programming

[opaque wharf]

Still on those slow coax?

[sage crag]

vdsl

[opaque wharf]

Noice

[sage crag]

the 8 bit model is 65% faster than the 4 bit model

the translation is better too

ok

[opaque wharf]

I wonder if LLM for translation have a different architecture than other more general LLM

[sage crag]

the translation is nonsensical

ow

[rough bloom]

they do, if you want a "natively" sequence-to-sequence model then you want a encoder-decoder transformer like T5
though decoder-only models (like most LLMs) also work, so in practice they win if you just use a larger model

[sage crag]

i tried to make it translate something longer and its taking its sweet time

also my pc sounds like its gonna explode

[opaque wharf]

Interesting

[sage crag]

all of my cores are pinned at 100

[opaque wharf]

But yeah, in hindsight it does makes sense it would need encoder-decoder

[patent shard]

You can't compare crf levels between encoders, svt-av1 and x264 are VERY different
Yeah, higher bframes won't always make a difference
I think qcomp has something to do with how much it compresses with motion? Or maybe more accurately how variable the bitrate will be depending on content

[sage crag]

oh, it started

it sucks

[opaque wharf]

I heard they are unveiling a "handheld" game console at CES with unified memory and Ryzen AI CPU. I wonder how viable it is to (ab)use those to run some LLM lol

[sage crag]

at 27 tokens per second its getting a little difficult to follow

that's something like 1000wpm, which is decent reading speed

[mighty thorn]

wrrrr wrr

[echo eagle]

Hi

[sage crag]

27 tok/s is 80% faster than 15 tok/s

i wonder why the 4-bit quantised model is so slow

[opaque wharf]

You tokenize by word?

[sage crag]

its probably because my cpu doesnt do 4 bit natively

[sage crag]

this is the simplest inference i could put together, its like, 10 lines, i have no clue if it goes word-by-word, but its approximated that every 1.3 tokens is a word

yeah its 74 tokens and 54 words

[trim valve]

my uni explicity allowed vibe coding on a group project

[hard delta]

ai won't replace my job
i will be a "look into the actual code" expert

(agents replacing me in 3 years: )

[trim valve]

I am hopeful that ai plateaus

[hard delta]

i tell every young person that they should vibecode as much as possible and never waste time understanding what they do

gotta thin the competition

[left solstice]

Yeah, qcomp uses less bits for low complex scenes and more for higher complexity scenes (stuff like grain, more motion, higher detail scenes etc if I understand correctly). Increases size overall but I still increased it slightly so I can use a higher crf (27,25) instead of lowering it while still having it look decent. Since crf is different between av1 and h264, If av1 was more reasonable in terms of time to encode and filesize then I probably would've done some trial and error testing to find a good balance of crf but I don't think it would be worth it on my hardware so I'll probably stick with h264

[sage crag]

output = ""
for chunk in res:
  output += (text := chunk["choices"][0]["text"])
  print(text, end="", flush=True)

i like that python supports inline declaration in the same way c does

it lets me write bad code

wrr

wrr

[fickle rain]

I mean Strix Halo has been a thing for a while now

It even comes in tablets like the ROG Flow Z13

[opaque wharf]

Yeah, and I want to know if some madlad has already used it for local inference

[fickle rain]

https://www.reddit.com/r/LocalLLaMA/comments/1m6b151/updated_strix_halo_ryzen_ai_max_395_llm_benchmark/

OFC handheld will be more thermally constrained prob

[quick condor]

Meanwhile I'm running 28t/s on my slow reasoning model, 35t/s on my core model and 40+t/s on my gate model... I don't run 70b simply because it's only 15t/s

[sage crag]

this model is

8b

https://cdn.discordapp.com/emojis/1455273822625333375.webp?size=48&name=mlntcandy

[feral cobalt]

peak llm usage (am making neuro integration for funsies)

[sage crag]

https://cdn.discordapp.com/emojis/1421763041074217092.webp?size=48&animated=true&name=MyHonestReaction~1

[trim valve]

redacts name
has vscode rich presence showing folder name

[feral cobalt]

fudge

welp

too bad

[trim valve]

[quick condor]

To be fair the 40t/s model is an 8b on an rtx a4000

Which 8b

[sage crag]

this model is 8b, 1.9gb in size, quantised to 8 bits, gguf, and its cpu inference

[feral cobalt]

just gonna uhh, quickly turn that off

[fickle rain]

I get the feeling Neuro doesn’t really need that

[quick condor]

I'm running 8b 6BPW HB8 (EXL3 quantization) it's about 7gb for the model but that's also from a pretty large 8b

[feral cobalt]

this was more or less just a joke to test that my sdk impl works (it doesss!!!!)

[fickle rain]

Well that’s great

[frozen hollow]

I want to use a series of agents with persona files to play test my board game prototype by making it with the react-based boardgame.io framework and an MCP or something.. I’m gonna have to beg a friend with a fast computer to let me run it as a docker instance or something

[quick condor]

Q8_0 is 9Gb (closest perplexity) , Q6_0 is 7Gb, Q4_0 is 5Gb and IQ2_XSS is 2.6Gb so an 8 bit 8b at 2Gb is wild

[frozen hollow]

How much power is needed realistically to do a low-rank adaptation on a model that’s good enough to call tools reliably (ig that’s the core of neuro but I’m only guessing)

[nocturne olive]

1.9GB 8B? That'd be Q2

[quick condor]

Depends, with sharing weights across 2 3090tis and an A4000 I can do loras on 24b at full weight with ram offload (60Gb offload and 60Gb to write layers) but smaller models using qlora can use less

[frozen hollow]

I’ll check what he has It might be a 4090 and 128 memory

[nocturne olive]

If you can run a model, you can usually finetune it

It's all on the memory

Now pretraining is a whole another thing

[quick condor]

For neuro size (my best educated guess Llama2 13b) likely 4090 with 128Gb ram is fine

[nocturne olive]

That's like more than enough

[frozen hollow]

A prompt file might be enough for what I need it for this is basically extra credit

[fickle rain]

Would DGX Spark or the like run better

[nocturne olive]

24GB of VRAM can hold almost 2 13B models

[quick condor]

Usually you want raw weights for training not quantized

[nocturne olive]

For LoRA you can finetune even on a 4-bit loaded model just fine

8-bit base model is plenty

[quick condor]

There is a difference between "you can do it" and "what is ideal"

[frozen hollow]

Basically I’m using it for actual design work but making it fancy is just for me

[nocturne olive]

The final LoRA can be applied to whatever variant of the model anyway

You're just obsessed with wasting as much memory as possible on marginal quality increases

[frozen hollow]

Persona files (Timmy Johnny spike kinda thing) and a running agent loop for each on the same model is my mvp

[quick condor]

It depends. When doing individual tasks that are isolated and not reused, yes I can see your pov but I'm working on a long horizon project where outputs will be fed back in and as such marginal quality is actually exponentially stacking as memories build on memories

[nocturne olive]

Hem
Training LLMs on themselves

That's gonna be not great

[quick condor]

Something like ST could handle most of this and would just make api calls to the model

[quick condor]

Yes and no. I mean memories get injected as context, errors in memories causes defects in output which then become new memories. That's the problem.
For training I am using model output but that is coming from 50/50 model user interaction which is actually ideal for rlhf

[frozen hollow]

ST?

[quick condor]

Silly tavern

It's a lot to set up but solid once set up

[frozen hollow]

Ah I will look thanks!

I’ve heard the name but I didn’t know people used it for anything other than like.. roleplaying or something?

