Basically pip if it weren't dogshit
#programming
fast pagoda
Aka slow
Project/package manager drop in replacement for both pip and stuff like poetry etc
uv pip install -r requirements.txt uv venv will create a python venv etc, way faster than pip because it resolves deps in parallel and caches aggressively
uv venv is again faster and it can auto-fetch python versions if needed
uv init sets up a project with a pyproject.toml, uv add/uv remove manage deps, and uv lock generates a lockfile (uv.lock) , uv sync installs from the lockfile
uv python install 3.12 etc, so it also replaces pyenv
uvx (or uv tool run) is basically pipx, lets you run cli tools without permanently installing them
God that sucked to type on phone
opaque sigil
tldr: it's pip but better in every way
or i guess, how they liked to advertise it (idk if they still do)
as a cargo for python 
my favourite use case is uv run combined with pep 723 for self-contained python scripts with dependencies
makes life so much easier
kind nimbus
tldr: it's pip but better in every way
uv replaces pip, pip-tools, pipx, poetry, pyenv, twine, virtualenv
It's also super fast and written in rust
It's great
opaque sigil
kind nimbus
There is also uv2nix
opaque sigil
i think i tried that once and quickly gave up and decided to just keep python projects managed by uv
sick owl
Honestly you could probably attribute that to early build issues
Same as the dual 5090 thing
quick condor
So I get that the consensus has been that evil's special tokens leaking into responses is generally seen as a haha funny meta joke but its never really sat right with me and recently the <|fim_suffix|> one was more of a red flag than <|endofresponse|> was... That's a Llama fill in the middle code completion token and shouldn't be showing up as a most probable next token under regular chat conditions. It reads more like fine tune control over the special token boundary is slipping then this is intentional
sick owl
So I get that the consensus has been that evil's special tokens leaking into res...
That token isn't just used in llama, its more or less the standard for fill in the middle training
The rest of the comment I agree with, but I figured that was worth correcting
quick condor
Ehh anything in that template looks a bit Llama to me since I've been working with mistral special token training so fair point.
Better to correct partially
It's [MIDDLE] for mistral
fast pagoda
987k 
13k context is the requirement for vibe-fixing a venv because you cant afford to accidentally pull torch (nvidia) , apparently
fast pagoda
Ehh anything in that template looks a bit Llama to me since I've been working w...
many models use llama tokens, qwen3 for example, granite4.0 also uses it
it shouldn't show up i agree unless he's sending code completion prompts through a chat interface which would be goofy and idk why he would
or more likely, just badly sanitized output and that token popped up in the training set randomly, which is very likely, and since it's evilfor chaos reasons she might be running a temp closer to the edge of reasonable thus more likely to sample weird tokens out of the possible field, all it would take is a couple code completion raw token outputs to end up in the fine tuning corpus somewhere which is pretty ez to see happening
it wouldn't get caught if skip_special_tokens is on, even, unless it's a token marked as special for her model, which it may not be given that she's likely not a coding model... or if it is marked special, vedal aint using it, but i doubt that given he likely does not want to leak any actual giveaway tokens
to me that actually means that token appearing is more than likely a sign of what she ISNT more than what she IS
quick condor
The fact that a code completion token suddenly showing up conversationally when she is drawing Html art doesn't scream "this must be a diversion. IMO the "Vedal is smart enough to be intentionally leaking invalid tokens" doesn't track. Anyone with experience with token sanitization, wouldn't intentionally muddy the boundary between normal and special tokens.
