#programming
1 messages · Page 299 of 1
87.120.191.*** - - [06/Dec/2025 15:13:10] "PROPFIND / HTTP/1.0" 302 -
Bots scanners
They all trying send http requests on my vds ip, but get redirected on my domain
Okay, so I feel bad for the dev and all
But why did you run it in the command auto approve mode with full access privileges
cause he a ScriptKiddie
you have to explicitly enable access to files outside of the workspace
hard to feel bad
let me check my dns
makes sense, id ont actually have a website
its just a minecraft server that doesnt get used
i doubt the ai zerod the entire drive, so it should still be there
just need to let a program scan the entire drive for some hours
i used to do that to hard drives form old bussiness laptops, the shit you find on there is crazy
LLM moment
having used C++ every day this month due to aoc I'm impressed with how far it has gotten (if you ignore the whole object oriented part of it)
I know your name is meant to be a play on santa, but in Scotland it means something so different that I nearly spat my coffee everywhere 
40 years of adding features will do that 
what do yall think abt my custom made catpuccin theme for mint
start menu is abit more rounded
looks nice 
should i release it?
if you want to improve something, this looks a bit imperfect to me
like?
uh its not my fault actually since i used original mint's css
its mint mistake
cant change that?
i just modded colors
ohh
i will
the colors are nice
i modded libadwaita, gtk4-3, cinnamon (mint as base)
bottom panel should be distinct but im not sure this probably depends on the taste
idk actually i made it to look like that since uhh it looked cool for me since i wanted make it like start menu is connected to bar
its early version tho so its possible that stuff will change
ok probably fine then
Mint !!
as i said its more of taste preference
but non aligned random size icons that i dont like but its minor problem
oh and text is too close to the edge
it will blend in with things behind it
for example if you open 2 of those
i have no idesa how to configure alacritty 🙁
will look like this
i used it since its easy way imo for me since i can easily just lookup for hex colors in text editor and use sed to replace it all
since idont have even to open stuff to edit it
just sed and done
i fixed that start menu corners weren't rounded at all
tho it looks fucking bad with bottom bar
either i will make it to float
or remove rounding
on bottom
mhm
yo im trying to start programming is python a good option to start with ?
Eh I would start with something more sensible, though Python is a common language to start with
I personally started with Kotlin
Python has a lot of weird annoyances
yes
i recently discovered https://nim-lang.org/ and i felt in love with it
oh god
it compiles to both C and js 
honestly this just looks like python with extra steps
compiles to c++.... what does that even mean
I personally hate Python, it sucks to write anything complex in it, the only thing it has going for it is PyTorch for ML
YEARS OF PROGRAMMING yet NO REAL-WORLD USE FOUND for going more underground than ZIG
L opinion
python is great cuz of how fast it is
to write in i mean
Eh, it's not very fast when I don't know what type something is
it does its job, its faster than python
but is it worth it? you can write fast in a lot of languages
wrong mindset, python doesnt really do types
I hate dynamic types they're super annoying
just wait until you hear about Haxe 
lets not hear about haxe
python can be made much more tolerable with a type checker
you are now hearing about Haxe
Yeah it isn't a sensible language for sensible people
it is
based
i feel like as a beginner not have types is the worst choice ever
if you cant figure out what type something is based on the variable name, you have bad variable names
imo
ok mr hungary
?
look up hungarian notation
win32 the pinnacle of api design 
that isnt even making the variabelnames sensible so you know what type it is, that is jsut straight up putting the type name in the variable name
i mean, python has types, you just dont need to specify what type a variable is
are you even programming if your compiler isn't showing you your types via a text file
im saying not to use hungrian notation
?
what im saying is, for python. if i have a variable named "distance" what could the type possebly be?
its not a fucking string is it?
it could be
Well the code editor doesn't know, so the name is useless
if you genuinely name a variable storing a string "distance" you're dead to me
float,int maybe a vector
float and int could both be an answer, but python doesn't do any distinction between those so having them be different wouldn't matter. vector tho, you're dead to me
idk a string could be reasonable
what if you're writing a cli tool that takes in distances like "3km"
on a 2d grid a vector can be like x and y scalar distances. for like pathfinding or something
Python does differentiate int and float
It just doesn't enforce it anywhere so you end up mixing types and potentially causing code flow that wasn't intentional
then put list in the name, cuz its gonna be distance in the x axis, and distance in the y axis
"distance" by itslef should always be 1 number
this is what stuff like pylance is for
list
if i see a variable called distance_list i would think it has like 50 distances and not xy
Yeah variable names don't mean anything
The actual type of the object does
then call it gridDistance is dont care.
gridDistance can still mean both tho
not in my code
and if you have it as a vector you'll know 
why not distances
luckily python doesn't have vectors
+1 reasons python is bad 
Python is just a pain

