#programming
1 messages · Page 409 of 1
do you mean "this" or "quantizations of this"? coz that's a little too big for what (most) people have at home.
are you into ai?
Quantization naturally, FP16 isn't exactly the norm for on-device inference
27b is up there for sure but not thaaaat big
little bit. i run some models.
This fits on a 4090, 3090, 5090 or 7900XTX at Q6
out of everything i tried - best performance i got is cloude 4.6 thinking
it actually solves problems, not only give direction to search for
is it actually best we have now?
it have reasonable thinking
it write something and then say that it is wrong, and accept it
Its the best coding/STEM model we have, and Gemini 3.1 Pro is the best generalist
i'm curious to see the next gemma
can it answer bread question?

it just implemented an SVG avatar for my agent which lipsyncs to the audio file with visemes.
that was a broad question.
bread question is: how did neuro think when she was bread?
other than that, i don't know bread question.
The issue is that Opus costs a bunch
I find GPT 5.3 pricing more agreeable
Opus is probably really good, but it costs more per query than Gemini 3.1
idk, i have it for free
i not use ide solutions or anything so idk about rate limiting, but (lm)arena,ai offer them for free
i mean no one takes these models seriously
they was promoted as russian, but actually just translation wrappers
Also has higher rate limits
all lm still bad at code generation
small yandexgpt models are just fine for what they are (8b-ish?), everything else is just bad
why would they be translation wrappers when learning a language is like the easiest thing for an llm
have you read actual note (post)?
the blog post?
ye
not true. like i said: Gemini wrote me an animated avatar for my agent which is actually not terrible. twice. Once using ASCII art, and a second time using SVGs.
and the code was readable.
i have yeah хлопок (cotton) / хлопок (clap). i think it just indicates exactly what the post says, a lack of intelligence, not that it doesn't know the difference
and because of the requests, i'm pretty sure nobody has done what I asked before. so it wasn't just copied.
like, if you told it to translate it in context, it 100% would just fine
i also got working script for lua, that says nothing, they are still horrible
while sanity checking current code, cloude for some reasons created example that looks like this
learning a language is like the easiest thing for an llm
you can finetune a language into an llm pretty easily
it really depends on how you ask it.
especially considering how much of the internet is in russian
it's the latter
i know
that;s the point
The Yandex models are jokes honestly, go for Qwen if you want the best Russian performance
aren't they proprietary anyway
Or Mistral if you want less guardrails
Does it do the classic "The vodka was good but the meat was rotten?" (The spirit was willing but the flesh was weak)
i havent asked
it just generated
and it cant work in any way
starting on that there literally no way to obtain coro_handler, ending with actual function being unused
nice, ds bugged
i'd say go for anything for "russian performance" whatever that means, the only llms that would fuck up languages that aren't english are gpt-3.5-turbo era
if you didn't ask for anything and it generated garbage, i'm not surprised it doesn't do what you want.
lul
that's why "prompt engineer" is a real job
friend saw an "ai vibecoder" position yesterday
i could do that. i'd have to leave some of my brain at home, but... ;]
guys, for real, how can i find job online now? linked in seems like crap, havent tried to write directly to any HR tho...
-# i got card, i had personal vacation, now i want to look forward in my life 
networking is the way i found jobs. people i knew told me about jobs where they worked or put my name forward
This
nepotism
i wonder how much of that is due to the fact that c++ coroutines are one of the more obscure features that few people even use so naturally there would be far less training data on it 
i definitely got one job because the boss knew me from a previous job.
that reminds me i actually have to answer someone pretty important that might get me a cool job like right now but i'm here instead
i prefer copoutine. mmm
yeah like... i thought that coroutine handle destroy itself always if you use trait, that it create owning shared_ptr-like variant
but turns out that it just call destroy if suspend_never 
(tho i only had to do this instead of clearing stack)
how
discord
while(not stack_.empty()){ // fix memory leak with correct destruction
stack_.top().destroy();
stack_.pop();
}
what isn't a clusterfuck about cpp
Magic
there are levels to the madness 
nah, api is nice to use
the only reason i used it: std::hash
it just poorly documented
The syndrome is kicking in
man, i just created software stack and i have coroutines that perform sync+preemptive task without state machine on same thread as realtime performance task
so i have natural priority + preemption just by using co_await co_yield
that's the big barrier for agents, and i guess humans 
Bear in mind as well that they advertise these models as well suited down to q4 quantization
The Qwen3.5 series maintains near-lossless accuracy under 4-bit weight and KV cache quantization.
In terms of long-context efficiency:
Qwen3.5-27B supports 800K+ context length
Qwen3.5-35B-A3B exceeds 1M context on consumer-grade GPUs with 32GB VRAM
Qwen3.5-122B-A10B supports
just read compiler code, and you'll figure out what it to 
Holy AI slop replies
"We’re entering the era where model design is constrained more by IO than FLOPs"
Bro, that's the era we've been in since llama dropped
Qwen is on my list of models for when i get a 2nd GPU/new machine.
But we as a human can at least do test and infer from result
bet that ai cannot write anything similar just by description what you want, tho...
https://godbolt.org/z/8MKWMPETc
gemini ran some tests yesterday when planning code changes to determine undocumented behavior of an API
yep thats ai. not just the em-dash but also the classic "it isn't just - its that"
i know that there some key flags
but i usually just read something, and notice that it is ai, only if it feels off
i also cant quickly differ ai from human art
its fairly easy to tell nowadays if something is written by AI because all of them are just not how humans talk
i cant 
art is harder but text in my opinion hasn't really evolved atleast in the spammy way its used like in replies or comments
(neurodivergent not AI)
tho here i noticed it on curious part, that feels really off human
(read img before solid's text to it)
But not yet, since apparently there's a bug with it in llama.cpp 
looks quite dead
at least nobody can read it over your shoulder so it's safe to use at a cafe.....
unfortunately you can't read it either but that's the price you pay for privacy.

