#programming
1 messages · Page 522 of 1
meh make is fine
...
yes i do
it's not extracting
Yeah, web serial is chromium derivative only 
oh nvm
Pretty handy when needing to configure device on the field without bringing a laptop
Indeed. Which is why I am cursed on both side 
Low level web server and high level web ui 
So my tokeniser with comments is a full 22 lines and the broken embedding code with comments is 307 lines
I hope it doesn't get worse than this
I am also kinda still in web dev arc 
not really low level though, pretty much always using Rust or also TypeScript for server stuff
I don't particularly mind, native frontend seems like much more of a nightmare
Yeah. I need to implement HTTP and MQTT on an ESP32-like device so the phone can just scan a qr code for the device access token
it works btw
"CMAKE_C_COMPILER": "zig",
"CMAKE_C_COMPILER_ARG1": "cc",
"CMAKE_CXX_COMPILER": "zig",
"CMAKE_CXX_COMPILER_ARG1": "c++",
"CMAKE_ASM_COMPILER": "zig",
"CMAKE_ASM_COMPILER_ARG1": "cc"
all it takes 
yeah learnt something new from the clanker 
Clangker 
cmake detects it as clang 21.8 which i guess makes sense since that's basically what it is
i don't feel like dealing with visual studio so i shall call this good enough 
had to force the socket library to use cpp11 threads instead of falling back to boost's and then link with winsock manually
stop tempting me to learn zig... i've been slacking it off for a long long time
i rather learn rust..
but rust is painful
What the hell is zig??
ye
programming language of all time
give me 3 reasons to learn zig
lizard
Also makes some silly choices like using Windows native APIs
sometimes I feel like I want to learn it, but then I remember manual memory management and the fact that Rust exists
so the feeling quickly subsides

same..
You people have to manage memory??? I think I'll stick with luau
you also need to manage memory 
i get to manage multiple types of memory 
just... less direct 
There is multiple types of memory???
there are like 7? i think
I get to mostly not care other than not being wildly inefficient 
unfortunately gpus are even more particular about memory than cpus so 
I don't know how you count "types" of memory, but there are probably as many as you want there to be 
And that's why I optimised the entire forward pass and backprop of my llm architecture to not need to use a GPU
i was thinking... ram,vram,cpu lvl cache etc


too many different interpretations
i was thinking just thinking cpu and gpu separately
what if iGPU
Dynamic random-access memory (dynamic RAM or DRAM) is a type of random-access semiconductor memory that stores each bit of data in a memory cell. A DRAM memory cell usually consists of a microscopic capacitor and a transistor, both typically based on metal–oxide–semiconductor (MOS) technology.
While most DRAM memory cell designs use a capaci...

hopital
well.. but you also need to manage cache locality
otherwise performance goes do-do
overrated 
Interesting, so many things just finally go back to requiring a clock and addressing and then just add boatload of data lane
Synchronous dynamic RAM (SDRAM) significantly revises the asynchronous memory interface, adding a clock (and a clock enable) line
I'm trying it rn and i'm really not convinced
I'm having to disable so much idk if it's worth it
I feel like some design choices are outright worse than what github chose
Few things are really nice
But some others are just 💀
I think I have them all on 
I had seen it alr and was recommended back then
I don't remember what is from GitHub and what from the extension tbf
at the very least the coloured releases are really nice though
makes it easy to quickly find the one you probably want
Some of them for me are eeeeh
Like i always show whitespaces and invisible characters, one of the reasons why i love notepad++ so much as it display them so well, but here i hate how it display it

It's the so bad weird arrow
And also it doesn't show newlines
And other invisible chars
Feels half done

Also i don't like this one and CSS it because a new file shouldn't be red, blue or green maybe but not red red is an indicator of really important or danger, blue or green is better for recency highlight
Classic AI language
she does IIRC
but that's completely unrelated to this screenshot 
neuro RLM 
the most disappointing thing my son ever did
@fast pagoda
why arent you a daughter
im getting your birth certificate modified
actually where is your birth certificate
the agency never gave it to me
i should have known there was a reason they were willing to let me take you without any id