????
what vectors are we talking about here
python
python lists are the same as c++ vectors. What is your point?
i mean... most languages don't have vectors in the sense
its a math library thing
we were talking about a Vector2 basically
idk i can't really use a language anymore if it doesn't have types
the point is that earlier we said lists, and then someone said what about vecotrs, and now im saying that python doesn't have vecotrs cuz it has lists
thats the point
you can argue to me about this all you want, im still 100% convinced if you cant at least guess what type something is based on the variable name, then you have a bad variable name
if you yourself wrote it within the last 30 days maybe
but what might be obvious to you can be a riddle to someone else
but isn't it still better to just have another thing looking over it? like something to enforce it regardless
I have just come to the conclusion that inlay hints work for me
it's a shame that there isn't a super practical way to get them in stuff like GitHub but meh
also type hints + type checkers
requires extra setup since type checking is not part of the language itself, but the tooling does exist
if its a riddle to someone else, and they cant figure it out form context clues either in the code, then its a bad variable name. idk what to tell you.
"size" should always be a number.
those are the most basic of examples. Just doing the small aoc tasks would already break the whole thing because you have to transform between data types
this is why functional programming is superior, no need to name as many variables 
what about person.gender? is it a string,enum,number how can you guess from that 
but with types you'd atleast know what it is and how to use it 
If you wanna be inclusive string if not, just do like
-1 = male
0 = nb
1 = female
2 = other
id do enum if we have presets
But you could also enum those
ye
-1 
Nah bro i identify as the number 1
i could make a joke rn but i dont think i should
Good call
Almost went to crush up some belgian waffles infront of you
Or to straighten out your roads
thats something we want tho
do you think we have bad roads for the love of the game?
Of course.. belgium has more than enough money to slowly begin fixing them, yet they don't
we're actually in a fuckton of debt lol
So are your neighbours up north.. yet they somehow still manage
(For any bystanders.. this isn't serious)
we're roughly 650B in debt, yall only 490B
Okay, so 160B difference
That's nothing
i think to repave an entire country properly would be 50B or so
Maybe your government shouldn't let the dutch people build one of the biggest music festivals on your soil
but it brings business to local companies, and its just the government that gets the debt
we're all wealthy
we actually dont even like our government
Yeah i wouldn't really want tomorrowland to relocate anyways
We need a different strategy

amsterdam dance event?
Yeh
i like how you felt the need to clarify we're not actually arguing
fijne sinterklaas btw
Tanks
Yeah idk, i had multiple occasions of people actually trying to break the "argument" apart while me and the other person were just having fun

I am False
in other words, I am not real 
numbered genders when NaN walks in 
hi proframiing
can be a dict or a list
that’s what. nonbinary means
by definition
i think ???
i thought non-binary just meant not male or female cuz they don't feel like either
idk
im interpreting NaN as not a gender
but NaN is a number so it has to also be a gender
That one is NB, idk why it was zero
Infinity is gender fluid (real)
This enum sucks
NaN is specifically not a number tho
don't make me pull out the ieee spec
NaN is a valid state of a float tho
i'm lazy
but it's defined as a number named "not a number" 
im gonna kill the ieee
Surely after the millionith floating point standard it'll work 
it should be a value that the bits can be, without being a number
That is literally NaN
toast said it was a number tho

