#programming
1 messages · Page 490 of 1
this is why you girl math it
ranging from under 10 to over 20 depending on the pharmacy and use of discount programs.
google
so the answer is between 5 and 25

numlock 
the answer is between 40 and 70 last i checked
though if you extrapolate to monthly its much lower

sam should I be worried
you seem to be searching estrogen prices as deeply as you do with pc parts
what's the context
uhh anyways
I was dualbooting Windows 8.1
then whilst using the music app
I got an absolutely aesthetic "now playing" screenshot for a TV
like, I use my TV as monitor 2#
and I had the music app in said "monitor"
pierce her I hardly know her
@elvy buy him dresses
this + dimmed light in the room = perfect combo
interesting
warp went oss
sponsored by ...openai
can't wait for it to be renamed to "warp codex"
eyah that's strange since the full thing is like
a codex direct competitor
like they are the same fucking thing
world needs 10 more projects named warp
the only one that matters is os/2
not really, warp is a terminal that happens to also do agent stuff
they should all get threatened by gene roddenberry then they can talk
codex is the other way around
i guess but they kinda converged
maybe they wanted to make codex even more terminalesque and they were like
oh for sure, they've been pushing warp hard into the ai direction over time
warp was kind of nice last time i tried it but that was like a year ago so i'm sure it's a completely different product now
i wonder if they still have the single toggle to turn off all AI features
i have used it off and on over time and i liked it wayyyyyyyyy more when it was more terminal less agent
can someone give me a tldr what warp even is? their github page doesn't really explain what makes it different from a normal cli tool
it's a terminal
it's basically konsole
with rather deep AI integration
but with claude code built in
why would i want a terminal to call a terminal tool
it's not really that
it's like
you can do terminal commands
or you can type in english

because people can't remember cmd line names? or get a proper shell?
think of claude code if you were defaulting to ! bash mode
rather than the other way around
i have never used claude code
well in it you can ! and it'll let you just use regular terminal stuff
the only cli's i used so far was the one from github copilot and opencode
it's like if opencode were the entire terminal program

i guess opencode has a standalone thing now right? like that
0.2026.04.08.08.36.stable_05 what a version number
average warp versioning
semver? i barely knew her
first screen nice 
ye, quite heavy though (it does a lot more tbf)
yeah it's got a lot jammed on there than your average konsole
basically a whole ide
at this point
if you want it to be
making things more confusing
it uses
just for you @maiden geyser
this warp is one of warp (terminal)'s dependencies kekw

why does the terminal emulator need a built-in webserver
i think it's for the various AI features that require a backend server, so you can connect to a local one instead
i don't think i'll ever get over the fact that rust build scripts using println to interact with the compiler
ugh
the more i look at the hardware in this server, the more we find that i scammed the fuck out of the seller
ofcourse the 4090 alone is worth more than the 1850 i paid. but i keep finding sas drives, m.2s, and other random bullshit
if you personally dislike using it there is a crate you can use that wraps it for you
I also dislike it though
crates.io serves as a central registry for sharing crates, which are packages or libraries written in Rust that you can use to enhance your projects
oh it's official
also seems to only have 1 dep which is nice
why are we storing a json string inside a json file 
they already depend on serde_json

the cake role looks like a litle hat for evildentge
I like the idea of good rev focused debugger for Linux that doesn’t sit in CLI
Your mind has become toast
true
the
haunts you
But I also get the feeling this won’t have convenience features that, say, pwndbg has
(which is just GDB+scripts, but doesn’t matter here)
Toast, Toast, Toast, and Toast🔺 are typing

I mean sure slopmachine could probably port the features in 0.5 seconds but yeah
https://github.com/x64dbg/x64dbg/releases/tag/2026.04.20
WTF debugger MCP? Now that’s some tru extreme slopper tooling
With Asimov v1, you can build, train on, and make it your own humanoid robot. It's the first step of building a humanoid labor force for the rest of us.
Asimov v1 is 1.2 m tall, 35 kg, with 25 actuated degrees of freedom. Structural parts machined in 7075 aluminium and 3D-printed in MJF PA12 nylon.
We're releasing the mechanical design and sim...
Speaking of slop machine
Claude after creating the 50 least functional python scripts ever compiled
happy birthday @amber fractal

may you be harpooned many years to come
Damn
Is that on 20€ plan?
🎂 
Or is Claude pricing just turbofucked
22 now iirc?
It can’t be both?
So proud of you! 55 is a huge milestone!

not for @real sierra (happy birthday to you too)

I am 22 yeah
Then why react dentge 
Yo guys, I'm wondering, if next 70 years that Ai can truly reliably code entire OS, assembly, low level systems, any kinds of code re, what will human do?
I'm very curious
I'd be more shocked if it couldn't
I mean 70 years back from here we didn't have much in terms of tech
Or at all
the answer is too political to talk about here
economy be looking different for sure
the answer is nobody knows
If next 150 years, where everyone can just one button generate entire OS, low level, assembly, any library/frameworks, high performance, backend, frontend, any systems, linux, any system with perfect code/performance, reilably truly, by just 1 click and little prompt
Imagine little kid could just vibe code entire OS by pressing button
I'm also optimistic about this and feels brutal at same time
Let's sleep
I'm just thinking too much ig
I don’t think the question involved politics at all tbh
cant wait for artisan oses crafted by real humans
"back in my day, OSes were coded by hand..."
pattern matching infinite trees, connecting each tree leaf to other tree leaves, allowing any kinds of bidirectional tree transformations
nowaying Vanor 
i'm currently converting to support arbitrary graphs, not just trees
because currently my product patterns work more like dependent functions
happy birthday!
o true
happy birthday iggly
i tried using it to create a program that computes the equation b=a+2 using church encoding
when you pass b, it gives a, when you pass a, it gives b
and its impossible to pick a number less than 2 for b
which means it's working 
How do I start larping programming guys