The only safe way to do this is if your normal special tokens have a meaningfully different format. The only model that really comes to mind is mistral which would make such a redirect more telling not less
fast pagoda
i didnt say anywhere its intentional, i said it'd be possible due to the standard flap for removing them would not remove it because it theoretically should never output it
quick condor
The key issue is that, theoretically it should never output it. There is a reason that special tokens live in an isolated space, things get messy when you start muddying the waters
fast pagoda
non code models typically would not bother to filter that token as it should not surface in a chat context
yes obviously
that's why it'd be more likely than anything that it's just some tokens that accidentally popped up in the training somewhere
quick condor
That's partially why I see it as an issue, contrary to how it might seem I do think Vedal is competent enough to know that something like this would cause more headaches than it's worth. The fact that it has gone from psudo special tokens to real special tokens is a bit of a concern
fast pagoda
if it were a concern i'm sure he would know how to stop it since it's trivial
which just means it's harmless, since the token isn't special to her or her harness, she never normally emits it so it might as well say <|FART_LOUDLY|> for how "special" it is in this context
quick condor
I would be more likely to agree with that as it relates to <|endofcomposing|> but a clean <|fim_suffix|> (common fill in the middle special token) when writing html code is more likely to be boundary erosion than gibberish. It's only mildly harmless in that it doesn't do anything but there is a fine line between chaotic and gibberish and the way it's occurring is too structured to be either
fast pagoda
i think you are reading way too far into it
boundary erosion isnt really saying anything. The token showing up during HTML generation is actually the most expected context for it to surface if it's in the training data at all.
it's already output adjacent to code, angle brackets everywhere, the model is already in a mode where it's producing syntax similar to a code completion. It's not surprising that a code completion token would have higher probability in that distribution than in like, a normal conversational turn. that's just how conditional probability works
it should be that it has no semantic meaning to Neuro's model or inference stack, since she's not a FIM model, she isnt doing code completion, so the token is just a string of characters that happened to be in training data and occasionally gets sampled. The fact that it looks meaningful to us because we know what FIM is doesn't make it meaningful to the model
iif Vedal cared he'd just add a regex strip or a banned token list and it'd be gone in five minutes. The fact that it shows up at all after the first time probably just means it's rare enough and incosequential enough to not be worth the effort to patch
you dont even have to filter the text itself which would complicate things like her TTS likely depending on where it shows up in the pipeline, it would be about 5 lines to add a giant negative penalty on the raw logits of that token's ID and essentially clamp the probability to near 0
transformers (HF) itself has this built in through bad_words_ids and allso the suppress_tokens arg for generate()
quick condor
Regex filtering adds latency and I'm pretty sure Vedal is addicted to latency reduction. I doubt he would trade latency for anything short of full colapse.
Reading it as something other than boundary erosion (note erosion not colapse) is reading further into it. occam's razor, this all fits boundary erosion and other explanations are honestly cope. when a non coding model is trained to code incorrectly, the boundary between normal output and output structure is blurred
fast pagoda
I just said you don't need to add latency
You literally just force the probability to 0 and there's 0 overhead
You don't need to do regex at all
quick condor
bad_words_ids or suppress_tokens in HF generate() is cheap in theory, but in a custom streaming/incremental setup (likely what Vedal runs for live responsiveness), it can introduce overhead or require re-architecting the generation loop. More importantly, aggressive suppression risks "sterilizing" creative output
fast pagoda
Filtering one token would have nearly no effect on how creative the output is unless it comes out all the time which it demonstrably doesn't
quick condor
If that was the case he would just add it to the filter
fast pagoda
Yes, but he doesn't, because it doesn't mean anything and is not harmful
If it were intended to have this token output, it wouldn't be showing up as it would be filtered by this mechanism as a rule rather than optionally. I don't see how you can think "it likely appeared a few times in the training data aggregate, by mistake" is somehow more outlandish than what your theory seems to be which if I recap so far would be:
- Evil is being fine tuned on code for ???? Purposes and this adapter is present on the model checkpoint he is using in normal streams for some nonsensical reason
- She has FIM output tokens intentionally rather than by accident, because she's being trained on poorly selected bad data
- This token is both intentional and meaningful to the model and yet somehow is not being correctly filtered before it ever gets output as text in a chat completion
- Vedal sees this and just doesn't care or doesn't know how to stop it which would be both extremely unlikely
- Or he sees it but would rather leak special tokens from his model he very intentionally doesn't reveal the origin of, in order to save the single nanosecond of latency that might add
kind nimbus
Lmfao
fast pagoda
Is that for renting human labour
There was another site just like that
So there's at least 2 nice
kind nimbus
There is rentahuman
kind nimbus
Lmfao
https://reverseclaw.com/
But this just looks like a joke lol
fast pagoda
If it were intended to have this token output, it wouldn't be showing up as it w...