This is why we have to have a function to check for NaN instead of comparing to it
Thread said it was for debugging that never happened 
use the rest of the useless bits as a trace
honestly at this point it feels like so many things that are bad design have just stuck around from old times
and we cant start again cuz than we'll have a situation with the extra standart
xkcd 927 
man how do u guys do allat while ur eyes are getting cooked
wdym?
my eyes are already fucked 
the monitor is bad
i just blink on occasiuon to moisten the surface
i too like to consciously blink 
i like to use va or ips monitors instead of infrared lamps
i use qd ips instead of laser projector directly into my eyes
i'm looking at my monitor for 12-18 hours a day for the past 15 years. and i got perfectly fine vision 
hello test test
Need to look away for a bit, can't stare at the monitor all day without rest
test
test recireved
my vision is not 20/20, but its like 18/20
are u familair with blue filters
yes. and i don't use them. i don't like the color shift
its just hard to read text at a distance of 40m cuz my eyes' focus is slightly off
I feel like the monitor is dying 
im red green colourblind, if i then also get rid of the blue i have nothing left
might aswell jsut turn the monitor off at that point
no contrast 
check out this ruby i got
This monitor setup is so jank
software jank only this time
I do not recommend overlapping monitors, I'm just not bothering to move them due to scaling bug
overlapping monitors is such a Linux feature I love it
wayland scales coords, removing them causes these two to overlap
x11 just straight up displays stuff in both due to x11's all desktops only being a single buffer
this is not that
If I spark the x11 vs wayland fight over infomation again this week, I am just vanishing
Tiled windows display only on one monitor, and floating windows display on all of them because of the dragging between monitors behavior
simple as that
i dont get it but im sure its cool
I'm still fighting a hyprland update to fix the inital cause of this
Might as well continue trying to get the damn thing to update
Don't need to deal with these side effects more than I have to
First time seeing those warnings on overlapping monitors, must have just been added
updated 
(now to see if it solves the bug)
Finally I can see again
I no longer need the laptop to be a 1x scaling
due to certian toolkits doubling the scale again
so would make sense to map not a number to “not a gender” in a table of gender which nonbinary by definition is
so 2x scale became 4x but the UI input was still rendered as 2x
and only the top left would be rendered
fair enough
meow


I have traded a scaling bug for a ghost cursor bug 
Man, I've gotten so used to 1x scaling that it is hard to go back
Honestly though, I think I might just not
time to experiment with other scaling factors
I at least know I ain't going back to 2x on a 13in 1440p monitor
Thank you me from an hour ago for switching the last of the configuration to variables
So anyways, I am much better off 
(one of these days I will make a monitor mapping script for the damn touch input, today however is not that day)
when people ask if the chicken or the egg came first, do they mean a chicken egg, or just egg as a concept?
cuz eggs pre-date chickens by a lot
egg
not mammal
I think it is chicken egg, because otherwise chicken believers have no legs to stand on (they are inside the egg)
the first creatures that went on land used eggs
which came first, the egg layer or the egg
an egg contains a chicken, so it would have to be at the same time
i think thats how it works
where do you draw the line of chicken vs other animal
what if the chicken has a genetic mutation 
That was also my own line of thinking
i think bird fists
it isn't even a biological problem, it is just an identification problem
first
i think its something genetic differnce where you need to draw the line, so the mother would be a different animal and thus that egg would be a different animal's egg containing a chicken
thank you google
so the chicken came first
a
I think your final answer is inversed from your logic
-# The place of phylogeny and cladistics in Evo-Devo research - Scientific Figure on ResearchGate. Available from: https://www.researchgate.net/figure/Ordering-the-evolution-of-characters-Setting-the-chicken-in-its-phylogenetic-context_fig5_8892372 [accessed 6 Dec 2025]
checken egg would have to be an egg laid by a chicken
so if the parent of chicken laid an egg, it would not be a chicken egg
it would be a non-chicken egg containing a chicken
if it was just any egg, then it would be egg first


The whole debate is just because people are not specific enough about the criteria so both eggs are talked about as if they are the same 
(here before the christian answer of god created a chicken as is so chicken was first in that case)

computer
this is the correct and objective truth

so if yorue talking about chicken eggs the chicken is first but if youre talking abotu eggs in general egg is first
yes
cuz eggs were used by creatures that weren't even able to walk yet
fish and stuff
i always knew that but i never knew about teh chicken egg thing
I successfully build Linux from scratch 
how long did it take?
... lets not talk about it 
the answer is that there is no line that can be drawn and there was no generation that suddenly was a chicken
because that definition of species doesn't work to answer that question
oh so that's why you were running the gcc test suite, i see
I can usually build a whole distro in under 2 hours
lmao
was it worth it
time for BLFS
yes. I feel like I understand my computer so much more after it.
it never ends does it. (I hope to get to GLFS eventually)
to be exact, the difference between the modern chicken and its ancestor the wild chicken, is domestication. it has a different phenotype too like reduced flight ability and increased egg laying. also has a smaller brain
there is no exact line we can point to, but there will have been a specific chicken that was genetically different enough to be called a new species
we just dont know who and when
well, there was no chicken that was like, suddenly more genetically different than its parents
enough to be called a chicken
like, there was no chicken suddenly unable to breed with its parents' generation
yeah but there's still no real line otherwise, hence why the whole idea of species classification is criticized
and why the species problem is a thing in biology
the change in genome is gradual, but definitely quantifiable
with some outright declaring that species aren't real
quantifiable but it's an issue of classification
i can agree with that viewpoint, but id still say chicken eggs came after the chicken
it mainly depends on if you see the egg as the one of the layer, or of what it contains
I'd still say the answer is that a chicken is undefinable for this purpose, like how a river is obviously a river and an ocean is obviously an ocean, but where exactly does a river become an ocean