[quick condor]

It's solid for building persona profiles and can handle multi persona switching in a single conversation

[frozen hollow]

I also wanted to look at fine tuning a vector embedding model (for a diff project) but I found such a good one it’s hard to justify the effort realistically

Considering I can run both the encoder and decoder models on my 16GB MacBook m1 it’s probably more a question of data curation

[floral wraith]

i want to run AIs at work the only problem is that we dont have gpus at work we have like 4tb of ram and 6 servers but 0 gpus and we cant get gpus

[floral wraith]

servers hpe proliant gen 8-10 and others

[wary mauve]

probably been asked a bunch of times but im getting into models, LLMs, RAG and the whole works and ive hit a bit of a problem

so to be blunt im attempting to create something similar to neuro, not neuro as a personality but the architecture she encompesses just for friends in a discord server

ive currently got it working so its a bot that joins a discord vc, ive hooked up each input stream from each user into a joint whisper queue ad ive got a decent mvp

now im trying to have the ability to allow it so the llm itself / the bot, is able to perform "actions"

such as, send_discord_message, play_sound_effect, leave_call, join_call, mute_member, unmute_member, etc etc

now i initially added a prompt injection that explained "actions" and how they should be used and when to use them, in the system prompt, and also examples and then the list of available tools

now this was good but my llm output is straight text so it'd be

"
My speech is here and this is what I say

[ACTION:name key=value key2=value2]
"

Which was fine but sometimes the action was malformed, which is okay I know LLMs arent guaranteed and giving you exactly what you want or what you specify

so I changed to using "tools" which some models are trained on and it seemed like an easier system because it was json, I could be more verbose without adding more noise to the system prompt

but ive just ran into a BIG issue, my assumption was the LLM would respond with text content + tool_choice (whatever the property name is) so it could kill 2 birds one stone

but i was wrong and it can only either

A) give a response of what the bot wants to say
B) tell me that it wants to run a tool

now I hate this because say I want it to react and run a tool, for example

"ive just sent a message" -> also runs a tool

anyone got any good recommendations / techniques I could use to improve this? been non stop workingn on this bot since he was born last saturday so slowly losing my will to live but its a fun project

tl;dr -> need to figure out a good way to run code that the LLM can tell my system to run and also respond with what he wanted to say too

SORRY FOR THE DUMP

thought this was the best place to ask

[stuck ocean]

Ig u could run sum small models with just the cpu but thers nothing much u can do without a gpu

[wary mauve]

ive heard of mcp too but didnt give it too much of a read ngl

[frozen hollow]

Think you should be giving it more a read here lol

[wary mauve]

i got suggested it yesterday by a friend ill give it a looksie properly now but i thought id get a general concensus from peeps who are 100% more skilled in this field

[floral wraith]

i mean the problem is that i could run a few okay models with my own gpu and 3 servers in a cluster but like i want to make an ai that you can talk to like to neuro but it has informations about our clients and some more things

[stuck ocean]

It's never a good idea to give a llm confidential/ private info but ig u could try fine tuning a 2b model or training one

[floral wraith]

i mean why not when we only have it intern and it has no Access to the outside

[trim valve]

[stuck ocean]

Well in that case giving it a copy of the database would be fine ig

And what do u need the bot to do exactly

[floral wraith]

yea that is not one database it would get access to a copy of a few databases like the internal wiki usw we have a few years of work

[floral wraith]

respond with logical things from all the internal things so you dont need to look it up like you cand find things in our database

[stuck ocean]

Ye u should be fine as long as the llm dosent have explicit perms and can't leak anything

[proven merlin]

You could run the biggest models.. just awfully slow :D

[frozen hollow]

You’re basically gonna be running a little service with all of those functions exposed as endpoints and then give your llm access to those functions through the protocol so I think it’s basically what you’re looking for here

[stuck ocean]

In that case u should be fine with a small model depends on how big the database is ig

[frozen hollow]

Also I guess I should buy GLM4.7 access for my little persona project? Has anyone used it for that kind of work?

[stuck ocean]

What's that

[wary mauve]

makes sense, ill give it a play with now, was also just thinking of using a model that supports json output and forcing JSON output, my only concern is, one malforfed bit of json and its just silence, which "feels" like high latency if I run a retry

[floral wraith]

i mean would be nice when u help me a bit but i ca ntry it

[stuck ocean]

I should prob learn more abt llms i barely know anything and everyone seems to be using it now

[stuck ocean]

Sry but I don't really know much abt llms myself lol

[floral wraith]

perfekt but do you know how to set them up

[stuck ocean]

Not really ig I only tried a few models with ollama

U should look up sum tutorials or ask someone who knows more

[frozen hollow]

Like.. Chinese knockoff Claude - but specifically you can get banned from Claude and I assume Google trying to do what I’m doing and the GLM stuff supports what I want to use it for

[floral wraith]

i mean i work with servers 24/7 i should get it so yea tutorials and than freee

[stuck ocean]

Oh lol and what are u trying to do

[stuck ocean]

Good luck ig

[floral wraith]

thx

[frozen hollow]

Playtest a board game im designing with different “characters” (play styles) for each AI player so i dont go crazy having to take everyone’s turns

Idk if it will work but that’s what I wanna try

[stuck ocean]

Oh that sounds cool

[frozen hollow]

Can’t get them to test my rules sheet for readability.. LLMs are too smart

They can read terrible instructions and figure it out anyway

[stuck ocean]

Lol

Well uh anyway I gtg config my waybar for 6 more hours lol

Cya

[nocturne olive]

Try returning something from the tool immediately for the model to "react" to, then if the result is latent give it the result info later once it exists

[wary mauve]

yeah I just ended up doing that, I specify a "say" parameter on each tool with a description that basically says put whatever in here that you want bot_name to say

and as for latent info ive got that too with my tool setup, like for example if its a web search, my generic Tool class has a returns_data property, which i check for, and if it exists I dont emit my LLM_RESPONSE event I just rerun the prompt but with the new context that it was asked so and so and you did it and here were the results essentially

well thats how itll work, ive not done the reprompting or context part yet

[nocturne olive]

I wouldn't put what to say as a tool parameter, rather give the LLM a turn after executing the tool

[wary mauve]

Well that was the second option but that introduces more latency as it has to go through a second LLM request which is slower than the tool parameter way even if its a bit hacky

Still need to dive into latency as my model isn't local but using openrouter for now as Im running faster-whisper (distiled model) + a vad and thats already using like 5GB vram out of my 8GB

Secondn LLM request I'll do if theres data that needed to be fetched by a tool

[nocturne olive]

8GB is very not ideal indeed

And I guess with a cloud model high latency kinda makes sense

[wary mauve]

yeah exactly, i mean avg response with deepseek 3.2 1.8s so not good, not to mention I check for silence and such

3070ti, looking to upgrade to a 4090 this year but with the costs of RAM and GPUs rn might not happen

or i bite the bullet

[nocturne olive]

For ML just get a 3090

24GB VRAM, inference performance almost same as 4090, and prices as low as 500€ each

[wary mauve]

had no clue might look into it, i do game still so a 4090 would still keep me in the loop for the next few years

well youd hope

not with ue5 games 😏

[nocturne olive]

Yeah 4090 is faster in gaming than a 3090

In ML though unless you're doing training the 4090 doesn't have significant benefits

[wary mauve]

well i do eventually want to fine-tune models for personality traits similar to what i assume vedal does just to learn and trty it, but i can always use a service like brev from nvidia

[wary mauve]

most i can see for a 3090 is like £800 😭

looking at pre-owned i lie

[nocturne olive]

There's this one triangle in #programming that knows where to get really cheap ones

[wary mauve]

how does one get in this triangle 😏

[floral wraith]

real need that 5080 for under 1400€

[amber fractal]

The triangle only deals in 3090s because all of the others are inflated to hell

[floral wraith]

real tho but for how much could someone get a 3090

[amber fractal]

600 euro last time I asked

some 500 ones exist

[floral wraith]

i mean gobbo you can see there is someone that knows where the triangle is

[amber fractal]

(triangle offline )

[wary mauve]

dont

come back

might have a friend who can get them cheap

but need the funds first

so waiting to start the new job

then when i get paid

its so over for my 3070 ti

[amber fractal]

I'm not the triangle man, just a friend of his

[floral wraith]

ikik

[amber fractal]

My forcast is if you get paid in a month+ you'll be paying 600 for the 3090 otherwise might be able to get 500

[wary mauve]

but the real question is from who

but i assume u mean

cus prices will go down

[amber fractal]

@olive sable gpu dealer

[wary mauve]

oh hey haha...

so uhm

got a gpu?