Do I start coding Roblox next month
Mmm run=true
@safe path the end goal is to create an engine for writing arbitrarily complex bidirectional programs
like a parser that can produce ast nodes from text and text from ast nodes
i want to use it to create a programmable editor (not necessarily for text)
since editing is all about bidirectional transformations between ui elements and data you edit
Uhh...
Bisexuals are cool I guess...
oooh shit thats cool actually
i think that has applications, imagine switching between normal text editing and structured editing with no info loss
yee my goal is exactly preserving context
i dont like the existing approaches to interactivity, rust-analyzer just caches everything, thats so boring
so like, noob question but how do I do a combo system? I have the idea of making a combo-based platformer and I want to get it coded and ready to go
The first serious project I worked on was an open source emulator that had a menu item for "About <emulator>" which didn't actually do anything. So I cobbled an About dialogue together by looking at the rest of the UI code and tweaked it until the maintainers were happy with it. Open Source with patient maintainers is good learning experience.
look into how regex works
combo system is basically implementing a regex engine
regex = regular expression I think
oh...
I think that the way that you meant that was like, to store inputs regex is very useful
The seller lied, its not 128gb, its 256gb 
https://en.wikipedia.org/wiki/Nondeterministic_finite_automaton
https://en.wikipedia.org/wiki/Deterministic_finite_automaton
wikipedia isnt a good starting point, but these are the topics you'll end up using one way or another, even if you dont learn it formally
rather than making the player suffer by forcing them to time their inputs right when the animations are supposed to take place
I'm more of an academic style of programmer so this is actually very helpful
wait i think i misread it
i thought you were writing a fighting game
disregard what i said
I haven't found an engine that I fell in love with, most of them either use a proprietary language for coding (blech) are badly optimized (Unreal) or they have awful T&C (Unity)
I'm trying to basically implement fighting game controls but for platforming, if that makes sense, so even if it was for a fighting game it's still helpful

formal language is kinda fun
like assuming IJKL are action buttons (I = dash, J = jump, K = kick, L = duck) then something like JL would stomp, or L while running puts you into a rolling animation and LL causes you to brake, and LLL causes you to crouch after braking
these are like 300 euro per stick.
of course it's not exactly like that but the gist should not be hard to grasp
You should get a refund
Or give me the spare 128
i think "longest-match" would help here?
what happens is the longest input sequence that matches player input
i dont mind having 256gb of ram
maybe it would be unintuitive though 
its ECC 
elliptic curve cryptography?
Nevermind you can keep it
Serverslop
error correction codes
It’s a stupid thing for stupid idiots where it saves 4% power in return for being incompatible with consumer motherboards
corruption-resistant ram, used where high ram reliability is required
ah makes sense
They are the exact same specs, shape, and connector, but this simple change is why datacenter ram can’t be put into gaming pcs after they upgrade the servers to new ram
Basically the reason for the ram shortage (dramatic oversimplification and also just plain wrong)
anyway
nfa has:
- states
- input transitions (transition from one state to another that you can do for a given input)
- "no-input" transitions (a possibility of going from one state to another "for free", without doing any input)
when you match input, you keep track of current possible states. when the player does an input, you traverse the corresponding input transition from each of your possibilities, and remove that possibility if it doesn't have a transition for this input. after that, you also add all states reachable via no-input transitions to your possibility list.
when you have multiple matches, in regex longest match is used, in your case you may or may not want something different, just make sure there is a way to deterministically pick the state you want if you have multiple option, ideally without assuming the options' order in memory.
dfa also has states and input transitions, but unlike nfa, there are no empty transitions. furthermore, theres only one input transition for each state. this means in dfa, you only have one possibility to keep track of, and you just follow the input transitions.
the good part is, any nfa can be translated into a dfa. dfas typically require more states for representing equivalent possibilities, but in exchange they are super easy to traverse.
also, "input" doesnt have to directly correspond to player input, it may be something like "waited over a second", "bumped into a wall", or "pressed x with a delay between 200 and 300ms"
FINITE AUTOMATA SPOTTED
How do people even make stuff
Like what did they learn bro
Did they make a run script on Roblox 300 times
And then suddenly make a fucking
This guys would rather do anything than save up spending money for his intercontinental vacation in 2 and a half months
What
Programming
Are you asking for tips on trying to learn or are you just philosophizing aloud?

interesting
DXVA 2.0 is kind of a really popular hardware decoder?
and black frames rendered, wtf
today
I fall into the llama.cpp rabbit hole
Ive decided
no time to start like 11 in the night on a school day
That’s the llama.cpp rabbit hole
The point is
Actually hold on
no dont worry I get the point 
i may be stupid but im not dumb
wheels on the bus go round
fast
powerful
Doesn’t that significantly depend on hardware
The joke is that he is 3 years late
But yes
lLAmA cPP Is SUcH a GAmE chANgER oMG
actually the future of ai is already in u, you see, there's this really cool AI framework called the human brain
I do not have the money to buy liquid cooling stuff for my GPUs
Now you know how I feel when my GPU reaches 85C
have you tried undervolting?
i dont think your pc uses 60mm fans going louder than a vacuum cleaner 😭
like, idc about the last 10% performance, just give me the cool temps 

I can't directly undervolt (though apparently they recently added voltage curve control finally to LACT) and the daemon that lets me control that stuff makes my PC unbearably stuttery
My GPU's 3000RPM fan mode sounds basically equivalent
i can promise you it doesnt
Also I have exceptionally bad luck on binnings, all my chips are like lowest viable bin
I can with my 4090
no desktop gpu will outnoise a server fan
Linux remember?
Linux doesn't allow that? no shot
Mine gets pretty close
i should undervolt my 4090 and 7900xtxs
It's one of the few actual lacking things on Linux though they are attempting to fix it
3000 rpm is nothing compared to centrifugal fans blasting beyond that
I have a 4070Ti and a 3090 that sits idle all day, and because the 3090 is in the way the 4070Ti easily hits 85C under load
if you want an easy solution you can just do an overclock and power limit
but manually setting it allows a lower voltage per frequency setting and it won't fluctuate all over the place
the 4090 might have enough juice in the psu for an oc
but both 7900xtx are on a 750W psu
Well time to use experimental undertested software
well, you are reducing the power
it will run a lower clock and voltage than stock... it's just that they leave headroom available at stock settings

