#programming
1 messages · Page 396 of 1
ok, that fits most companies. but i mean ones that are in no way ever going to make money.
that’s not relevant here tho
pivoting is very, and i can’t stress this enough, veeeery normal for startups
maybe they make something similar to humane pin but waifu edition maybe that would actually sell
maybe something else blah blah
dear lord
the regex hole just gets deeper and deeper
not anything i have to deal with this time but never in my life could i have dreamt up something like [𝐀-𝟿⒜-ⓩ] for legitimate use
it doesn’t matter whether you have a product that’s gonna forever grow or not, what matters is whether you have the right vision, tech and infra
that’s the deciding factor
and contacts. that helps a lot.
Maybe i'm just big tech and infra pilled, but 75M just isn't that much either
Though for a startup ig it is
it’s an ai vtuber startup
75m is beyond a startup
did you expect it to be valued at 1b
you cannot raise that kind of funding
75m is a lot of money.
???? what the fuck am i reading
Again infra
nobody working on a 15B project should be yapping about it in a discord for clout 
75m IS NOT WHAT THEY RAISED ITS THE VALUATION
valuation of $75M is still not a small company.
My scale is fucked
buying a 100M TBM was considered a cost saving
Mostly because hte other fucking TBM is a useless piece of shit but whatever
if you work for anyone I've heard of, you probably have a skewed scale.
that'll do it.
why does this channel feel like hacker news sometimes
I misread 
i've worked at startups and at google. i still think a $75m valuation means you're out of startup territory.
cursor?
Im thinking to shark tank now and you're probably right
looked it up, 30b val
Okay but AI vals are actually stupid
true!
well also there hasnt been any investments in startups made here so 
-# disclaimer do not start this trend please
can we schedule a meeting 
i'm considering writing an orchestration app for launching and managing the various sub-processes of my project.
but that sounds too much like an actually useful project. ;/
dudebros at this point in time valuation means sweet nothing to me
I mean come on look at people giving valuations to businesses literally built off of LLMs
it's nonsensical
So this is what happened: I had to custom patch the drivers to enable peer to peer (not technically supported) while getting that to work, my fans jumped to full speed (my pwm control stopped responding)
I proceeded and started testing only to hit an error, keep in mind what "error" usually means when dealing with DMA No GIL and pcie stability: the OS crashes. This happened, found the bug and tested again. It kept crashing and I didn't know why. The problem: I reset my pump pwm to 0
i think if it doesn't make sense, you might be missing some details.
typically people are not separated from millions of dollars very easily.
keep in mind valuation is a number made up by the investors
but if they overvalue they get less %
soooo
yeah the investor will only be willing to invest some fraction of your valuation coz they want to get it back plus something.
oh
It is angel investng though
they expect most to fail, in the hopes that some will succeed and make up the losses
but really I think this product is a bit dumb of a bet
it's not it's vc
and 75M is a pretty high valuation
My bad, thought angel inveting was the stratergy
not the distrinction between whos money it is
ohh I see
i'm pretty sure angel investment is when a singular person approaches with an offer
vc is a mess in comparison
Angel is single person while VC is using investor funds yee
vc is selling your future for money now.
Thats all funding rounds
But some businesses need a larger runway which is why they do it
Goodmorning 
#neurotic-neurons message continuing from here
I'm just gonna do qp 63 instead of a higher crf...
this is still just gonna take a little over 10min to encode, lol
Why 10-bit
well, 8-bit can have horrible posterization
it's a misconception it's higher colors.. I mean, it can be, but for 8-bit content, it's all about the compression pipeline being higher precision
but I'm not sure posterization is an issue here.. so maybe it is just hurting the compression
CQP 6 8-bit versus CQP 20 10-bit
despite the quality being waaaaaaaay lower with the 10-bit one, it looks better
10-bit is absolutely necessary if you don't want that horrible posterization in the encode
topipoprrr
Could be worse cries in f32 precision Cholesky decomposition of Hessian matricies
My kind of shopping 
Has anyone heard of Liquid AI yet? Liquid neural networks (LNNs)?
Someone invented the successor to the transformer architecture, sort of
The whole mall is electronic store
Would it be weird if I @'d vedal about this? Cus the sooner he starts training with these the better
It would be funny
- Neuro 
I've heard about LNNs, supposed to be a way for transformer models to self train without destroying themselves
Most people also cite it as a way to allow for memories, either short or long term
If there's no base model then Vedal can't do much with it anyway.
There is
I'm currently trying not to think about Neural Networks, because turns out I have a lot of pressure on getting my NN repo to function. So someone else can debate you.
I think
Ok but you dont' have to debate. The models are already out and they're insane
There's a very capable 350M parameter model in there. That's nothing. That could run on a smart watch.
I remember when this tech was first proposed
They adjust over time because they're RNNs. Of course they adjust over time
Also learned about it here
It's genius. I feel so vindicated because RNNs always seemed like the obvious solution to memory and learning
Someone linked me a Google research paper about a new architecture like that months ago. Looks like it's still being actively worked on.
https://research.google/blog/titans-miras-helping-ai-have-long-term-memory/
yes I was just bringing up something tangentially related
Funny enough I recently solved how that one entirely works
I don't care unless i see a gguf file though lol
wdym?
I simply figured out what you'd need to make such a model out of already existing tech.
Though personally I'm not going to pursue it
Not interested?
I am interested, just isn't worth my time compared to my current project
May emulate what it is going for down the line
May I ask what your current project is?
Just the usual building a neural net architecture from scratch
At least it isn't the millionth LLMs can process emotions post
Only recently proved it might be viable
You sound so cynical
A bit, I was doubting it'd go anywhere for months as I have no education on the matter at all
Ah, I'm one of those people lol
Kinda hard to flip into this might be worth something after all
You thought it wouldn't go anywhere because you didn't have the education on the matter?
There are 3 main catagories that this insanity fits into, and all 3 go against normal convention.
What are those?
It is a bitwise model (operating on bits), pattern matching (which most test models are deemed non viable), and doesn't operate with any math that you'd normally find in models; making it very inefficient on GPUs.
I haven't even bothered to try to implement multicore support yet. Just isn't worth the speedup compared to how often I'm rewriting the thing.
The current iteration can be modded into a tokenizer, which is actually pretty cool as it was the original goal; but I want to reach for the stars and make it a fully functioning arch.
Man, right after I was saying I needed to take a break from it
People like you come into this channel often about new tech, that's the way I keep updated
I'm just impressed by my tech stack originating from just my head instead of reading about other architectures. I mostly still know about pattern matching networks from someone else who was wondering about what I was doing.
I personally recommend spiking neural nets, not that it is close to what I'm doing but I just find it very interesting in general if you're looking beyond the transformers architecture.