as you do
if you 3D model an entire human you do eventually have to make the feet you know
i mean
are you doing the internals too?
no
not if they're wearing shoes
I was gonna say I had to 3D model a whole bunch of a specific organ
but the server thought that organ was NSFW
im making the base for a bunch of different characters who can be wearing different outfits potentially, so i need the feet to exist
apparently a certain internal male organ (not the external one) is NSFW
erm
I do cancer research so it involves modeling the organ that men get exams for
I get that a lot

scope creep has once again made an idea dangerously unviable before even writing a line of code
jsut dont creep your scope
i say on month 5 of what was supposed to be a week long game jam
it went from "I should stream programming" to "I should make an AI play peggle, but I'd also move my PC to linux and have to write a mod for peggle and also learn how to do machine learning"
have you considered not doing al that
just vibe then
leave the regrets for the future
i feel liek this is jsut
although learning to train a model is actually an afternoon's time if you have basic python knowledge
"thinking baout programming" :D
"programming" Xd
😭
this is a dangerous statement for me
you can get something workable in 50-100 lines of python
I probably need to sit down and do a basic ass tutorial
because every time I overcompliate and get nowhere
I'd honestly just ask an LLM to put together a very simple training script for a model with pytorch
for a basic task (like classifying images between two categories)
like, they're neat tools
I don't like using them
well you could also find an example on github for one but
I've never had an llm that has made me want to use it
they don't really mess up something as simple as that
it either:
- makes me feel like I am being lazy
- generates useless code
it's when you try to do the entire stack using vibe coding that it all goes wrong
i use LLMs as debuggers, not coders
^
I tend to use them by isolating them to very specific tasks that I can unit test
I can trust them to write a function because I'll be able to tell if it messes up
but I won't trust them to string functions together
i still dont really get what a unit test is tbh
LLMs are idiots when it comes to literally anything
just testing an isolated part of a program
like a single function
so you can treat it like a black box after
Hi
Hi
How are you doing
idk
Ohh sorry
I want to know what the code I'm writing does
ye
if I know how AI training works, sure
yeah but you do read any code that is generated
I am not writing a magic ai training function
Alright what are you doing for fun

i feel like giving inputs the code was never designed to handle will obviously have the chance of producing errors
I would really rather not add unchecked stuff to my code
also like, I have been gaslight one too many times by LLMs to trust them to teach
well, you would be the one doing the checking
i make games. or well, that is the end goal
idk I sorta refuse to use LLMs for most stuff out of principal
I do projects because I'm interested in them
but in any case this is an example of a very simple classifier: https://github.com/pskrunner14/resnet-classifier
not because I want clout
you just need to look at train.py, model.py, and predict.py (to actually run it)
(or well, clout is probably there, just secondary)
while debugging ai has given me like example code to use instead, and i have used those before
so i guess that is technically vibe coded code
but i wouldnt consider myself a vibe-coder
eh, I only really consider it vibe coding if you never look at the code
fair
having cursor write you a regular expression or a simple function because you don't want to do it yourself isn't really vibe-coding to me so long as you look at what it wrote
I treat it like having a really knowledgeable but really dumb phd student working for me
doing intern work
uhm, this kinda looks like a bot
@safe path their messages seem suspiciously repetitive, kill time
even if it isnt a bot, seems to be artist scam
instantly friend req'd me
ogie
Ohh okay
How
post mortem investigation, maybe i jumped the gun, but it is still suspicious
ah yes
if they appeal their ban or use an alt, my bad if you werent a bot
just kinda looked like chatgpt generated text with all the emojis, and sometimes there were just the same message
what happend in here
person that acted kinda like an npc, but may not actually have been a bot
needs further citation
eeeeeeee
their only mistake was having their 2nd message immediately be promoting themselves as an artist that does 2D and 3D and takes commissions i guess
i’m currently sitting uhh like
erm actually
beside an apartment door
lmao what
and im in the middle of connecting toes to the main body
because a friend forgot we were supposed to come back and closed his apartment

bro is drunk enough to not care about sleeping outside 
i wanne have what he's having
oh
just not in the apartment
dont worry about it
erm
erm