[opaque sigil]

[amber fractal]

still offline, will be on in maybe 3 hours?

[wary mauve]

naw but ill ask my friend before i pester peeps i dont know or dont know me

ty

[opaque sigil]

The absolute state of this man's sleep schedule

[amber fractal]

What sleep schedule?

I only know he sleeps

[opaque sigil]

True

[amber fractal]

To be transparent, the used market in my boi's area is just cracked

[floral wraith]

how hard is it to get only the vram

[wary mauve]

not hard at all

and its cheap af

but

you need to solder it on ( heat gun ) which isnt the worst part

the worst part

is that

the firmware

wont support it

ad writing custom firmware / patching ca lead to bricked gpus pretty sure

and its generally

super hard to do

so

so it is hard

[floral wraith]

ik but i think i know someone that has custom firmware and custom pcbs

[wary mauve]

well the firmware is possible the pcb part no clue sounds pretty op

[floral wraith]

they can make them and he has the model for one

[opaque sigil]

it's doable provided someone at nvidia has previously leaked a compatible vbios

[floral wraith]

but idk for what board and vram and idk if the firmware fits the pcb

[floral wraith]

i mean i cant do it alone but prob with help from other ppl it is possible

[amber fractal]

That reminds me of the 48gb 3090 cards

I imagine that is how those came to exist

[floral wraith]

i would love to have something like that

[frozen hollow]

yeah that sounds sick

[fleet jungle]

Man, GPUs these days have become lards

Oh. I'm starting to sound like a cynnical boomer. I'll just stop myself while I'm ahead

[frozen hollow]

idk lol it's clearly gone a bit nuts when you look at the port and the assumed support structure and the voltage of the PSU

I'm surprised they haven't gone up in voltage to supply less current or something to reduce melted connectors? I don't really look inside my computer much

[rigid snow]

yo programming chat

it was too far to friends home and too cold and dark and late

he's staying over at mine

he wanted to test the 5070 in my pc but

case WAY too small

[nocturne olive]

You mean 4090D?

[rigid snow]

it's regular 4090s too

they mod

[nocturne olive]

Either way I want one
But expensive

[olive sable]

https://tenor.com/view/azumanga-daioh-gif-2910224911933025207

[olive sable]

[olive sable]

[olive sable]

3090 for 500 euro + shipping

3060 12GB 110 euro + shipping

3080 320 euro + shipping

[rigid snow]

[olive sable]

that 3060 is actually a good deal

[fiery anchor]

sounds pretty cheap.

[olive sable]

i can also deal cpus ai guess

7800x3d for 280 euro

[amber fractal]

tbf you were eeping

[olive sable]

ye

i can also do ram i guess with curreent pricing

[wary mauve]

😏

oh really

now

[olive sable]

4x16Gb=64Gb PC3-14900R 1866MHz (Serveurs/Mac Pro) € 70,00

ddr 3

[wary mauve]

ddr3?

😭

nno pls

does speed matter though for LLMs / AIs

[olive sable]

for ai i think you'd want to use vram

[amber fractal]

honestly, not the worst

[olive sable]

its sad that ddr3 for 70 bucks is not the worst

you used to get 2x32gb ddr4 brand new at 100 euro

[amber fractal]

compared to ddr4? that is a sacrifice I'd be willing to make

[wary mauve]

yeah ofc but if VRAM isnt available it goes to RAM but i sort of also meant if ram is useful outside of that

[fleet jungle]

More vram the better for LLMs, but if you have even just an ordinary dedicated graphics card, you should be relatively good.

[olive sable]

if it flows ovcer into ram its gonna be really slow, no matter the type of ram

iirc

[amber fractal]

That is correct

[wary mauve]

fair ive heard so

[amber fractal]

the process of fetching is heavy

[fleet jungle]

I've run gemma 3 12b quantized to some degree on a 1070 and it works alright.

[olive sable]

HP 4x 16GB 64GB 752369-081 774172-001 726719-b21 ddr4 ecc € 140,00
cheapest ddr4 64gb i can find

its ecc even

[amber fractal]

[olive sable]

seems to be 2133mhz

[amber fractal]

decent for ecc

[fiery anchor]

and not necessarily a lot slower. For general compilation and stuff more than usable.

(vs ddr5)

[wary mauve]

we love elevenlabs

https://medal.tv/games/screen-capture/clips/lUhNObkmN14NPDuRz?invite=cr-MSwxdmwsNTgyOTU4MjA&v=32

shame its so expensive

[olive sable]

the cheapest ram i can find is all ecc

i guess not a lot of people have ecc support

[wary mauve]

pretty thats close to mine rn anyhow

[amber fractal]

better grab a xeon then

[wary mauve]

i need to enable xmp or whatever its called i bios

[fiery anchor]

prob. salvaged from old buisness pcs and notebook that don't support win 11

[wary mauve]

forgot the name

i just got my new work laptop thats got 32GB DDR5 but its a laptop so its probs soldered

[olive sable]

i think overall its still a better deal to buy a 2ndhand prebuilt pc. before the ram crisis it was only slightly more expensive, but now even getting ram that fits is pricey

[wary mauve]

which i have the tools to remove but id feel like even if u ca make it compatible for desktops it'd be aids

[amber fractal]

[wary mauve]

i tried to get a framework

[olive sable]

it could be soddim, but its doubtful

[wary mauve]

but framework 16s are all on batch

or well

decent ones

https://www.laptopsdirect.co.uk/lenovo-thinkpad-p14s-intel-core-ultra-9-32gb-ram-1gb-ssd-14.5-inch-windows-21g20030uk/version.asp

it is sodimm

[amber fractal]

framework 16 didn't exist when I was hunting, so had to go for a levono boi

[wary mauve]

not sure wehat that means

[amber fractal]

I do have a 13 tho

[fiery anchor]

bought my kit for 140€. Same kit Today is 612€

[wary mauve]

sick

[rigid snow]

wait we spent how much i wonder

all new parts

[frozen hollow]

are the higher end macbooks any good for this? I need a lot of memory anyway just to run stuff but if I can justify 64 on one that'd be cool. I guess they suck for training?

[olive sable]

these are available for 130 euro in my school's discord, i jsut dont have a ddr5 laptop

[frozen hollow]

Sodium ram!

[amber fractal]

ddr5 as well

[wary mauve]

goated

[rigid snow]

$1.7k

[olive sable]

https://tenor.com/view/jimmy-neutron-skeet-sodium-chloride-dude-actually-dude-its-salt-nickelodeon-gif-15948213436140545826

salt

[rigid snow]

not counting the fucking case which he already had

[rigid snow]

$1.7k build factoring in ssd and ram being like $250 more expensive than they should’ve been

[olive sable]

who's pc?

[rigid snow]

in a LIAN LI O11 DYNAMIC EVO XL

fuck my life

[amber fractal]

only ddr5 device I have is a N100 mini pc.

[rigid snow]

my friends

but i’m also pretty excited to see how the 5070 performs

shits supposed to be crazy

[fiery anchor]

for cheap parts always check what local businesses are throwing away.

[rigid snow]

same price new as a used 3090 here

[olive sable]

i got mobo + 5950x + 4x16gb 3600 ram for 300 bucks
3090 for 600
psu for 50
case for 70
cpu cooler for 50
and i recently bought new storage so 600 euro in that

[rough bloom]

speaking of 50 series performance
6x frame generation soon

[rigid snow]

ok so the funniest part about all this is

bro plays esports titles and fucking valheim

[olive sable]

totals 1670 i guess

spent as much on storage as on gpu

[rigid snow]

but then he’s maining deadlock which should probably benefit a lot from a nice pc

[olive sable]

[rigid snow]

for context i have like 150 fps stable and drops to 30 in intense battles

fights

“battles”

[wet monolith]

@burnt marsh someone asked about this before and vedal's direct response was "you can't rn" #programming message
so it may be the case in the future, but for now this is how it is.
also,

No chatting here
per the channel topic in #merch

and technically it wouldn't be reverse engineering if the app was open-sourced. lol. the protocol would be right there for you to see

reverse engineering would be e.g. capturing packets to/from the lava lamp

[burnt marsh]

wrong word, mb. i could easily make it myself and take colour from stream without swarm sync, but that app would be nice to use

thanks for the answer!