Buzz online is that the 35B MoE is better despite benching worse, but I'll reserve judgement on account of the whole bug thing
I will say I've gone off unsloth quants recently, I couldn't tell you why but Bartowski's releases just seem more stable
Might compare for this release once they fix the 27B model
Ubergarm releases are best but naturally take a while ^
Ubergarm?
The guy who puts out all the ik_llama.cpp quants
resolutions being put together 
DLAA and no FG or gtfo 🫸
the gif does nothave DLAA. 3/5
remember when everyone was shitting on dlss like even harder than what the current sentiment around fg is
yeah it got better for sure but so will fg
just you wait

am i supposed to dlss + fg my compute workload
interpolation works great... for certain tasks.
yeah you should ai generate your ballistic missile trajectory
using an image gen model (important, will not work otherwise)
"Alexa, calculate trajectory for enemy position!" "I drew a fluffy puppy."
i'm memeing but can't current models generate the graphs they need with the llm part and feed that as input to the image gen step
like don't they already do that
i've seen AI graphs before but they were laughable
i would test but they do be expensive i'm not spending 20 cents per image
or whatever nano banana pro is
imagine using diffusion to generate graph line
gemini 3.1 pro can probably do a decent job given how hard they pushed it towards generating svgs
you could get it to do a deep research project and produce a graph showing SOMETHING vs SOMETHING_ELSE just using a free account.
(you need to specify the SOMETHING and SOMETHING_ELSE though)
oh god. gemini's got "create music" now too. ;[
please don't replace me, thank you.
https://gemini.google.com/share/7d44ddbdd146
I'm so sorry.
i wrote it some punk lyrics and it did about what I thought it would do, except it ruined the ending.
yeah, it sucks. that's 30 seconds i won't get back.
it's such a stupid prog it chose lmao
annoying but straightforward things i don't feel like doing myself
what do i use LLMs to help code? or what am i coding that uses an LLM?
ok.
i guess i did just use it for entertainment but that was also a "what does this feature do and how does it work compared to other similar features elsewhere" test with the music.
it would be nice if there was a music generator which produced something i could load in a DAW, with all the track split out.
then i could redo all the bad parts.
Therapy of course like a normal person
I use LLM for entertainment like stardew valley mod
For legal reasons this is a joke. AI is not a good therapist. They are programmed to be suck ups
Also, I wish there is a mod that provides arbitrary text input to chat with NPC in stardew valley to further extend the LLM entertainment value lol
so you can progress more than 5 hearts?
Nah, like make the NPC can be spoken to more than once/twice
i use it for coding but i never use llm code because its so bad, does that count
when i was off work for a month i watched a streamer play stardew valley with a mod so she could marry anyone.
Hmm, I don't think that counts. I use it for boilerplates for example. So it does get used
Because I am too lazy to write a proper tooling
Like defining a bunch of sql tables and their relationship for example. If I provide it a good enough prompt, some LLM is actually helpful and I only need to change like maybe 5% of the column name
i've been giving LLMs more and more to write, and with Gemini 3.1 Pro (Preview) as the LLM, copilot has been knocking them out of the park for anything i specified well.
the latest patch i asked for had 4 distinct parts (i had intended only to do one) but it has a checklist functionality that it used to keep track of what it was doing and what the plan was. and it asked useful questions.
and the code worked first (or at worst in the last few days second) try.
i've made maybe 10 changes in 1000 lines.
Some people are so socially inept that they copy and paste chat between LLM and their conversation partner. Like, I know a guy who does that irl to speak with his client
https://gigaio.com/products/fabrex-system-overview/
>Latency from system memory of one server to system memory of any other is less than 200ns
resources, such as accelerators including GPUs, DPUs,
TPUs, FPGAs and SoCs or storage devices, such as NVMe,
PCIe native storage, and other I/O resources connected to
compute nodes. In addition to device-to-node, FabreX is
unique in enabling node-to-node and device-to-device
communication all across the same high performance PCIe
memory fabric. FabreX can span multiple servers and
multiple racks to scale up single-server systems and scale
out multi-server systems, all unified via the FabreX
software.```
can someone loan me like a million dollars (give or take a couple milly)
that's the wrong direction of million :(
Rookie mistake using int instead of uint
i can loan you $1im
you did say give OR take
mfw too permissive as always
ALSO -$1im
a conjugate pair of millions
How much is $1j?
1 iMillion
roughly the same
1 jillion
i is confusing, we use j in this house
lil wayne is an engineer
"i AM confusing."
Jut je jusually juse jenglish jon jother jlace
jujutsu kaisen is good, i agree
j0seph j0estar
Believe it or not, I have not watched it
thats crazy
Cool project lmao
bbbbbasis vector
I don't remember who in vc did the hand recogniting thingy
have you ever seen a guy with 20 fingers, and were like, lemme eat one?
thats like 50% of the plot

