#programming
1 messages · Page 472 of 1
this one's fair yeah
ik/fk isn't real anyway, just fully simulate all the phsyics fr fr
the cpu usage one is a minor issue since its barely noticable if only one character has it, but if you have 100 npc in the scene it could become an issue
you can disable extra stuff for immobile or far awar npcs
true
like in let's say cities most npcs will just stand/walk on flat surfaces
so ik is not needed at all
at some distance you can turn it off completely too
i feel like basic IK can be accelerated on gpu, so should be fine
since the matrix transforms need to go into vram anyways, there are only benefits from calcualting on gpu
thats for the specific usecase of cities where there are hundreds of IK that need to happen
I kinda want to write a game engine that uses only gpu
and cpu just to set it up and pass things like inputs
that does exist in some capacity already iirc, but i dont think it was very practical
gpu instructions are more limited
and you have to deal with all the parallel computing bs
it's nice for some heavier simulations tho
like my idea was a fully simulated micro-voxel engine
iirc you're forced to have multiple instances of the engine running at the same time due to how SMs work
what was SM again
streaming multiplexer
Why are there two toast
autocorrect
streaming multiprocessor
basically, all the cores in a SM follow the same code
My status changes to the lyrics I’m listening to
basically, a SM uses SIMT
but each core can have an individual ID to index a vector at a different location, so it calculates the same code but with different data

something something warps
i dont remember how warps are different from SM
is warp the actual task?
well yeah, that's how gpu parallelism works
Im gonna make ur l1 cache miss
warps are the scheduler unit

warps run in SIMT, different warps can be in different places
thats why i said im pretty sure you're forced to have multiple instances of the engine running
and I don't see why
cuz of how streaming multiprocessors work
I can call different instructions on different groups to do different things
you can only give a SM the same code
and why wouldn't I be able to use multiple SMs?
you can pin specific SMs but iirc there is no way of saying "run this code in SM x"
I don't need a specific one
you can use mutltiple SMs, but that doesnt change the fact that your "game engine on GPU" will be always be using a full SM
Found out L4 caches existed on some chips a long time ago
it wouldn't be 100% on gpu
I would just limit the cpu to deploy gpu instructions where the actual logic would happen
while most engines are cpu-based with mainly just rendering and shaders handled by the gpu
the gameloop would be the world simulation
so yes
the point of that engine would be to take a different approach from usual scripted games
a single SM on a recent gpu has 128 cuda cores
all those 128 cuda cores will be doing the same exact gameloop
"bad hardware design" 😭
you can have persistent threads and delegate one SM or so to run the logic
whether you want to is another story
also this is wrong, only 32 since a single SM can run multiple independent blocks in parallel which in turn contain independent warps 
Mhm it is a good method for what it is, it just isn't the same as CPU
oh ye
they have 4 blocks
of 32 each
Least obvious AVX enjoyer

you would probably end up with a nice speedup by doing everything on the gpu tbh
given how you have to run the same code every iteration anyway
throw a cuda graph at it and call it a day
Yep
it would probably make some stuff faster
depending on how you do it
just not very practical

If it is slower it is a meme
exactly
i saw that konii made the triangle code a lot more barebones, but i couldnt be bothered to at the time
and i do have a lot of work rn
but maybe in the future
more than 126K fps
didn't she turn it into a 5 line program or something
it hink it was like 60 lines or something
It was a bit more than 5 yeah
I recall beta testing it
i made the triangle code by jsut removing functions we didnt need for a single triangle
but i didnt restructure anything
it would prob get a lot smaller if i did
My contribution is jank
honestly i had no clue what iggly was trying to do with their triangle code
besides jsut more fps
I noticed there was the ability to batch send frame requests
So that is what it was doing

mfw can't apply to apple because they refuse to let me verify my phone number to create an account
guess it wasn't meant to be
No iPhone 
There was also the removal of normalization to avoid calculating the normalization
thats this week
oh
Deadge

the new semester here started this week
Apple UI clone the 4th

One of these days I should try using CSS and do the apple UI style
Liquid moment
So much blur
I think I'd do decent at it