[wet monolith]

i agree though, people could probably make some really cool stuff with it

[burnt marsh]

N E U R O T O I L E T

[wet monolith]

lmao

[burnt marsh]

/jk

ill reply to vedal in like 6 months and see if the answer changes lol, [unless that would get me banned, if so then nvm. appears to be fine in rules, as long as i dont send notification to him]

[wet monolith]

can't say for sure whether or not that would get you banned lol. you'll be trying your luck anyway

[burnt marsh]

ill ask the mods beforehand, but thats a job for august me

[wet monolith]

i recall he said that it's technically just a smart light. the real question is the firmware.
i'm not sure of the specifics, or whether or not the phone is still required after pairing. it obtains the color from a remote server, that much is obvious. and with negative latency, might i add; although that's probably just a side-effect of streaming (an action performed irl always appears on stream a few seconds later)
as for the protocol, no idea. it could be repeated http requests or it could be a websocket. it could even be a raw socket for all i know (that is, i don't know.)
anyway, your best bet if you really want it is to find someone with a lava lamp that also is willing (and knows how) to capture traffic between the lamp and the internet. from there, you read the traffic using something like wireshark; something that can recognize patterns in the data. then you can really start figuring out how things work from there

[burnt marsh]

i was planning to do that but i dont know anyone with a lava lamp lol

im pretty sure im one of like 4 people in my county who like neuro enough to buy one

[real sierra]

[burnt marsh]

hi shiro

[wet monolith]

lol hi shiro

[real sierra]

hiiiii

I have programmer's block

[burnt marsh]

oof

[real sierra]

I need to just write code

[wet monolith]

doesn't have to be anyone irl
can be someone in this server. they can just send you the pcap file

[real sierra]

too much worrying about if it'll work or not

[wet monolith]

too real ^

[burnt marsh]

ooh fair point fair point

[azure lynx]

write code to help you determine which path of pain and sorrow you should spiral down into: gambling, drinking, video games, Netflix.

[wary mauve]

fuck it, if doesnt work thats half the fun, figuring out whats the problem, yeah it can drive u down a rabbit hole for hours but youll learn from doing so

[fiery anchor]

if you ship something, you can fix it. if you ship nothing, you have nothing.

[wary mauve]

^

[real sierra]

I get stuck in this loop a lot

most bad ideas are quickest spotted by good planning, but some are quickest spotted by trying them

[wet monolith]

i just have no motivation. not sure where to pull it from

[rigid snow]

NOT listen. .shipping garbage make your image garbage

[wet monolith]

last time i wrote code was probably like a week ago lmao

[wary mauve]

im a perfectionist when it comes to my code, but ive learnt to loosen up a bit, not everythig i write needs to be perfect on the first try, get somethig basic adn workig and iterate as you see fit, if you dont like looking back at some code u wrote yesterday, you refactor it, boom progress

my n key is so fucked

[wet monolith]

holy i relate to this too hard
i need to learn how to avoid making everything perfect the first try

it puts a real damper on my motivation i think

[frozen hollow]

What actually was the question? I assume you asked it in the merch channel but deleted it?

[burnt marsh]

just asking if vedal could open-source the swarm sync app

[wary mauve]

yuh its a horrible habbit but just tell yourself you can always implement a harder feature later, youll always iterate even subconsiouslly

just let it happen over time

[frozen hollow]

Are you just wanting to jack like a nano panel into the neuro light system

[sage crag]

https://cdn.discordapp.com/emojis/1154436961277313126.webp?size=48&animated=true&name=evilSlam

https://cdn.discordapp.com/emojis/1154436961277313126.webp?size=48&animated=true&name=evilSlam

https://cdn.discordapp.com/emojis/1154436961277313126.webp?size=48&animated=true&name=evilSlam

bang bang bang

[wet monolith]

[fiery anchor]

always depends on the application and risks connected to those. iirc is it a game in this context. -> risk is almost non existent. A pacemaker would be a different story, or llm training destroying data and progress.

[rigid snow]

as long as it actually does what it’s supposed to and if it involves ui it doesn’t make the user gouge their eyes out with different types of cutlery and workshop tools then yeah ship it. maybe it doesn’t do it well but you can fix that. can’t fix bad impression that’s forever stuck with you

[sage crag]

https://cdn.discordapp.com/emojis/1154436961277313126.webp?size=48&animated=true&name=evilSlam

[tender river]

[sage crag]

https://cdn.discordapp.com/emojis/1154436961277313126.webp?size=48&animated=true&name=evilSlam

bang bang

[burnt marsh]

i want to make a neuro lava lamp because i missed the preorders

[rigid snow]

omg that’s me

[sage crag]

https://cdn.discordapp.com/emojis/1154436961277313126.webp?size=48&animated=true&name=evilSlam

bang

[real sierra]

i think at the end of the day, my developing woes are self-inflicted

[rigid snow]

after walking the electronics market for two days

[wet monolith]

knock off lava lamp producer xdx

[burnt marsh]

id just make one, hell, ill pay vedal for the code if thats what it takes

[olive sable]

hi shiro

[rigid snow]

step 1: obtain official lava lamp to reverse engineer

[wet monolith]

pay vedal for the code and sign two NDAs

[burnt marsh]

if i could obtain that i wouldnt need to make my own

[burnt marsh]

if thats what it takes, sure

i just want my neuwo lava lamp

[rigid snow]

that’s the joke

[wary mauve]

https://medal.tv/games/screen-capture/clips/lUapufr6fatjS0TC5?invite=cr-MSxJOEksNTgyOTU4MjA&v=22

see this is the pay off for writing code instead of not, get to experience e.t learning how to type 😏

[real sierra]

im completely new to unity and i'm trying to make a melee attack for a multiplayer game, and while i can kinda apply experience and say "okay, so i'll need some kind of animation for the attack, and then a hitbox of some form for the weapon, and i need to make sure these calculations are run on the server", you could watch a 4-hour tutorial video on any single one of the components involved before getting a good foundational understanding

[tender river]

you cant get lava lamp data without a lava lamp

it because

it uses

tuya servers

[olive sable]

i wonder if they would ever make a second edition of the lava lamps

[tender river]

for synchronization

[real sierra]

in lieu of that, i've been trying to trust my instincts and go with "yeah that sounds like it should work intuitively" and applying the same paradigms i have for other engines

[wet monolith]

lmao yessss

[burnt marsh]

this is confirmed?

[real sierra]

and it always leaves me with some element of "idk how any of this works but if i wrote the engine, i'd imagine it would work this way"

[tender river]

yes confirmed by me

[real sierra]

and i think thats why i have programmer block and cant trust my own plans enough

[fleet jungle]

now shipping with explosives so you can skip the stove heating step

[real sierra]

[wary mauve]

dont go that far down, learn what you need to learn to get your prototype done

[burnt marsh]

[wary mauve]

then if u are really interested

[wet monolith]

what is a

[wary mauve]

read about it

[wet monolith]

tuya server

[wary mauve]

and explore it anothe rtime as another project

[tender river]

tuya is

iot thing

[wet monolith]

[sage crag]

https://tenor.com/view/borzoi-snoopa-lord-foog-dog-funny-gif-20125052

[tender river]

https://tuya.com

[real sierra]

even what i need to know is quite a lot

[burnt marsh]

i know what tuya is

[wary mauve]

it is but do it system by system

first

[real sierra]

well that's what ive been trying to do

[tender river]

you werent the one i was answering to

[wary mauve]

do it in singleplayer

first id say

[real sierra]

too late

its already multiplayer

[wary mauve]

😭

[burnt marsh]

and that would make it very hard to make a custom one because those servers are encrypteddd

[wary mauve]

okay change it, or make a new unity and learn how to do the parts that make up your system, break it down

[real sierra]

see that takes too long tho

my motivation is gone in like 1.5 weeks

[burnt marsh]

oh fuckfuck mb

[frozen hollow]

Yeah this is a lot at once even though it’s “just a weapon swing” can you break it down into smaller testable units

[real sierra]

thats the deadline

[real sierra]

not really, thats why this is hard in particular

ive already tried to do that and so there's existing logic for characters' health and hitboxes

[frozen hollow]

Well for one you don’t need hitboxes for the animation right?