this had some sort of telemetry on it that is telling me some account named cecilia shared it with me, idk if that's
or what but just advising
Is that base64url encoded?
$1€m
nrr
Path of least resistance
Also, $1m is 1 milli dollar
you can take the stuff after the ? off from the instagram link and it's gets rid of the tracking part. (all they know is you saw the content, not how you got there)
Ɛ
i was more concerned that it tells you who shared it with you so if you didnt want to show someone your own account when posting an ig link you'd have to remove it
$ echo MWl2enYyejZpbnZlOQ== | base64 -d | xxd
00000000: 3169 767a 7632 7a36 696e 7665 39 1ivzv2z6inve9
looks like it?
so idk if they just told me what their ig is basically by visiting the link lol
my favourite thing when i encountered dual numbers for the first time was defining e^ɛ as 1 + ɛ via intuition, and then proving that it was true from only this statement
TIL about base64 utils
used to be a hacker. ;]
yeah Instagram Share Hash is what it's called and it does leak your instagram to the person you share the link with
sheesh
actually it relies on another informal axiom, that you can reliably swap lim(x->0)f(x) with f(ɛ) to get a result in the same form (in terms of epsilon)
that's the igsh obv
What's the difference between informal axiom and a formal one? The suit?
well, not for e^ɛ itself, but for arbitrary e^(aɛ + b) where a, b, are real, its required
formal axiom would be defined in a way where i didnt think about it for 20 seconds
like you can move epsilon some way and it doesn't change it much, but it does change it usefully
for e^ɛ you can use the maclaurin series for exp and pull out the correct value
without any limits required
god i miss when i would've considered this a bad deal
you can prove it over an arbitrary input with the limit form for exp
same
havent looked at the chart of despair in a while. 64g ddr5
and then you can use it to prove the derivative of e^x is e^x in a much simpler way than the method from first principles
i havent used dual numbers but i saw a gjs talk where he mentioned modifying a lisp interpreter to automatically differentiate all expressions using dual numbers
good news! It's going up less fast!
it's holding steady at <total nightmare> price!
that's coz the entry fields for price must only accept 999.99 and then you have to hack it

(e^(x + ɛ) - e^x)/ɛ is the derivative of e^x via the first principle of calculus and the substitution of ɛ into the limit
is that central limit theorem?
No
like looking at anything which expands out with epsilon, i just 0 that and put what's left into like mathematica or wolfram alpha
it usually means there is an error when i see it, not an exploration for calculus reasons
Central limit theorem is about probability density function that converge to normal distribution
= (e^x)(e^ɛ - 1)/ɛ
e^ɛ = 1 + ɛ
= (e^x)(ɛ + 1 - 1)/ɛ
= e^x
if you look at the standard proof for the derivative of e^x, its a lot longer
well, in truth, you can do this same proof with limits too 
ɛ doesnt really enable anything except breaking division and multiplication on zero
but its a fun concept
and also useful for computer differentiation
take a look at forward mode differentiation on wikipedia
wrr
epsilon in the real world is because of rounding errors mostly. ignore it when calculating, don't when comparing.
double a,b;
...
if(a==b) { abort("this never happens. well, hardly ever really."); }
}```
stuff
i wrote a wrapper for python functions that does AD using dual numbers, and supports imaginary numbers
and hooked it up to some graphs on matplotlib
theyre pretty interesting

actually i wrote some desmos implementations of dual number functions too, but they were a bit crumbly because i had to use a sentinel value for ɛ
i derived a lot of results for functions of aɛ + b
including log_x(aɛ+b), log_(aɛ+b)(x), (aɛ+b)^x, x^(aɛ+b), and more
sentinel as in
desmos doesnt support symbolic computation so you had to pack two numbers into one? 
mhm
i replaced i with ɛ for convenient graphing
though, you can do operations directly on tuples of up to 3 numbers
naturally this means you can find the derivative of functions like ln(x) trivially
also of trig, inverse trig, and hyperbolic trig
i spent a long time trying to define the integral in terms of ɛ but never thought of a way, i couldnt be bothered googling it
ɛ plays nicely with complex numbers too
you just treat it like a real value
funnily, division by (a+bɛ) probably caused me the most trouble
since division by ɛ is only valid when the numerator is a multiple of ɛ (including zero)
oh ɛ can also be used to justify small angle approximations of trig functions
i.e. sin(x) where x is small, is equivalent to sin(kɛ), where k is real
then using any valid numeric form of sin, you can make a small angle approximation
wouldnt you just use this property
yes, i did
but
i had to derive all of these properties myself 
there are more ways to do it
but that one works
the problem is the different methods have different domains
fucking Epsilons

from brave new world?
the simplest one is the exponential form
The definition of an Epsilon in Brave New World is a member of the lowest caste who performs the most menial of tasks and lacks intellectual capacity.
sin(kɛ) = (e^kiɛ - e^-kiɛ)/2i
= (1+ kiɛ - (1 - kiɛ))/2i
= 2kiɛ/2i = kɛ
pretty much
holy math, someone tuned her to opencourseware lectures
nrr just trivial substitution
honestly im content of watching this going on
that looks like if algebra learned greek, so probably.
sin(x) = (e^ix - e^-ix)/2i
e^(a + bɛ)
= (e^a)(e^bɛ)
= (e^a)(1 + bɛ)
from these two
conjugate pair of imaginary millions.