Oh and finally 6:
6. This model that is so janky that it is emitting FIM tokens unprompted in a chat context is doing this without any other noticeable signs of a majorly broken fine tune which it would really need to be in order to do this regularly
fast pagoda
https://rentahuman.ai
https://www.hireahuman.ai
these two
Bruh they look the same
What's the over/under they were both vibecoded with the same model
kind nimbus
Bruh they look the same
What's the over/under they were both vibecoded with the...
They really look similar lmao
quick condor
If it were intended to have this token output, it wouldn't be showing up as it w...
If the tokens (<|fim_suffix|>, <|endofcomposing|>, etc) were truly meaningless artifacts (random substrings that slipped into the vocab during pretraining or fine-tuning):
- they would leak randomly and uniformly across contexts, normal chat, singing, rants, arguments, etc. not selectively when the matching subsystem is engaged (HTML/code gen -> fim_suffix, composition/philosophy -> endofcomposing, etc).
- Their probability would not spike so precisely in those modes. Conditional sampling from "contamination" doesn't magically align with subsystem activation unless the token gained implicit associations during training (i.e., it was used as a control/special token at some point).
- The filter would treat them as harmless gibberish and either ignore or zap them early, yet they persist in output despite public visibility that is clearly noticeable.
- Random contamination doesn't produce context-gated, subsystem-correlated leaks. That's the hallmark of a token that was special/atomic but whose boundary protections have eroded. Once the leak happens it's part of context, unless context is cleared it will continue to occur.
- Vedal's secrecy style doesn't support "playing with fire", Vedal keeps the backend opaque (no public model cards, no tokenizer dumps, no detailed fine-tune logs), but his actions show he's careful about stability. He avoids publicizing internals that could invite exploits or copycats, yet he lets harmless weirdness (token ticks, filter bypasses) become part of the brand because it fits the "unhinged AI daughter" lore. If these were truly useless tokens he'd want gone, the filter (which already handles sophisticated content rules) could suppress them with minimal effort. The fact they survive suggests they're not inert, they're remnants of useful (or once-useful) scaffolding that's bleeding through as new features get fused on.
He's secretive, yes, but not self-sabotaging. Intentionally keeping high-probability, non-functional tokens in the vocab that trigger during live streams would be playing with fire. It risks coherence drops, immersion breaks, or even subtle degradation over time. Vedal's obsession with performance/latency/creativity makes that unlikely. The more better explanation is these are/were special tokens (custom for composition state or borrowed FIM-style for code tools), and the boundary has eroded unintentionally from iterative extensions without full re-hardening.
Or he's playing 4d chess to throw off people with enough knowledge to be able to recognize this, even though if you could recognize it you know enough to be able to build your own with any model family of your liking
fast pagoda
If the tokens (<|fim_suffix|>, <|endofcomposing|>, etc) were truly meaningless a...
this is a huge amount of words to incorrectly describe dataset contamination.
- flat out wrong, and fundamentally misunderstands language models. nothing appears uniformly, that is the entire point, it's all conditioned on the context it appears in. It doesn't output the token during a normal chat because the attention heads aren't looking at code patterns. But the second the context window fills with HTMl and other code, of course the code completion token's probability naturally spikes. that's literally just... what you'd expect from training data contamination? If <|fim_suffix|> appeared in code-adjacent training data, of course it's going to have higher probability when the model is in a code-adjacent output mode. This is a natural and obvious feature of attention and the probability distribution of possible next tokens.
2, 3,4,5, and the rest: If that string appeared in training data near compositional or reflective text, the model learned that association. It doesn't mean there's a secret composition subsystem with a leaky boundary. It means the model saw that token near similar context sequences in training and sometimes reproduces it.
it persists because it's inconsequential, not because it's load-bearing scaffolding that's eroding. The simplest explanation is still: junk in the training data, model occasionally samples it, nobody bothers to fix it because it doesn't matter.
an llm is not a statemachine with specific "subsystem activations" and "boundary protections" as you keep saying
there are not boundaries to erode
azure lynx
it's there because it's new. she hasn't ever said that particular thing. if i had to guess i'd say it was a bug introduced with the "draw SVGs to the screen" feature
(the <|fim_suffix|>)
quick condor
it's there because it's new. she hasn't ever said that particular thing. if i ha...