[real sierra]

but the rest kinda has to come all at once to be testable

[wary mauve]

yeah thats the same with me i do thinks for like 2 weeks then dont touch them for months or forever again it sucks and it is the cycle that i cant break but keepingn yourself entertained with a project that may die isnt the worst thing in the world, you still gain from it

[real sierra]

true...

[wet monolith]

oh. you mean the lamp might be using a custom keypair or something?
that would indeed make it pretty hard to reverse engineer lolol

[real sierra]

oh god i forgot i need to add inputs too so you can actually use the attack

[burnt marsh]

if it runs on tuya servers id assume the web traffic is encrypted

big iot company

[fleet jungle]

wow

[tender river]

you can probably get the data but you'd need an actual lamp to authenticate with the servers

[fleet jungle]

All these third parties really gives a scale of what the neuro osu project has become

[olive sable]

[fleet jungle]

One would think its some company or something

[olive sable]

the 3 seperate conversations happening at once is confussing

[burnt marsh]

and thats where i run into a brick wall

[wet monolith]

vedal ai moment

[burnt marsh]

welcom to #programming, this is normal

[tender river]

well i cant help you there they werent shipping to my country so i gave my lamp to quack

[olive sable]

you act like im not a regular here

[wary mauve]

with unity the best way in singleplayer is to use

animation events, where you fire an event when an animation gets to a certain part

for example a sword swing would be

click mouse1 -> at 0.3s of the animation, this is the part that ca hit the enemy, so create an animation event that is like enable_hitbox_detection -> at 0.7s turn that detection off

and so tyou have these frames where your sword is active, and if you get any collision events or however you detect that hitbox collission when its enabled, you hit em :)

[fleet jungle]

hey there's still more room to discuss security

[fiery anchor]

[burnt marsh]

thas the joke

[fleet jungle]

Haven't seen that happening since joining 2 days ago (?)

[wet monolith]

making a game sounds super fun but i don't think unity's ide or whatever works on linux sad

would have to resort to like

godot

or lwjgl

[wary mauve]

pretty sure unity works on linux, i mean it compiles mono games so

[olive sable]

it ussualy doesnt happen because of not much actual programming happening at once

[wet monolith]

or raw c++ sdl or something

[wary mauve]

yeah it does

[wet monolith]

yeh unity itself works, just not sure about the ide

[azure lynx]

i'm going to imagine it's possible to open the lamp up and dump the contents of any storage on the board and maybe also work out which processor it's using. if you're lucky there may be some pads for a serial port and that might give you console access to it. once you have the filesystem you can probably dump whatever config it uses to connect.

[burnt marsh]

i might just not use swarm control, just HA and just run twitch 24/7 with a colour detecting script running

[wary mauve]

you can use vsc

[burnt marsh]

but thats annoyyinnnngggg

[wary mauve]

you get to choose ur ide

:)

[wet monolith]

how to 3d

and assets

and etc

[wary mauve]

blender

[olive sable]

iirc there's a unity AUR or whatever

[fleet jungle]

honestly security programming is probably best left to specialists

[wary mauve]

free assetys all over the intnernet

[fleet jungle]

unless thats not what you mean

[frozen hollow]

When I used to program Unity VSC’s debugger was useless

[wary mauve]

Debug.Log is your debugger 😏

[wet monolith]

oh i know. i'm thinking about how would one manage the actual project.
but anyway, doesn't matter. i'm not really thinking about making a game too much rn
kind of burnt out

[fleet jungle]

I worked a little with python's cryptography lib just to try getting a https server working. Terrible headache that was

[olive sable]

i thought you were refering to the multiple convos at once

[fleet jungle]

oh

Well that was just with my joke about having room for more :p

To make it even more tangled

[olive sable]

not enough yellow

[real sierra]

good enough

[fleet jungle]

I will respectfully (fearfully) resolve myself to my current colour (I don't wanna join a cult sounds scary)

wait hold on

this is vedal's server and he's not using U in colour??

[tender river]

it say right there

NOT cult

https://tenor.com/view/ai-denial-no-head-shake-nope-gif-18396288939731808167

[real sierra]

vedal didnt have a hand in setting any of that up

[wet monolith]

the project of mine i'm most impressed with and proud of to this day because of its complexity, is for a website that i've completely lost interest in. so the project is basically dead, and i've thought about deleting it before but i haven't because i'm proud of how far it came.
does anyone else have a situation like this lol

[real sierra]

as a canadian im also sad there's no U

[fleet jungle]

oh

[wet monolith]

fellow northerner 👀

[unkempt citrus]

Americans

[fleet jungle]

As a fellow member of the Commonwealth and even a child country to the Brits, this makes me sad.

[olive sable]

shiro is a canadian?

[azure lynx]

the u does nothing. the word is still distinct.

[olive sable]

the u is absolutely vital

[burnt marsh]

shiro is canadian, its their birthday, we just keep learning shiro things

[olive sable]

her?

[burnt marsh]

typo mbmbmbmb

[frozen hollow]

I’ve been bitten by <colour> tagging before

[fleet jungle]

My man!

[azure lynx]

my favorite wourds don't exist.

[amber fractal]

bwaa

[olive sable]

nah it was a genuine question

[burnt marsh]

favourite* /joke

[burnt marsh]

unsure, they dont list pronouns on profile so i just assumed

[olive sable]

pronouns are happy/birthday

[amber fractal]

[azure lynx]

maybe they're american... did you check under pronons?

[burnt marsh]

yes

[olive sable]

canadian

apparently

[sage crag]

me realise i dont know shimo pronouns

https://cdn.discordapp.com/emojis/1439394374877450273.webp?size=48&name=sognog

[olive sable]

ye same

[wary mauve]

uuhhh too many modules

[olive sable]

[frozen hollow]

Good thing you got brain module

[amber fractal]

Who has that these days?

[olive sable]

not me

[wary mauve]

thank god what will he do without it

he did just say he was being held hostage

[amber fractal]

I feel like the correct move would to repack in an __init__.py file and do from modules import *
That's just how I'd do it personally

[frozen hollow]

Whaa why would you do that

[amber fractal]

in internal repos I tend to be unpredicatable with imports

[azure lynx]

why would you only expose the things that you want to import and then import *? convenience and brevity?

[olive sable]

me when i import

[wary mauve]

i do have the init but i like to bit a bit verbose

[olive sable]

the amount of imports we had in the python engine wasnt even that bad

from PIL import Image
import numpy as np
import asyncio
import pygame
import zengl
import sys
import platform
import time

[amber fractal]

I wouldn't do it normally, but in this case I feel like the repo is more stuctured by directory than API, every module having close to one main module. I feel like it'd make sense to remove the exact directory stucture and let the API not depend on it as much

[olive sable]

fair enough

[feral cobalt]

[burnt marsh]

maybe ill see if vedal could just get the current hex of the lava lamp and put it on a website, that doesn't seem too difficult, but hey idk.

[wary mauve]

anything discord js related it super easy to make messy its impossible not to

even if i use something like discord x ts

[azure lynx]

that's just asking for people with badly written webapps to pound the site.

[burnt marsh]

true that

i remember when swarm control got ddosed, probably similar thing would happen, whether intentional or not

[azure lynx]

especially considering the use cases is that it doesn't change very often but when it does change, you want to know immediately, it seems like this would be better off being push notifications instead of client polling

[burnt marsh]

maybe block ips that make too many requests?

[burnt marsh]

thats what i was more talking about

[wet monolith]

yes, it makes more sense for it to be something like a websocket, with continuous back/forth (or just one way?) io, rather than repeated http requests

[burnt marsh]

i guess its up to vedal what he does with it

or even if he does it

[azure lynx]

also if it's opened to everyone then it would have a lot more consumers than it currently does. the only people using it currently are the people who have paid for it. but if you don't need to pay anything to use it, it's the sort of thing people would have a software version of it just because it's cool and free and that'd put the server under a lot more load. so i wouldn't be surprised if it's not opened.

