#programming
1 messages ¡ Page 428 of 1
pretty much C++ or Python, yeah, so that part very likely isn't getting rewritten
naturally the implementation is written in c


c++ addons for js 
đ¤ Program in C
you can use llms capably in js if you use ffi

node
emug
or even better
native js llm
i mean.. nothing is limiting you to write it in any language really... python just has a lot of llm libraries that others have written that make it easy to do inference and basically put pieces together for an llm.
write your llm in c and compile it and use it via bun's tcc 
i hope vedal atleast uses a good TS runtime.
surely he won't just use node right?
who needs optimisations anyway

I see
ts runtime 
honestly writing your llm in c sounds like a fun weekend project. kinda like "raytracing in a weekend" but like "LLM in a weekend" 



i should look into how llm's are architectured. maybe i can reimplement something like that for like enemy ai in games or level generation.
bro what other runtime is there
V8 is the best engine
that bun thing is just the Apple javascript engine wrapped to behave like node
i know that arc raiders uses machine learning for their enemies. tho i would be interested to see how difficult it would be to implement yourself
do not the transformers
but... funny
bun, Deno for example
in what way
Microwave transformers you generally shouldn't touch
rust is so out.... we write everything in Zig now
Not unless you know how to safely discharge caps
I am sure itâs a good idea to rely on Windows native API
Yeah
But this is specifically designed more for SVC and TTS they said
As it's noisy and not ideal quality for listening as is
There's a lot better models out there to use the output as is
why windows native api?
actual cope language for people too intimidated by C++ and real RAII
i'm learning odin. fun language so far
IDK apparently they have native fetish
So theyâre using NTDLL for most things now for whatever reason
Why not just use stable, documented APIs? Fuck if I know
performance
and win32 is less flexible
though iirc you can also just use win32 apis
To be honest, fuck ntdll. Shit's too high level.
What you should be doing is reading ntdll, enumerating functions to extract syscall indices, and then fuckin' slamming the syscall yourself.
This might trigger less EDR shitstorms than calling the ntdll functions directly tbh.
Ignore the konii and do the funny world gen
I want to see it
Actually
You should do CNN / Conformer so you can have the entire rest of the map be used as context for each new chunk in order to get worse and worse terrain with the added benefit of worse and worse generation times
simple, put screwdriver inbetween terminals, and either the cap will discharge safely, or it will not be your problem anymore
He said he is rewriting the orchestration part of the project. I think TypeScript is a good choice for that part especially if he uses EffectTS (which i doubt but why not)
Skip the concerns about which languge to use, he should write it straight in binary. /j
vedal to neuro:
https://youtu.be/9CyC8OD8NyY?si=lZHxUIHhOn4aLUMz
Thinking about switching to tech? TripleTen helps people go from zero experience to real tech jobs. Book a free career call to find out which role fits you best. https://get.tripleten.com/currentconcept
Human speak to robot in robot language.
Want more? A lot more? Two videos a week found here:
https://www.patreon.com/CurrentConcept
Join the ...
we can have two AAAAAAAers on stream now
Something flying, don't really know my dinos
pterodactyls are not dinosaurs đ¤ âď¸
from stream today:
I'm gonna rewrite all of the orchestration stuff into Typescript
vedal... as a person who writes typescript code...
please don't put yourself through that hell 
^
probably lmaoooo
unless he's like.. rewriting a web orchestration tool in typescript. that's based
thats location gonna location be location a really location funny location bit
i say keep javascript and typescript to the web unless you're writing something realllyyy quick and dirty and for some reason python doesn't cut it
discord bots too cos discord.js is for some reason the best discord library rn
it might be for more discord support?
I thought he wanted to give neuro a presence off stream though
oooh, intresting thought
what is orchestrstion?
like the conductor to an orchestra
it controls all the smaller ais together
to make the illusion of neuro
okay i was gonna be like "oh why is my loadbalancer using so much ram" (it's written in typescript cos i like being a hypocrite and torturing myself) and it's not even using a gb
suprising
can't wait for something to break catastropically cos he accidentally turned a Date variable into a string
(i totally didn't do that a few days ago. nope. nuh uh)
Yay! I just did an Open Source and submitted a PR to a thing!
Meanwhile Discord and its 15GB of memory usage
At this point it's slightly absurd how badly it leaks memory
-# meanwhile my nginx instance is using 8.4mb of memory
As a person writing TypeScript aswell I agree that there are still a lot of pain points with typescript but once you gather enough experience to know how to avoid the pitfalls its honestly a joy to work with
oh absolutely
alas i have to move some stuff from typescript sometime
im gonna do that and proof it might not be a stupid idea
besides its probably would be closer to diffusion based than transformer
nah but it'd be funny

well.. i was looking for a video. but i can't find it anymore. someone wrote a small neural network to generate a 2d world map
has macbook chargers always been this slow? dawgggg
no, chrome is just eating all the power
Chat, I'm making a discord bot, thoughts on the calculator command
dont run thast
that is a very bad idea
like a very very bad idea

someone can just send import os; os.sys("rm -rf /")
basically a goldmine for anyone looking to perform injection attacks
never use unprotected evals
i would heavily advise against using the eval function at all unless strictly necessary
The execution probably wouldn't work anyway since message.content includes the command (!calc)
Obviously not running it because I'm not stupid enough to do so but 
acktually you can't run multiple statements in an eval
you'd have to do __import__("os").system("remove the french language pack")
close enough yeah
I can't find the calc command for the bot i wrote a while ago .-.
probably got purged when oracle terminated everything for no reason
wait what happened