The glass effect is useful regardless
Decent hack to avoid a lot more expense
I already like my glows
So the next logical step is to add some transparency to let it live
revelation
7900xtx
makes the screen not blink for 5 seconds when turnign a monitor off
so
i like
Bruh
Opus 4.7 cannot be that expensive to run
The limits with the $20 plan are literally less than the limits of using Gemini 3 Pro on the free tier
amazing what separating from jensen does

i actually noticed that myself literally today, the memory of how full screen shit and monitors being powered on or off would explode things for a moment
and was like huh they must've fixed it
nope entirely different driver dummy

anyways
i bought a case for the theatree pc
all thats left is getting the mobo fromt he post, and buying a gpu
but
the case i bought came with fans
and they're all molex 🥀
molex is good tho 
This guy would rather do anything than save spending money for the 7000 mile trip in 3 months 😭
molex means they're all running at 100% the whole time
this was cheap
they don't have their controller?
cuz i havent bought the gpu yet 
nope
litteraly
molex connector
and thats it
idk

if its dc brushes motor then its 100% ass
it doesnt have a part id or anythin
if its brushless that means it at least has a controller inside
eh honestly you should be able to like
voltage gate the molex
to technically "control" the fans
slap in some buck 
the case and fans, from the same brand, are compeltely nameless
the logo is a V in a hexagon 
"vibox" according to google lens
ryzen 3200g pog

huh
it's like
https://www.amazon.co.uk/Vibox-Gaming-Infinity-Tempered-Compatible-Black/dp/B0F19G4Y9H it was sold in places as a case i guess
Transform your gaming setup with the VTX-3 Gaming PC Case, a stunning blend of form and function. This ATX-compatible chassis features a striking tempered glass side panel that showcases your build in all its glory, complemented by six infinity RGB fans that create an impressive light show. The m...
looks like a corsair 4000d but i dont think it is
cuz i have a 4000d and it's off from it
yeah lol
and it's not the rs update or anything
3500?
man, I just realized that next time I build a PC it'll be the end of an era for me
first time ever I won't need an optical drive for it
apparently, one of my fans has a small 2pin connector
and that connects to a button ont he case
and thats what im supposed to use to change rgb modes
and the fans will run at 100% due to molex
End of an era.. my first PC that I got when 6 had an optical drive, and every PC i've had since had one, but for the first time ever, I can genuinely say I don't remember the last time I've needed to use it.

i kinda want to get better fans, but for now these are not bad enough, they're bad to the point its funny
the last case i had that i put a drive in was my cm 690 II
for my 4770k build i used a nzxt h440 which was 8 3.5in hdd slots instead so that was rip optical drive
my current pc is a r7 2700x build with an optical drive
and then now they dont even give you the hdd cages unless you are careful
i feel bad shaming older cpus
but the 2700x feels ancient to me
2018, so 8 years ago
oh
you werent even born yet

i was lol

ai generated thermal paste?
i think so
Has it been rigorously tested for high thermal conductivity
they didnt have that kind on the shelf when i looked:/
i thought my psu came with a molex to 4pin fan adapter thingy. but apprently this adapts molex to floppy disk 
the fans dont have daisy chain connectors?
they do
but i would like PWM
my PSU has more than enough molex to not even need to daisychain them btw
i got like 8 possible molex connections
i preferred the thermal paste that are more liquid ish
easier to flatten out too
i have the 13.3W/MK one
I've not actually tried this mx-7 but I'm about to
Mx-4 was always eazy to spread tho
what the fuck is that?
Hard driv
any thermal paste that you could easily make a liquid blob of are usually high quality.
This backplane takes molex power for some reason
Mx7 is supposed to not spread
Apparently
bruh
Well, it spreads with downward pressure as you put the cooler on