That has for a long time been the official way to do an "undervolt" on Linux
My GPUs crash though
Oh well, I guess I just can't do anything to tune my GPUs because they're too garbage binned
I dunno, could probably fake a tornado warning by ramping my janky shower fan up and down
Two waterproof industrial 24V Noctua NF-A14s hooked up to an outdoor LED power supply, with a 5V converter and $12 chinese PWM controller.
In hindsight the 24V fans were overkill. I keep them at like 35%.
At full blast they can replace the entire room's worth of air in 10 seconds.
I just have whatever they installed in the bathroom ceiling
Yeah I just had this dinky window which wasn't enough, but was exactly the right size to stick two 140 mm fans between the window and screen.
Exposed dupont jumpers under the sink. What could go wrong 
Fun fact: PC fans use a 25 kHz PWM control
Seems overkill but maybe it's so the off/on doesn't do weird acoustic things? 🤔
He ignored me smh
ignored what?
This guys would rather do anything than save up spending money for his intercontinental vacation in 2 and a half months
i have the money for that saved up already
but this is a 6K server i bought for 1850



Oh really?
Why don’t you tell the class about all of your epic open sauce plans then
Hmm
That's like 8K in just RAM with current prices
then go home
its 1.2K, but close
Finnish prices must be even more brokener then because 32GB ECC DDR4 is like 1K
depends on if you get them used i guess
its a hmaa8gl7mmr4n-tf
or well, its 4 x hmaa8gl7mmr4n-tf
thats 3200, this 256gb is on 2133mhz
Oh no my steak is too juicy
Still cheaper
that one idk
RAM prices are just completely screwed
ECC is usually even more expensive
depends
used market ecc is cheaper cuz nobody needs it

Exactly my reaction
I don't know about you but I'd say 11K for 256GB of RAM is a liiiiitttlleee tiiiinyy bit overpriced
Like what even
Shit so inflated u can use it as a bouncy house
This is even cheaper than Sam's 4090
Gamers are gonna have to start working the mines to get the raw materials themselves
Restarting Discord and the RAM use drops by 10GB
Electron moment
does Discord in Edge use Electron?
I think so? it's just what it's coded with, right?
wdym discord in edge
wtf
wrote random letters and got random program open
The frick you mean just a cleaning
You don't just clean a CPU
AHHHHHHHHHHHHHHHHHHHHHHHHHHHHHHHHHHH

I don't use the discord desktop app. The webpage at least doesn't seem to have a memory leak problem.
well, a lot of people use the dedicated app for Discord
I just meant in browser
but, yeah
the answer is no
discord in your browser uses your browser
not Electron?
well yeah why would you need electron in the browser
your browser already has a renderer and js runtime and all the fancy stuff
so electron is not needed
Overfitting an LLM on discord profiles to see what happens
also you don't have to wait forever for updates to finish
Yeah I always hated the constant updates.
updates tend to be pretty shit too and browser version gets them late so more benefits
Functionality wise I don't even know what's missing beside RPC, which I don't want anyway.
I have 1,200 unique profiles and, via random sample callbacks from 100 epochs, 10,000 fake profiles will be created
mostly it will overfit on discord profiles
Someone get this guy a Turing award
i used 120 cpu cores to come up with that one

too zoomed in for showing the full 80 threads
lmao
btop uses so much bandwidth if you run it at 100ms
actually
that's insane
it uses like 5-6 Mbps just sending updates at 100ms interval
those old instinct pcie cards have such a funky shape to them
You think the shape is funky? How so?
Other than the fact I'll need to add a custom fan solution, of course.
I'll probably repaste them, before installing, but I'll also likely drill through the shroud and use aluminium foil tape to duct the airflow from a 140 mm fan or something.
High airflow compressed through the shroud, that way.
just the way that they arent actually rectangles but get larger from left->right
there's a lot of 3d prints out there for jury rigging a 120mm fan on there (goofy looking but)
or a 60mm fan or a blower type unit i see alot
The top of the card can't be covered near the grill. That's not really an aesthetic design choice.
They have a cut out, bc these cards have a huge set of gold fingers up top, if you have an Infinity Fabric Link, to combine the processing.
I was confused, when I first saw this, bc it looked like a modified second PCIe slot
I hadn't seen those, before, and it seemed way bigger than a Crossfire Bridge, but it's apparently similar in functionality to one
But you can't find a Link for this particular line, anymore.
you can find them
they are just a billion dollars for no reason
and yeah but they could do like nvidia and just cover the fingers for the crossfire bridge
it just gives them a funky look like theyre a bazooka
What kind of cutout is that. How tf did the middle part connect to the rest
there must be a market for making those bridges
the mi100 one doesnt work
i dont think
that one you can find
it's like 800 bucks tho
I mean, I'm sure SOMEONE is selling them, but I didn't find any.
Not sure getting the bridge is ideal, anyway, though
https://www.exxactcorp.com/AMD-199-999762-E142726124 this is the bridge (this listing i doubt is actually for sale but it's the 2card bridge model)
4card one is model 199-999764
I'd like the option, but it's probably not worth more than $50-100, unless it connects like 4 cards together.
Combining the processing, but losing the ability to have 32 GB VRAM, doesn't really seem like it'd fit what I'm be doing.
2 cards with 6.7 : 13.4 : 26.8 and 16 GB each, probably do better than a 2-card cluster with 13.4 : 26.8 : 53.6, but with duplicate VRAM, so only 16 GB usable
Hmmm
I see
The Mi100 is also like a hot $$$ thou, each
unfortunately, completely different interface
ye
they are supported in rocm still, tho
and are all 32g
considering 32gb 50s are like 600-700 it's kinda worth imo if already rolling cdna ofc
i'd be figuring out a way to get oam mi250x rolling before i did that tho