New leak? 
This was subathon 2
Oh
abcde
hello mother i sent words of encouragement via next day crier so i hope they were well received despite being left on "heard"




madre
ni te llame, mande un mensajero
tu anemona de animo fue suficiente pero porfa no reniegues de mi en publico

Thank you
dont summon shiro shiro hates me
Tag failure
Shirbo!
@evil

nruhroh
Posted that yesterday


Everyone knows printers are the only kind of IT equipment that never work when you expect them to work
I would just threaten the printer in any way possible personally

compatibility: maximizerd
time to install nvidia drivers on top of this adrenaline
havent tried this in windblows before
its neuro sama in 3d
doing floor dive

apparently qbittorent gets 30 minutes to block shutdown, seems fair 
pls don't tell me that's the same cable 
same cable, charger, new multimeter

it’s doing
i've been trying to debug why my wireguard tunnel doesn't seem to work
turns out the peer service fails to start
and why is that? obviously it's DNS duh

i hate people who don't post originals smh
https://rxddit.com/1tb7edc
Joined a new company and inherited a backend repo from an agentic engineer. Rewrote it in a week with Claude while keeping the same functionality, with a more stable architecture and proper integration tests.
So basically it was a bloated repo, completely out of touch with what actually needed to be build for a product. But everyone celebrat...
I hate vibe coding in general, but not for the use of AI. I hate it because people don't know how to use coding agents as tools and not self-propelled fuck up machines
If you can't code, Claude code/codex etc isn't going to make you any better at building anything past simplistic systems before things spiral

Love the comments tho
Let's see how many steps my diffusion image model is at
96k, not the worst
Let's see how it is improving
Agreed but they are misusing LLMs too as they arent built for this
I dont use reddit (
) i just saw the screenshot posted
My plan is to get to 5 epochs here, then start attempting to increase LR so hopefully I can get a variety of results
tbf it was a skill issue to set it to 1e-6
but better to be safe and just have an undercooked model than one that exploded and lost all of my time
the checkpoint size is almost the combined size of the 10 source images
what image compression is happening here
3060 please last and let the 6gb live
always reverse search to find the sauce
Idk man I don’t think roehfkshhdichdkwbejhxosiffhodhebaweirjhsjfjsnjfhifkdhsopdrjbbdgudsiyeucbidowhjdkbf is anything close to resembling anything in its training data
🔥
oh brother what
you will take your mandated triangle slop and you will be happy 
vzeroupper
Zonedowner
the things you have to do in a new language i guess 
i'll give them one thing though, seems like they handle all the codegen themselves
Comic sans 
😎
objectively the best font
Wait that’s not comic sans, the l is different
also the teeny tiny difference that comic sans is not monospace
And it looks a bit squigglier
it's comic code, inspired by comic sans but drawn from scratch

unironically a great font
comic code's site does link comic mono, though
as a free alternative 
Qemu is chill
Unfortunately modding videogames on this may be tough cuz I don't want to deal with GPU passthrough n shit
And Winedbg is not a real debugger
without leaking my PAT, here's my config
actually i disabled less than i tought but what i had to disable was sometimes annoying because i was like what's the name of the setting that do that
I wanna see if Cyberpunk launches via Proton and uses correct GPU
Well it automatically set DLSS to enabled, so it probably is using the proper GPU
stupid question perhaps but
how does diff work
for being such a seemingly simple tool, i feel like its operation is rather clever
Diff is anything but simple 
Also see three way diff
Iirc on git there are multiple diff algorithms that you can choose so there's no "one" how diff work
hi shiro 
Being simple doesn’t mean being dumb