were you on the free plans
noit suprised
yeah but unless you linked a cc and switched on billing (despite staying on always free stuff) they just randomly term you
for that to work I'd have to be doing something similar to the command above, this wouldn't work
get stars on my github projects
idk just dont use eval if u can help it
async def calc(ctx: Context, *, content: str):
result = eval(content)
...
not that you should be using eval
no the ctx.contetnt is right
This command isn't going anywhere, I just did it as a joke because someone in another server is struggling to make a calculator
i made one that had buttons n stuff
it was super cool

ok but you throw away all the parsing on stuff
also bot.command_prefix is such a crutch
unless its a different type to what I remember
tf is that
ideally you should be using ctx.prefix
discord.py supports more than 1 prefix
it always makes me sad that not a lot of people know about interactions.py
it's a really great library
ok, but also rewriting this shit would be a pain https://github.com/Firecrafters/FirecrafterBot
calc('__import__("subprocess").Popen("cmd.exe")')
I check and I see my preferred library being shilled. Nice
thats not even a big codebase
Still takes longer if I'm changing to something new, the original typescript bot took an hour to make, the python rewrite because something with the typescript one wasn't working took a day
its not that big bro
But I don't know how tf integrations.py works
THEN READ THE DOCS

THE DOCS ARE ACTUALLY READBLE, READ IT
you CAN and WILL read the docs
they're good docs
anyways how do i get more stars on my project .-.
One of my favorite things is Annotation based filtering
i want stars
make it really good?
make it really good?
rewrite in Holy-C
that alone wont help
if you make something that people want and it's really good, people will use it
not if thhey dont know about it
if nobody wants it, there's nothign you can do
you were helping a bit in reverse engineering the cdawg canvas frontend before it was open-sourced, that requires reading docs, so you should already have the skill, no?
name it better then.
naming alone wont help
What docs
being easy to find will. i'm not using your project coz (a) i don't know what it is and (b) i don't know where to find it. (and probably (c) have no use for it)
bru
Looking through minified javascript is not reading docs
Wait, whats the repo?
I was never able to find where they published the source
i can't remember but it's in dms somewhere
You said that was a fake one
update on load balancer
it has apparently decided that the whole logic i wrote to figure out what server region a request came from to determine what servers to forward them to is overkill and is now assigning every single user to NA
why? idk apparently it can't get the country code for users anymore
Hi cloud 
i think i'mma just go to bed before i lose my mind trying to figure out what the fuck even happened here

no it's node.js because i hate myself :3
also cos i haven't looked at bun yet
i like it so far

wrote my site in it and a few bots for streamers
seems to be a drop in replacement. i'll look at it
its drop in but has some extra stuff like built in unit testing and some database types

i will look at it once i've solved my whole.. "why can't the loadbalancer balance load" problem
welp ig everyone lives in the usa now
I will say that the only thing worse then node as a package manager is pnpm
welcome everyone
thats why i like how bun does it
u can configure how it does packages for the project
nothng i can do without restarting the load balancer which i cannot do when about 400 users are actively connecting to it
bun is the best one
what is it for if i may ask
stream distribution
for hosting ur own live streaming?
yeah. runs as part of an event branch of an online game for streaming content into the game
yeah idk why u went with typescript for that
i assume ur using something like round robin or least connections to actually sort out traffic
C# or lower level yeah
i think rn it's using the least loaded server
since i keep track of bandwidth usage across all edge servers both for load balancing and also to determine when to add more capacity
???
and ur also trying to sort based on location?
yes. stream loadbalancing requires region-based balancing for latency reasons
that was about something else iirc
yeah if ur trying to do it in a game i imagine that will be a pain in the ass unless ur given region info
(though there are painful methods to check through ip lookups)
each individual user requests the stream client side. it currently does it via downloading an ip lookup table locally upon build
Read it then, I canât find literally any other mention
I donât think the canvas was open sourced
i assume what's happened is it's tried to update that table, failed, and stored invalid data so now nothing works
if its a big table u could hash it
I mean yeah I dont think so 
use that for updates
i currently rely on a 3rd party library for it tbh
I told you that someone said that was the repo, and youâre response was âthey liedâ
i didnt know there was one for offline use
if it keeps pulling this shit i might give up and look at it's source and write my own implementation
because it breaking in this fashion during a stream is just a no-go
is it for a game like vrchat?
based on the time the messages were sent, it was for something else
similar
if its a game ur coding u might as well prompt the user for a region
i don't have control of game code unfortunately else i would
is it a public game?
cloud why does your git url not work
what
oh that's outdated
leads to git.cloudburst.lgbt
classic
There is no context, also the contents donât line up with things that were in the cdawg canvas
i've been meaning to work on a new site but i can't be bothered tbh
yeah it seems they had a different version because no such things as powerups were in the oss version
i remade mine recently after putting it off for months
honestly the current setup has worked well for ages. i've had a few hiccups lately because i rewrote the stream and server management implementations but i've never had the ip lookup table break this badly before
i don't even know how it's broken this way
dependency update?
completely normal
interesting timezone 
unlikely, this broke midway through a stream
is it opensource or na
the ip lookup library is, the game and the distribution are not
if the game is being made by the crew here ur better off getting them to bite the bullet and add a region setting
what's said lib here
not looking at the code rn but i think it's ip-location-api
ill give it a look in a minute
i really dont trust offline ip lookups personally as i used to experiment with some for bot prevention back in the day and they get dated fast
this one seems to pull every few hours, i assume one of these pulls fucked everything up
Ohh it uses an api
yes
finally
yeah absolutely not lmao
Lmao good
it pulls updates by itself
i legit don't know the cause of this problem tho cos all i output in logs is the country code and the region that my code decides on
Is the information just broken when stored or is it like failing to call the api and falling back to US?
and country code is apparently currently undefined
Yeah but it might be grabbing an undefined
it straight up makes itself a database on startup
honestly i think the only way i'mma figure this one out is to add a bunch of debug info
Honestly yeah
Might as well
If push comes to shove u can locally fork it to patch it in the meantime
this hasn't happened before so it might be a one-off failure for all i know
Might need a failsafe
it could actually be.. oh that miggghtt be a problem
i believe each import of the library starts a refresh task
nvm youâre right, a request I saved has something similar to the repo name in it
when triggered it'll refetch the database
Like importing it as a dependency?
yes
i kid you not the script i wrote to download the db on docker image build is just the require
that's it
it just adds it as a js node.js dependency and it downloads the db lmao
but if there's 2 instances.. a race condition could theoretically occur where one tries to update and potentially fucks up the other one?
I hate blackbox bs like that
ikr
i will tomorrow
it's nearly 2am
honestly i'll probably just write my own version from scratch atp
cos then at least if it fails i can fuckin debug it
and then at some point eventually rewrite it in c#
but for now i'mma sleep and deal with this tomorrow
Yeah I wouldnt trust something that dated and designed for something important
actually update the docs use import
still sucks tho considering it downloads a db on import
i hit a GitHub rate limit
amazing
lmao how
Someone's selling me a Toshiba laptop with i7 CPU and and amd (idk what the GPU is yet) for 24$ if I convert it to dollars
I did try it and only thing I need to change it buy it a ssd and change the os
Probably nice for some thing ?
Portable small game setup? Like 2d games
Gonna put steamos on it lulw if I bought it
(I have no idea what you're talking about kinda)
finally managed to fix most of my driver issues, I can finally use the penguin!
I see no shiro here
shiro is offline rn
Dont hbs gor the shiros that are here
Hbs for the shiros that will be
And for the shiro in our hearts
Idkwhat im on about. Waking up at 4am is weird
@rare bridge what's wrong with the orchestration layer using typescript
It'll probably be fine, it just odd to rewrite from C# to a less strict language when he's presumably trying to fix reliability issues.