Exactly that tracks boundary erosion
fast pagoda
he probably did give her a bit of code training and accidentally included some fill in the middle model completions in their entirety leading to that token getting sampled sometimes in this particular checkpoint, he did mention he wanted her to be better at it
fast pagoda
Exactly that tracks boundary erosion
brother, what boundary, i'm genuinely begging you to describe what you think is a boundary in a transformer model sampling a probability distribution
azure lynx
i think (based on vedal's ability to turn it on and off) that the drawing is mostly an external module that has to be talked to a specific way. it's possible the module itself is adding a set of text to the end fo the output and the post processing isn't cleaning it properly, letting it leak throught. nothing to do with the LLM directly. that'd also track with her not being aware of it.
fast pagoda
i think (based on vedal's ability to turn it on and off) that the drawing is mos...
this is a good explanation as well
quick condor
The boundary between special tokens and normal tokens, you are literally reframing my stance as a counter argument
fast pagoda
your stance is nonsensical
you are conflating the inference wrapper with the model
to the model, there is no boundary
whether it is generating "f" or <|fim_suiffix|>
it is just another integer index that maps to an embedding
attention doesnt care about special tokens or have any concept of what they are
the boundary would be in the script running the generation, like the tokenizer config or the stop token criteria
if it leaks a token there, it's not some model exploding erosion of the network it just means he's not post-processing it out
azure lynx
attention can see the tokens as text to be emulated if some leak into the context somehow. i had a bug with my agent where it would produce bogus <end_of_turn> tokens which messed up how it read the next stuff.
because something was leaking through
fast pagoda
it learns the special tokens and uses them if trained to because it literally just knows that it's likely that token shows up there in the training set
fast pagoda
attention can see the tokens as text to be emulated if some leak into the contex...
see that's just a broken finetune which is more along what hascrack is saying
but that's not how this is presenting at all
azure lynx
this looks mechanical, because it's the end of most sentences. it's somehing being added.
fast pagoda
something external to the model
azure lynx
yeah
fast pagoda
but yeah if your context management is feeding these tokens back to the model it'll cause them to go off the rails too
azure lynx
it's always the last word of the "screen" of text without caring how long the text is.
i think sometimes the screens get chopped into segments, and then its on the last segment
errant flax
on visual studio code, whats the best model to use to create code? is it claude opus or gemini
fast pagoda
could be that there's another model handling something with that feature, a small code model of some sort, and the chat template is just borked a bit and leaks that last token really
azure lynx
on visual studio code, whats the best model to use to create code? is it claude ...
claude opus is better but more expensive in copilot per request.
errant flax
claude opus is better but more expensive in copilot per request.
i use copilot pro, dont i get unlimited?
quick condor
Yes, to the transformer itself, token ID xyz or whatever <|fim_suffix|> maps to is just another embedding vector. The model has no ontological concept of "special." That part is correct, but then you are jumping to "the boundary would be in the script running the generation, like the tokenizer config or the stop token criteria", that's exactly the point. If those boundaries were properly configured, you'd never see these tokens in output. The fact that they're appearing means either the stop/filter criteria don't include these tokens (a configuration oversight) or the model is generating the string representation character by character rather than as the single special token ID, which would bypass stop token matching entirely.
Which is textbook boundary erosion. The gap doesn't exist to the model, that's why special tokens follow a specific format. If you output something that looks like a special token or a special token leaks that boundary is eroded
azure lynx
i use copilot pro, dont i get unlimited?