[amber fractal]

I don't have any projects with massive imports I'm working on. At most it is numpy and typing at the moment.

[olive sable]

most of the imports i have to do are code i have to write myself

the amout of external stuff im importing is small

[amber fractal]

smol

I honestly don't even know why I'm bothering with a base class tbh, I only have one other one

so I just keep altering the base to match the only other one

[wary mauve]

i do the same thing dont worry 😭

you always think u nened another when one class is fine

[amber fractal]

The previous iteration worked with two, so it was a healthy assumption. But I decided to not continue on that path.

so only one remains

I might reintroduce it later, but it'll probably not share this base class anyways

[obsidian mantle]

my compiler says i have something redefined but its not redefined what do i do

[amber fractal]

The only base class I'd actually use is just some random one for post class init, similar to how torch does it with it's modules

[olive sable]

panic

[frozen hollow]

That’s really common in big codebases - restarted ide?

[obsidian mantle]

also vs studio compiler and xmake give different errors now

[olive sable]

[obsidian mantle]

i guess i fucked up my project

[obsidian mantle]

xmake says it

[frozen hollow]

Oh well

[obsidian mantle]

i just tried to add some new repo as module and it all went to shit

i mean i have backup i hope it will work

[amber fractal]

I'm wanting to build out a system to where once I start to build out models I can just shorthand them and construct the expected model from just a list of args.

dealing with all of the shape casting and stuff

[obsidian mantle]

i mean maybe it is redefined and i just dont see how

its actually link problem or how is it called im not good at it

i dont know where these things even are

ok they are in build directory

[fiery anchor]

same symbol in multiple translation units.

i guess?

[fleet jungle]

Just found out numpy isn't a standard library for python like I was thinking. That's a surprise, considering how many people know and use it.

[obsidian mantle]

i have main.cpp and socket.cpp
both have #include socket.hpp

can i not do that

the thing thats conflicting is inside of socket.hpp

[opaque sigil]

not without an include guard

[azure lynx]

#pragma once as thefirst line in the .hpp file i think will fix it

[jagged turtle]

advanced users can probably ask to register webhooks or smth

[obsidian mantle]

i think i understand it now.. initially this thing was in the socket.cpp file
i moved it into .hpp file because i needed to access it from main
its a member of some namespaced class and i couldnt access namespace without doing it

[frozen hollow]

Does the header file have pragma once or ifndef

[burnt marsh]

yeah agree you would need to pay for access

[obsidian mantle]

does this cancel pragma once

[fiery anchor]

ye. remember includ is basically a copy and paste of text. if you don't have include guards, the symbol gets duplicated onto both .obj. later when combining those it sees double and is confused.

[olive sable]

i got a wrapped form nintendo now too in my email
everything needs a wrapped these days i guess

[obsidian mantle]

it has pragma once.
but it had this #pragma endregion right before this thing thats conflicting

[frozen hollow]

Not that I think it’s your specific issue but do you also have a pragma region higher up

[obsidian mantle]

no its lower

[jagged turtle]

why not

people just want to hop on trends

[frozen hollow]

It’s like for making foldable segments of code in IDEs I don’t think it interacts with the compiler

[jagged turtle]

[fiery anchor]

not used to working with pragmas -> banned at work by coding guidelines.

[azure lynx]

#ifndef __filename__
#def __filename__
{rest of include file}
#endif

then this instead?

with actual filename of the .hpp like __socket_hpp__

[rigid snow]

ssjdjddjfjdfhkjhksj

[fiery anchor]

sortof.

[obsidian mantle]

so __socket_hpp__ will point to Socket.hpp file

wait i think i got what you mean

[rigid snow]

I Need Sleep

Send Me Sleep

[fiery anchor]

only got sleep debt, want that?

[jagged turtle]

why is queen from deltarune in chat

[rigid snow]

me too but still not sleep even though tired and broken

sleep not work

valve pls fix

[sage crag]

wrrr

wrr

wrr

[rigid snow]

[sage crag]

https://cdn.discordapp.com/emojis/1154436961277313126.webp?size=48&animated=true&name=evilSlam

[obsidian mantle]

so i have this in both files that include it

and it doesnt help

[opaque sigil]

no that's for the socket.hpp itself

[rigid snow]

what am i looking at

[opaque sigil]

not that it matters if you already had a pragma once inside

[obsidian mantle]

yeah i guess

so wtf does it want

maybe im blind

[azure lynx]

maybe need to change to order you include things?

[obsidian mantle]

surely VS search works correctly when i search across whole project

[opaque sigil]

well it's complaining about socket.cpp

[rigid snow]

meow

[fiery anchor]

doing any #include *.cpp shenanigans?

[obsidian mantle]

no

i think its fucking xmake or cmake or clang or whoever the hell is in those modules

[azure lynx]

i think you said you put stuff in the .hpp file so you could access it in main.cpp?
is it saying that stuff is redefined?

[opaque sigil]

i can guarantee you this is not the build system's fault

[rigid snow]

header fiel

[opaque sigil]

headers were a mistake

[frozen hollow]

You’re learning!

[obsidian mantle]

i have 3 files
main.cpp
socket.cpp
socket.hpp

socket.hpp has pragma once
both main.cpp and socket.cpp include socket.hpp

the ActionList object from error is defined in socket.hpp

[opaque sigil]

personally i'd cut out the socket.cpp but that's just me

[obsidian mantle]

yeah every time i deal with this shit i have an urge to slam everything in one .cpp file

its so stupid

[fiery anchor]

try making the variable static.

[opaque sigil]

do you have forward declarations inside the cpp files

[azure lynx]

what is the actual error: what does it actually say is wrong?

[obsidian mantle]

full error

[azure lynx]

so you're doing includes wrong. you should just declare that there will be a function in the include file, and have the actual function in your .cpp file

[obsidian mantle]

then i will have to do stupid copies in main.cpp file because i cant access namespace of socket.cpp

or can i somehow relay namespace from cpp to hpp

[opaque sigil]

that's what the socket header is for

[obsidian mantle]

i tried it didnt work

[opaque sigil]

it contains all the necessary declarations

and the cpp file only fills those in

then you messed up

[azure lynx]

just declare the function exactly like you have it but when you get to the { change it it to ;
the move the original code back into the socket.cpp

(deleting it from the .hpp but leaving the declaration)

[fiery anchor]

declaration: void foo ();
definition: voif foo() { /* body */ }

no extraneus {} anywhere?

[obsidian mantle]

i know i can put empty functions in header file which are actually defined in .c file
but what do i do if its not a function but a struct within namespace

the namespace is accessible from header file

its not accessible from other .c file that has this header included

[frozen hollow]

Woah interesting font choice

[opaque sigil]

[obsidian mantle]

i need to relay this into header from cpp

its not a function

i make class and define function to access it?

only way?

[azure lynx]

if you are doing C++ it's the way you should.

[obsidian mantle]

this is bullshit i will rather make fucking manual copy it will have like 10-20 strings anyway

i see i see

[fiery anchor]

depends.

[obsidian mantle]

unless..

[opaque sigil]

you can do extern neurosdk_action ActionsList[]; i guess

or static i don't remember

[frozen hollow]

Does the header declare this as external

[azure lynx]

extern in the .hpp and actually declared in the .cpp

[obsidian mantle]

i actually dont even need this structure in another file i need namespaces from that file

uhh

so i put namespaces into header and this structure into cpp file

this is just a mess

[opaque sigil]

why do you think people like rust so much

[azure lynx]

the pretty color? ;]

[fiery anchor]

propaganda?

[frozen hollow]

Are you putting it in the namespace inside the header as well

[obsidian mantle]

yes

the cpp file will have this structure without namespace

[frozen hollow]

Hol up

It should be in the namespace inside both no?