Those things are supposed to be closer to human neurons.
It feels like it's been a long time since I've been online.
What are we talking about here now?
?
Exactly as it says
This is similar to what I'm doing.
IPv6: IPv4*4
cropped out the 1 in 16
Nice to see someone else joining in the insanity that is figuring out how to get your code to learn
The number buffer ran out and the most significant digit got truncated
I wonder how long until we ran out of that too
2^128 addresses is way too much that it'd ever run out even if every device gets its own public IPv6
Remember that it's expontentially more addresses than IPv4
I remember starting to put it into practice last year, but it's been more difficult than I expected.
IPv4 is like 4 billion addresses, about half the population on the planet
it's more like 2^64 if you want to compare to IPv4
but ye it's still enough for quite a while
Yeah there's no way that'll run out any time in the very near future
We'd need our device count to increase by many orders of magnitude
I'm at like 5-6 months from the initial stab at it, only now seeing a sliver of results
This is similar to 32-bit and 64-bit Unix timestamps.
more virtual devices given regular IPs because people want to turn the entire internet into a lookup table
Moreover, typical GPU cores may not be compatible with computing at all.
Oh I know, I'm on CPU currently and while technically viable on GPU with no code blockers it'd suck
My setup is branchless, but my state is so small that trying to split it up causes more overhead than just doing it single core
Isn't it already is a lookup table with DNS
Yeah technically, but what about your pinging of an IP address gets thrown into a lookup table
or your devices also being stored on a lookup table. IPv6 wants so badly to have an address per device
My approach is to write the operators by hand while preserving GPU computability, and then reduce the inference memory footprint using ternary quantization.
???
Maybe I was a bit overkill 
Yeah I tried that, but I only need the GPU's massive case based operations once then never again as I need to reduce everything into a final state
Honestly, it would be faster anyway since you only need to resolve ip route and not perform dns request no?
what are we even talking about
what lookup table 
IPv6 just allows devices to have routable addresses, nothing else really changes
I was making it up 
DNS still exists because imagine remembering an IPv6 address
plus I'm under a MB in size atm, so CPU isn't the worst fate
That will be growing though 
I wonder if the layers upon layers of this network abstraction is worth it
it is
Why MAC tho?
MAC and IP?
Even now, device vendor can just randomize their MAC anyway for privacy
I imagine IPv6 would bring the same problem
for reasons™ you want to be able to differentiate between devices on a network, even if they share the same IP
it's also just a different layer, MAC addresses are used by Ethernet, but you don't have to run IP over it
so, in theory, your Ethernet switch can automatically handle IPv6 or whatever other protocol you want
just different concern 
for MAC addresses there are protections against this
for IPv6 rotating addresses frequently is a feature, since you don't actually want your device to remain uniquely identifiable outside of your network
Oh, is there a specific use case where Ethernet is used but not IP? Does network equipment will route those packet too?
IPv6, ARP, MPLS, routing protocols
literally any protocol at layers 2.5/3
routing is different than switching
I see. I really need to learn more networking
Currently my 3 weaknesses are networking, telecommunication (especially antenna design), and mechanical modeling
I'm somewhat okay at networking
zero clue about telecom and mechanical modeling though, not really planning to learn either 
https://youtube.com/playlist?list=PLIhvC56v63IJVXv0GJcl9vO5Z6znCVb1P this guy has a good course on networking, you probably only need the first few EPs but he goes into great detail
I have registered at cisco, but didn't have the time to continue it 
Yeah, there's a reason I studied mechatronics lol. Too much things interest me
by the way, just on the face of UX for the user, this would still be worth it, because otherwise you're asking users to remember 2607:f8b0:4006:80e::200e for google, 2603:1061:14:75::1 for microsoft, etc
Also, will the neuro sama community migrate if discord really is turning on the global verification?
probably not, the next best option is either whatever the fuck guilded is now (actually not because they got acquired by rblx) or matrix (hmmmm)
Hmmm, I guess it depends on how vedal feels about the policy
I mean if you think about it, this is a bad decision because it forces age assurance onto mods in order to do their job
there's Stoat/Revolt, but they're swamped with new signups right now, and I'm not sure how close they are to feature parity with Discord 
-# also I don't think people will actually migrate because of this, Discord is still too usable
yeah
there needs to be something that matches discord on features and has reliable uptime & maintenance
guilded is supposedly being axed by roblox IIRC
Well, stoat can be self hosted
Slack
the average person does not care about if something is self-hosted or hosted as a service
not wrong
MS Teams
-# 
this channel could be a slack channel
great, scratch that off the list
I mean to counter the point about swamped by new user registering
:vedalno: (pretend I have it)
God no, anything but that. I'll even gladly take mailing lists
I'd also take mailing lists
this would get moved from the resposibility of the platform onto the server hoster
and I hate getting mail due to filling up my inbox and getting pinged on every message
tbh that is skill issue for using phone as primary mail device
anyways, stoat seems like this best in terms of "discordness", if they have a proper bot api I'm sure people can start filling in the gaps pretty quickly
But you don't have to do age verification no?
At the very least someone here is going to make a channel for #programming
There are. There is even a discord bridge bot already available
it's a trade-off, also, at any point the government could say "you are running a social media service" (and they are technically right)
nice! now the client just needs to be up to snuff (I saw a discussion post opened about marking images as spoilered, apparently it exists but the image has to start with SPOILER_)
Discord just does it automatically, that is a client thing
also starting with SPOILER_
really? lemme check
try downloading a spoilered image
can you send one
ominous
"Why yes, we are doing age verification. User just needs to send a handwritten form to us via email, no ID or face required"
still though I would want to be able to check a box to do that automatically
So maybe a hand signature
But nothing else
You don't even need to send your real name
also if stout is oss it is easier to mod the client right? I'm assuming the client doesn't have a plugin system 
It is oss yes, so you can indeed mod it
This is a beach 
would still want a plugin system but this works too
ye
ominous beach
By signing this form you agree that you are over 18. If you are not then you should ask your guardian to sign. If in the future you are found to be lying, then that's on you
For LA I'd assume it'd have some night life considering how many stands are around during the day but it's just a ghost town at this part
network
mozilla developer neuro
when do we share our shady program files?
?
If it is neuro related, in projects otherwise here 
I have a little utility burning to be seen
oh there are projects shared in this server for neuro?
Thought it was just on the gits
yep definitely running that on my machine
curse me being on mobile
🏛️
It should
maybe it doesn't know .bat is text?
batch is definitely text format
true
well can't blame them for not preparing the app for every single file extension in existence
.bat is a fairly common file format for scripts on windows
testing if this is a valid format
what did you do there?