fancy 
i havent done enough math to be used to using imaginary numbers for trig
good opportunity to derive from trig form of e^xi
via
j

can also be used to prove derivative of sin
Epsilon is always a wild one
its a variable representing number so small you could consider it as 0 but its more like something really really close to 0
just like a number that is technically infinity but isn't and just really "close to" infinity
honestly Epsilon and Infinity is similar in that way i guess in a sense that 1/x for x being a number really "close to" infinity is Epsilon
oh god, j has made it appearance once again in #programming
evacuate all braincells 
can be used for arbitrary derivative proofs of functions that can be defined in elementary algebraic functions
including ones that are written in limit, sum, or product notation
e^itheta = perfect.
þ
i learned cos and sin when i was 9 because it meant i could draw circles on the computer.
d/dx sin(x) = (sin(x) - sin(x + ɛ))/ɛ
((e^xi - e^-xi)/2i - (e^(x + ɛ)i - e^-(x + ɛ)i)/2i)/ɛ
= i dont want to keep typing this on my phone

but if you simplify and extract like terms its very easy
just a bit of crunching
= ((e^xi - e^-xi)/2i - (e^(x + ɛ)i - e^-(x + ɛ)i)/2i)/ɛ
= ((e^xi - e^-xi)/2i - ((e^xi)(1 + ɛ) - (e^-xi)(1 + ɛ)/2i)/ɛ
= ((e^xi - e^-xi)/2i - ((e^ix)(1 + ɛ) - (e^-ix)(1 + ɛ)/2i)/ɛ
= ((e^xi - e^-xi)(1 - (1 + ɛ))/2i)/ɛ
= -(e^xi - e^-xi)/2i
uuu
that's not quite right
phone maths
impossible to read
anyway derivative of sin is NOT -sin
it -cos
so me dentge
is it not +cos
uuh
yes

how did i find the derivative of cos using the derivative of sin
idot
this type of thing wants a pen and paper
with paper it takes no time and is much easier 
im a bit out of practice though
Brute force it like ye olden days mathematician did
Make a table of log
no copy paste on paper 
Carbon copy 
Easier to slip a carbon paper and trace since light table would need electricity/light source
That's cut and paste
no... there is a copy first
photoshop has paths which are vectors... i think (i have never used paths.)
i'm thinking Lady Away of Wrr probably likes math channels like "Another Roof" and "Mathologer" that do proofs of stuff.
Another Roof has made videos on "What IS a nuber As Explained by a Mathematician", "How to Count", "Defining Every Number Ever", "How To Add" .
then he did some combinatorics and recently did a 2h42m video on the new NIST encryption standards tournament process. (including describing how these post quantum crypto work).
quite good.
Patreon: https://www.patreon.com/anotherroof
Channel: https://www.youtube.com/c/AnotherRoof
Website: https://anotherroof.top
Previous videos!
What are numbers? https://youtu.be/dKtsjQtigag
How to count: https://youtu.be/QO9a7h87DbA
As previously explored, we can define the natural numbers in terms of sets. We can even use them to count. But ho...
and he just did one about using the determinant to count spanning trees.
(do not judge them based on how bad my skills are ;] )
That moment when you realize that you have gone beyond normal hoby level fine tuning and are precomputing teacher soft logits then slamming 3 Gpus with a 3x ddp qlora training run
What's really mind blowing is that even having a computer, let alone one with a gpu, is hobbyist stuff. The broad general public does not have that in thier lives at home.
Does anyone have data on the neuro lamp on how it works
tuya
if you just want to hook it up to something like home assistant you can use tuya integration
if you want to unofficially sync something to neuros lava lamp you wont really be able to do it unless you own a lava lamp and want to figure out how to extract the auth keys and use tuya sdk to get what you need
I'm trying to build a virtual clone
mhm you wont be able to reuse any of that
your best bet is creating some system for finding the lava lamp on the live video feed and extracting the color from the video
https://github.com/alyfrillnya/NeuroSync someone made this a while back, could be useful as inspiration 
ye can probably be used as is
it it's a hack which reads the screen from the video and gets the max color from a region and uses that as the color
it uses OpenCV to read the remote video URL and then parses the frames
maybe you could set up a fake server to steal the credentials? or have IoT devices suddenly learned how to check certificates properly in the last 2 years?
yeah. someone with the lamp would need to know python. slim chance.
it does it through your phone
if it's through your phone then you can route it to anywhere
just lie
or listen on the phone
i dont know what you mean by that
anyway as i mentioned it uses tuya
uses your lamp's hardware key to authorize with tuya servers
If someone had a spare lamp you could also splice the cables and read the input and upload it to a pubsub server
porting something to windows and ran it on wine
something doesnt seem right about that read throughput to me 
I found a message from 2023 where vedal is disagreeing with shiro's amazing take
vedal used to be capable of sending messages in discord
Magic
this is how you find out your drive is dying
i have the fastest 980 pro in the world apparently
can't wait for liquid cooled storage servers
Damm that's smart
me when i delete the code to make it faster
My idea is since it uses ble provision protocol to connect the the app and sends the API key over
I just wish to get the API key that way and act like one lamp
Like official app controllable
getting compilers to NOT delete benchmark code is a struggle in itself sometimes
You would still need the endpoints no?
This is why I ask if someone reverse engineered it before
Mostly all on GitHub is the tuya part
No one has the app provision part
many people have
its not really hard but you wont get anything from that
unless you have a lamp yourself
and have enough knowledge to extract its credentials
and are willing to reimplement the parts that use tuya sdk
That only covers the tuya
Guess I'm out of luck?
Neuro merch store doesn't have the lamp listed anymore
has anyone dumped the app?
ye
theres nothing other than tuya
What about the stream link?
I thought by default all lamp has a persistent wss to Vedal's server?
no
tuya servers
not vedal server
Then how vedal send the color on stream changes
he sends it to tuya servers
That simplify my job
But I still need to find the API key that vedal put in each lamp to associate it with each user on tuya right?
bzzzt bzzt
you need one.
zap
no, each lamp has a separate key
physical keys are not affected by quantum hacks.
False!
How hard is it to find someone willing on this server to record the whole ble interaction between the app and lamp and all the tuya commands and response
they're too big
mine is a lump of copper with some grooves and angles