[nix-shell:~/Projects2/LLM]$ python -m train build --config config/config.xml --dataset-file ./outputs/mixes/mix.jsonl --output-dir ./outputs/models
build completed: model_file=outputs/models/model.json
Codex
There is no universe where this is the correct thing to be outputting
its a canon event
I'm still stupid enough to be using it so not yet
idt he's rewriting it because reliability, I think he's rewriting it because spaghetti code
and figured if the orchestration layer has a web panel might as well rewrite that part to be typescript
Cloud was in here earlier talking about it
oh
[2026-03-15 01:56:17.370] [info] RagRuntime: WordNet seeding complete total=5000 ok=5000 skipped=0
so far so good. now the RAG has something in it to return. hopefully there's a moderately good match for just about any likely thought.
next: actually testing that it works.
What's an orchestration layer
a layer that orchestrates things
Bro what
Think of it as something that allows a lot of projects to properly commincate
A program that contains all of the glue code between programs
yeah it's a thing that allows two or more modules to communicate and coordinate with each other
I wouldn't necessarily say glue code but more like a tunnel
That entirely depends on how compatiable the parts are 
tunnel at best, hefty conversion at worst
I see I see like the spinal cord of the whole program

The two a gonna have a new spinal cord yippee
I guess spinal cord makes sense
I would argue closer to nerves personally
but spinal cord also makes sense
Codewise an orchestration layer is normally made in a higher level lanuage some are in python; if things only need to be alive and that's it, most people use batch/bash for it. This is mostly due to it being reposible for bringing up indivual parts and communication is not where a lot of processing should be happening so the lanuage tax isn't as high.
(I love typos, discords autocorrect is broken here)
not quite bindings and not quite normal code
HOLY CRAP
do NOT drop a TS file into Discord 
For good or bad reasons