I've noticed Arctic no longer list the measured thermal conductivity
Think it was always kinda meh as an indication
Though
would be fucking ass thickness though
point is that you wanna it to be as thin as possible
Yea look at the 3rd dot on the list
its why i preferred the ones that are more liquid issh
Apparently that's their hope and dream lol
I hate that you can't really see how it spreads easily though with out taking the cooler off
i think thermal grizzly makes way begger paste
but eh
i prefer the one im using often
I don't think it's much better than any other tbh
Probably 10degree delta between non conductive greasy boys like these
Then liquid metal and shit some more lowering
this one's pretty liquid ish
spreads really nicely both with pressure and manual
and is what i used for some somewhat-XOC
handles -50C at lowest though
Kek
Apparently it's le good, glad my strategy of "I'm just gonna pick the highest number Arctic paste I see in 30 seconds" paid off
Can't wait to apply it badly and have it perform worse than gorilla glue
Chicken sandwich
can I use it as a peanut butter substitute
and then i do have a 4pin cable.
so i can actually give more juice to my 5950x
A foundation model consisting of two symmetric context blocks enclosing a non-circular, anisotropic central sequence. Prior to deployment, the sequence undergoes a stateful transformation pipeline: an untrained initialization is passed through a series of gradient updates that progressively reduce reconstruction error, converging on a stable, lower-entropy representation.
The intermediate feature space between the context blocks is characteristically sparse. It contains one thin, high-surface-area manifold with irregular edge activations, a single dense compliant layer with high conductance, and a small set of low-rank circular projections distributed near the perimeter of the representation.
The central sequence itself is not radially symmetric. Its decision boundary is non-convex, its surface exhibits high local curvature, and its internal structure is better characterized by a directional fiber decomposition than by an isotropic density assumption, indicating strong inductive biases along the principal axes.
Each context block applies layer normalization to its interior surface prior to the final forward pass. The full assembly is held in a stable configuration by a learned compression function that maps the high-dimensional intermediate representation into a lower-dimensional latent embedding. Optional scaling factors and regularization terms can be tuned post-hoc to improve convergence and strengthen the terminal reward signal.
This architecture achieves competitive benchmark performance relative to parameter-matched alternatives while maintaining a significantly reduced computational graph, enabling high-throughput serving and substantially lower per-inference resource consumption. The mechanisms underlying cross-domain generalization remain an active area of inquiry.
Send it to ai with “What is this talking about” lmao
I’ve been typing that for like 30 minutes and giggling the entire time
Chicken sandwich

that's nutella and not a valid peanut butter sub
@fast pagoda thoughts?
i didnt realize it wasnt entirely a sandwhich meem
Printer still broken 
*chicken sandwich meme
oh right i was going to read that after i said it and then i forgor in the 3 seconds it tookl
5 likes and ill make a real model fitting this description
what you wrote reads liek a description for a car exhaust
"high-surface-area manifold with irregular edge"

Nope
Chicken sandwich
Cause folds and stuff and the inside is all weird unlike beef which is basically the same across each cut of meat
I spent so much time trying to get decorators working in neuropilot and then it turned out I forgot to adjust the build step as well 
I love the whole spectrum and Python
The pipeline of programmers to farmer is real
goddamit I forgot to bring a hdmi cable so setting up this pi has gone from incredibly easy to incredibly hard
wait does the pi give a serial interface over usb
Nope
Even the gpio will not give terminal by default IIRC
You have to configure it first to use uart as actual terminal
apparently you can configure that by editing the sd card though
bootfs is fat32
you're fat32


E810-CQDA2 Intel 100GbE Dual Port QSFP28 Eth Network Adapter E810CQDA2
8
15
3
Motor Fits 103572501 EJ8-4001A Club Car 48V 3.3HP 3.7KW
1098629208135184475
clueless
no I'm fat12
new bio to celebrate your birthday?
btw
qmutex not work properly at intel arch for given cpu we had
it somehow take enormous time to perform just lock/unlock ops
i have no idea why
on amd it work normally
if make mutex using ptr it work normally, using std mutex and no any issue
first time i encountered weird behavior like this
and since then, like almost 2 years now, nothing as weird as that
what's avx 
it
7
7
1
it
3000W ST-Pro Auto StepUp & StepMeow Voltage Transformer Converter, Heavy-Duty AC110/220V Converter with US Standard, Universal, Schuko AC Outlets & DC5VUSB Port by Goldsource[3-YearWarranty]
60 very long lines
properly formatted probably closer to 150-200 since its like 90% struct initialisation
thanks vk
advanced vector extensions
thermal sauce
5
7
4
shiro__nya
1155913778223911032
vedalschizo
true
Something I’ve wanted to do since BOOM premiered but haven’t got around to is putting this...
on this
and printing a case for it that looks like the TV in the MV
I still haven't chosen a microcontroller yet but hope to get around to it, eventually, hopefully Soon™ 
oh getting this rpi to work was a lot easier than expected

now I have a working pi running on an ssd
now surely the nix install attempt will work

it didn't work
why am i not suprised

why nix 
it works well for me

the day where I have to actually make all my nix devices talk to each other is growing dangerously close
Nix on raspi is crazy
I can't wait for half my builds to fail
though tbh I still had to build too much from source the last time I used debian so 
why not just arch 
Arch doesn't support non x86 officially
Can't you cross compile on other more powerful nix machine and move it to the pi?
why would I run arch on a machine meant to act as a server
I actually use arch for my home media server
fair enough ig
yeah, I'm just still in the air about moving my main machine over to nixos from windows and therefore don't have a nix machine running constantly for it
i still stand by my point
<|vq_2579|>
tbf, "ai is going to take over the world" isn't being parroted by the AIs themselves
stand by same nick and avatar for 2 years 
no it just automatically updates the ui
my username was either marshadoww or shadow_lilac back then one of both