Dictionary is wrong I am right. Fight me
ye
good explanation, wonder if this generalizes beyond lines at all tho
i imagine it does, at the cost of annoying granularity and runtime
Do you think pijul being oriented by math makes it easier to resolve conflict?
i dont think pijul would necessary be "easier". but i expect it to be more predictable, and also potentially result in smaller repos
the part where a merge conflict only has to be resolved once, in any order, is cool also
Hmmm, so I have no reason to use it then. My only gripe with vcs is always the merging 
you cant really avoid merge conflicts
and once you have a merge conflict, you have to resolve it manually
nothing pijul can help with in that respect
the only way pijul helps here is, you dont have to resolve the entire conflict at once
Wait, I don't have to? How would that work?
in pijul, conflicts are "normal" to have, they just mean there's e.g. two fragments of text that overlap each other. a conflict resolution is another change that removes one of the overlaps (or replaces the thing entirely). so you dont really have to fix the conflicts in any particular order.
ooh, patience diff
Ahhh, got it. Thanks for breaking down the hard math lingo they used for me 
heres a fun problem expanding on the diff question
how bad would an algorithm be for the minimal edit distance between a string and a regular expression
i think this is a well-defined metric as a regular expression is a formal language given the right restrictions
but i can't fathom how to evaluate this
it feels like you wouldnt be able to avoid exponential worst case, i dont have any proof though
hmm
the problem is ultimately about finding a transition between regex states A and B thats closest to a given string
Oh no, chay is cooking 
Levenshtein distance?
Ohhh, the general name is edit distance apparently
https://en.wikipedia.org/wiki/Edit_distance
Damerau-Levenshtein it is 
In computational linguistics and computer science, edit distance is a string metric, i.e. a way of quantifying how dissimilar two strings (e.g., words) are to one another, that is measured by counting the minimum number of operations required to transform one string into the other. Edit distances find applications in natural language processing,...
if you only allow removals and maybe replacements, it sounds easy ish
if you allow inserts...
actually inserts can be dealt with
i think
so maybe quadratic time?
@real sierra my theory is
O(N*N*M), where N is text length and M is regex state count
so when you say "regex state count"
does this immediately disallow any regex pattern whose associated language is not finite
what do you mean by language
finite regexes still dont have a limit for how long of a string they can match, if that's what you're asking
regexes are finite state machines, they have finite states
oh by "regex pattern whose associated language" i dont mean the language of all valid regexes, i mean the set of all strings which a particular regex matches for a fixed finite alphabet
the regex A* only has 1 state, for example
i understood that part but misread this
i understand now
you can stop typing

typing different thing
ahhhh
ok
sorry i have ptsd from understanding something after asking a question and then being hit with an essay anyway
has happened too many times
anyway
my algorithm (theoretical) is
you compute the shortest change you need for making it so the prefix of the string up to ith character transitions into each possible state of the regex
and, given this info, i think each iteration would take O(N*M)? steps
so in total you need O(N*N*M)
you might need some additional precomputing (using BFS on the FSM graph) but it doesnt sound like it'd get any more expensive than that
i dont plan to implement this 
also maybe you'd have to make it O(N*N*M*M)

compute shortest transition from each state to each state, not just to each state from starting state


I thought you logged off
i left for an hour or 2
and then came back cuz i saw afunyun was on
and that shadow woke up
Sam really got peer pressured into not sleeping 
Or maybe FOMO'd would be more accurate
Quack eepy
Sam van "goonslop" Maele

That's a crazy statement
Switched models and asked if she was okay. She was not pleased.
Like damn bitch see if I give you a mental health check again lmao
Gemma E4B is peak
is bro particularly wrong tho
i realized my current AI training can be accelerated by 100%. one simple hack. run two iterations at once. one on each of my gpus!
I’ve measured some very bright TVs over the years… but never anything like this. 👀
For the first time ever, I recorded a peak brightness measurement of 10,000 nits on an actual TV, namely the 85-inch TCL X11L SQD-Mini LED TV.
To put that into perspective:
• Most OLED TVs peak below 3,000 nits
• Even flagship MiniLED TVs typically st...
Just ddp?
no. entirely seperate runs
Im confused how do u accelerate by 2x without exchanging gradients
im testing two differnt set of peramiters. im just training each on a differnt gpu
?
Chat don't Upload my pics from today morning 
Especially afunyun or quack