Average #programming conversion
you see, if they dont know what im wrong about i dont have to explain myself
i don’t even know what is happening anymore lol
this is genuinely a fever dream experience
we have asked a random convenience store worker to boil us some water to make instant noodles
and it worked

you're getting the main character of an rpg experience
walk to convenince store, press a, get boiled water
stay safe

nixos dual boot installed
now uh, to figure out how to make it boot into windows by default
grub?
yep
On the same drive? masochism
different drives lol
does os-prober even detect windows on a different drive
oh then you can just set boot.loader.grub.default to whichever index it is in the options, i assume it's probably 2nd right
it's at the end of the list
does the index remain stable as more nixos revisions appear?
or is it different in grub idr
i'm honestly not sure, since the previous revisions get put in their own submenu
o yeah they do on grub
may as well just try i guess, worst case it doesn't auto boot
oh you can just set it to the title/id of the entry apparently too
assuming nixos just passes it as is to grub via GRUB_DEFAULT
hm
I tried that (at the suggestion of Google ai at first)
no luck on my first attempt 😭
there's always the option to disable os-prober and create the windows entry yourself and then point it to that one 
did you use the entire entry name as it's shown in the grub config or just part of it 
(ignore the double quotes in the configuration.nix, I was testing something - it still doesn't work without them)
wait 
why is it a different nvme drive
ok after fixing that it now boots to the correct choice selected in grub
I sure hope I don't have silly nondeterministic drive names 

extraEntries = ''
menuentry 'Windows' --class windows --class os --id windows {
search --fs-uuid --set=root 0CF9-FFEE
chainloader /efi/Microsoft/Boot/bootmgfw.efi
}
'';
default = "windows"
this is what i had, might be worth if you're worried about the drive names
hm
I can just copy what my grub has
I'd just like something that's slightly less reliant on it idk
something something reproducible builds
though I guess its a tradeoff between hardcoding my grub entries and autodetecting them
I'm not even using a vpn
yeah nixos.wiki is like the wrong site or something
don't ddos the wiki 

yeah, just got fooled by first result
I recognize what it is now, and yeah the unoffical wiki is pretty useful
ok sunshine is having some slightly funky behaviour
programmer goes outside
"omg are you ok stay safe"
i’m home btw
for everyone concerned
so true
currently it won't let me connect if the pc is locked
which is funky
...I don't actually know how to approach that one
maybe I can just be lazy and make sunshine force unlock?
but even then it seems to refuse to let me do anything with the device from the lock screen
hmm
I am very sus of it being a user service
despite me not wanting it to run only when I'm logged in
ok it works fine if I lock, just not sign out

there's got to be a good way to handle this
on the other hand
its nearly 5am
reddit has the wise solution of "auto login then lock"
That's non-binary
lebron++
Okay

bro wiped but left not
@opaque sigil
yeah nixos keeps switch around which nvme is which
gonna just haddcode the entry at this point

I am not getting the authentic linux experience
how do I wake my display on wayland, from the terminal?
"don't use wayland"
"run <command>" (doesn't work)
"don't let your screen lock"

why is the seemingly best solution "simulate the mouse moving"
because just turning on the displays breaks kde power management
i dont understand
wake display from the terminal? you cant see the terminal if the dispaly is off
I use remote desktop a lot
oh, ok that makes sense
so I need my display to turn on
or well, I can make a virtual display I just want an MVP first
guys what should i run on my homelab
protein folding simulations? idk
Invent a new vocal synthesizer
Who up gramming they pro right now?
im 3D modeling fingies rn
220 is trying to tempt the lion
Why does ChatGPT seem better to me than Gemini, yet score of it much lower?
hi guys. Does anybody have an info about Neuro's params?
give us a sec
we have a graph
i guess this isnt really exactly what neuro has, but for most of it we dont know
interesting, thank you!
I created an multi-agent system locally and simply wanted to compare my AI to Neuro at that point 😄
iggly 

does anybody here also run their own AI's?
some do ye
fck yea, glad to see it
I just make vocal synthesizers
i'm interested in how did they integrated long term memory (like SQL database or weights alteration in real time?) and other agents (like "vision" agent, STT agent e.t.c.)
about vision:
the best I can propose is to use VLM agent (like qwen3-vl-4b) and connect it to LLM as an agent with tool call. LLM, while getting an image [absolute path] calls VLM agent to analyze the image. Then proceeds the output


how the hecc does one make an LLM that holds conversation
like
i tried ollama and was like 
you mean to make it say things by itself?
honestly idk. i guess have some sort of input stream still
now that i think about it she gets responses from chat all the time no
ye that too
there's probably a pretty thorough system prompt or memory corpus in the first place too tho
An LLM is a text completion machine, nothing more, you just trick the text completion machine into completing natural conversations
👍
so basically just train a model for that instead of using a pretrained one
Eh
No you can use an instruction tuned stock model
You can't really pretrain your own model on your own system
Training an LLM takes stupid compute amounts
You just need a good prompting format