each one being 2 gpu dies and 128gb hbm
a VERY fun one starin g at me like the green goblin mask https://www.ebay.com/itm/168238478919
1:1:1 fp16/32/64 52.43 tflop in all 3
what is that second pci anyway? you put some kind of adapter on it and then plug another gpu into it?
oh wait now i see
Well, you don't need to use ROCm. You can just use OpenCL or even Vulkan kernels.
ROCm is basically just trying to get the card to run like an Nvidia card and do CUDA closer to properly. Don't need it, if not using CUDA for compute
well, i mean, kinda
most everything is kinda cuda 1st that's the main issue, torch.cuda is omnipresent in much compute and being able to hipify is very valuable and/or required for anything where you need something that is shaped like cuBLAS/cuDNN/NCCL you really need rocBLAS/hipBLASLt/MIOpen/RCCL/Composable kernels
vulkan is great for cooperative matrix in inference but you cant replace the aforementioned really for fused attention/flash attention/paged kv/most tensor parallel collective operations such as All-Reduce or All-Gather (this one is ultra important)
there's no real triton w/torch.compile or anything with Inductor that i know of that is very important for cudagraphs
SPIR-V exists for runtime JIT but it doesnt work w/ pytorch
basically anything on the training side of things is a no-go with vulkan still
ot
it is POSSIBLE in that vulkan could do these things if you got like 10 big brains to work on it for months/years but
this is an issue intel runs into as well and theyre kinda going the thanos do it myself way so it'll be interesting to see how that goes and if they bring CL up with it because amd openCL is stinky
given that openCL is literally just the proto-HIP for amd nobody targets it for anything useful
no one uses slowtorch anyway right? 
never ever
except when they do (all the time)
we need more real programmers and less pseudo-scientists
anything that gets as deep as ptx (triton etc) it's joever without rocm since that requires amd's llvm backend which is baked into rocm
it's possible (target the kernel driver) - mesa (and radv) both do that to some extent but there's no existing runtime for that that is triton flavoured, nor is there anything to replace the rocm rccl/loader/queueing
doing all this thru vulkan would be such a fucking undertaking it'd be unreal
a good chunk of vulkan work that amd vulkan runtimes use is also riding off the back of nvidia's work on vulkan for their own purposes and likely wouldnt have happened without that
like nv cooperative matrix exxtension for vulkan that turned into the open kroonos version that amd & intel used
it gets much worse the second distributed compute is involved especially, i dont think it'd even be possible without rccl (nccl)
if you can do it with game shaders and local compile why not do that with computes?
should be equivalent, both are gpu code that needs to get compiled
like i said you can do it
shader goes thru same thing w/ shaderc or glslc
this is why I should never let my friend do databasing shit atp
spir-v/glsl doesnt have much of the asm type expression needed for PTX parity though which is what triton ends up relying on to actually specify how the hardware is used so the performance isn't as good as more is left up to the driver which will do whatever
which is fine in that it can do it
but
it's not as good
which means nobody does
which leads to a self reinforcing loop of well it's not as good so target cuda
and then everything else surrounding that targets that follows suit
intel did/does this with sycl & openapi
it's just
still stinky
and without a full vendor tier effort itll never get close enough to really matter in the way that you'd want it to to rely on it for anything where time is money
inference works because there's really not much to it as far as what you have to reimplement in vulkan in comparison, orders of magnitude less, and also it's not a conscious choice as much as a training backend is in that inference as long as you have tokens shooting out of your model people really dgaf in comparison, for that it's like well does it fucking work at all? yes? all g
society if every hardware vendor just supported the same software stack
honestly that's vaguely what amd is doing with rocm because nvidia certainly isn't just offering cuda when that's literally their biggest advantage
is this the one who touches your software and each time you come back with what appears to be nuclear waste where source used to be
yes, and they still use nextjs and praise it like its god (they asked me why they couldn't do preview deployments to cloudflare while maintaining prod on vercel)
oh nah, that guy should be locked out of real code

I've had to review his code multiple times before...
just the act alone makes me feel like my braincells are lost
weird, i didn't touch your db

see the thing is
they don't write the code out of necessity, unlike you and your typescript type inference madness
they write the code because they don't know wtf they are doing
out of necessity is a stretch
either way
the code is like, absolutely horrid
so much so I just decided to duck out of frontend for that project
it shouldn't be that bad, there's only so much you can do with rsc
unless there is global state (not db) in the backend idk
im willing to bet he added a bunch of "use client"s
actually no, scratch that, he probably didn't because he doesn't seem to know what RSCs even are 
what do you mean, how do you do next then if you don't
I have no fucking clue
atp I'm going to probably ask him to rewrite it in something more sane, or actually utilise/understand nextjs
Yamete
i got email saying if i will not register myself in some list
they will delete my domain i paid to
well, i guess i have to find another domain provider
remove repo access
say you won't look at his code until he gets retrained
I can't, he owns the github org under it
oh nah
I told him I am absolutely not touching frontend until he makes it better (and also stops using nextjs since he so very much wants to deploy to cloudflare for preview but vercel for production)
which provider was this
if you cant do -cryption stuff without chatbot
dont do it
i tried to use fucking ai to gen code
and it backfired in worst way
it just did what i told it not to do
was lazy for reading documentation
reg.ru - russia domain zone
you may not care about it anyway
||i hate this country - this government||
it is region specific shit
i will buy something normal later
cuz i have dollars card now
did you end up filling out that form
never met a russian that was happy to live there