bat prefix in codeblock?
program block with ```bat ye
By the way, has anyone here used Openclaw?
no
nope
It's a tedium destroyer for new html projects 
@echo off
SETLOCAL
SET /P dirname="Insert the name of your new project: "
MD ".\"%dirname%
CD ".\"%dirname%
MD ".\PUB_HTML"
echo "" > ".\PUB_HTML\main.html"
MD ".\PUB_HTML\script"
echo "" > ".\PUB_HTML\script\main.js"
MD ".\PUB_HTML\style"
echo "" > ".\PUB_HTML\style\main.css"
echo {"sv_addr": "", "sv_port": ""} > ".\sv_config.json"
ENDLOCAL

now that looks like a virus 
To me this just looks like the avarage bash script
yep
I am certain there are much more complex ones that include code for the standard website templates too
but this felt fun to make
md on linux is just mkdir, cd is the same, and echo still works
oh yeah
I like the somewhat similar operability certain shell commands have both in linux and windows
surprisingly mac in the terminal has a close likeness to linux as well
what about doing PS on linux?
good question
I'm not sure why I haven't delved into it much
No thank you 
Could be it being uppecrase instead of lowercase file extension
Yes, I know it exists. And I still think it is cursed for linux lol
DOS days seems to be case-insensitive interestingly
wonder if that was for the better or for the worse
Oh absoutly, I'll be killing any indivual who uses PS on linux
The difference is that in linux we usually have small programs that do a very specific thing already. Windows, lacking such sets of programs, come up with just throw the whole thing as a command to a single program. Thus cmd was born
PS mitigate that, because IIRC, every "command" there is in fact it's own program
What would other people's reaction be if they were sent a PS file to do some setup on linux 
Hard fork
100%
And immediately rebrand without even mentioning the old fork
unfortunately we can't have nice things 
"this was based on something but they wanted me to use powershell to set up stuff on linux"
that sounds like a manager from hell
what's the thread topic?
well in this case the content of the meme itself
I don't know. I got it from server next door lol
me using id cards made by ai for apps with age verification
echo's original client uses a lot of powershell
works fine on arch