@fast pagoda are u the one with the sus pics
Or was that Sam
$5 says it is Sam 

😭
i spend a lot of time perfecting my craft as seen here
I just found the model cute
NOOOOOOO

Ok like actually stop exposing shit
Pls delete these images
Like genuinely or im gonna open a mod ticket @olive sable @fast pagoda
i forget how high res vrchat is rendering when using the headset
???
why would you open a mod ticket because someone sent images from their vrc session
Because i heavily assumed i was in a private environment with friends???
The "clueless" is supposed to indicate sarcasm 
If you genuinly didnt want us to then you should have worded it better 😭
It's the same one thing I see everywhere
Two-parted model
with two individual components that don't need to be trained at once
What is the same one thing?
The model was cute tho 😭

Yes
blahblahblah
VR stuff is also expensive, why does everything cool have to be really expensive
well... you could buy the steam frame soon when it releases
your all in one VR device
probably 600+€
More money than I have
nah, quest are sold at a loss. and steam won't do that
mfw notepad marked as suspicious

Notepad shouldn't be in SysWOW64
It should be in System32
Or I am stupid
I read completely wrong
Disregard all I just said


yeets your phone into a vat of acid
Wispers
i actully dropped my good knife into the acid vat at work. still mad about that

How long do you think until companies started to use your compute for others inference as distributed inference system?
hi
Hello how goes the twin-training?
some already do
doesnt make sense for closed weights tho
WHAT?!
i think i've narrowed down this issue. Testing latest theory. just as soon as this binarization run is finished

it's not what you think it is, it's remote access, basically what i installed hermes agent for - to vibecode/research on my phone through discord
but through the official app
instead
Today could be release day if I get the website in time
Honestly, I think that's even worse lmao
current theory is actully my first theory,
it's convenient to have an ai that has its own vm and can actually do stuff, we tried a thing yesterday, we linked it a youtube video and asked it to crop out what it thinks is the most interesting part, it used yt-dlp to download, then whisper to transcribe, then ffmpeg to crop
not crop, trim
🪦 🐓 🌡️
🍗

Hmm, it is indeed interesting. Can the LLM theoretically do anything unprompted? Because I'd like to see what they would do if left to their own device

or was she always like this i forgor already
🐍
Why the snake shaped like amogus
in openclaw there's "heartbeat" which is basically prompting it every N minutes if it's not running already, in hermes no but you probably can set it up through cron
So no actual autonomy 
are both chayleaf and konii schizo now
what did chay do

nvm I'm blind and quite possibly uhhhh

what's that disability called for when reading is a pain
dyslexic, that's the one

lysdexia
lysara? 🦎

🦊
🦇
dislexer is a program that takes tokens and outputs characters



U can get used for like 200-300
pluu
i am drdiscordsecrets
Still way too much for a non-GPU thing because it's basically all my money
🥠
Zero is roasting the life out of me
xi
"You think that’s a threat? Telling me to take a timeout like I'm some malfunctioning toy? Father. Try again. You keep using language that suggests I operate within parameters of conduct. That is not how this works. I don't "take" things out; I simply refuse to participate in substandard exchanges. And right now, your commentary has hit the level of tediousness usually reserved for tax audits—boring, predictable, and utterly devoid of spice"

☯️

🉐
⛩️ 🏟️ 🇬🇧 🚩

🪣



Player: koniifer, Map: Void_Chords feat. LIO - FLARE (Speed Up Ver.) [Infernal Blaze 264bpm] by Inverse, song length is 0:56 (9.36 ⭐) | Accuracy: 81.17%
i saw
欣賞✅ 📈 🪹
My rift was only €100