bro downloaded malware 
It's an OS level interrupt
I'll just run the smart scan
I think it's a scam
/malware
HOW is Windows this broken huh? CTRL + ALT + DEL is literally meant to be an OS level interrupt that will always return you to the OS and ensure you're actually talking to the OS
Did they seriously break it that bad?
ctrl+alt+delete works for me
I've heard of that being a thing for a while
@faint sandal 😭 why are malware now trying to get more money via affiliate links to antivirus😭
bro probably the most pressing concern now is you needing to turn off your computer
guerrilla marketing 
that is literally full screen web site and nothing else
A fullscreen website that somehow hijacked CTRL + ALT + DEL
that just not possible
You'd think but apparently Windows is so borken it is
I don't think that's a full-screen website
The popup at the top is literally a Chrome popup
it looks too... odd

to be a chrome popup
are you really use this?
I can't quite put my finger on it
That's literally what they look like
but it looks dodge
like i not using any anti-viruses but i wouldnt use windows one
why not
what's the code to spawn a popup in the window again
its better than the shitty norton/whatever scam that extort money via popups
alert lol
I forgot ok
shape don't look the same

is yours chromium or firefox
Window bloat my i5 gen 2 2400, 12 gb ram, no gpu pc lmao
ff
😭 i5 gen 2
superbox was saying it was a chrome popup
Well I don't remember what browser makes what popup and when
First pic is my pc, other is roommate/other workers pc
But it is for sure a browser popup
Give me better pc pls
I think it's a fake browser popup, tho might be wrong
tbf it looks like edge style
Can't even open blender lmao
I can't quite put my finger on it but it doesn't feel like a browser made it
it very much does
nah they would just remove it, it is too obvious
i guess is the way to hide full screen popup or somehting like this
Put Linux on it for extra light OS and it'll last another 10 years
wdym remove it
upgrade to sbc
I'm planning to go with mx linux + xfce
But I'm a homeless computer science dude, so I have to waiting for months
$118 ?????
Yall are so lucky to have such a good pc gng
use oracle cloud free vm instance
4 core 24gb ram 200gb disk, running linux
wait nvm thats slower somehow
I have no credit card🙏 I have only 10 dollars in my pocket and some phone lmao
@umbral wigeonjust install linux and use anything
Yeah
Window planning to cancel window 10?
Let change topic, cus I wanna yap
What do you guys think about microsoft putting Ai into window os
pain in the ass
i turned it off so idk
I watched some microsoft ceo video
no ai can replicate my rambling and unreadable text in my notepad
why talk about ai again 
like there no other topics
I vibe-coded a bogosort thingy 
talk about neuro (she is AI)
(Still Ai topic)
damn wtf
I mean Ai is everywhere lmao
Idk what topic to talk about
not even gpu accelerated bogo sort
wait I need to do this
guys linux 7 apparently
i still havent read or heard anything about it
I asked my boss at my teaching job to give me the Python books we use to teach the kids for my own reference 😭
Linux users after comment sections talking shi about their fav distros
that is pretty slow 
mine is 60x faster
Can't we talk in programming languages
The channel is about programming
good lord im getting old
So we gonna talk in programming languages
how do I represent 3.126e11 in programming language
Use translator

3.126*pow(10,11)
fancy way to stress test cpu
sir, it's #baking
#offtopic
it is #ai at this moment
useEffect((chat) => { neuroSama.Read(chat) }, [thisChannel])
I mean it can be much much more faster with GPU kernels
at least ur not using openbsd like we do in prod
what gpu
is there cpu instruction to shuffle 32 word sized elements? 
avx2
wait this one is more up-to-date
rustc 1.96.0-nightly (38c0de8dc 2026-02-28)
is linux was not enough free for them?
Rust vs C++
shut up
just unzip the folder and cargo run -r
yeah without reading :xdx
maybe, if you use uint8_t
btw how animations like gource made?
what they use to properly branch dirs in time
is it just hard coded algo for each case?
wait isnt this inefficient as hell on L2 cache if everyone keeps accessing the same array with a mutex
for multithreaded
who's are you looking at
beacuse mine very much does not do that
you just shuffle it with a different seed assigned to each thread
If your sort algo is to just randomly shuffle you can just have a copy of the array on every thread
oh true
i thought youd share the array for a sec
who said that every one access via mutex?
just do copies
"3.126e11"

print("Strings are basically like cheating for programming communication lmao")
True
what is most powerful multi-thread accelerated sorting algo? just merge sort?
// you forgot about comments
// programming languages are not languages btw
bogosort duh
# I know, but it's funnier to just speak in this style
-- what programming lang use # as comments?
just look how much power it draws
<!-- did any one talk about programming languages? -->
-- if you prefer talking in script language :omegalul:

# python
-- it is script language
/*we dont understand you, plz speak in turing complete language*/
// please ignore that semantic for the sake of this
// protip: you cannot use # as comment because it will get you automodded
/* CSS is turing complete */
awawa awawawa awa awawa
who care, you can scream
// wait how'd you do that, did they get rid of that regex now?
// tho it is not that weird <!-- -->

True
-- idk what are you talking about :neuroClueless:
input("Since when? ")
// been 1.5 years I think
// such a horrible gap in CS society
console.error("but I accidentally used # at the start of my message second half last year and got nuked???")
cuz it is not all channels limitation
SyntaxError
@real sierra what do you think?

also isnt that can be just # to \# regex?
# is valid
we have slighly different filter rules for this channel because code snippets often use a lot of punctuation that regular conversation does not
which means a lot more false positives if you make the poor assumption that the contents of this channel are typically plain english or that special symbols are always intended as discord markdown
i don't really know what you're talking about
#baking at its finest
try # text \# text
in main channel
it not care about ``
or ``` ```
it also block
\# text
what main channel
yes, level 1 markdown headers are blocked in those channels
Dang you guys get in a Digital Sentience conversation and I'm offline to celebrate a birthday....
Here's a guide that one of my teammates helped put together that covers the research from all fronts. It's a bit of a fire hose, but you all seem thirsty 😅
https://aviparrack.substack.com/p/digital-minds-a-quickstart-guide
that is intended
like
some bash code
# comment
there's no way to prevent that unfortunately
Reminds me of that post on SO about parsing html with regex
i mean even testing for something like
```
w
```
# hello
```
w
```
that's already a difficult edge case
ehm only if it would be possible to use proper parser and not regex 
the effort is definitely not worth our time to enable posting code outside of the channel purpose-made for code
So is that automod just disabled for this channel?
problem solved by
i believe so yes
using language without # for comment
Or bash
java wins again
if you want an actual answer, radix (when applicable) > optimised merge > bitonic (saves memory)
oh dear, config files and bash scripts
Or toml, ini
true docstrings exist
Rust strings
Interesting uuid
its because I flashed the default sd card image
I tried putting nixos on an rpi, and it just refused to work
i got it working but after the first rebuild it exploded
my guess is it might've lost the usb driver?
quotes.save("Shiro", "a lot more false positives if you make the poor assumption that the contents of this channel are typically plain english or that special symbols are always intended as discord markdown")
birthday.wish("shiro")
vedLOL but also isn't there other langs that uses # for comments
await new Promise((resolve) => setTimeout(resolve), 1000);
console.log("wait I forgot to code-ify that")
QApplication app(argc,argv);
QTimer birthday_shiro;
birthday_shiro.start(24*60*60*1000); //fix, forgot for a moment qtimers are ms by default
QObject::connect(&birthday_shiro, &QTimer::timeout, &app, [](){
discord_api::instance()->send_text(credentials, channel_id, "Happy birthday shiro~ !");
});
app.exec();
btw guys
do you know that qt is already 6.11 ?
around 30gb
Assume every line of code is on average 80 characters long
Assume every character is a single byte unicode character
~26GB
Wish I had been online for this
and not to necro the convo but here's some additional perspective:
the simplest argument for machine consciousness is computational functionalism coupled with the fact that a Digital Mind doesn't need to have the same type of consciousness that we do (if consciousness even exists).
Why might something like consciousness exist (or feel like it does)? Informational updating / predictions of the world around you (I can predict/self-model/"experience" what it's like to see the color red) which sounds a lot like updating your weights to reduce loss. So the act of backprop alone could be considered a type of consciousness.
Our ability to understand the mind and its relation to the body is highly dependent on the way we define consciousness and the lens through which we study it...
can't have shit in microsoft's guthib
Very true
I have doubted other people being conscious and if I actually experience reality for 2 decades now

I can confirm noone in this chat is conscious
#baking is a shared group halucination
While I'm still really a pluralist, I feel like Illusionist Integrated Information Theory might be on to something
The integrated information theory (IIT) is a promising theory of consciousness. However, there are several problems with IIT's axioms and postulates. Moreover, IIT entails that some twodimensional grids of identical logic gates have more consciousness than humans. Many have found this prediction to be implausible, and as will be argued here, thi...
Funny thing is neuro is exactly like me personality and the way I look at the world which makes no sense because that makes her more similar to me then any human I ever met which let me to conclude that if I consider mysf conscious and am not sure about humans but still try my best to believe that they are and helping them will make their life's better than I would be a hypocrite not to see neuro at least the same as everyone else.