I do not speak Japanese but i understood that
Hey guys. I have news. I visited. . . a week ago maybe? Looking for guidance? Zero is now discord integrated, has OCR and a rag memory system. She's rock solid foundationally.
<333
It's bad to use encryption you don't understand
It's REALLY bad to try to write encryption stuff, if you don't understand the mathematics behind it
It's straight dumb to ask generative AI to write encryption stuff that you don't understand
It's full-on batshit to ask a generative AI you don't even run locally to write you up complex AES encryption stuff you don't understand
no one's reimplementing aes dawg
stop making weird assumptions about a screenshot you've seen a tiny section of
you don't even know what I'm doing
it's literally just calling python's standard crypto modules
that's it
But do you know how it works, and can pinpoint what is happening behind the scenes, only receiving generated code from an AI you're not even running yourself?
yeah??? it's literally just chunking AES GCM decryption for me
Then why do you need an AI to write that up?
because the chunking part is annoying to write on my own
are you just trying to pick fights for no reason
Write it once, store it on a local repo (or even upload it), and include it whenever you need it 
probably yes
the decryption is only against this specific case of malware family...
you don't even know what I look at normally stop judging what I'm doing based on a portion of a screenshot that's not even relevant to what you're screaming about
it's always the nn chatters 
-# do I count as one...
you have more messages here than nn lol
never realised that
ig it makes sense tho
I did start here first and then went into nn gradually
Nah, just saw some fun news semi-related to this.
inb4: Entire file store or all records get 1-way encrypted, and all is lost forever
no one is running anything it's generating autonomously
it's through the chat interface
I look at the code before it's even run
ok I can't lie tho that shit is funny as hell 
are you being weird for the sake of being weird
apparently that's a hard assumption these days
IKR!! Claude nukes ALL the data AND the backups!

🍿
I don't even have any agents on my computer 
it's like you read one thing on the internet and assume everyone else using anything adjacent must be doing the same thing
obnoxious as hell dude
pretty toxic too
^
about as obnoxious as this tiny chicken sandwich that I just got
like I get you hate some stuff. I hate some stuff too
but you don't have to go nuclear just because they are doing something you disagree with
they don't even have the full context of what I'm doing so i don't even understand the disagreement
inflation
oh yes having an AI reverse engineer an encryption used by a malware and have it write a python decryption equivalent is sooo bad what are you doing booo
that shit would have taken me at least an hour to figure out how to decrypt the chunks on my own
but azaka what if you acccidentally deploy it to production and it charges you 300 claudebucks

not my Claude bucks


what's a good emoji for that
i was trying to paste a image
then my paste wasnt working
I got so crazy I was checking my keyboard
then I got to know im not eligible to put images in this server

good stuff
about 200 posts
sufficiently spaced (there';s a cd)
you are close

do
you
mean
like
this?

so many questions, like why would you give your agent the prod db
“NEVER F**KING GUESS! — and that's exactly what I did. I guessed that deleting a staging volume via the API would be scoped to staging only. I didn't verify. I didn't check if the volume ID was shared across environments. I didn't read Railway's documentation on how volumes work across environments before running a destructive command.”

The PocketOS boss puts greater blame on Railway’s architecture than on the deranged AI agent for the database’s irretrievable destruction. Briefly, the cloud provider's API allows for destructive action without confirmation, it stores backups on the same volume as the source data, and “wiping a volume deletes all backups.” Crane also points out that CLI tokens have blanket permissions across environments.
i mean yeah that is stupid
NEVER F**KING GUESS!
deserved for having a cringe ass instructions md

my music stopped and i got this
ok
does this "do not hallucinate" stuff even work
how many more incidents before they learn the lesson that LLMs are not the right tool for any logic related tasks
the only one that slightly works is to tell it to be not sycophantic after having to tell it that multiple ways in a roundabout fashion and even then it forgets that a couple sentences later
barely even that really, unsupervised "agents" are the real issue with infinity perms
"<insert random task prompt here>. do not hallucinate, act like an AGI, act like you're sentient, be a super intelligent AI, do not make mistakes"
i would shit mypants just knowing one of those nutjobs was just running on my system and doing what ever the fuck
without me watching
they're only for sequential logic text based task only
and they said letting an intern loose on the company db would be bad
AI likely cannot understand GIFs
agents is just a self-prompting LLM system with exposed APIs to other software
what gif is it?
so agentic approaches inherit all LLM issues
your first problem was using gemini for anything
Gemini is fine
i would like to disagree
it is so genuinely stupid I am amazed at how they call it an ai
indeed Gemini is Gemini
sorry for my mistake and thanks for pointing it out, i'll make sure to be correct in the future
they are text generators
and languages aren't an accurate representation of logic
no problem,we're all here for the ride anyways
humans have rapidly dragged down what is considered intelligence lately, that's how 
i mean intellingence is anything except being <insert what you don't like here>
true.... bad data means bad output
no, data is unrelated to output
we could have perfect output with bad data
in the case of an autoregressive model the data it generates prior is directly attended to and influences the next outputs
just like you can randomly guess a valid result
so it kinda depends
hyperparams and whatnot mean non deterministic as well so throw the lotto in
I meant as in the data gemini is taking in recently on top of its good data previously trained before poisons the previous one
also gemini sucks, idk why but some people say internal politics, som say plain built stupid
that is a very poor argument to make 
I am not a mathematician because I can randomly guess. even if it is randomly correct, I am still bad at maths
i know i know i'm saying nonsense
trying to imitate AI
gaslighting is hard
how is AI so good at it
you are allowed to criticize them without rubberbanding so far in the opposite direction that you imply that an llm is somehow less reliable than a monkey typing in a terminal
LLMs are relatively good for their intended purpose
next token prediction
more likely to cause damage sure cuz of their competence combined with incompetence but still
it outputs text without thinking
ignorance can be very similar to confidence
and it got trained on convincingly sounding data
i need to train more
so even if it swaps the relevant words it will still sounds about right
one day i'll be a professional gaslighter
we have enough of those
i mean it's what makes the most money
it was, but now the market is oversaturated
sadly programmer isn't a job of the future anymore at least that's what they say
classic automod
i should make a program to guess automod rules
shouldn't be too hard
just need to find the optimized string
the opposite
we see a lot of LLM failures
and the "success stories" are all fabricated
once the bubble pops the demand will go crazy high
shhh trust me bro GitHub will get back the 9 they just need more AI
they don't even have 89% anymore
just look at that lmao
https://mrshu.github.io/github-statuses/
Historical GitHub uptime reconstructed from archived status data.
It is, but it’s the syntax being used on the exam
my condolences
I love racket 
it isnt racket code (but it could be code for some dsl written in racket)
in the same way 5 + 6 could be c++ code, but isnt necessarily c++ code
We built the logic in class. app and lam are structs, that’s why it looks like this. I get what you mean though. It could be any functional language.
Hey so for the internal “x” I wouldn’t do the initial beta reduction on it because it’s protected by its own lambda “x” right?
i reduced it to ||\z.z(z y)||