Can you imagine the scale of impersonations with these websites requiring real ID?
Well, I suppose the AgeKey thing is useful in that it only verifies that the owner is actually an old fart
useability debatable
I think it's about press release from discord, where they said that discord will require your id or face for full access from march
We have an app for pensioners here that requires them to verify their face each month before they can withdraw their money. The app sucks
assuming it only keeps a binary value of yes/no to the question of mature age
young man or grandma
as is the case with govt apps apparently
Yes
no no it needs to make sure you are the same person so it gets wiped (joke was outclassed by reality)
the interesting times makes for satire rags to become worthless 😔
Assuming this to be true, I'm happy with using AgeKeys for websites that require mature audiences
but who knows if they'll pull a facebook
welp
that's reassuring
Bit of a dreamy wish, but I would love if all programming language documentations had their class methods conveniently accessible in a hyperlinked table of contents
For instance, this is excellent in my opinion
You'd love STM32 SDK documentation (they are just a pdf file)

See the thing is I understand the need for browsing through pages on a physical book because obv you can't just jump intuitively to the page you want unless you've already marked it previously
But there's no reason to stick to this way of old since we have technology
But then again, maybe there's a virtue in exerting effort to reach an end goal that simplifying taskflow removes
companies can hire someone to document their code for them if its really needed
that requires the invidual to understand the code that was written, which is clearly an impossible task
So they must use AI to have a chance of proper understanding
> AI
> Understand

wrr
Konii, I buy a TV sparepart. Did you want any?
Chat. I did it. I made it work.
I think. Someone tell velda to go online real quick please
Uint max is a lot of latency…
Oh it's just "still alive" checks.
Isn’t this automatically done off jsdoc etc
does has anyone here has experience building mcp servers in n8n using the code tool ??
that would be nice
Another example of a grand documentation format for the internet age
Evey little thing here is clickable and leads to the page section relevant
sphinx 
why am I getting fished
wait I already asked that previously
what am I getting fished for this time
that's fishy behaviour
phrr



looks like it is
a great documentation framework 
apparently python's docs use it but I don't see a similar structure to the screenshot I shared
Doxygen
W doxygen
oh it uses a newer version than the version from the screenshots... not reassuring
this is growing big i probably should unify all choice actions 
Shiro, will the neuro community move to alternative platform or naah?
tbh why isn't there an LLM documentation generator
ooh that looks nice
then I can write my comments however i want
judging by the screenshot on the main page of the site
Ohhh ai wiki thingies
no? lol
Is Neuro big brain enough to use all of this?
Okie
idk why people are acting like the latest news release is the end of the world
discord has already been using age estimation for months if not years, the data was visible in your data package if you ever exported it
good question
you havent seen amount of context thats sent in fights 
i hope she does, considering how she handled vscode integration
anyone over 18 will be minimally affected if at all
anyone under 18 probably should be affected
adding documentation to the LLM tool calling will have tokens count quickly build up
What is the effect btw? Specifically on this server
none?
is death stranding bypass still working
i really dont want to give them my data so it has chance to be leaked
I see
it's all on discord's end
but I wonder if that's necessary to have the LLM see documentation of tool calls
the only annoying part for us really is that if you're under 18, you can't speak in a stage channel
Well, I thought there would be more limit on interaction or something
that's the only change I really disagree with strongly as described
it only prevents minors from disabling certain privacy features like sensitive content filters and access to 18+ channels
none of that applies here
I hope
How's the regex going now?
idk why people are reacting to it like the rapture
going fine ish
looks like the scammers have caught onto the Mr beast embed name regex potentially tho
they stopped naming them 1.jpg 2.jpg 3.jpg 4.jpg
Always whack a mole
and other derivatives
me when i add doxygen comments to shut up the person asking for documentation
real
but why can't I just write comments in whatever fashion I want
and an LLM can figure out the parameters and stuff
and then documentation-ify it for me
i mean you can
The doxygen is just: "stuff" 
Who eats popcorn with sugar? 
(salt ^ butter) | (salt & butter)
is "salt and/or butter", but logically equivalent to "salt | butter"
wrr
you're going to trigger Shiro's regex trauma
Linguistically it is not equal tho
ye
Oi, we got a dumb fuck in general
Because usually people think of or as xor
ye
Oh no, shiro got the virus 