well... yeah
who uses .ts audio files in the year 2026 
https://github.com/PrismaticDepths/neoprisma do you guys like my readme
Twitch
impressive, very nice
why
yt-dlp downloads the ts files and then muxes them to mp4
But there's an ffmpeg but where you can't download sections past about the 26 hour mark
So I just downloaded the 10 second TS sections of some stuff I really wanted for some of the VODs
ffmpeg butt
lmao
anyways yeah .ts is more commonly(?) used for typescript
you can download the whole 48 hour vod.. but I didn't feel like doing that
I've got to work out how to stop the agent from reading out DMs they get while talking.
hide the dms from them
i mean DMs like "it's time to change the subject". I expect they'll also read the tool use help messages to anyone listening too.
Welp, i submitted the fraud thingy
Next time im taking superbox's money beforehand
It shouldnt be mine at risk, im the middle-man
wow i fixed some issues in my zsh and starship config and now in combination with Ghostty my shell runs super fast and exactly how i want it
Thanks again @fast pagoda for recommending starship
what did you do
eeyy
nice
Superbox wanted another 3090
international GPU smuggling ring.
Basicly
lovely
it's so nice when terminal is set up the way you want it to be working lol sucks so bad otherwise
that reminds me to buy v100
are there cheap v100
near u
in ur area
single v100s looking to be swept overseas
ii just want something with 32gb that doesnt also require me to use my entire gpu to train things and then rdna4 is still unstable with certain sequence lengths because there's a specific amount of vram that if touched will try to access memory that doesnt exist, explode the driver and crash the training for an illegal access
Except for buying a whole server
lol that would be the dream
I cant afford to ship a whole server
shipping a whole server would be difficult... you'd need to know what everyone was into and all their interpersonal relationships... or you could just do it by feel. (i know you mean "sending a server computer by mail of some sort" but "shipping a whole server" sounds like the mad rantings of a fanfic writer who hasn't taken their meds today.)
Y'all seen that Nvidia godot rtx fork I'm tempted to see it looks nice
wtf happened
Fraud
got scammed?
Ye
Calling it fraud allows me to take legal action tho
(to clarify, the fraud was the seller, not Superbox)
idk which one of you the one that found the listing and agreed on it
Dont matter anyways
also don't be so damn spoiled you also had to look for it aswell
the one looking for the deals is usually not the middle man's job
middle man only distribute
that is how it is in my case of smuggling dealing chips from china
Banks in NL are forced to freeze the account and give the personal info from the guy.
Just need the "procces verbaal" documents
most middle-men i know of are people who wanted 1 or 2 of something but could get them in batches of 30 or 100 or whatever but really cheap so they sold off the rest to others in the same hobby
The only issue with potential money is if it already got transfered to another country
But even then i can just make the person pay me back wuth the threat of legal action
that's called as group buy if agreed upon or redistribution if sold to randoms
done a group buy before for an enterprise SLC nvme
w0w i missed this context earlier, da h ell
that's annoying
sorry that went that way big dawg
I love my i5-13500
Laravel my beloved
cough cough Iâm back
finally found a package more annoying to update (due to compiling for ages) than llama.cpp-rocm
has to hipify like all of vllm
lule

youtube feed is interesting today
Dang... how is my country dropping this movie at the same time as the US? lmaooo
I think my agent had a stroke while writing my code. ;[
You should be good to go and context crashes due to scaling will be automatically generated thought loop back to growing history overflow should be eliminated.
wut?
does anyone here use Linux, I need help with something 
depends on what the problem is. there are some ppl who do.
I just wanna know if I can install stuff in custom directories
like installing stuff on my 1tb hdd, instead of using my ssd
that's a filesystem organization problem mostly. you can mount partitions wherever you want and even over top of existing ones if you want to extend the drive.
not an expert on it though.

but the short answer is "you can do it. somehow."
me just silly, dunno
for normal packages that get installed through apt or similar, no, not really
the right way to fix the problem you have though is to fix your filesystem so it uses the right drive for the right stuff
not to change where it installs things
change what storage is where it installs things
To be fair, I can still use Linux just fine like that but, my daw for some reason doesn't detect my hard drive within it's file Explorer, so I can't use my sample folder
typically that'd be /usr on the big storage drive

that can be solved easily
you just mount it somewhere
the best way to ask a question is to say what you are trying to do, not how to do a particular task. it may be much easier to solve the original problem than the way you are trying
it's not a feature of the DAW, it's a Linux thing
oh, I see
instead of being assigned drive letters like on Windows, filesystems are mounted as directories at certain "mountpoints", usually under /mnt
so you could mount your HDD at /mnt/hdd1 for example, and that directory would now have the contents of the HDD
Fair enough. So I just installed Linux mint on a new ssd, cause I don't know how to partition and I didn't want to wipe my windows one. I eventually got everything working but, when I open my daw, the file Explorer only shows files inside the ssd, and not outside of it, unlike it did on windows.
I forgot to mention, this daw is for Linux aswell, so I don't need wine or anything like that
oh
some file explorers do this for you to make stuff like using USB drives much more convenient, but that's a feature of the file explorer
your DAW doesn't know about any of that, so you'll have to navigate to wherever it is mounted manually
typically for fixed storage linux won't automatically mount non-linux stuff. but you can tell it to. (not an expert: last way I did it was editing /etc/fstab but that's probably the wrong way)
depending on which dekstop you use there's a gui for it
it won't automatically mount anything 
I'm kinda clueless 
man, this shit mounts fr
unless configured in /etc/fstab or as a systemd mount target
but that's not really "automatic" in the sense of mounting external storage
USB hotswapping is different. i meant "automatic" as in "defaults to mount" rather than "mounts when available"
you just want to see how the mounting works you can probably just open the HDD in your file explorer and then see where it's mounted in the directory properties or something, it should say a path somewhere
then you should be able to access that path from your DAW too
I'll try
so you'll need to mount the drive (which probably requires sudo access) and then tell the system to make it permanent. perhaps this is something gemini or claude or chatgpt could explain?
if you always have that same HDD attached the same way then it would be proper to add it to /etc/fstab, more convenient long-term 
but that has a chance of breaking stuff if you do it wrong, so can probably put off until you're more confident with it
manually editing the /etc/fstab file would be a last resort. it's possible to make your system not work by adding or deleting a character from the wrong place. (/end of scary warning)
btw silly fact: my pc fans have to be poked to run, otherwise they just stay still every time I power them on
mine sound like jet starting up for the first 2 seconds until the BIOS loads and they slow a whole lot.
oh, lmao
so at startup they get like the full 12v. if they're not working, they're not getting the voltage or they're very very dead
I used to type man "program" into Google to get the manpages.
But that stopped after I typed "man strip" and got traumatised for life
Aw
Now I'm sad
on my box pc that's what happens atm
got em set to turbo mode half the time anyways
@chrome rampart make sure FUSE drivers (package is usually fuse) are installed in your system, that lets the system detect and mount storage devices at runtime with appropriate permissions for your user. also NTFS drivers (usually the package is ntfs-3g) if your other drive comes from windows
thank you!
fuse drivers for what?
manually mounting (whether though mount command or fstab registry) is usually the wrong thing to do outside of system partitions, as it won't show up how you expect in your software, and it won't have the right permissions and will need tweaking
oh wait nvm I misunderstood
FUSE lets storage devices besides the system one show up in your file explorer and other software like this, which is more or less how people expect it to work (gnome's file manager nautilus for example)
yea yea I got it confused, for some reason my brain only thought it was used for stuff like s3fs
yeh just explaining a bit more for buhspoon since they're new to it
yea yea fair enough
linux mint should have both fuse and ntfs out of the box, so it's odd that it didn't work right away
why would you need FUSE for that
fuse is what makes it possible at all
usually comes out of the box nowadays though