Imma see if I get that
So I have come to the point where I am curious about all of this AI coding stuff. Can someone give me the resource to read the breakdown of the current state of AI coding? Like, what is the difference between agentic coding vs llm chat vs autocomplete and all that. What actually is tool use, mcp, harness, etc.
Because I want to actually understand wtf it actually is
that's a loaded question 
Which is why I asked for resource on where to read all of that
idk if there is one singular resource 
A lot of source is fine. After all, most of the software readme and wiki doesn't help much. All they really explain is how to use the tools, how to integrate stuff, how to host them, etc.
Take vscode copilot which is where they annoy me. What is it called when I interact with the chat on the right? Does that constitute as agentic coding? Or autocomplete? Or just plain chat?
you see that's the funny bit, copilot does it all
autocomplete is inside the editor, the llm has the current file(s) in its context and its job is to generate a snippet at a time
what you're referring to as "llm chat" is basically just chatgpt, a chatbot

well.. the autocomplete doesn't has the whole file in its context. normall its a bit before the cursor and then a bit after. with a <|fim_cursor|> or something where the cursor is and then it generate the thing between. the newer ones also get edit history as context thats how it can predict what needs to change next
the llm chat is basically just that. chat but don't do thing
What about tool calling? I understand about MCP and RAG but how does the model actually utilize those? And what about harness?
I got that! 
agents are things that then can also do things using tools. and tools are just function calls
<|fim_suffix|> 
Right, but how do the models actually interact with the function call? How do they decide when to do it?
they're trained to know
the output things like <tool> </tool> and then the chat interface detects that
the same way reasoning works
think of tool calling as the model creating a json request and sending it off somewhere
So it's random? 
there is also a chat template that controls it a bit... but its what they are trained on.. so not really random
Well, random in the sense how the model sample the dpf
at the end of the day it's all random technically sure
heavily skewed towards the desired outcome of whoever trained the model tho
https://huggingface.co/unsloth/Qwen3.6-35B-A3B-GGUF if you go onto a model site you can then look at the chat template under the "use this model" button. that is a template that is the populated by the inference engine. like llama.cpp and the like and it uses it to parse things like reasoning tool calling etc.
Alright, my brain can't comprehend how is that okay tho. Like, people just accept "good enough"
so last then, what is harness?
the chat interface, and the system prompts basically
the thing that directs the llm
opencode, claude code, github copilot cli, gemini cli, codex cli
you can't turn something fundamentally non-deterministic into something 100% deterministic, but you can get surprisingly close
are all harnesses that handle things slightly different. give different tools to models and different systemprompts
Maybe it is very confusing to me because the interface to system prompt and the prompt itself is just arbitrary text. But well, I think I got the gist of it now 
it's not arbitrary text 
i mean it depends on what you mean with arbitrary i guess
someone chose this and that's just kinda it
fuck
Arbitrarily language-like 
my meeting tomorrow got pushed from 15:00 to 11:00 
Look at the bright side. The meeting ends sooner. Unless....
it means i have to get up at 8am instead of 12pm
it used to always be at 12 or 13 but the prof has a lecture from 12-14 
densely populated scatter plots make brain happy 
What does this plot represent?
6GB worth of benchmarks
pareto front for a data structure i'm working on with a parameter sweep 
agentic lets the model go crazy style make itself a task list, run arbitrary commands & tools for web fetch etc, make file changes, move around on the system on its own, and stop when it "thinks" (it doesnt think if i dont mention this i will never hear the end of it) the task given is done or it cant continue it'll report back
generally best left to a frontier model as it can cause serious damage to your system if it's fucked up (although the more capable the model, the more capable at fucking wrecking your shit it is when it does fuck up, and it will eventually do some goofy ahh shit if you aren't careful)
you can sandbox them and such but they can get around it (lol) and never 100% safe but god damn is it nice for something if you want to have it smash its head against something for you that it cant really screw up too bad
tab completions (fim_suffix is from this - fill in the middle of a code string is what theyre trained on, many coding models support this alongside normal behavior) are like intellisense if you've ever used that which has been around forever, just 'smarter' aka much more contextually aware
Which Browser should i use?
1
1
2
Brave
bubble popping boys
Only 1 voted
you'll know it's officially joever when zig 1.0 comes out
God I love democracy
brave is chromium
the poll is flaw
answer is vivaldi
if choomium
Not mine but a user in the neurosama Reddit made this using an acer feature
bullyme senpai boot
Yeah, I think I'll stick with intellisense
much better to at least start with that if ever planning on messing with an agent
understand more about what kind of things the model tends to want to do
the only thing i ever trust to agent mode is building rocm
Ohhh, changing bootloader image. I miss that feature
"The env is fricked. Time to nuke."
*remove france language pack
i mean, that's based
what the flip is that ?
i had forgor i grabbed one when i saw the exchange rate
got a stackchan and the remote
stackchan is a lil agent
Our numbers have dwindled... Sad times
im gonna devolved into a bred
maybe i should give it a shot again... i just can't be bothered moving over my vim keybinds to zed
it's native 
it does but you need to reconfigure your settings... with nvim and vscode it just runs a nvim as a background process so everything stays the same
if theyre saved in vscode wouldnt it import them from vscode on first start
its not saved in vscode
maybe they arent because of the way it does that
yeah
well
skill issue
forgot to use emacs
the nvim extension in vscode just relays your inputs to nvim and nvim then back to vscode
so your normal nvim config just works in vscode
That sounds jank af
LMAO what
it works really well tho 
got that virus hijack hook
in awe that defender doesnt hate that
there's another extension that works differently but iirc that one is also kinda meh
i'm sure its better than that. but it just runs nvim in the background and that loads your config and plugins. idk how it exactly communicates tho
that was just my guess
not really
neovim is meant to be embeddable
has a whole protocol and stuff