[opaque sigil]

[fiery anchor]

[frozen hollow]

So I think you’re importing the declaration out of namespace then redefining it which is why you get that error (I have very little view of what you’re doing)

[fiery anchor]

"something"

[obsidian mantle]

i think it worked

at least the build is successful

[patent shard]

@left solstice I should of also mentioned, you hit diminishing returns very quickly
I'd do "slower" preset MAX, as it's just kind of silly to push it any further
really medium or slow is already really good

you also can't just look at file size for same crf to determine if it's better, the subjective quality will change drastically when settings or changed even if the crf is the same

[obsidian mantle]

nice now it all works
now i only need to define name of action in only one place no need to make manual copies when i add something new

thank you guys

[sage crag]

im more concerned about the font weight

reading in "Extra Light"

[opaque sigil]

only normal light

[sage crag]

[fleet jungle]

This is gonna be a incredibly stupid question because I'm not well-versed with LLMs or NNs even, but does the LLM decide the "emotions" the twins should have? Or is that a separate model? Further, could it be possible for such a "emotions" centre to control the LLM instead? I apologize in advance for evaporated brain cells.

[quick condor]

Most likely either a separate model or the main model adjusts perameters that affect various aspects of the prompt generation that affects emotional expressivness

[hearty notch]

emotion is semantically embedded in language itself and its just another aspect of meaning thats embedded in the weights tbh you can get any llm to simulate any emotion with just prompting

[fleet jungle]

I'm guessing if it were possible, it'd require far more training data that includes things that have never reliably been measured before (what was the speaker feeling at this moment?).

[hearty notch]

if you mean the facial expressions for emotions

i think they used to be directly linked to certain phrases

and they seem to have a lever to pull to change their own expression sometimes

[fleet jungle]

Hmm. That's true, but arguably there's a lot hidden that words alone just cannot represent.

[hearty notch]

but e.g. using the word embarrassed regardless of context would trigger the embarrassed blush for quite awhile (i havent paid tooo much attn recently)

[hearty notch]

words represent a ton in the implications and connotations and tone and word choices

also im pretty sure the text to speech input from collab partners also now tries to embed tone to some extent (and like parses noise into sfx tokens)

[fleet jungle]

True, but people do hold back expressing emotions in certain contexts. That said, considering how vast the database of training data is, I guess there's already a model that covers basically every scenario.

Not that I would have any clue as to what exactly the training data contains.

[unkempt citrus]

My guess its a combination of keyword detection and sentiment analysis

you can see how sometimes the displayed emotion when a specific word appears even if the actual sentence itself is more neutral

I think scary/creepy does that someitmes

[hearty notch]

ye i alluded to that here with embarrassed too it sticks out

[fleet jungle]

That's true. 🤔 Still, the vibe I get is that a certain word gets picked up which effects the expression display, but then its gone by the next stream of text.

[fleet jungle]

I'll have to see to confirm whether I missed the expression staying, even with a new stream of output beginning.

[quick condor]

yall are really overthinking it

[opaque wharf]

At least he uses space for indentation and not tab

[real sierra]

i find it hilarious how many people are clearly doing the exact same things as i am

my mmorpg needs a way to save account data to disk

and read it from disk as well

and i figure json is nice for this so i wanted to know if there was a nice built-in json serializer

and the first result is

[thorny axle]

whats wrong with tab?

[opaque wharf]

Why not flatbuffer?

[real sierra]

flatbuffer?

[opaque wharf]

Nothing is wrong with tab

[real sierra]

(hi uncomputable )

[olive sable]

hi shiro

[opaque wharf]

Yeah, flatbuffer is a good format too IMHO

[olive sable]

hi uncomp

[real sierra]

idk what flatbuffer is

which part of the process is it meant to replace

[opaque wharf]

Happy birthday shiro! Yesterday was actually my birthday but even I forgot about it lol

[real sierra]

[olive sable]

[real sierra]

happy birthday uncomputable

[opaque wharf]

Serialize and deserialize data

[real sierra]

its a day late but im sure its fine

[olive sable]

happy birthday

damn that was fast

[opaque wharf]

Lmao, thanks! It's the thought that counts right?

[real sierra]

id like to be able to read and modify the save data outside of the game

[opaque wharf]

Thanks Sam

[real sierra]

y'know, for reasons

[opaque wharf]

Ahh, then yeah, flatbuffer is not the best choice if ease of editing is one of the goals lol

Hey, my birthday is yesterday

[real sierra]

but i didnt get to give it to u yesterday...

[opaque wharf]

Why am I purple

[olive sable]

bday

[opaque wharf]

Lmao, thanks

Anyway, I think flatbuffer has JSON mode too. But at that point yeah, just use JSON

[orchid crest]

black-magic

[olive sable]

no

purple magic

[orchid crest]

so evil magic?

[olive sable]

evil is red and black

[opaque wharf]

I am fine with that. Evil is my oshi after all

[orchid crest]

XD

[olive sable]

[real sierra]

https://discussions.unity.com/t/unity-serialization-or-json-net/928261

bro...

you're telling me my options for serialization are

• a good package that does binary only

• a terrible json package

• a json package that isnt supported or recommended for use

• a serializer that works on only very specific platforms

am i to believe everyone making save files in their games is running their own serialization

[opaque wharf]

Unity is C# right?

[real sierra]

yeah

[opaque wharf]

Yeah, you're on your own

[real sierra]

do u know what im making

[opaque wharf]

MMORPG?

[real sierra]

ok i told u already

[opaque wharf]

Why not use System.Text.JSON?

Isn't there .NET for linux too?

[real sierra]

that could work

nevermind

its not in unity

for.. some reason

[olive sable]

[sick owl]

So about the cocoon inference network someone posted here the other day

Looked good, but it just clicked for me

https://vast.ai/article/Absolute-Security-with-NVIDIA-Confidential-Computing

Its just vast with crypto tacked on

[opaque wharf]

The reason is that they use 2.1 dotNET compatibility lmao
https://docs.unity3d.com/6000.3/Documentation/Manual/dotnet-profile-support.html

[real sierra]

wait the unity serializer can do json serialization

that other forum thread lied to me

[opaque wharf]

You think someone would just do that? Lie on the internet?

[real sierra]

never

[patent shard]

I keep hearing pops in the audio on the recent neurosama YT vids
I think the audio is just super compressed, or something
I'm pretty certain it's not YT causing most of this...

and the voices just sound kinda like a bunch of frequencies were taken out

[olive sable]

if i procrastinate till the last day, i only have to work on it for a day

[idle dune]

I wanna get into running a local LLM for funsies, is it worth grabbing a used Tesla P40? They are like 200-300 bucks. I wanna make one to help my legal document wirting and creative writing.

I don't even know where to start either.

[burnt marsh]

no! work NOW

then rest

[olive sable]

p40's are fine afaik
mainly the vram

[real sierra]

im liking this

it feels very java-esque all of a sudden

the unity specific stuff has all felt super foreign

but now that im working in my little pseudo-singleton box again

it feels like home

[idle dune]

I know they are old as fuck and a modenr GPU would crush it

But 300 bucks for an AI accelerator card is not bad (imo).

[true hemlock]

3060 level performance

[real sierra]

im so exhausted

[true hemlock]

except it can't do fp16

[real sierra]

i find myself going back to code i wrote 5 mins ago and being totally confused at why i made it so bad

[burnt marsh]

https://media.tenor.com/jksbEFnU0k8AAAAM/book-of-boba-fett-the-mandalorian.gif

[idle dune]

I can get a RTX 3060 for the same price and probably have a better time then.

[true hemlock]

if you don't really need the vram yeah but for 200-300 bucks a 3060 is overpriced imo

[flint dome]

@opaque wharf happy birthday 🥳 🎂 🎊 🎉

[true hemlock]

like really overpriced

[idle dune]

I'm trying ot think of what I even need. Because in Criminal Law I can't even use Claude, ChatGPT, or literally any AI to correct the grammar of what I draft let alone help me plan documents.

[opaque wharf]

Yo, thanks!

[opaque wharf]

I think correcting grammar is best left to some other program and not an LLM

[flint dome]

make sure to get one cake for shiro as well

[opaque wharf]

Of course, birthday boi will always get cake after all

[idle dune]

When you get into the 100s of pages for a legal analysis its easier to throw through an LLM for aid.