you wont find anyone because you need lots of technical skills for that and if someone with that kind of skills wanted to do it theyd have done it already
i wanted to do it but they couldnt ship lava lamps to russia
wobbly
its a cool colored cat. phantom-cat of the opera.
website stupid
what website

shr
did you know butterflies drink blood
and sweat
and some butterflies have a wingspan greater than 20cm
larger than some birds

I like punycode domains

practically free
i also have a wingspan larger than some birds
pet butterfly
Do you know domains can contain non ABC and symbols
cyrillic is popular. but don't they look like "xn-ABCDEFG.something?"
xn--w8jw34jnha.dev 
They get punycoded but yeah
largest = 30cm
birdwing butterflies

what
30cm is the size of a ruler
terrifying
butterfly larger than your head
butterfly larger than a plate
butterfly larger than a small dog
it's a potential security issue when people see them decoded and think that the Н is a Latin H or whatever
xn--w8j0k5eo57knmh0oa887p8qwc.dev
idk about potential 
It's cheaper than my main lol
thankfully every usable browser will at the very least warn you
but there are small dogs larger than a small dog
and they're small.
Oh yeah I need to pay soon
Before they start moving my domain in June to that half suspended state that costs money to recover from
also: Lepidopterophobia 😮
there are also large dogs smaller than a small dog
for sufficently large values of small, sure.
and small values of large.
1+1=3 for sufficently large values of 1.
noice
I like to see things obviously semi-break.
just spamming Lalalalala in Greek makes it first refuse to actually transcribe and then complain.
Omg after 5 hours I have made my own AI vtuber. (not as good as nuero tho)
Almost perfect clickbait title
lol
Big pretzle **hates **her: 5 hours to build an AI vtuber!
That's one small dog
less than a footlong. so a hot dog?

rude
Also, new tsoding video just dropped
https://www.youtube.com/watch?v=gtk3RZHwJUA
Support: https://github.com/tsoding/donate
References:
- https://github.com/nothings/stb
- http://github.com/nakst/gf
- Source Code: https://gist.github.com/rexim/48d7087bfc8ba2c28a8b0b3aa3183558
Chapters:
- 00:00 - stb_ds.h
- 02:23 - Shadow Header
- 03:50 - Dynamic Arrays in C
- 04:37 - Initializing the Array
- 07:24 - Pushing into the Array
...
this is about storing data in front of the pointer isn't it 
okay yeah it is
useful 
Well he is the dynamic array guy
That video makes me think of unholy circular buffer by linked list 
https://github.com/google/cwisstable also reminds me I wanted to replace the stb hash table with this one for a project
Single-header C11 port of https://abseil.io SwissTable - google/cwisstable
So only one way

i'm sure it's a perfectly serviceable file manager.
with the advantage of version control
needed a new battery for my UPS. bought the knock off with 4.5 stars (1 star review complained it took forever to arrive and when he cancelled it they took forever to refund him. nothing about the product.)
but the top two reviews were soooo ChatGPT. em-dashes and one had green checkboxes.
Genuinely how did you manage to do that lol
probably clicked "open with" on a folder and accidentally said "make this the default"
The cooking never stops
have no clue
i figured something was horribly broken, nope vscode was file manager for literally everything on system, it's been like 2 days doing that kek

typically after you sneeze
also wtf
If there's someone around that is. And it's better be a priest
I'm an idiot and think I'm incredibly funny so generally I will reverse any given common social convention if I can get away with it, like for example I will usually say 'good evening' when greeting people at 8 am while at work, and then at 6pm I'm handing out 'good morning'
If I know the person and that it won't upset/offend them, that also means I will typically respond to a sneeze with something more like 'excuse you'. Extremely loud burps get 'bless you' from me
I just say “bless you” for everything from hiccuping to walking into doors
I think my friends think it’s a running joke atp
it’s not.
'god have mercy on your soul'
genuine concern
'good evening' is really good close to noon because it causes people to hallucinate like an LLM
"oh yeah it is like 2 already huh"
no it's 10 am

oopsies (that's like $4000 of RAM)
Fuuck
Looks like killer claude drones and mass domestic surveillance is back on the menu
kegseth got to 'im
they must've had a hell of a party
“We felt that it wouldn't actually help anyone for us to stop training AI models,”