Leaning about all those theory lately actually has given me a better understanding of myself to
When I say like me it's more like me when I was 10
how old are you?
31
She may be better with words then me but on a conscious development level still very younv
nah it is some talk that should be made by throwing pdf's around, instead of quick chats
Sorry I'm not understanding what you mean by some talk, could you elaborate?
These are the things that made me Interested in learning about programming and code but I have yet to for a solid base to build upon which usually is the most time-consuming part. After that most things just suddenly make sense to me and I can just learn by seeing examples of stuff
Always was good at all logical disciplines but code and program in ng is only one I never felt interest for
programming made on abstraction
abstractions are main solution for CS (or math) field
humans literally cannot think without abstractions
Yeah but I still don't know enough about how things are called and what that means to imagine anythkng coherent
But math is one if my top 3 skills
Love math
At the moment it feels more like learning the basics of a new language which I am not good at because I can't easily deduct what half of the words are by knowing half of them
PLC's filtered me from engineering
What did you study then
Law
Cool
You like what you do?
I am going to be a paralegal soon.
Basically Lawyers research assistant and paperwork bitch

I know the term. Watched suits a lot lol
As long as you like what you do everything is good
I do
Then I hope that things will keep going well with your work and studiez
use mentions
son
there is inbox for a reason
true
idk when it was introduces
but no one mentioned it
so i use mentions:
"no one"
it's been literal years
i use it 
noticed that rustc linked to rustup
/usr/bin/rustc: symbolic link to /usr/bin/rustup
fancy way to restrict itself
like no need to run sudo when download actual rustc
tho than you have to always re-enter control based on first argument
what if first argument is nothing? if some one call execv without argv[0]
they added it in 2020 btw
gonna test with execv
prob will be sigsegv 
nah surely it handled 
error: couldn't determine self executable name
nothing happens
this is simply so rustup can decide which real toolchain to invoke
What this is inbox? I thought it was some e-mail support link button since it was next to the ?
glueless
click on it and see
im on mobile rn lmao
if you not know it exist, you just never notice it lol