Neuro simply has chat input
nah. Without input model cannot proceed output
besides chat input she has vision and vc tho
cannot llm by default emit infinite stream of words
simply agents (vision one and STT + TTS)
i thought thats what it does. just wire its own input back so it proceeds?
oh you mean chat as in text, not chat as in the twitch chat
or maybe it gives her some prompt like "go on"
nope. If that was true, then Neuro couldn't have any conversation with Vedal
Nauro has agent that scraps chat
maybe this?
she doesnt need chat to talk ininitely
autoregressive
the input for the calculation of the next token is always just the stream of preceding tokens
no matter the prompt, LLM cannot generate outputs on its own
huh
it literally generates words based on previous words
(tokens)
i mean im not an expert but that was my impression
I mean LLM cannot give output without input. It can't just treat its own output as input
why not
because arch limitations
what if i feed random noise into it when there is no input 
last 100 tokens for example
so it continues the flow of tokens
- some kind of memory he made
anything on top of that is just post-training, usually with a chat template
you can force a post-trained model to continuously generate new outputs by itself without human input, but it'll usually quickly go off the rails relatively quickly because that's just not in the training dataset
then it will proceed output. Even noise is considered as input
what limitation exactly?
transformers one
i pressed bake in blender
now the entire window is black and my ram is suffering
oh maybe if its fine tuned llm she gets that massive "you are neurosama twitch ai streamer ... (500 words essay) ..." before each prompt
maybe not before each
any LLM today is based on transformers
then no, no such limitation exists
transformers sort of encode current environment/topic, right?
so u wanna say Neuro can go into loop with her own outputs?
it will not be repeating loop (maybe it will if you leave her for 100 years like that)
if the training data allows for it, sure
oh its not happy 
training data only gives weights to the model, but autogeneration of outputs is archetecture thing
unrelated to weights
imo Neuro uses some scrapper agent for chat while streaming, so she has permament input
hm. Then it becomes interesting
yeah, but the training dataset limits what a particular pretrained model is useful for
then it's weights + Vedal's own arch for autogeneration of outputs
because making LLM generating something on its own is not something generic?
but it is though
a base model does precisely that
oh actually sometimes she falls asleep.
vedal said its when "there is no output for a long time" and "she gets poked when it happens"
never seen a thing with LLM which can generate outputs without any input
that capability is just usually somewhat trained out of the model during post-training because it's not very useful as a tool
then Neuro can have inputs from agents. Like, OCR agent autodetects -> gives input -> LLM (Neuro) generates output
something like this scheme
only 145w at 100%? do you have it in eco mode?
I alredy did that with my LLM (agentic input)
tdp is 105W, this is already PBO. this is quite literally already 1.5X as much power as it normally uses
if you have a problem with "no input" you can probably put there some observer who gives it "go on" input when it stops
do you know how an LLM is trained? as in, without any base model, only from data?
I'm not gonna explain, just curious
never had enough resources to do post-training, but as far as I know we simply give model dataset, some control agent (another LLM, for example) and model alters weights based on reward function
if I misunderstood question - sorry
deepseek v3.2 when given a completely empty string as input generated a chat template assistant message with this as content:
你好!👋 看起来你发了一个空消息,可能是误触发送了,或者还在思考要问什么? \n\n我在这里随时准备帮助你,无论是: \n- 解答问题 \n- 写作辅助 \n- 学习指导 \n- 日常聊天 \n\n如果你有任何想问的,随时告诉我吧! 😊
which is something about me sending an empty message
which is weird because no message is there
maybe, but it's difficult. Agent needs to autodetect no input for X period of time and send "go on" to LLM. Too diff imo
seems pretty simple to me
i mean .. architecturaly
in vacuum
i found new shit to mess around
you will need to put this agent between you and ai, so it can see entire flow of input/output
orchestrator of some sorts, yea
it relays your inputs to ai and then relays outputs back to you
and can poke ai on its own when needed
post-training is only one step of the process, and what you're describing is specifically reinforcement learning, which is basically another sub-step of post-training
idk where your confidence on this topic comes from when you seemingly don't know pre-training or self-attention 
with common timer or whatever logic you want
simply never seen something like what you're describing, that's why i'm confident
I mean yea, it can be, but autoregression makes model kinda stuck in the loop
nvm this is wrong it did wrap the input in a message
it's not effective
i did figure out how to disable it though
here's what deepseek v3.2 has to say with an empty string as context:
oops bad formatting
ye, this is what's likely to happen if your model isn't built to work without input (which is practically all non-base-models, because it's a stupid feature), it's a training data issue
though you also have stuff like presence or repetition penalties to help with this
its just basic loop. for neuro it works well because there's constant input from people and twitch chat
seems to be hallucinating a prompt
glueless
well, that makes sense. But idk if Vedal has such a thing on Neuro: to train model for that (a medium one at least, 15b for example) needs absurd amount of compute powers
you don't need to train from scratch, there are pretrained base models available for free
it's also not that expensive as long as you're not training GPT-6
there should be a lot of auxiliary stuff around neuro which makes her look not repetitive
memory and stuf
oh i know why its takign ages to calcualte, its doing 250 frames while i only wanted 1 
still frame video 
and as others already said, Neuro doesn't actually work without input
quite the opposite, there's always input available from chat and other sources 
also people kept misunderstood reinforcement learning
the model never learn something new with most popular RL method, it just learns to favor specific outputs that is more preferable from the feedback most of the time
then I was right and Neuro simply has constant input from chat at least
there's absolutely no reason to pretrain a whole raw model yourself unless if you're a giant tech corpa with billions at hand
pretraining just form semantics in the model. pretrained model gives pretty good enough starting point for finetuning
it does have large benefits for math and alignment AFAIK
but yeah, you can argue that it just brings out capabilities of the pretrained model (as does most post-training in general)
vedal sometimes disables chat input.
and then if she has no voice input from whoever shes talking to, she is left with no input (i think). but she still goes on
doesn't eliminate the fact that you can actually inference an LLM from starting tokens that technically isn't an input
although the llm will start outputting random topic that way lmao
That's literally how LLMs work
Neuro doesn't do that
no, you said that a model can't work without inputs, which is false
Neuro obviously has some input all the time, but debating Neuros architectural details is stupid, because we just don't know any (other than relatively safe assumptions, like being some combo of LLM + TTS)
doesn't have to be "neuro doesn't do that"
any llm including neuro's can be just inferenced with no input token
are you sure
and it will spew out random topic that it auto regress through
possible but very stupid, waste of money unless you're training a SOTA LLM with capabilities beyond what's freely available
I'm abserving Neuro from YT for a good amount of time
and even then you're gonna be better off with continual pretraining or other augmentations
function calling on pretrained models are pretty well generalized aswell so finetuning works well for that either
do base models have function calling now?
they shouldn't even have the special tokens to do that
huh, iirc they have at least a few special tokens
you're about tool use or what?
or im confusing it with formatting token because im pretty damn sure i remembered seeing basic format tokens on pretrained base models
Deepseek V3 has just bos, eos and pad (and a whole bunch of placeholders)
the placeholders work well as function calling tokens for finetuning ig
i think
used for fine-tuning I think, can assign them a specific purpose then
how long have i not been working on any llm 
my memory about it is sort of fuzzy rn, or maybe because i haven't slept since 2 or 3 days ago
LLMs silly
I haven't touched them in a while either
last time i touched ANN models, it was just pretty recently when i was working with alphafold
might be the second one 
sleep more 
insomnia
Without fear
Insomnia