[flint dome]

@real sierra HEY thats not funny you changed the solar and gas generator thingamajig to a transformer

[idle dune]

Dramatically cuts editing time down.

[real sierra]

it was time to change it

[flint dome]

why though

also wheres the nya at the end

[opaque wharf]

[olive sable]

[real sierra]

maybe i'll sneak it in at some point

you'll have to check back later

[idle dune]

I think my highest so far for just legal analysis assignment sis 54 pages with 4 of those being sources.

Its a whole lotta words.

To say very little.

[flint dome]

you have no idea how much my father loves using perplexity pro to draft financial reports for a BANK because the interns are worse at writing them (he's smart enough to ask him for public stuff only, no confidential)

[burnt marsh]

sneak it in unyaversal

[opaque wharf]

Well, isn't that to avoid ambiguity and alternative interpretation?

[flint dome]

ogey

[opaque wharf]

Well, at least financial report still have the data. So the error in words is hopefully rectified by the graph/table

[flint dome]

its just hes smart enough to edit them quickly but i feel uncomfortable giving these things to a software to do it especially one which is not a mature enough platform for all this stuff especially for loans involving millions

[quick condor]

P40?

[real sierra]

why do ceil and floor methods always return as floating-point types

i dont care about the edge cases just be convenient

[opaque wharf]

Until you have NaN / Inf

[idle dune]

Yeah and I rather not give confidential client information to these companies. Regardless of if they would make use of it, my Bar Association has rules against it.

[real sierra]

if someone gets NaN or Inf exp then im just deleting their account

non issue

[idle dune]

[quick condor]

A local model trained on legal English prose would be a decent first pass for editing. I wouldn't have it write things but I'd imagine it would be a good way to avoid dumb mistakes

[idle dune]

Your right, it wil make tons of mistakes and interprete all kinds of shit incorrctly. But if I can make a lawyer court binder in thirty minutes rather than three hours I rather that.

[opaque wharf]

Why not pass it to hunspell or something?

[quick condor]

Well like I've seen enough reviews of legal briefs where there are basic mistakes that a model could catch. If it can reduce time spent looking for minor mistakes it would let you focus on the important things

[quick condor]

Can't really look for issues that are beyond sentence level

[idle dune]

Its not just grammar and spelling its formatting and logic. Its looking like for my career I am heading into criminal legal assistant which is going to require me to aid in the creation and documentation of my lawyers court proceeding. Many times the process is I research for them, they present my research, if I fuck up at any step then its a pathway for the opposition to argue and then its on me to fix that.

[opaque wharf]

I sure hope the logic part is triple checked if you use AI. Because that is what I am worried about the most

[amber fractal]

ready to grind for NaN exp

[amber fractal]

Happy birthday!

[unkempt citrus]

just use floating point exp 4head

[opaque wharf]

Thanks!

[unkempt citrus]

allows you to implement exp multupliers too

[idle dune]

Yeah it has to be. Even before any of documents are handed over theres meetings to read through them and discussions.

I'll give an example, while a rather extreme one, of the lengths that these write ups can get too.

https://www.canlii.org/en/ca/scc/doc/2025/2025scc39/2025scc39.html

[amber fractal]

I would honestly love to deal with all of the damage formulas for an MMO
-# This isn't my first showdown with this type of formula (the minecraft Armor Curve mod has done so much to me)

[unkempt citrus]

I did genshin calcs for awhile

numbers wao

[opaque wharf]

Yeah, I am just worried it will ended up like valve v rothschild

[amber fractal]

My goal was to use that mod to actually not have the armor value be hardcapped by the formula, I'd like to think I suceeded in that regard.

[opaque wharf]

The formula:
dmg = base × 0

[amber fractal]

[real sierra]

nah its ok i kinda have a plan

exp scaling will be exponential, with it increasing 10x every 20 levels

[opaque wharf]

Pick a curve. Any curve. And then just throw a bunch of parameter mixing into it

[real sierra]

this will make sure people dont grind early-game areas for xp

[amber fractal]

the secondary goal was to make it entirely possible for a person to visualize the formula.

[real sierra]

and also means big number go up for flexing purposes

[olive sable]

Pick a curve. Any curve.
that specific wording makes me think of how magicians do card tricks

[amber fractal]

XP boring, I want to deal some damage

[real sierra]

save file important

[amber fractal]

Where is that desmos page

I already have one up there

[real sierra]

but i think save file done

sooo...

now we do ui work

[amber fractal]

[opaque wharf]

The bane of backend engineer

[olive sable]

[idle dune]

Thats a pretty valid concern ngl. But this is not going to be used for finding my case law and hallucinating case law that never existed. WestLaw, LexisNexis, and CanLii are pretty much the only supported digital law libraries that courts accept. Also I was never going to use it to "Find me x for citation." its a "Does my research and conclusion on X make sense.". Also as a sidenote I find there is no good AI for finding citations, even Claude, an AI that is marketed towards legal professionals is incredibly ass for finding citations and aiding in research.

[amber fractal]

The curve I went with was an inverse of the triangle formula
Not exponetinal but good to deal with less damage reduction as more armor was added

[idle dune]

I would also need to find my firms LLM policy becuase thye might jsut say "No we don't allow it" and all this convo and planning is basically usless.

[amber fractal]

Why did past me allow for armor to go negative

This was balenced around minecraft, so that is why it became what it is. As one can see... this thing has a certain quirk if you can call it that.
https://www.desmos.com/calculator/f49357a313

[unkempt citrus]

Pull a pokemon and get a little funky with it

[simple sky]

So, tf is lambda calculus?

[amber fractal]

in hindsight I could have done a better job labeling

[olive sable]

not for mortal comprehension

[patent walrus]

any pc building nerds around

[unkempt citrus]

I know a bit

[olive sable]

[real sierra]

no you cant eat thermal paste

[patent walrus]

FUCK

[simple sky]

Oh I’m taking a programming languages class and it has lambda calculus as a unit.

I’m screwed aren’t I?

[olive sable]

as long as the teacher explains it well it might be fine

[simple sky]

“We expect you to learn while doing the assignment”

[real sierra]

gg

[simple sky]

Word for word on the syllabus

[real sierra]

you're cooked

drop course

[unkempt citrus]

Welcometo university

teach yourself

[simple sky]

I mean. In the past 2 years no one has gotten less than a C

[real sierra]

if the university wants my money they can do their job

otherwise i will take my time and value elsewhere

their loss

[olive sable]

aluminum oxide and zinc oxide

[idle dune]

A little bit

[olive sable]

[simple sky]

How cooked am I, here’s all the topics:

lambda calculus, debruijn indicies, computation passing style, monadic programming and proofs

[amber fractal]

For 2024, I accept this. It does it's job of being an alternate damage formula for the block game

also for my 3rd attempt at it

[olive sable]

if you got a whole year you're only mildly cooked

[simple sky]

16 weeks…

[olive sable]

ok you're fucked

my discord reloaded for some reason

[patent walrus]

okay so
broad strokes: i've got a huge ass PC and i want to downsize
i have an ATX motherboard (GIGABYTE X670E AORUS MASTER) and everything's stuffed into this case https://phanteks.com/product/evolv-x2-black/
when i first heard about the steam machine i thought "wow... it's such a cute cube, i wish my PC were that small... so i can carry it around places..." because sometimes i'm going between places for weeks at a time

[unkempt citrus]

Memory leak reload

Smaller than ATX are miniATX and ITX

ITX also gets some custom case form factors

[patent walrus]

i'd like to keep my CPU, GPU, RAM, and PSU (if possible)
something something
AMD Ryzen 9 9950X
PNY XLR8 RTX 4090 VERTO EPIC-X RGB
64 GB - 2x G.Skill F5-6400J3239G32G (DDR5)
Corsair HX1200i

so i'm tryna think of a motherboard and case combination that would support those components

[idle dune]

Whats your CPU? You cna probably get a Mini ITX board and case if you only have a GPU. In you need more expansion for a sound card or any other baords you cna grab a Micro ATX, a bit bigger but not much.