but it has kernel drivers
no need for usermode fs driver
so FUSE unnecessary
FUSE is mostly for when you want to expose filesystems that aren't "real" filesystems backed by a block device or that don't have a kernel driver
still convenient to have so that you can mount ISOs and such, but shouldn't be necessary just for mounting normal devices
a lot of things are built around fuse, particularly the permission-appropriate detection and auto-mounting
stuff like ntfs and many others just don't work well without it

uses udisks2 apparently
no FUSE involved
-# which makes sense since FUSE would be really really stupid here
ntfs uses fuse
well uninstall fuse and let me know how that goes for you
I don't think I have it installed 
cough cough s3 buckets cough cough
(apparently there are people who actually use s3 buckets for personal storage)
the kernel driver has never worked properly for me, it's always corrupted my ntfs drives
dunno, ntfs-3g has worked for like a decade for me
the old horrible kernel driver, the FUSE ntfs-3g driver, and the new now-recommended ntfs3 driver
so you used to need ntfs-3g and therefore FUSE for decent NTFS support
but that's not really the case anymore
nvm it's not ntfs3 
that's the Paragon one
which is also relatively new but turned out to suck
new driver is just ntfs which replaces the old horrible NTFS driver
so not yet released
old article, pretty sure that quite a few patch sets have been merged since then
not out yet, most recent update https://www.phoronix.com/news/NTFS3-Linux-7.0
NTFS is a mess
Thus why I'm definitely gonna get rid of it once Xod sends me that 10TB to shuffle files around
best solution is
not use NTFS at all 
and if you do use NTFS you're probably fine with whatever driver the ntfs mount type gives you
that's the one that's always corrupted everyone's shit
should update automatically to whatever the distro maintainers decide is best
udisks2 was the detection and automount service i forgot, thanks for correction, fuse is the other half which allowed the weird filesystems to mount correctly
0.8b when messing with backens
if only it was a poem, or about the 5090
it vaguely has an idea of what a 5090 might be
maybe