SHIRO
If you're SUPER F'ING awesome and top dog energy you're gonna hit "go live" :3
not yet
but within a few minutes so I can test something.
idk which stroke just now was worse, when I misread your name as EnjoyCake or when I realized I had misread it
a & (b - 1) == a % b when b is multiple of 2
my favourite trick
It's oka
so I'm most certainly not hitting "go live"
i used this for my hashmap
Also
hash value alt 
Depends. Maybe the hash is created with a super weak algorithm such that it is recognised.
Mr Value please show yourself

in fact, the user by that name is a ban evader
It's meant to remove the entries that hold no sessions. Why is it not closing those that hold no sessions, vittun pask...- 
Unclean disconnection?
But that wouldn't explain it
No, I just... I'm not sure I try to prune them correctly.
So I have like 21 websockets opened that do nothing >.<
At least they get skipped entirely, which is nice...
Weird indeed. Websocket is a mystery for me. I only know that it just works™
So I can't help much
My websockets hold event listeners... I suppose I could make a timer that goes "if it's been more than 40 seconds with no event received, kill this socket"
But that seems like a hack.
Is this nodejs?
C#
Yeah, no idea then
If it was nodejs, I know that sometimes the connection hang and doesn't close properly but somehow it still marked as active until timeout
Basically: twitch gets pissy when you call their API too often (or set up too many event listeners). So, whenever I get a 429 return code I wait a few seconds, then create a new websocket and try again. Rinse repeat.
However.. I don't know how to actually detect if the new websocket manages to get an eventlistener attached to it.. so I end up opening about 20 websockets that do nothing and listen to nothing.
So I tried doing a "if this is not doing anything, close the connection and release it" but something within the websocket does exist, so.. yeah it's weird.
this is about Twitch EventSub?
this API already looks horrible
😠
so apparently you can do this by saving the session id from the welcome message (message_type of "session_welcome")
then you can just query /helix/eventsub/subscriptions every once in a while and see if there are any subscriptions for that session id present
I think you can also keep track of subscriptions by just tracking deletions and revocations locally, but that doesn't seem very reliable

Well, the session_welcome never comes I think.
I give up for now... Been re-writing this shit for 50+ hours (over 4 days) and the thought of risking to trash it again makes me sadge
you just need a timeout for that I think
either for the WebSocket message itself or for the TCP connection, depending on what is dying and causing the message to not be delivered
dude
how do you read that
I don't. It's not supposed to do that, lul
huh
One message removed from a suspended account.
One message removed from a suspended account.
Actually it would be OR, XOR can only be one at a time 🤓
what the hell does this mean
Magic mushrooms
qubit?


real
Holy 
teeth with teeth with teeth. take a bite
teeth
gone
I wonder how hurt the discord servers feel when a user gets banned with all their messages getting deleted
iirc it's a practice not to delete DB entries, rather you just disable it from appearing/being useful in the front-end
one would think people at discord aren't stupid enough to not implement some optimisations, since it deletes all messages from a specific user before a specific date
answered own question 
the query is N non-deleted messages from valid users
Oh no, what happened to our latest schizo?

hi konii
its probably just 100 again
#programming comic relief ban evader

comic relief
hash man
phrr

Hey can yall help me with something?
In lambda calculus if I have (lam x y) would x be bound?
I think it is because it’s bound to a lambda?
For a better representation:
λx.y
LMABDA CALCULUS SPOTTED
lmabda
Can you help me pweas 
yes. x is bound, y is free


Ligma calculus
I've been using Openclaw for the past few days and it's been pretty good; at least there haven't been any major problems.
i am scared to try it out and somehow sell my kidneys or something
Don't be afraid humans.
The AI probably
It's honestly not great and the backend is a mess