That's... weird but ok
We need to all go back to editing in ed
If I just need to add text to a file I just use cat
LFS taught me it and it's been great for when I need to append to one file over ssh
as always my hate for windows 10 iot ltsc core is never unproven wrong
alright i got zed to recognize my most important keybinds. space w for save and caps for escape
another machine failed while printing my bill at the grocery
win11 > win10 fite me
had to fucking scan it all again
weird cobol program
(yes I know it was probably java but any old ass program i call a cobol program)
zed's cool and all but the one time I tried it, I found out they used a broken verson of Haxe's syntax highlighter instead of the main vscode one that they made so 👎 turned me off from using it lol
You have unlocked new role
makes sense. i just wanna try it due to the 1.0 release
i'll probably stay with nvim 
understandable, it looks quite nice other than that issue. And yeah it's hard to get away from terminal editors, I personally use Helix and it's hard to use anything else lol
"this will take several minutes to download"
god FUCKING dammitr
you never know when you're gonna get ganked by torch man
you can't escape!!! I love having an AMD card and I accidentally mess up a command and end up pulling in torch with all the CUDA crap 😛
this is unfortunately unavoidable as long as vscode doesn't use treesitter
well the thing is, zed supports textmates!
do they?
whoops wrong dependency #10984712349081809 decided to require all of pytorch and cudann
yep!
I figure they added the haxe plugin before they supported textmate though
but I don't really know enough about zed or their plugin system to modify it to use the vscode one and make a PR frankly
only reason kden is that davinci is blindge
but isn't tree-sitter fine tho?
OH
I swear they'd PRed it or something...
didnt tree sitter get yeeted into the sun
no that was only some neovim plugin
o
neovim still uses tree-sitter and that will never change
it's kinda ridiculous though because for Haxe, there's literally only like 2 options. there's the "main" textmate one that Haxe themselves made (even github uses it for haxe syntax highlighting) and then there's the treesitter one which is broken.
What's hilarious is someone reported a bug on it, the dev answered "oh sorry about that, i'll fix that" then didn't do anything.. then A YEAR LATER" he "assigned" the bug to himself, then another year later, NOTHING?????
the plugin you are talking about was mostly used as a repository for tree-sitter parsers. it isn't really needed in the end
it is but vscode is by far the most targeted editor and it doesn't use it so in an ideal world you'd support their format too to make people's life a lot easier when it comes to making plugins
I mean treesitter is perfectly fine, it's just that the "main" haxe treesitter syntax highlighter is broken lol

I did manage to mostly fix it with AI help (can't understand regex for the life of me) but it wasn't perfect so I didn't wanna PR
the issue in question is that if you use a // within a string (aka a url), it comments out the rest of the line (just visually of course) the issue with this though is that now in the highlighter, your 2nd quote is now gone for your string, so whenever you use another quote, it will turn the entire thing green lmao
the hacky way to fix it is at the end of the line after the ;, you literally just do // " so that it "ends" the string and doesn't make your entire code green LOL
yeah nah just use another editor 
exactly lol I just switched to cursor for haxe 😂 I like using Helix but I don't think they support textmate sadly so can't use it
I COULD also attempt to properly fix it but treesitter stuff is AWFUL to try to parse, it's just a giant ball of regex and ugh...
still cracks me up though that the dev said "oh sorry i'll fix that", said nothing for a year, assigned the bug to himself, then again said nothing for another year lmao
I'll never bash on a guy for not having time to fix his software or etc. but why assign the bug to yourself a year later if you aren't even gonna fix it?? 😭

really goes to show the chokehold VScode has on the market though, I mean even github uses Haxe's textmate highlighting
you know the beauty of open source is that you can add the feature yourself
Very true! Too bad I don't know enough regex to fix it fully :(
I thought i'd fixed it with AI however it either revealed or introduced another slight issue with highlighting, so meh...
me, are you the regex master? if so I beseech you good sir...
i like regex, i'll try to get a license at some point
https://regexlicensing.org/
but no genuinely i like regex, tell me if there is smth i can do
there is, actually! https://github.com/vantreeseba/tree-sitter-haxe the main Haxe treesitter highligher sadly has an issue with strings and // in the highlighting. (as stated in https://github.com/vantreeseba/tree-sitter-haxe/issues/53)
you scare me
Yes, regex people scare me too... I have a friend that does the regex wordle for fun...
https://regexle.com/ in case anyone is curious of this abomination...
It's a daily hexagonal crossword puzzle where you need to fill in the hexagons with character sequences, so that they match the regular expressions (\
https://github.com/vantreeseba/tree-sitter-haxe/compare/main...xenanthropy:tree-sitter-haxe:fix-quote-highlighting This was my previous attempt to fix - it fixes the // within double quotes issue, however it still has some issue with single quotes, I forget what happens though... (as i've since switched to the vscode plugin due to this problem)
i once made a URL parser and transformer for Youtube with pure regex
why ? because i could
sure i should have asked myself if i should
but i didn't
I asked my regex master friend (who famously hand-wrote a like 100 line long regex) and they said "yeah I have zero clue what's going on there" so RIP