it could be some shitty poetry slam type thing is how i read it
i think "misconfigured" and <number here> sent it into a generic tech direction
the numbers 5090 on the keyboard
poems need structure and rhyme which this doesn't really have
yeah that's why it was amusing more than anything else
gets up on stage and just talks complete nonsense and then leaves
it's pretty neat that the 0.8B model is coherent at all though
SMOL
Sounds like something I could put together at home
1b used to be coherentish
although i was not using instruct version
so i am not a great judge of that
it was only able to talk in a thread of any sort of coherency thru finetune giving it a bit of that
dunno, they're distilled from a huge properly trained one
it's not like they train each size specifically
i'm very happy with the 35B-A3B variant though
i still prefer ol 27
27 is slow af, i dont have the power to run it
that makes sense
still, 35b has been more than adequate for me, it's smart, even if it tries to gotcha itself for 5 minutes when reasoning
i never have good luck with qwen models & reasoning
probably because most of the time ive fucked up their ability to be consistent thru abuse
for me it seems to be fine if the prompt is long enough, and temp+repetition tuning
also that
i don't try to retrain mine, aint got no time for that
last time i did a sweep of all the hyperparameters i randomly found that temp 2.0 with sensible top-k and min-P actually was one of the most stable versions of it
slap on a system prompt, tune temp and stuff, that's all i need
temp 2 is pretty high, wouldn't it be fairly neurotic with that?
i run mine at 0.4
it's very high but all it does is make it more likely to choose a less likely next token as the temp increases
flattens out the softmax distribution of possible next token
yeh, which may be garbage though
usually is garbage
2.0 being randomly good was anomalous but it worked
but only with min-P and top-K set and no penalties at all
a certain frog's workstation has its gpu ramping up even more and i think i need noise cancelling headphones holy fuck
wtf is she doing
:V
echo (smart version) has control of the box in the corner, i just got him set up w/ that tonight
waiting for him to kill me somehow
who tf is echo
the stupid bot in my disc
which frog!
this must've been too traumatic
thought for one sec you meant amazon echo
I think you need a tractor driver's headset
Awesome
This is an idea of what I have been up to.
AI IShowSpeed Artist/Rigger: https://linktr.ee/M3nn05
Support me on patreon and join the discord!
https://www.patreon.com/SchizoDev
https://discord.com/invite/chWagUEHb3
Ishowspeed clanker is utterly crazy
even the top streamers wanna be neuro
| <|say|> BBQ sauce? Seriously? Okay... let me think... I was talking about pepperoni... and then you said something about... oh, wait. I think I was trying to remember what *you* liked. I'm so confused. <end_of_<end_of_<end_of_<end_of_<end_of_<end_of_<end_of_<end_of_turn>
i think the fix to get rid of spurious "<end_of_turn>" messages didn't work.
the end of the end of the end
ez fix, just preg_replace đ
here /(\<end_of_)+(turn\>)/g đ
Are you not using a special token for "end of turn"?
it actually means it's calling the step function and the function is returning part of a token for end of turn and when it comes back after it should have been done it just sends it again... the last one gets added by some other process i think.
it's easy to get rid of them afterwards, but each end_of_turn is actually 3-5 tokens generated, and it's generating lots of them.
Gemma doesn't have a token for it.
nobody's gonna know, just shove it under the rug
Are you using API? Maybe it has "end of sequence" you could use?
i'm using llama.cpp directly to load the gguf.
Cause if you can I would absolutely recommend adding a special token to the tokenizer and train with that
this is version 0.1, with the first proper testing (as in not a script which sends inputs like asking them to count or something). fine tuning is absolutely needed to make it work with this model size.
it's also having trouble understanding how to use the tools. it's ignoring the parenthese it needs around the command... i thought that would be the easiest to understand. ;[ /dm(username blah blah blah). but it keeps leaving the parentheses out.
yeah, if you are doing fine-tuning then you should add a special token to the tokenizer, use that tokenizer in training and then save it with the adapter or merged model, or whatever method you are using
it's defintely too small for what's being asked of it at the moment.
custom token(s) should work fine independent of model size no?
Then it's easier to teach model to do specific stuff
since there's no way for it to be generated as a "half" of what you wanted
yeah
the fine tuning is independent(ish) of changing model size. the model size is more about the instruction following.
mainly that + from model's perspective it only appears in situations you intend the model to use it for, so it can't get confused with pretraining data tokens
yeh
they sure did something interesting with their token choices for gemma.
also this specific issue could be caused by training data padding issue, or maybe something related to End of Sequence token
there's a bit i read also that with llama.cpp you must have the correct chat template or it will end up in garbage
<- amateur llamacpp user
it works other than that, and it respects the "tokens"
it was originally matching on <end_of_turn> but then for some reason occasionally it would output <end_of_turn<end_of_turn> so the 'fix' was apparently to change what was being searched for to the new string... which resulting in the message that I posted. ;/
it's probably adding the token in two locations and only the last one is doing it all the way.
I've had all the chatbots look at it because it's such a long running (but able to be ignored mostly) problem but because the codebase sprawls now, they generally try making the string that's matching shorter. ;/
actually wouldn't the fact that it can output a partial end of turn mean that it's gotten trained on the literal string <end_of_turn>, which should have been a single non-text token
I'm kinda confused by the wording, is the model generating the <end_of_turn> or is it something you add?
yep. you'd think so. but here we are with the tokens "<" "end_of_", "turn" , and ">" it seems. ("end_of_" might also be split but i haven't been stupid enough to keep trying to make it shorter.
(the early ones might be different)
i don't think it's a length issue, i think it's that the signaling is being done within the normal text content, the metadata should be their own special tokens
the literal text "<end_of_turn>" shouldn't exist anywhere in there
well yeah, tokenizer will tokenize based on what characters are behind and after it as well, so you can't guarantee it to be always the same tokens unless you specifically add it as a one single token to the tokenizer, which then will make the tokenizer prioritize in tokenization so such a string is nothing else than that one special token
i think it should stop polling the LLM until the context has changed after it's output one <end_of_turn (even if it hasn't finished it)
it's probably a state machine type thing pumping out the tokens making up <end_of_turn> at the end and my 'clever' logic is messing something up
what does the end of turn actually signify there?
is it end of the llm response?
it shouldn't be multiple tokens if so
<start_of_turn>model
<|thought|>BBQ sauce... okay. That's... unexpected. That changes things. I need to re-evaluate everything. This is getting weird. I'm going to try to remember what I said before.
<|say|> BBQ sauce? Seriously? Okay... let me think... I was talking about pepperoni... and then you said something about... oh, wait. I think I was trying to remember what *you* liked. I'm so confused. <end_of_<end_of_<end_of_<end_of_<end_of_<end_of_<end_of_<end_of_turn>
<start_of_turn>user
i feel like it's been trained on the metadata representation text by accident
yeah probably they got to a certain point and were like "oh... it's still working, i guess. ;/"
what are you using <end_of_turn> mechanically for?
i'm using it to know when the agent has finished so I can stop pumping output. trying to.
because there shouldn't have been multiple or anything partial about the turn tokens, they should have been separate metatokens from the training data, which never appear in the training data
but most models already have their own "End of sequence" type of token, which they always generate and it automatically signals whatever software the model is running on, that this marks the end of output
check your training data and vocab methinks
that also
if you're fine tuning you shouldn't need to manually add your own right? it should be part of the vocab already
blame google. this is using a base Gemma 3 Instruct thing
hmm. could be how i'm loading the model. I see something saying it should be a single token... but this is what the model is producing. hmm.
it definitely should be a single token
the metadata signaling tokens should be singular ones, the "<end_of_turn>" is just a readable representation, but it's never part of the actual source data
load: control token: 1 '<eos>' is not marked as EOG
load: control token: 0 '<pad>' is not marked as EOG
load: control token: 2 '<bos>' is not marked as EOG
load: control token: 105 '<start_of_turn>' is not marked as EOG
load: control token: 255999 '<start_of_image>' is not marked as EOG
load: control token: 256000 '<end_of_image>' is not marked as EOG
load: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
load: printing all EOG tokens:
load: - 1 ('<eos>')
load: - 106 ('<end_of_turn>')
it's loading it and has an <end_of_turn> token that it's just not producing??
you shouldn't see them normally
It seems like <end_of_turn> is a special token in Gemma 3 so there seems to be an issue with tokenizer you are using, cause for Gemma it should be 1 token for the model to understand it correctly
those are for the internal signaling
nah you should see them unless you are doing any logic to hide them
but they should come out atomically
ah i guess depends which end of the api you're holding
also this says <eos> and <end_of_turn> automatically make the model stop generating
Maybe perhaps whatever data you used for training had its tokenization done incorrectly, and model learned to output <end_of_turn> however training data had it. For sure not as one special token it should be
there is no situation where you get a series of tokens to make up the <end_of_turn> token normally that I'm aware of unless there is something weird happening.
i've added an explicit check for the current token matching llama_vocab_eot(vocab) with the vocab for the model.
There was an explicit EOS check but not EOT.
Absolutely
Well, for sure that series of tokens existed somewhere for the model to generate end of turn as separate tokens. Don't see any other reasonable way for that to happen
are you able to inspect the raw untranslated values?
check what the partial <end_of_ is in the vocab
perhaps if you keep asking after it sends an EOT token it generates the EOT message by hand so it produces it?
I'm rebuilding now. have to cleanly shutdown the agent.
coordinator app is starting to look cluttered.
what are you cookin anyway, your own neuro?
best 30m vibe coding i ever did.
the easiest check would be to see what the token id for End of Turn is, and then compare that with your tokenized training data
never thought about it but imgui might actually be perfect for vibecoding a dev ui đ¤
something like that. not 100% sure what the path will be.
nice
i've used typescript and i feel like i'm losing my mind every time i use it
anyway
@silent cloak found why the library reloads the db on import
as for how it works\
it pulls GeoLite2-country on reload
downloads the data, unzips it
and uses that
last time I tried to use typescript I had to make a manual wrapper around websockets because I couldn't stand callbacks
yeah thats dumb
if u can workaround it good but u can always just fork it too and expose a public method to do it yourself for when you want to
i might just rewrite the entire thing lmao
monkey patch reload 
honestly if its that dated might as well
if they dont update often then u probably dont need to worry about api changes
typescript rewrite đĽ
use async 
yeah that's basically what I did
i like typescript besides how it infers stuff sometimes
though that just could be an ide issue
promisify my beloved 
just got an email about a subscription i thought i had cancelled increasing in price.
NOW i am unsubscribed.
yang
u subscribe to a lotta 18+ communities eh
ive never had that be forced D:
actuallynvm
fucking
gym
that reminds me of my funny hetzner bill
what like a server?
ÂŁ0.28 bill 
cpu time on a server
i had a box over off in the netherlands for like 7 years for $5/month and i can cancel whenever
that sounds horrible
it is
im currently paying for my life lesson
i havent been able to cancel it for ages
i should be the one thats sad, my money is gone
i was supposed to cancel it in february bybphone but i forgor
during covid i had a 24hourfitness membership i couldnt use because covidf
my service would have automatically stopped at the end of may anyway when the credit card expires.
and they made you either go to branch (closed, confrontational)
or like
mail them something
i used whatever that first "robo lawyer" service was i cant even remember what it was called
and they mailed a certified mail demand to cancel my shit and give me a refund
which they did
so glad i'm in europe and corps get sued out of existence if they try to make cancellations hard :V
favourite subscription service is AWS, if you forget to meticulously clean all services out of your account they will keep charging you after closure with no recourse
landlord behavior
they make bank on that shit. i hate *aas, but i'm making one hah
296/574 on a programming competition where the cutoff to the next round is top 50% 
oof...
he's passing no?
it lost by 9 points/spots
welp, thats unfortunate lol
better luck next time andrew!
turn it into a CTF, put yourself in the winning half
top 50% of all scores or has to be over 50% of the maximum possible score?
"; DROP TABLE competition; --
oh... ranking.
i forgot about ranking as well ha
whats so confusing about the rank
i though it's score lol
well, at least you know almost all of of the competitors who were eliminated were worse than you
why would the score be out of 574... usually scores are round, like 500 (although rare) or 100
you know that you're perfectly average
could be that's how many questions they had?
idk, wouldn't be the first weird scoring system i've seen 
many CTFs have weird total scores
574 questions? what would all those questions be about, for a programming competition, im sure theres better ways to score than Q/A
by question i mean "challenge"
or how many points there were coz some might be worth 15 or 7 or soemthing
oh you think you wrote a correct implementation, here's a stupid edge case