great timing after seeing this earlier https://www.newscientist.com/article/2516885-ais-cant-stop-recommending-nuclear-strikes-in-war-game-simulations/
I doubt we'll putting it in charge of nukes any time soon if that calms your nerves
full speed ahead woo woo let's see where we end up boyos
Still a grim situation though
claude -p 'please review the national security situation' --dangerously-skip-permissions
real and accurate
but substitute i am alive with use the nukes
Too much Civ 5 in thhe training data
Yeah I'm not shitting bricks over that article either, still probably shouldn't put them in charge of nukes but not like that's a realistic prospect 
ghandi mode
do we honestly believe anything is unrealistic//impossible at this point especially if it sounds uber absurd
i dont think it's something that'll happen any time soon either but i think the main point is more of a preview of where the decision making abilities are at this time, lil sneak peek if it were to somehow occur and signal that even "aligned" models are likely to go that route if given ultimate free reign which isn't ideal
i dont like the implication that it would be trivial know if i'm using opera GX, a totally separate application, to facilitate this quest
True, autonomous killer Claude drones would have sounded absurd just last year
I still think anything nuclear is stretching it a bit even then though
I mean how do you think discord activity works
I'm confused, why does Discord randomly pop these up sometimes if I'm following the post (like I commented on it)
nothiing is impossible but part of the reason or maybe most of it was that the department of war was pushing them to remove safeguard
yeah i know what it is and how rpc works :| but i do not give it that permission and would have to opt in from both ends to do the quest myself
Fair enough
i guess opera GX specifically might already have a discord rpc functionality baked in since it's
uh
'gaming' browser
whatever the fuck that means
imagine giving somewhat probabilistic models such access to weaponry
the thing that would doom us all is the lack of knowledge from people in charge/power ig
i tortured my 14900HX hard
6.8GHz holy
also
not a godbin anyway
poor computer 
Programmers be like:
heh, somehow now i liked the 5060
oh hey this thing is somewhat perfect for my tiny briefcase
classic phishing mfs selling their domains for 1 dollar after use
does anybody know if this is too many individual pieces of context to send to neuro? I'm also wondering if it should be silent or not, it is dialog so it probably shouldnt be so neuro doesn't talk over it, but I also don't want her to get interrupted
fun fact: numbered .xyz domains, whether for registration or renewal, will always be cheap
lol, discord is also spyware, like any other client
just use browser web app, it at least isolated, unless there CVEs
.online domain
$2 -> $2000
?????????????????????
what? on 8+ (real) cores?
has unsloth updated for it yet
time to echo
arent pp.ua domain just free in general atm
it's like co.CENSORED (scam site now) way back in the day
might wanna remove that domain from ur message
i believe registration is around 1 doller
but i was told before that renewal is also 1 doller, not free xd
aint complaining tho
is it hobby to collect domains?
theyre so cheap if you want to park one why not ig
maybe
neuros.click
website put together by a real moron (me)
why? since when?
domain by me
frontend website by you
ftp and server backend by superbox
-# not even mentioning subdomains
absolute cinema
i just downloaded qt and mozila(firefox) latest several weeks ago
ftp has been out of ff since 2021 same as chrome
so it is just website wrapper? https://ftp.mozilla.org/pub/firefox/releases/
the subdomain is not indicative of the protocol in use in this case
the url is legacy
see how it says
https://
at the start of the url
if it were ftp it would just be ftp://
What about sftp 
not in browser either
i know, so it is just their side wrapper? i thought it resolve ftp silently
nah it just resolves https index like any webserver
regular directory listing, files served over http/https
ftp was the domain because it used to be ftp since mozilla is ancient
dunno but you can probably still connect up with like filezilla ftp as well but ive never tried
just cannot do it in a browser
without extensions
if the server actually sent ftp protocol traffic at your browser it would simply say it's not supported
and/or pass it off to ftp client if you have one configured to open ftp:// links
just one more domain
i wonder what is used for this ui
cuz it just the same as nginx (or apache? or both?) file navigation
both are displayed https:// as i remember
tho i thought it may be not server-side website, but firefox ui for underline data
tho actually simple http may make more sense..
webserver sends this
it's just a basic directory listing
simple as possible easy as possible to serve with as little bandwidth as possible
apache docs on its version
there was a standard ftp LIST protocol that kinda looked similar back in the day
idk who created this but almost nothing work
https://neuro.us.to/homexp/
also, i can just do this ||(she not die, she just became frozen in time)||
man, i decided to look what on server and that... kinda fucked up
||anyway, just wget on given url||
I mean, if it is a story rich game then it will need a ton of context anyway. As for silent or not, that's your choice when you want Neuro to comment on something or not
its a point and click so yea, story is important. Probably going to end up making it silent manually on specific pieces of dialog just to let her get a word in edgewise, though this is a cutscene so it matters less
are ftp servers online?
yes
make sure it put the username in properly
RandomNeuroStuff
no password
it should do that from the config but sometimes filezilla is stupid
or whatever client you're using
i just use fz and it does that
how do you sign cert on specific port?
idk why config not work, but with manual it warn about cert, seems partly successful?
it will warn about it yes
how do you sign for port, that weird