why I can't send a screenshot 
once i got the time, i'll work on diffusion models for real time robotic applications
I can′t get no sleep 
you're new
you need to level up to earn embed perms
ewww
gpt-oss is shit for its size im ngl
for STEM it's quite good
moe moe kyun
it can properly solve tasks from quantum mecanics, so it is good for me
image generation disgusting slop
it can definitely use more
maybe with custom bios profile
im jsut usign PBO advanced 4 or whatever it was called
ppt for 5950x is 142W
im on that as well, what motherboard do you have?
but pbo do sometimes push it quite far
msi x570 tomahawk
ye
its literally the ppt on most agesa preset
I'm not sure why mine is so high then
i don't think 200W would be too healthy for my setup, specially with me only having 750W psu and my 3090 still needs most of that
MoE is good though 
fast
some pbo set higher current limit
so you tend to draw more on full saturation even though voltage isn't above stock
my psu is 700w, but my gpu is only a 6800xt so I have a bit more headroom
pretty tight headroom though. what model of psu?
gotta ensure voltage transient is at very least ± <100mV on 12v rail
oh nice
it might be cooling actually
i never reach more than 70C with the nh-d15 so i doubt its cooling
im lying i got 75C here apparently
sam, just do advanced/custom pbo lol
^
increase ppt limit, increase current limit
i could i guess, i dont think i need it tho
set pbo clock offset limit as high as the board allows
preferably you can also undervolt
i hit 340W ish on my 7950x engi sample before with custom pbo
could push more tbh
but the ihs is shit
and the die package is also shit
i can cool it fine but the thermal density would always cause hotspot spikes
im more comfortable pushing intel chips to above 400W for this reason