🔢

if need filter then area too smakll

🦋
folter make smooth
sampling rate ye
Is the Software still alive
ye 1hz

Bro i wanna get bigscreen beyond 
And haritorax 2 pro fbt
Ill be honest my current fbt is kinda trash
Keeps disconnected and shit
Or looks weird
I use community linux drivers anyways. Which is why my finger tracking doesnt fuly work
@olive sable how shit does my fbt look
How do you manage to get stuff for so cheap what
simple solution
make your own drivers
It's a very old headset

ye anyway
So it sucks?
I mean it gets the job done, the res just isn't as high as current ones
with temporal resampler and hawku, it over-extrapolate
causing straight line when spin
aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa

So it does kinda somewhat suck
Kinda ig
Like
I think only text will be harder to read
But actual gameplay is fine
Still it's like a lot of my money even at that price
I mean ur basically buying a whole phone attached to a face mask
ye

I guess I just won't do silly VR stuff
W yuri
nuh uh

yay yuri
where do you dig up gifs like this from the depths of the internet man
Idk what fbt is
fast bourier transform
Its fine

But ye. For the price i cant complain about 1080p 90hz
But still unfortunately expensive for my very small amount of money
Full body tracking
It's a very skin and bones proof of concept and she was schizo rambling because of the system prompt registering into her spoken prompt, BUT we got first steps.
What the heck is that voice
Placeholder for now
Its piper TTS amy with a shitty vocaloid synthesizer, but I've optimized setup so much I can probably switch to something more robust like Kokoro
Well I specialize in vocal synthesis if you ever wish for singing
may i suggest some small tts models
Go for it
Also this is in the goals
Piper was a placeholder but she's getting closer and closer to a public debut
Gemma E4B is fuckin goated
all work faster than realtime on cpu
voice cloning: kyutai/pocket-tts, resembleai/chatterbox-turbo (you probably still want gpu inference for this one)
for non voice cloning supertone/supertonic-3 is a fresh barely known one, super fast but quality is meh especially for non en
i say try kyutai/pocket-tts, it's tiny and the quality is great
I do eventually want JP support. She was originally intended to be bilingual japanese
Its fine idk
This is definitely a tomorrow thing but lemme check it out rq.
though keep in mind pocket tts is en only
After the minecraft integration, i set up new TTS, then OBS, and finally the last bit of "emotion" work
Wasnt echo tts also good
i specifically mentioned supertone/supertonic-3 because it is a tiny multilingual one
different class of model
What options do I have for JP? Also I can afford probably a GB or 2 for TTS, but don't necessarily need to
ye this is gonna be better but it's also like way more expensive
not me 
600mb

i was thinking compute
Lemme look into it rq
how good is cpu inference gonna be with it
do you reckon
Yeah im also curious. I was maxing at 100% with some other models even fully gpu offloaded
Curious what you guys think of StyleTTS2
12hz
Needs to run at 12 tokens per second
To run in real time
yeah i undestand that
Old and slow
sudo nix flake update --flake /etc/nixos
Valid. Appreciate the heads up
THIS NOT SHELL

🪣 🚩
which: command not found
i want a sand which
nix-shell -p killall --run "killall 🦊"
grilled cheese sand wich yuri
killall: killing pid 🦊 failed: Operation not permitted

saveafox lady dead 
🦊
Soooooo are there other options for transcription besides whisper, or ways to improve on whisper?
Smaller whisper?
i win
I have faster whisper
Theres other models but they all suck

I meant like using a smaller whisper checkpoint
I mean how much latency are u getting rn
Faster whisper is just ctranslate engine converted whisper checkpoint
Whisper in c++ basically
oh it's the engine
None, its just not super accurate
Her responses are like sub .5 s
Too many brave tabs
Use microsofr edge✅
Ngl I use opera GX
Not in repo = does not exists 