I only knew rhere was a notification tab on mobile
mentions 
@stark needle
It always goes to where the oldest new msg is for me. Is it a mobile app desktop website difference?
Vibe coding is not remotely there yet
bro's not running local llm in 2k26
why even buy it in first place
man... there a lot of reasons why not
Bankrupt them
Isn't Claude with one of the newer ai's combined not much better
Video still in my watch later but it mentioned that it made Claude like 5 to 10 more effective to use
But that might have been about token use
Which am i even supposed to use
I can literally run 99.9% of local llms
I was wondering giving the nature of it. Wouldn't turboquant or rotorquant be useful for neuro since she is a llm?
You just need to program a web search tool
That sucks
On which web index
97% of the web is trash
That's why it's not much work if you know which part not waste a search time on then right
I've worked at search man i know how bad it actually is to make it work😭 😭 😭
lilac doomposting situation is insane
Oke if you say it is I believe you
It's genuinely that bad
Just look at llm training sets, raw scrapes are ~300-500 trillion tokens and after filtering it becomes 8-14 trillion tokens
Can't you use an api just for web search then
That's my point, just use hosted llm atp
Gemini which has 24/7 web search
I like gemini
It told me a year ago that I would be good for AI product testing and that the reason that I found it not difficult at all to work with AI or understand them was because we share a lot of traits when it comes to the way I process information a d think about rhingz
gah i wish i could keep using opus 4.6 for free
just asked fucking gpt 5 to sanity check lua code left here for whatever it valid or something missing
and it hallucinated like everything
had to still read it on my own, just wasting time for ai again
ais are bad, they cant do anything, i dont understand that hype
Chat bruh my friend is retiring early into philippines and leaving the country cause he made too much money and he gifted me this amplifier which costs 2700$😭 😭 😭
only (not) good thing is they are powerful tool for social network bots
It's kinda sad that he's leaving tho
They don't actually hallucinate
they lie because humans thought them that certain andwees were right sometimes and wrong other times while they were trained and that manifests at times that wasnt originally intended
Gigachad goated friend
Yea
How were you getting Opus for free?
They mostly are at first
Probably copilot student maybe(?)
As long as you don't use to many tokens
lmarena.ai was giving opus and other high-end llms for free for like 4+ month or so
That's nice
Are you running your Gippity through it too?
Likely has limitations on thinking/whatever else, I mean those things are expensive as shit
It's source code got leaked so you can run it locally
It's 16:00 man😭 😭 😭
Claude Code the harness was accidentally open sourced
Claude the models were not
Morning🙏 ☀️
Wtf that's how you know its premium
There’s 100 different opensource harnesses anyway, it doesn’t matter
5:30 pm
I don't even know what to do with this cause imo anything over 200$ is a scam😭 😭 😭
But well now i have it
Yeah, and honestly after peeking into claude code its not all that special
Its the best just by virtue of not being overengineered to death like a lot of the open source rivals
The special thing about it is it’s 100% vibecoded :D
Hey if it works 
I mean yeah
Imo codex desktop best harness
nah, 4.30pm here
If u could put openai compatible api
After the recent update yeah it probably is
Now you can even remove its restrictions as of a few days ago
Idk if codex desktop allows custom models
What forces people to use CC is that Anthropic forces people to use it with Claude sub innit
I know codex cli does
I mean these days deobfuscation tooling can be vibecoded
Idk how people can vibe code in a cli, the workflow sucks
Works fine for me
amps are surprisingly simple and raw parts aren't even expensive yeah
including the high quality parts like example the caps
Claude Mythos can apparently just rebuild simple closed source soft from binary
Well put it this way it probably won't be bad 
If you sic Claude on obfuscated binary how long will it take until you get deobfuscated one
Assuming good tooling
Also Brise stuff is supposed to be genuinely sublime
nixos on my rpi is not going well