u promise want overclock 13980hx together
tomorrow maybe
I hate windows, for some reason windows explorer crashed and didn't start back up so I had to open task manager and run it myself
I'm thinking about it, another bug happening recently is that windows gets unfocused or something and the only way to fix it is with the actions you can do with keyboard shortcuts, sometimes opening a new window fixes it so I can just open task manager, but sometimes I need to use keyboard shortcuts to restart the computer to fix it
Windows is borked for good
Yeah, it's getting worse with every update
I think they said that like 30% of it is written with AI at this point? Surely that couldn't be the problem 
Yeah they did say they use 30% AI code
Which is fine, what isn't fine is when they don't fucking test that everything the AI does actually works
guys is that suspicous that my site got request with "GoogleBot" useragent on my site with mullvad vpn ip?
Yeah they seemingly just go with "if it builds it goes into the release"
from Netherlands
Yeah, it's not great, also, the page for the changelog of a recent windows update was completely AI generated, Including the screenshots, there were no real screenshots it was just fucked up AI generated ones that anyone who looked for half a second would realize were completely fucked up
Yeah they've lost any chance of actually making a decent product they just push slop
Funny relvant photo
Yeah, they've fallen far from when they used to have a good product
windows always alternated between good and bad releases
Windows 10 was good, if they had stopped there, with only bugfixes and actually useful features I think people would have been happy
Windows 10 was the final usable release
personally I liked windows 8.1 over windows 10
moved to linux shortly after windows 10, not because of problems with the os itself, I just liked the start menu of window 8 more
I can't remember a single problem I ever had with windows 10
I rember having a few, but those are my issues
Also, Windows 10 ran really well on pretty much any device, Windows 11 just runs like shit unless you have a very high end computer
Mine went to completely garbage, after running for a bit it would just become unable to open any programs, or the start menu, or the right click context menu
except for the one where an update completly bricked the system, wonder why I moved that one over to linux 
Yeah an update bricked mine too and I swapped to Linux for good
That happened to me but with Windows 11 although that might be my fault since that was the update that would bring me to a windows insiders version
I'm lucky I already swapped to another PC before it did any real damage, but yeah that sucked
Although it's really stupid that they don't give you any sort of warning that it could have any negative side effects, you just have to agree to something that probably makes Microsoft not liable if anything goes wrong
my discord bot got DDosed
Classic Microsoft behaviour
Some updates i recall made my laptop blackout once in a while
windows always ran like shit for me
then i moved onto ubuntu
best decision of my life
then i moved onto fedora
bester decision of my life
fedora on a high-end pc from 2018 is lowkey peak i'm not even looking to upgrade for a few years
except maybe my ram whenever that drops in price
same witha rch for me
when i use windows my pc makes a lot of noise, on linux i forget it even has cooling
I wonder how Microsoft can become so incompetent at managing their own operating system?
pure carelesness and greed
I'll add laziness as well.
And a sheer nonsensical inability to see to their customers.
if they were lazy we would still be on windows 10, they do put a minimal amount of effort in to try to look novel
It is almost cartoonishly evil the way they operate. Its like they got their ideas from a cartoon. This isn't even how a decent criminal/evil person would operate.
And they only make their OS look good to appeal to their investors, otherwise, they wouldn't even bother with that.
Its like they think their customers are filth.
Or vermin.
They show who the real vermin are by their actions.
Swap around some computer hardware
Attempt to install Linux on spare drive
Spend an hour figuring out why the live CD grub keeps going back to the menu with no feedback when trying to enter the OS
Disable secure boot
Install
LCD monitors are flickering
Can't sync my KeePass database because proton drive has no Linux client
problems for another day, back to Windows
I don't really use cloud services, and I think life is better that way.
Other than that, what graphics driver were you using?
Dunno. I just assumed CachyOS would set the driver up correctly.
I don't really use cloud services, and I think life is better that way.
Well I do need a password manager and I don't want it tied to any specific cloud service. Hence the keepass database in cloud storage.
That and email are too critical and high-maintenance for me to trust my own servers
there's the option of connecting to proton drive via rclone i guess but yeah, not ideal
i'm both surprised and not at all surprised that they don't have a linux client
they're all the same anyway, pick whichever one seems the most interesting to you


LFS 

IDK I am only looking at the top ten of Distro Watch and I know already I like Debian and Debian based OS's.
I'm looking for osmething that i lightweight becuase my mac is
and something that has support for the software I need
Everything else i can run through Lutris, Wine, Heroic, and WinBoat
xubuntu 






i also have the motivation