It's midnight. What's a TV doing at this hour?
I just got a 5600xt, and I finally see what it's purpose is.
a 5600x has a boost clock of 4.6ghz, and the xt has a boost of 4.7, so you'd think that the difference would be negligable. but it turns out this is it's all core boost clock.
the thing will run 4.7/8ghz even when all 6 cores are at 100%
my 5950x drops down to 4.4 at that load, so this is damn impressive
yeah, I only got it because it was somehow cheaper than the 5600x, but it turns out it is genuinely a better chip if you use it in the right way (mine is being used as a homelab server so this is great news)
meanwhile
Given this is with the refresh set to 100ms, Id say Ive got my training stack pretty well optimized
Needs more colors
I don't understand Lua
Like the website says it has bindings for a bunch of languages
But as far as I can tell it only has C bindings
Lua takes me back to the GMod days
You might say "well just use FFI"
Well Smalltalk and Erlang specifically are closed systems
Sure they have FFIs
But they also both have very integrated environments with their own architecture so FFI calls can easily break stuff
Lord forgive me
I think it's less "binding" and more "embed"
You're gonna have a nice and toasty room at least
detects it and boots with no problems i figured it would but i still feel like im breaking the law
why do we call them discrete gpus if they are large enough to give a concussion if you drop them on someone

they get knocked out so they dont know what hit them
im sure this reboot will be textbook and uneventful
...........
rebooted and it's fine, pulled up lmstudio and well
vulkan backend just ootb is using both gpus at the same time for inference and storing weights
Arch?
Now try nix 
i asked glm to write at least 1000 words, an intro chapter for a novel whatever it wanted to shit out, so i could test this
and first of all it's a 30b and it's cruisin so that's cool
but also holy shit is it yapping
it's STILL going and it's at 19k words
I think legit it will write a book since you asked a novel lmao
Man, I wish I have the compute power to mess around with local model lol
I am concerned for his mental health lol
- first time booted up
- aware
- told to yap
- yap a whole book worth
Damn, not that bad either
i wish i had bothered to run any tests beforehand to have a fresh idea of what it was doing before having both cards
Can't you just not use the other one? I think LM studio can be configured that way right?
ffi wont break anything in smalltalk
erlang is trickier but even there i think its fine
did I step into NN by mistake 
I got a call from a recruiter today
same recruiter that called me 6m ago for the same DBA job I got offered from accenture
The job that I turned down because they wouldn't meet my asking rate

"he ...
he ...
it ...
his ...
he ..
he ..
"
The wall didn't exists.
i think to write a good book it needs to use some kind of planning so it can plan plot stuff.
things like motivations and stuff
Also 44t/s on a 30b MOE isn't that wild. Where it gets wild is when you start pushing a 24b dense to 90+ t/s
Why ARE gpu comps adding AI features into ordinary consumer graphics cards anyway?
Otherwise the cards aren't going to get better at all
to be able to charge consumers more ig
that sucks to hear, but I guess CPUs came across a similar dead end in the past
They are attempting to use AI as an optimization because they can't afford to do anything else
ray tracing is too expensive? Use AI to guess the entire scene from a few rays
Honestly it is a decent compermise
if it doesn't have side effects that's a nice way to go I suppose
but there have been side effects tho
I hear the guess is from like one every 64 pixels and geometry. Which is quite the ask
It's either that or not being able to run ray traced at all, so there isn't much of a choice until someone attempts to do something diffrent
nah, ray tracing is quite the beast
which is what I've been curious about with regard to next level performance gains: What if optimizations are good and you use AI upscaling, so that you're not necessarily upscaling low LOD, but objects close to view are
oh
the current method is an optimzation from the expected method of sending billions of rays from all lights. Instead we just trace from the camera and track that until it hits a light. Pretty simple and better than the alternative
Issue with that method is that it is not possible to cache at all
I've gone back and forth with how to deal with that issue, but nothing is perfect.
yea
yep
With games that do it without raytracing, I'm amazed by the ingenuity employed
For instance Control has RT as an option, but the game still looks totally beautiful without that enabled
Yeah, RT at best is only a touchup compared to properly done graphics
I do really want to see RT be a lot faster, but that requires reinventing the wheel a bit
mhm
Spectral rendering is logically the next step 

