoonisacheese

no, and jira beautifully rendered it in-browser :D

so it wasn't even that big of a problem but i've always heard of this happening but never seen it myself

it also used to be that if you had an image in your clipboard from snipping tool or win+shift+s, you couldn't paste it in a web form, and you couldn't paste it in explorer.exe to save it either, but nowadays jira allows you to straight up paste images

New Posts

12:04

0

Q: Is it France, Netherlands or Poland?

Below are simplified versions of the flags of France, Netherlands and Poland respectively, as 9x6 pixels: (Many other flags use this color scheme, but I think a smaller number of flags allows for more creative solutions.) Task Given a pair \$(x,y)\$ of coordinates within the flag matrix and a co...

noodle person

13:38

@NewPosts This should be classification, right? But the tag info says "grouping input data by properties"

Simd

16:51

Hello!

New Posts

18:50

0

Q: Index into a Fibonacci tiling

The Cartesian plane can be tiled with increasingly large squares like so: This tiling can be generated by starting with a square of side length 1, placed at the origin 0, 0, and then repeatedly adding a square to the side of the resulting rectangle clockwise of the previous square, starting on t...

code-golf grid fibonacci tiling

Seggan

19:22

@Themoonisacheese i ran an n body simulation once, was way faster on my integrated gpu than my cpu

mousetail 'he-him'

19:35

You can write python with braces now:

# coding: braces_indent

def main() {
    print("Hello, World!")
    if True {
        print("Finally, sanity!")
    }
}

6

bitecode.dev/p/change-pythons-syntax-with-the-coding

hyperneutrino

20:11

oh that seems cool. given that i went through the effort to make a whole programming language just to get python but braces i should try this

Seggan

20:22

205

A: Hidden features of Python

ROT13 is a valid encoding for source code, when you use the right coding declaration at the top of the code file: #!/usr/bin/env python # -*- coding: rot13 -*- cevag "Uryyb fgnpxbiresybj!".rapbqr("rot13")

Redz