but ye it's obviously ranking, i'm just cooked
make me a selfhosted basic websearch mcp
so that i can plug it into my local llm for extra smarts
+1, -1, 0, maxint maxuint maxu?int[+-]{1,2,4,8,16,32,64} things. oh and also for multiples of 10 +/- 1
^ some magic numbers for breaking bad code.
did you know that fgets() doesn't always read an entire line? it's true!
it's interesting how many people think 1024 bytes is enough for a line.
or some other fixed constant. coz bad guys can sometimes count higher than the maximum size and trick the reader into doing weird stuff.
and if your program is reading partially evil guy controlled data and thinks the middle of the line is the start of the line, bad stuff can happen.
self-hosted as in, the search engine itself is self-hosted?
nah just the mcp server, it would call out to google, etc
well hey you can do that too :V
the internet is big. you are not going to crawl much of it.
also do it in PHP
well with enough time I technically could
but like
I bet the internet is growing faster than a crawler could index
So it would still be behind..
technically you can't crawl the whole internet because it's full of private, hidden and restricted subnetworks
yeah? they're hidden for a reason
some fraction of the public Web is best anyone does, even the giants
also what about a random ip address
thats part of the internet too
but
lets be real no crawler would index that...
the internet technically speaking is the physical network
the Web is the content on it, public side anyway
because it has no way of finding it (for the most part)
the Web is some of the content on the internet.
remember that servers aren't just webservers.
yuh
I find it funny how search engines use the same algorithm used to solve a maze...
you can actually hit the entire ipv4 space in a fairly short amount of time nowadays
like under a day
Hope you can get it back
that's because there are only really a limited number of effective algorithms.
But crawling ips isn't crawling all hostnames
yuh
under 574/2th place
the ip space itself is fairly boring
the real meat is hitting whatever's behind the billions of NATs everywhere, for which you need hostname + port for any response
the internet isn't very well utilized in ipv4 space.
it's incredibly well utilized actually, we ran out of addresses like 20 years ago
im not sad about the comp I got the same score as my friends who are national olympiad level 
it's been densely packed through layers and layers of NATs
occupying basically all available addresses and ports
IPv4 should only be used for development. ;]
hah well, that's ipv6 for most of its lifetime
did you know we created ipv6 in 1995 because we knew we would run out of the v4 space quickly
and it's still not properly adopted
only like 40-50% of traffic can be handled fully by v6, aside from mobile networks
I was partially responsible for the internet running out of space. i had a /24 for a while.
you must be wealthy
a whole 256 addresses!
ipv4 is the next bitcoin
it's true! people minted it, now it's rent seeking, it only goes up in value as there's more participants
it wasn't mine personally unfortuately. but my name was on it. would've been nice to keep it but it's long been reassigned.
darn
damn
for the SDK for the lightcrafter
i need to promise i wont use it to make missiles