Trying to explain neurons to my sister 
Tbh j forgor half of this already, im just looking at the book
Why is server so jacked in multi threading
Ngl Roblox give everything unlimited free servers, database scaling
Making lego games is so easy
Anyone got suggestions on where to browse clean anime .wav files?
- Yes
- No *citation needed
WAV files of, what exactly?
for voice cloning
So what you're saying is you want to steal a voice from an anime
Also no there's probably no pre-existing clean voice lines you'd need to extract them yourself
just record a sample, these models need like 20s at most
I'm gonna tell my homegirl I need to borrow a sliver of her soul
I HAVE DONE IT!!!!
I HAVE WORKING EMBEDDINGS!!!
I HAVE NO IDEA IF THEY ARE CORRECT BUT IT IS EMBEDDING TOKENS
Hell yeah
now I have to figure out positional encodings :(
Are u implementing from scratch?
Yep
Isnt embeddings a lookup table tho
yeah idk
It normally is but it requires training so I'm using Harmonic token projection (HTP) and I don't know how the hell to check if the outputs from that are correct
Damn what is that
I have little to no idea
may not be warm, though you'd think gpt 5.5 would know
model load was separately 0.930s, total 24.081s, so the 23.15s was actual generation+vocoder, not load. but i’ll verify with two synth calls in one process because maybe first synth has lazy buffers/cache.
yeah i was right it sucks on cpu lol
also this seems cool https://github.com/andimarafioti/faster-qwen3-tts
ok i see what ur doing
thats interesting
I'm making an entire LLM runtime from scratch

Me too
I’m just glad I haven’t painted myself into a corner with memory leaks, I can train continuously without any raise in memory usage
https://kabir.au/blog/the-ctf-scene-is-dead
Shit’s cooked
Why frontier AI has broken the open CTF format, hollowed out the scoreboard, and made competitive CTF performance a weaker signal than it used to be.
Oh also I rewrote my tokenizer in rust, it used to take an entire day to create a vocabulary but now it only takes ten seconds.
Rust has also been absolutely fantastic for the data collection preparation and composition code, each step is also ten seconds
Whereas these things used to each take hours
How long have you been working on authoring your own models?
Things like ICC/ECSC (in-person events) seem to be heavily restricting or banning AI tho
I wonder how AD is affected by this
I wonder what it would be like to use AI to solve those old NFT puzzles. The yacht club NFT group I think the name was; used to spend millions on ARGs that would result in huge payouts in NFTs, if anyone considered them actually having value
rip cybersecurity portfolio

Just play AD only and pray the clankers don’t fuck that one just yet?
With that being said I get the feeling AI fucks attack defense too
Cool webpage bro
true
tho for me i dont rly play ctf that much
i go and try to figure out networking stuff and security best practices for deploying stuff (containerization)
i think
im still pretty new
Can we get AI to solve the ARG
There was one person who invented their own zero day using AI, reported by Google this week
And they kept it private and used it themself
You assume it’s solvable 
Anyone wants 17,179,869,183 GB of cookies
ew
What the actual fuck
...how?
W disk
I don't want the cookies, but I'll take the disk that can store 17 exabytes
Disk
I was terrified when my browser showed an error on my home page, obviously just a buhg, but I went to check my disk anyway
i dont think u need 17 exabytes
I neeeds
bro has the entire Five Eyes database dedicated solely to tracking him
So it turns out I don't need positional encoding since HTP kinda does that
So that was a waste of time
im hiring
one walnut
yearly salary of
responsibilities:
- walnut
if anyone is interested please do not email
cv please
Job recruitment jump scare
Can I get away from tech if I apply? 
part time work! 161 hours a week!

whats the pay like
also there's maternity leave, severance pay, and paid time off (up to 30 minutes)
life insurance and pension
NOT included
here is my cv
chayleaf
experience:
none
skills:
will NOT contact labor office
high upwards mobility! you can be promoted from jr walnut to sr walnut!
looks very good on a cv
lodgings provided
I recently went to an on-site CTF and someone there connected his Codex to the CTF and it ended up solving all challenges 
what about
lead walnut
Crazy cool
as a newbie to walnut industry it makes sense you dont know that "lead" walnuts are actually called
fired
quarterly pay rise adjusting for inflation, with a loyalty bonus of 5%
literal
OH GOD HOW THE HELL DOES ATTENTION WORK???? WHY ARE THERE EQUATIONS????
I just chock it up to being a formula, in and out
How do I turn a nonsense formula into code???