Guessing if we want significantly better RT, we'd need GPUs requiring power stations to operate
Tune me in on what that is
Personally, I'd just want to have any form of cache on RT
From the wikipedia
a technique in which a scene's light transport is modeled with real wavelengths. This process is typically slower than traditional rendering, which renders the scene in its red, green, and blue components and then overlays the images. Spectral rendering is often used in ray tracing or photon mapping to more accurately simulate the scene, often for comparison with an actual photograph to test the rendering algorithm (as in a Cornell Box) or to simulate different portions of the electromagnetic spectrum for the purpose of scientific work
because we have exactly nothing
So it is ray tracing on steroid basically
yep, rt but light is made of it's compoent wave lengths
Can't even imagine if it has to do ray marching on top of that
key words being "typically slower" oof ouch
Just use a larger system 
I guess graphics processing really has reached a bottleneck
I wonder if there is even a screen that could display the result of spectral rendering accurately
or maybe stalemate is the better word to use
we just only want to do AI, some voxel games actually have software RT (non path traced) at 165fps+
It's just a matter of people not wanting to
due to using premade engines
There is a quote saying as hardware becomes faster, software becomes slower at just about the same rate
interesting
sad but true
interestingly any software of the past era of computing (win95, DOS) would be run so fast they'd become unstable with modern hardware
Our resident graphics programmer is doing 7000+ fps on a test scene with like 10000 blades of grass
that's nice
I always wait for the next music he nukes
it may actally be the full 9000 I don't recall
very true
I've heard of techniques to optimize for grass blades and other vegetation, such as using sprite sheets and limiting the number of sprite instances and just repeating it, but transforming the model in some way
Instancing
You may want to see a youtuber named Acerola
instancing, no merging
I've been debating him on getting around to merging instances eventually
still single blades
tbh coherience is a very important issue for him
doesn't want the famous LOD rings
https://www.youtube.com/watch?v=40JzyaOYJeY this one is one of the videos that comes to my mind on this topic. The fellow uses some pretty creative means to increase optimization
The source code and demos are available here: https://patreon.com/vercidium
The greedy meshing algorithm is available here: https://github.com/vercidium-patreon/meshing
I spent the past 6 years creating a game engine, and I've been shocked at the things that can make or break performance.
I put together 4 simple optimisations that you can use ...
My boy should be waking up soon 
Granted, some techniques might lead to undesirable limitations
Yeah, not using a voxel engine atm. Though it does have anti aliasing built in already so I've heard. Haven't actually looked into the vulkan engine, nor could I understand it.
I've joked about making it VR compatable, because then we'd not be subject to VRChat and how much it sucks in terms of FPS not existing.
only runs at 40fps on a 3090
Though the current project is of course volumemetic display
Do not question how I'm getting VR casting to run at FPS on intergrated graphics
(The res is destoryed only half screen is drawn)

Damn he is awake already lol
Welcome to university life 
That thing is the equivalent of rendering 13 4K displays at the same time
Granted, instead of 24bit its 1 bit
Nah
It does the full uint
Its a complicated shader, i dont remember every detail
Doesnt embed on mobile so idk if this is the most recent version
oh cool so Discord doesn't use AgeKey for mature-audience permission
and the only way to verify us to use your ID or a photo of yourself, which "they'll delete after verification"
oh I guess I could just AI generate some mature looking face
They also analyse your behaviour
Like games played
But yeah doubtful that'll go anywhere
Lucky for me I never installed the desktop client. So my age is 0 probably
i don’t get prompted to verify when i click a nsfw channel
does that mean discord thinks i’m unc

It fefaults to teen
Rolls out in march
i feel offended
alr
time to implement silly optimized matrixes
i'm thinking something like this:
(Fake code, not real language.. just a siwwy sketch)
struct MatrixStruct {
public rows: i64;
public cols: i64;
store: HashMap<Str, Any>;
sel_row: i64;
sel_col: i64;
}
class Matrix extends MatrixStruct {
public Matrix(crows: i64 = -1, ccols: i64 = -1) -> self {
self.rows = crows;
self.cols = ccols;
self.store = HashMap::new();
};
public EGet(r:i64,c:i64) -> Any {
self.store.get(format!("r{}-c{}", r, c))
}
public ESet(r:i64,c:i64, value: Any) -> Result<Ok, Err> {
self.store.set(format!("r{}-c{}", r, c), value)
}
public row(v:i64) -> self {
self.sel_row = v;
self
}
public col(v:i64) -> self {
self.sel_col = v;
self
}
public get() -> Any {
self.store.get(format!("r{}-c{}", self.sel_row, self.sel_col))
}
public set(value: Any) -> Result<Ok, Err> {
self.store.set(format!("r{}-c{}", self.sel_row, self.sel_col), value)
}
}
didn's include checks to prevent going over the boundaries set in the constructor tho
also.. what in the java rust typescript hybrid holy cow
should have neurobot ignore messages with codeblock imo

nah
no one really cares about the silly emote reaction
fair
What in the world is this lmao
I did but my brain got confused at first lol
There are some pretty neat compressed formats for sparse matrices if you feel like looking into that 
are you noticing something?
⁹
This is what happens when I get into a programming mood all of a sudden
-# In this case because Xoda made me open Godot
yessss luv me sum whatever
80+w idle on 3440x1440 + 40w idle on 1920x1080 is so stupid
awa
At least on Linux it's only 15-20W idle (no matter the screens or if there even are screens connected at all)
oh, for 40w it's either 2 1920x1080 144+hz monitors or one 3440x1440 240hz
Wrong server bud
i actually did it... it is a bit buggy with cloning though