I guess
I didnt sign anything for buying the hardware
And the sdk just sends usb commands anyways

can you make a drone instead
volumetric display missile
Probably
This thing is not heavy at all
The main purpose of these 4225hz projectors is 3D imageing
The DLP LightCrafter 4500âs feature set makes it an ideal platform for powering industrial 3D machine vision scanners, medical imaging devices, spectrometers, and much more.
i would like to add that this 20MB sdk is downloading at 30 KB/s
i saw some folks in the last week or 2 have open sourced a manpads as well as field radar
now it's your turn sam
i also had to say i wouldn't redistribute
its not hard to get your hands on this software yourself tho, except if you're in certain countries
is that for your code using the SDK's API or more specifically just "don't redistribute our SDK binaries"
yeah makes sense
erm
mods, if this is breakign rule 3 feel free to nuke the message



damn, there goes my plan to use a TI 555 timer in my bomb
i think I got a similar page when looking for dell manuals which was funny
you cans still use that one, just dont use the SDK
i dont condone the creation of bombs btw
Speak for yourself 
-# in minecraft
im in
đş 
lemme read the manual actually
good news and bad news.
good news: my agent was just playing Zork and finding it frustrating. that means the game launcher, the agent, the tts, the avatar and the dm client are all working together.
bad news: I had an unexpected power down while it was running.
edid reprogramming
why did e program

om
what are u using the lightcrafter for
volumetric dispaly
update on that programming contest:
top 50% was apparently including everyone with no correct problems, so I ended up making the cutoff by a long shot

the 912x1140 goes crazy
huh
show
if i had built it already i wouldn't be testing the lightcafter rn
On camera it has funky lines
I don't know, looks like just a normal day on Windows 11
can superbox go one day without throwing shade on win11?
this isnt even caused by win 11
not a single person besides me has had this issue before in the entire world
well, maybe a couple other people
looks like windows to me
its not a windows bug

I'm just joking around
fixed with longer exposure mabe?
Let me do some of that sometimes too, even if I'm not good at it
probably
im just getting tired of this "x platform is better" elitism
the linux people are not better than the apple people
they both have this
Meanwhile NeuroSynth has already escaped swarm
NeuroSynth now featured in a Deltarune community project
thats crazy
And this is just a beginning
NeuroSynth isn't targeting swarm
It's targeting the wider Vocaloid community
darn
i do have to agree tho
every platform has its strengths
none of them are objectively bad
they're all very very capable and have many many use cases
NeuroSynth also has plenty of internal progress
I call NeuroSynth/JA basically done
NeuroSynth voice bank assets are complete
We have a plan for NeuroSynth/EN
I have a song picked out and in progress for NeuroSynth's release trailer
Things are rapidly getting closer to release
at least every major one
idk how capable temple os is
cant wait to see neuroâs voice overlaid on some minecraft parkour while talking about the political state of the middle east /j
Sil
its a really cool project jokes aside
And it's quite some absolute tomfoolery
how exactly does the ml work? do you feed neuroâs official voice and try to make it as close as possible
Eh, kinda
It's a little more complicated than that
NeuroSynth uses synthetic data alongside organic data and cross-training data to achieve maximal results
could you adapt it to work on human singers too
like those ai sloppers are doing
or is it finetuned to neuro
This model is made for Neuro, but there's nothing stopping the same techniques from working with real singers' voices
Got it
maybe
random NTFS mount failures