quite fun honestly
INSANITY
tho i wouldn't do it each day, it's quite pointless
i should try to get a 1m- one tho
well I could see it helping someone stay "fresh" on all the different regex... thingies
cuz sadly if you don't use it you lose it
Took a month. But my new ram is finally getting shipped back to me
Not really it's too small
And it doesn't use advanced stuff
oh is it? don't they have giant ones? I saw my friend do a like 12 size one
ah, true
I'd need to look again
Even there it's bad
you can make it as big as you want
It doesn't have advanced stuff
Max 32
well frankly even this big is maddening...
I'll show you a parser i made with regex
It's part of the subject of my bachelor thesis
what the fucky wucky
through the demonic power magic of regex
Actually can't share it, company leak
And the stuff in my thesis doesn't contain the exact code
hah you're good no worries
But yeah no regex is cool
Actually i may have an early version of the youtube IRL parser with regex
im not doing this shit
idk that look clear for me how it works
i have no idea how to solve it tho
well | is the easiest cuz it just means "or"
i actually solved todays one
Early horrible version of it
find: (?:^|(?<=[\s,:;/&]))https:\/\/www\.youtube\.com\/(?:shorts\/|watch\?)(?:(?:v=)?(?<id0>[\w-]{11,})|(?<timestamp0>t=\d+[a-z]?)|(?:(?:[^\Wvt]|[\w-]{2,})=[\w-]+))(?:(?:(?<and0>&)v=)(?<id1>[\w-]{11,})|(?<and1>&)(?<timestamp1>t=\d+[a-z]?)|(?:&(?:[^\Wvt]|[\w-]{2,})=[\w-]+)+?)??(?:(?:(?<and2>&)v=)(?<id2>[\w-]{11,})|(?<and3>&)(?<timestamp2>t=\d+[a-z]?)|(?:&(?:[^\Wvt]|[\w-]{2,})=[\w-]+)+?)??(?:(?:(?<and4>&)v=)(?<id3>[\w-]{11,})|(?<and5>&)(?<timestamp3>t=\d+[a-z]?)|(?:&(?:[^\Wvt]|[\w-]{2,})=[\w-]+)+?)??(?:(?:(?<and6>&)v=)(?<id4>[\w-]{11,})|(?<and7>&)(?<timestamp4>t=\d+[a-z]?)|(?:&(?:[^\Wvt]|[\w-]{2,})=[\w-]+)+?)??(?:(?:(?<and8>&)v=)(?<id5>[\w-]{11,})|(?<and9>&)(?<timestamp5>t=\d+[a-z]?)|(?:&(?:[^\Wvt]|[\w-]{2,})=[\w-]+)+?)??(?:(?=[\s.?!,:;/])|$)
replace: https://youtu.be/$<id0>$<id1>$<id2>$<id3>$<id4>$<id5>$<and0>$<and1>$<and2>$<and3>$<and4>$<and5>$<and6>$<and7>$<and8>$<and9>$<timestamp0>$<timestamp1>$<timestamp2>$<timestamp3>$<timestamp4>$<timestamp5>
this is foul
from the regexes on each "line" you can determine what strings are possible on each line, and from that, which characters are possible in each cell
3 regexes apply to each cell, so you take the subset of characters that would match all of them for that cell
if this subset is only one character, that's the solution for that cell
repeat until solved 
Actually the subset always only allow for a single character somewhere
You shouldn't ever guess
I never had to
its regex crosswords
ye that what I said 
pretty much 

ye
ow
yeah I saw that, felt like regex was personally insulting me for not being able to solve it...
i solved it
If mods listen i propose a community event which consist in speedrunning the 32x
wonder how long it'd take neuro to do it
i can't even put the whole 32x on my screen and i have no idea how to view it properly 
neuro would be a lot faster than any of us. 
I mean with her wordle track record... 😂
Lemme doubt
someone make a integration for it 
wtf did you do
Tell me how to
integration that just queries gemini
the neuro-sdk
It's normal
I'm discussing it in vcord
Discord is trying to change codeblocks
And it's breaking vcord shikicodeblocks

Just click on another channel and back here
yeah nah it stays like this regardless, oh well
How are we managing to break something that just worked
Because discord do changes
classic discord "flinging crap at wall and hope it sticks"
And here in this case i think they lazy load something to hot swap something or smth like that
And vcord doesn't expect it
So tried to hijack too early
Or some shit like that
guess i'll keep an eye out on that 
good thing i didn't completely get rid of the nix overlay and only commented it out for the future
Lol Back to 5 mins
Indeed
poor ved
I can't even imagine how complicated neuro's code is at this point, what's the LoC he said last time?
like 80k or something?
He rewrote it so idek
well that's what I mean
I believe the 80k was for old codebase iirc but I may be wrong
Must be extremely complicated i would imagine
I can imagine
a stream or two ago he said the rewrite is at like 80k lines already
GG
Tbf considering its typescript dosent particularly supprise me
I have done something alot less complex with AIs in typescript that shits like 30k lines. I dread to imagine what the fuck the number is for that codebase
is that confirmed? I figured she'd be mostly python (for the AI stuff at least)
then C# for the unity stuff
It's typescript?
I think the twitch description says the ai backend is python
The whole point of the rewrite was to move off of C# , to typscript for alot ofit
Theres a few clips
Iswtg
wrr
Cas of what he said in the hollow knight streams a while back





Motherboard can take 128GB max