idk exactly what he has configured but i always figured it was a self signed cert
the port is just there cuz you're connecting to that port so that's the host url
most clients have an option to ignore the cert from the start
even cli ones
like this for example
i hate my isp and rkn
it is literally blocked by filters
i cant connect without tun proxy
kek
i cant connect to my main vpn server
current free proxy configs i use (nekoray), are slow as fuck
it also might be mad because it's forcing tls explicit so when it connects it immediately upgrades to tls and wants a cert
the cert
i start to thinking that they drop anything except 22 and 443 ports
i have actual server (+git) on port 443
ill try setting reality to 443, and maybe it will work now
yeah im aware
it should not connect to 443 on neuros anyways, he only allows connection on 21 which is nonstandard for ftp tls
i mean my proxy not work under this shitty isp and DPI
i cant normally download anything with 200kb/s
only text like discord
id pass away
you are welcome
i hate them all
i want to leave this country
charaverk | ~/Public ~$ ssh charaverk@charaverk.online
charaverk | ~ ~$ sudo tail -f /root/xray.log
2026/02/26 07:22:20.008340 [Info] transport/internet/tcp: listening TCP on 0.0.0.0:443
2026/02/26 07:22:20.008343 [Warning] core: Xray 26.2.6 started
2026/02/26 07:23:42.277478 [Info] transport/internet/tcp: REALITY: processed invalid connection from asd:54434: failed to read client hello
``` no use even 443, they do scan traffic and replace/abort it in way reality fails
lol i was about to mention
region ip just leaked
asked claude if there were any am4 apus with pcie-gen4 (they all have pcie gen3 only)
i think it has me figured out 
and no there were not which checks out
and is annoying
I'm a bit confused, every x570 motherboard supports pcie gen 4
you had the same confusion i did
that is true
but
turns out
all APUs (having integrated GPU) supported only pci e gen3
even with x570 boards
for ryzen
okay but you can just not use an apu
8000 series apus supported it
i want the APU to run with no graphics card
I thought you asked Claude about CPUs
technically i did
i just wanted to see if he'd pull some random special one off 5700gx or something that had pci-e gen4
because yeah the reason i looked into it was because 5700g: pci-e gen 3 max
||i changed ip of server just to be able to use it as website
and what now? it just not works without vpn
they literally blocked Amsterdam's ips||
I assume you already have a 5700g then?
internally ig the apus used some pci-e bandwidth on chip
yea
it was hella cheap tho
so oh well
i only need the gpu for a moment anyways
well technically not at all
just lazy
the thing will be headless mostly anyways
just wanted to not have to swap card
that's what you get for not buying a strix halo machine 
i would get one if it wasn't a billion dollars for a mini with the ai max 395+ (which is the same chip)
comparatively
would like a socketed one
why is this cpu have AI in name?
npu shit
you saying that they actually selling consumer NPU?
NPU in cpu have been around a while
i havent read anything about new hardware for 1+ year now
npu have been getting jammed into mobile stuff for what has to be 3+ years
they have just never been good for anything
still really arent
because that makes it sell better
the best part about the ai max 370-395+ line are that they are APUs so they have a gpu but not just shitty mobile gpu, the 8060s is an iGPU but its performance is around 4060 to low 4070 level
lately figured out that igpu are actually good
i play civ6 rn on intel11gen, on pentium gold without real issues
only issue is memory, but that like windows being shit topic
early igpu was only for videos or 2d, but even 2020 can run 2016 game with stable 30 fps on HD
their apus can also use up to 96gb (shares with system ram)
so the apu can also run bigger models locally and do some training assuming rocm plays along (in my experience so far, it's not happy to)
even the younger sibling to it i have in my mini that's not nearly as fast (still pretty wild for an igpu to me) is able to report and use 40gb of vram
it's not as fast as vram
I hope they make a rubin version of the gb10, given how much faster the bandwidth between cpu and gpu is there 
at least strix halo managed to have a better mem controller
if only its x86
or it was just the cpu ram I don't remember
I guess double the interconnect speed doesn't hurt
officially supported too
its insane
give it to me on my 9000 series amd 
instead it's a miracle on ice that i can do 4 slots 6000
lucky bastard yeah
256 bit of ddr5 instead of the usual dual channel (128 bit)
looks like it's a modest 33% increase in bandwidth
official specs at 273GB/s bandwidth
the shitty 4060 bandwidth is 288GB/s for reference
im mad that my 9700 loses to my 3080 on mem bandwidth lmfao
I keep forgetting to halve every bandwidth number from nvidia in my head 
point and laugh at this user

its ok titan v bandwidth also loses to a 3080
3080 listed first >_>
guys, do you need that speed? why im fine with 15w cpu and 8gb ram
real world so little lower than spec in this case
what cpu
this would cause me to be hospitalized
pentium gold 7505 if i remember right
and how much did you paid for it
for laptop? $300
bro got scammed
Its 16GB
Now its been a while since I cracked open a textbook but I'm pretyy ssure bazookas don't fit in PC cases
lul
it is russia
it even considered well price
KEKW
the third world country build
bet
I could afford a better PC. But not with current pricing 
there's a lot of older xeons and such that sell for way cheaper than you might think and perform way better than you might think
god i remembered paying Rp 20 million to build a low end pc in 2021
20 million kek
The problem is not the CPU. I need GPU. Games have got ridiculously unoptimized even on 1080p
i7 10700, 2x8gb ddr4-2400, gtx 1650, 500gb sata ssd + 1tb hdd
i love hyperinflated currency
about $1300
20 mill for that spec?
That's scam level shit
used for 5 years or something btw
755 (or 775?) socket
regret building during that time
i paid so much fucking money for my 3080 in the shortage i cant