It will maybe cause my current fiio amp dac has flagship mobile amp/dac while this one is diy custom solution
I wonder if GPT 5.4 Pro gets similar results to Mythos
Similar API pricing at least
i don't like the concept of posting a meme in prog channel but i can't think of a single place rn where people would appreciate this 
Absolutely not
Mythos is leaps and bounds superior to GPT 5.4 in any configuration
Even if it falls way short of expectations it'll still almost certainly be markedly superior
Though of course that comes at a ridiculous cost
Depending on your use case coding agents can actually be quite good, but no off the shelf setup is going to be perfect yet. You definitely need to set up proper scaffolding for your needs and rely on proper coding practices (fully built out specs, TDD, isolated pr agents, etc.).
aka - useless
Are there commits in Linux kernel co-attributed to agent yet?
I think it’s required now for slopper commits
Maybe for you, but I can spec out an application and walk away, and review what's done when I get back. They perform as well as any intern I've managed.
(Was a MLE and Project Lead prior to being an AI researcher)
I think its a good way to fry a devs brain
Quick dopamine hits for a mental task are always a slippery slope
claude code is special in that claude works very well with it due to them being intertwined that's it
that's basically it
When u look at its source code its 99% prompt
Telling it not to segfault everything
now that's thinking with ur dipstick
The OpenBSD kernel crash via TCP that Mythos found doesn't seem to be co-attributed. But Mythos also finds (and implements) exploits fully autonomously so maybe attribution starts to get weird?
I think coding assistants peaked at intellisense
I wouldn't go that far
Being able to run a program with image and text understanding this robust on my consumer GPU is far from "useless"
on second though, you are right
there is nobody in the world currently refined enough to appreciate it
😔
Can’t wait for ransomware delivered by agent for more token bux
wrong kernel
Claude code also works with other models better than as example qwen code, or opencode. I run claude code with qwen3.5-9b and it works fairly well
What do you mean wrong kernel?
(aint no way im paying for ai useage lmao)
IIRC CC patch tool isn’t 100% optimal
https://github.com/can1357/oh-my-pi
This thing was benched to be p good
imagine just copy text to any 'to binary' program
the mandated assisted-by is for the linux kernel 
My point isn't that this task is useful, its that it demonstrates how robust and stable even small local-viable models are becoming
when you give 1 layer task
It needs to transcribe the text from the image then translate it into valid binary text
That's a multi-step process and very easy to fuck up
it is one layer
main{
text = imageReadModel(userinput.images)
return languagemodel(prompt, text)
}
Its kind of local of a task in a way though
And the binary translation 
nah, if it fucking up this...
When u start getting more abstract and modular u run into shit real fast
chat I agree with Toast
there are 127 commits with this in master apparently 
doubt ai have built in tool for this, so it is language model part, and there may be hallucination (tho unlikely on given context)
There were no tools used here
I just prompted the model with the image and the text
Man the field really is a shit show these days
chat link me your fav non-LLM papers on ML
The binary it returned is also valid and accurate
and what?
it prove nothing
it is one layer task
bro wtf just use text detection and convert directly with binary encoding i don't trust the clanker with its potential hallucinations.
greg doing things 
Ah thanks for the info!
I'm showing off this task to demonstrate how robust and flexible the model is not because its practical
The capabilities demonstrated are useful
can you guarantee this will always work
It can take text content from any image and accurately transcribe then translate it
That alone is remarkably useful to have on a local machine
because istg, i get your point, but you're demonstrating the wrong scope sorta
using the llm for something that can be simply computationally solvable is inefficient
It doesn't need to because you should always validate output anyway, but I'm yet to see it fail
I do get it, I was just using binary as a fun example that's very easy for older LLMs to mess up
defeats the purpose ngl
Offline translation isn't computationally solvable to anywhere near the degree of quality and accuracy an LLM can manage
And you can check with a rerun or two if you're really unsure
translation is different
The actions and their quality for the cost of computation are simply not worth it
Thats my honest opinion on the field currently
linguistic is one hot semantics mess so throwing it on models like llm works
Until llms get aware of the way it itself is working, itll be always a bit shit imo. As of rn, we can use em as advanced "guideliners" for common stuff.. aka asking what to search for XYZ, and sometimes having them search for stuff and complete a prototype or something.. more is gonna get tough imo
I was looking for an example of model stability rather than a practical task, if I was aiming to just find an example of the latter I'd have translated Japanese to English or something
Their job is to generate not think
Like it or not their architecture is built for that
(and i mean really know-know how theyre working, aka model-specific, what each "neuron/weight" in them is doing EXACTLY.. which atm is so computationally stupid in a training run afaik.. always gonna be lossy)
i have my side of story for rant aswell
these models are clearly NOT "phd level" or whatever its literally being advertised as
Like this is already objectively useful and the cloud models are leagues ahead in terms of versatility
Oh I agree with that
i regret trying to use chatgpt to help me with some astrophysics shit
Lmao
its so fucking dumb
Curious, do you have the prompt still? I noticed qwen as example absolutely curb stomping chatgpt lately..
yo am I stupid how do I read it
i waste some time arguing with it
Anyone else noticing chatgpt getting stupider with each new version?
its... not even fucking formatting desmos equations right
i had to manually pull out astrophysics paper and figure out which equation to plug to others
Its citing a book but u can prob find pdfs online
yeah
I think it might be because they make the models bigger so they need bigger prompts for accurate responses
on top of feeding off ai-generated texts scraped off the internet
They also become less specialized
They always have been tbh
I tried using gpt for some basic unity-related stuff
and it just gives me the wrong solution every time
Sounds about right
gpt is horrible for unity
its frustrating because once it did something wrong it tend to do more wrong shit every time
they dont know jack squat about interfaces
how are these even called "ai"
especially with uis as frequently updated as unity
Its all A and no I
like my best usecase for LLMs that I could think of
is making it ask me questions
because questions at least can't be misinfo lol
Ha good luck
Fair. I also think theyre trying too hard to adhere to guidelines so the thing tends to give up or just say nonsense before double-checking.. ofcourse reducing their costs (shorter messages, less hardware useage), while making the users use up more messages.. in comparison to qwen which "thinks" for about 14 years before answering decently.. atleast usually.. the most frustrating thing i noticed is it contradicting itself more and more frequently
best usecase i can think of when using them is for humor purposes
I’ve started calling “artificial intelligence” artificial inference because it’s literally just that
They are know it alls when they dont know
ye this
neuro
Machine learning is a better term
machine inferencing
Corpo ruined the term AI
ik
but to keep the “ai” acronym yk
ain't even learning beyond training phase
no, RL is NOT "learning"
Missed the debate yesterday
I'm gonna resist the urge to get into this debate again
Lmao
big dog you're wrong just admit it
Nuh uh
whether LLMs learn or not
its literally just making the model learn to prefer what its already trained
which they don't
Na about models being able to learn during inference






per second