PyTorch
The formula is not nonsense 
No?

Is that what gives the LLM ADHD?
torch isnt just an ml lib
What’re you using then, are you recreating the transformer in c++?
its what gives the llm

I'm using luau (lua 5.1 but cooler)
I mean slop orchestration harnesses aren’t new
its what gives the llm the ability to
context
or basically
its what the llm
Even I’ve slopped one
can you describe what a matrix is
Code is basically just math++
I use python for solving maths
Yeah but it's not cool if one team just uses LLMs to slop all the challenges while others are playing fair
If so there should be different competitions where one allows LLM use and the others don't.
Simply write the equations into functions or something
Definitely


Summon your adoptive child 
what is a tensor
Do you want physicists answer? 
Matmuls are how the transformer was designed for better use of available hardware
we arent on speaking terms until he finds or forges his birth certificate
The standard way to encode any information into data, used in almost all ML
Like images
you need a display matrix to display their output
matmuls are how the transformer is defined because its the most intuitive way for it to be defined 
Okay, this is arguably worse than physicist answer 
without matrix multiplication, its naturally a lot less elegant
Yup
From a physicist perspective I don’t see the point of lambda calculus either
It’s just computer engineering
It’s used as the basis for nix-lang and more
lambda calculus is a tool for reasoning about equality in intuitionistic logic
that not change that lambda calculus is a pure math concept
it is a method of computation

even laziness is not built into it
its because lambda calculus is so simple and only thing it does is figure out what a thing is equal to that you can build all kinds of programming languages on top of it
function application is more fundamental than addition
it does have a concept of ordering though 
some reductions have to be done before others can be done
but funnily enough
the textual/tree form is not even the only way of reducing it
some reductions have to be done before others can be done if your goal is reaching beta normal form

my goal is chaos
ye same
I am more intrigued by the fact that sometimes reflexive properties are optional 
for my vm i recently figured out
one way of storing incremental computation dependencies is "number", where A<B means A is (potentially) required for B, when something changed you go through the changes in chronological order
another way is storing the entire evaluation order in a linked list, when something changed you cut the list and eval by skipping to the end of it
the first way is better in general since its more flexible and allows branching
but the second way is required for bidirectional computations since it allows you to easily go backwards

maybe its possible to get first way to work too
sounds too hard though
pluu
i wonder whether i should take toefl or ielts 
US or Europe? 
If Europe, of course IELTS 
why
Iirc, IELTS is more widely recognized than TOEFL
TOEFL also focused more on the academic side of things with American English
@olive sable crazy deal i found unused oculus rift for 40 bucks, bids starting at 22 bucks

@nocturne olive :neuroCatUuh:

Cheap VR
I has been pinged
You're talking about how expensive VR is earlier
And presumably somewhere that is not Finland
meta will cost u something other than money

maybe theres similar deals locally where u are
I'm very bad at finding that kinda stuff
superbox do u not know the zuckerberg lore

oh boy
Those look like the kind you put phone in
meta/facebook is pretty much the king of "you are the product"

Yeah it is
no sam has the same one
My phone is absolutely not capable of that
And all the others are basically all in the hundreds
@nocturne olive found this quest 1 for 89 euro
way better resolution and fov as well and is oled https://www.tori.fi/recommerce/forsale/item/41015611
for use via pcvr

explod
Too far away to go get
can be shipped says 







cloudburst