oop didnt mean to reply on that
but i regret it too

i dont even remember exactly how much it ended up costing because i bought a prebuilt because i couldnt find one and was impatient
and that thing was like
I was thinking at the peak point of gpu shortage to sell my gpu for higher price than I purchased it
$4300
Thank god I didn't do that
god fuck
considering that i can build a pc similar to my 2021 pc specs for just
$69
and i already did exactly that last year
actually
and he saying that laptop for 300 is scam
even better specs
yours before the shortage mf
Help me build a PC in Java man 
3700x, 2x8gb ddr4, gtx 1660 super
erm, i bought it 10 month ago or something
775 came out in 2004 for the record (pentium 4)
but pentium gold 7505 is a bga chip from 2020, it is not socket 775
i used xeon for 5 years, on something something should be 775 or first 10**, <- so i remember like that
i havent uncovered motherboard
i guess it integrated, not replaceable cpu for socket
yeah now you can get 3647 and 4677 xeons for very cheap rather than going back all the way to 775 era stuff
sockets 3647/4677 that is
and yes teh pentium gold if you had to name its "socket" (it's soldered to the board) it'd be FCBGA1449
....even fucking worse
idk what is prices in your area, but when i was searching for it was even low
some alts (not used) was 55k rub, while i got fully functional laptop for 24k, and i genuinely glad i tried laptop
looked for low power long time mobile working machine, just for coding and text/youtube
got machine that can run civ6
and on linux system everything loads instantly, also it is first nvme for me, feels faster than sata ssd (but might be just feelings)
now i not want to have stationary pc. and next time i will look for similar size (by size it literally magazine) and performance/power consumption

sanctions go brrr 

I'm so grateful I live in a country that isn't sanctioned or imposing heavy tariffs
That shit really stings
I let my monitor discharge its electricity(I did press the power button few times) for a day, still cooked

it really fucking looks like my gpu got cooked
how can a monitor issue look like this
have you tried a different cable yet
yes
unlucky
different HDMI cable and another DP one
tried a different OS?
I tried to boot on a clean installation of Win10 at least
it happens in the monitor's logo phase, so probably not anything about my pc but just the monitor

ui itself seems like ai markup output
i tried to use jippity it says my current quota was exceeded after 1 second
it only sent integration command list its not even that long
you get up to $100 worth for free per month is how I'm reading it
does it not work under vpn
could be blocking vpns for the free tier
what does it say about "allowed geography"?
says switzerland is allowed
and im there rn
oh wait you mean check if it says "vpn = bad fuck you"
Accessing or offering access to our services outside of the countries and territories listed below may result in your account being blocked or suspended. monka
if it's ip-based not blocking known vpn IPs would kind of ruin it for them
Checks out
ok no jippity for me
nice
wtf you mean nice i need it to test my integration 
whatever i still have things to do by hand
try the free openrouter models if you just need to test with an LLM
the jippity wants openai key
there is also Gary which seems to work with local models? maybe i can download some llama
ignoring AI finally hit me
i have no idea what im doing
does openrouter have free plans
openrouter doesn't have plans but they have free models
what was the message you ended up getting
info: WebSocketServer listening on port 8000
info: New WebSocket connection; there are now 1 connections
info: Set game to "Skyrim" and cleared all registered actions
error: Error calling OpenAI API with request ID SSDAKDASKDHAKSHDKAHDKAHSKDSH -> 429 You exceeded your current quota, please check your plan and billing details. For more information on this error, read the docs: https://platform.openai.com/docs/guides/error-codes/api-errors.
info: Main function completed
Yep looks pretty cooked
you ran out of budget 

When you realize OpenAI is not open or free
I've used that, it works but is quite strict on the limits
free is free, 70 billion parameters
70b means 2x 70gb
why 429? never used services like this but http have 402(right?) payment required
ok
i need something smaller
it 100% 40x
It depends on quant but APIs usually have FP16
maybe its "you are using known vpn provider"
You don't need to download the big model by the way, it's an API
429 Too Many Requests
OpenAI compatible API even
429 means limit exceeded of some kind so
ok i'll try set it up with gary
or wait can i use jippity? even though its not openai?
I guess technically "too many requests than your budget allows"
is the logic here
how do these tokens work
Depends if it supports setting a custom endpoint and model value
429 Too Many Requests
The user has sent too many requests in a given amount of time (rate limiting).
it shuld by now
perhaps you've already used all your openai credits for today
i literally registered on openai 5 min before that
.
I think OpenAI API from OpenAI just isn't free
maybe they need my credit card anyway
like, "its free but we need your card anyway"
so they can steal some money if i forget about it or die
Either way you can always use either a free API like you already found or just llama.cpp
Depends on the task but usually you can find a competent enough model
👍
429 implies it may work again after some amount of time
402 is basically only used by cloudflare for their scraping payment system
maybe i need to wait 2-3 days until it refreshes quota
no you need to pay for tokens
it's a prepay system
429 Too Many Requests
The user has sent too many requests in a given amount of time (rate limiting).
It's usually "you asked for too much too fast so we're slowing you down". but can also mean "you don't hav any credits left this month"
if it was coz of location it should be:
451 Unavailable For Legal Reasons
The user agent requested a resource that cannot legally be provided, such as a web page censored by a government.