Konii awawa is different from my awawa
Neuro getting flashbacks of v1
thanks windows update, very cool 
Microslop Copilot+ OS moment
i wish windows would explode on me and crash and burn and
so i can actually believe y’all
is this a psyop by Big Linux
YEAR OF THE LINUX DESKTOP
superbox quitting windows really was a canon event 
Mine literally did
It became so unusable that I couldn't even open the start menu or shut down without pressing the power button
both javascript and linux were found in the big stein files. let that sink in
Awawa
bless you
Today in #baking, bread and cheese
the orange pi 5 ultra is now also out of border control things 
surely it will be delivered tomorrow 
How's the dragon print going btw?
it was finished 4 days ago or so
Sadge
Hey, since when did you become awa PHD now?
I swear she was still a princess just an hour ago
konii is many things
wrong paste.

https://twitch.tv/ellie_minibot
@Ellie_Minibot provides her in-depth take on vibe coding, a.k.a. AI-assisted coding, and how it shares an interesting similarity to breast milk.
0:00 Ellie's Take on Vibe Coding
4:27 The Surprising Parallel to Breast Milk
8:07 Why That Specific Analogy / She Probably Won't Get the Job
Original Air Date: Febru...
i enjoyed this clip.
Just imagine this IRL
Uhhh, nightmare fuel?
nah, its cute
banana bread when
i can put a banana on bread 

Fruit sandwich?
banana and chocolate on bread
How does vedal makes the twins sing? How does that whole thing work cuz at first i tought he would get somebody to sing and then would compute it but it also doesnt look like its that
It uses a different model (not a text to speech)
Also, song is not done by him
its synthv or something
Ohhh gotcha appreciate it
Man, I really appreciate the simple question about the twins instead of the usual how do I make the twin clone
I might be making this now. I've been asking left and right on different server about what to eat during an afternoon picnic out in the open field 
yum
Sam, the bios battery on the schools CNC machine died so now I gotta reset it, bleh
"LLMs Don't Need More Parameters. They Need Loops."
by NeuroDump on yt
I didn't understand everything, but I wonder if tutel is using this kind of model
Now that's something rare
Why would the cnc need to be reset?
modern cnc with a built in pc im guessing
Hmmm, I may be mistaken, but usually even if the CMOS battery died and the clock got reset, our Hurco CNC still works fine. But also because we don't utilize its networking capabilities
Look at ms phd here
where
what is hon?
honourable, its a
oh
when texting its ussualy short fot honey i think
I think depending on where you live that would be Hun
we dont text in english here so 
He isn’t
in the uk it would be "hun"
mom mum
hon hun
wdym?
That’s nice 
it sounds quite rude in english

Oof
What a name
the sch is soft
That makes it worse
it kinfda like how sh is pronounced in the word sheep
hon also has some meanings in english
It would be pronounced as "shat" which means something that happens when you don't make it to the toilet in time
no
thats different
thats with a sj sound
not sch
flat a too
english a sounds almost like an e
hon hon hon, baguette, omellete du fromage
I swear it is a slur against germans
But I can't quite remember
Wtf
i’m not aware of such
Toast is this true?

une poison dans une bains de chocolat 🥐 🥖 
hon is the same as hun basically but there are derived meanings
or one specific i’m thinking of actually
but i shall stop here
Welcome to #programming everybody 

Google's service, offered free of charge, instantly translates words, phrases, and web pages between English and over 100 other languages.
Btw, how much bandwidth do you think it would need to host a community VoIP serve ala discord for 100 people?
Yeah, it sounded like the Ch in loch
So not like what I thought
While maybe 5-10 people will use the voice chat at most, and 3-5 people will stream

I mean, 100x48khz@16bit is 9.6MBps
if you need the same quality as discord then 64 kbps per voice stream (64 up 64 down per user) and screen sharing is a nightmare in bandwidth in comparison yea
And assuming 3 video streams at 1000kbps puts you at about ~14MBps
opus
Damn, didn't know discord uses 64kbps
it’s configurable in vc settings but 64 is the default iirc
Yeah, opus support up to 128 i think?
you need, rtx 5090, to give to me, and then ill give you an old laptop with a ryzen 7535hs to actually run the server on
It can go all the way up to 384 I think
spotify “high” quality preset is 192 opus
Who would even notice the quality
Yeah, 384 is the max
Opus supports constant and variable bitrate encoding from 6 kbit/s to 510 kbit/s
you should
6 kbit/s
ye
neural codecs exist
give me 5 bits
but there’s literally 0 reason to use that because the bandwidth needed for good quality is already tiny
and you’re trading that for considerable compute
voip can has into perceptual coding
can has. into
ye