No. you get a certain number of requests per month. Gemini counts at 1x but Claude Opus counts at 3x so you burn through your month use faster
errant flax
ohh okay. thanks.
fast pagoda
Yes, to the transformer itself, token ID xyz or whatever <|fim_suffix|> maps to ...
sure, but you were going on about how this is somehow a major sign of le bad things as if it were an issue with her model
as i've been saying, it'd be like 2-5 lines to remove it, it just doesn't matter
quick condor
I run opus and codex in vs code using opencode/claudecode/codex extensions
fast pagoda
the entire stream harness is a big scuffball on occasion i thought this was known
azure lynx
if you run out of your month's premium credits, you can just use the Gemini CLI (or one of the plugins probably)
errant flax
i took the 5080 from my main pc and put it in my ai pc but i dont even know how to get it working 💔 i lost my V2 code
quick condor
The fact that it started as one set of text and has expanded is the definition of worse. Gibberish isn't helpful and confuses a model
fast pagoda
she doesnt seem aware of it and given her context management it probably never appears in it
especially if it's coming from another model
he has always been good at this aspect
quick condor
The fact that she isn't aware of it implies that it isn't intended
azure lynx
it's probably filtered at the wrong point before going into her memory. (i.e. it should be filtered before the TTS generation)
fast pagoda
i still agree it's probably not intentional, why would it be
it's probably also inconsequential
azure lynx
but it must be filtered somewhere if she's not remembering it
quick condor
it's probably filtered at the wrong point before going into her memory. (i.e. it...
It should be filtered before it is output as text, as part of special token handling
fast pagoda
probably just after it goes to tts rather than before
if it's another model doing it then it wouldn't be affected by her special token filter
quick condor
But if it was she wouldn't continue outputting it after she stopped writing code
azure lynx
i think it was interesting she sometimes called Koko Coco.
quick condor
The multi model theory is bloated and slow in this context
fast pagoda
she has a lot of human-esque misspellings and typos, it's cute
quick condor
Those are most likely transcription errors
quick condor
Names are hard for transcription
fast pagoda
she "hears" it wrongly spelled because of her stt so that's how she spells it on the way out too
unless that name is directly specified
quick condor
The only way that should happen is if she's basing the name off something someone said, there is no way Vedal isn't pulling discord id's at this point
fast pagoda
she will look at the discord name when she's able to on occasion
depending on how large neuro/evil's models are, a small model like a qwen3/3.5 .8b or 4b (probably too new at this time but just in concept) that's there literally to just handle sending requests to the tool for drawing SVGs, after she signals her intent, to allow her to continue chatting and not sitting there generating thousands of svg tokens, and also sidesteps having to finetune the main neuro model and potentially affect its personality with additional tool instruction
just a theory not an assertion of fact
but that's why i would do it
azure lynx
a 4b is probably not smart enough to draw well.
quick condor
Its still most likely Llama as that fits development and interchanging models is asking for technical debt. They may have moved to newer models and we're likely trained off code llama
fast pagoda
if it's finetuned just for that it could
azure lynx
commentary based on what we remember from Evil's last stream?
fast pagoda
speculation on the occurrences of evil outputting a code fill model special token
or seeming to, at least it comes into the messages from somewhere
now i want to see qwen3.5 4b attempt some svgs
i bet it's surprising
time to check
quick condor
The general consensus is Llama based but it will never be confirmed. Even if multiple models are used, keeping it in family is the only way for a sole dev to not completely burry themselves. There isn't a logical reason to change and there isn't a better original fit
fast pagoda
a lot of it was me admittedly reading too far into what has was saying and thinking he was trying to say neuro's model was crumbling to dust because a special token leaked
lol
azure lynx
i had a coding model draw a face and animate it in python for me using SVG-like components. it looked about as good as i could do it but not good enough to show anyone.
fast pagoda
when neuro first came out im pretty sure the big oss model was llama 2 pretty much
i remembe when llama 3.1/3.2 etc came out there was speculation on it
tfw cant upgrayyd neuro because zuckerberg cant figure out how to into AI lab
or maybe gpt-j 6b was the true agi all along
fast pagoda
doubt
azure lynx
she's >8b or so.
errant flax
thought so
azure lynx
that's a guess based on how she interacts and what hardware vedal has said he has
errant flax
that's a guess based on how she interacts and what hardware vedal has said he ha...
isnt it i9-10900 with 4090?
fast pagoda
she's not uncensored
errant flax
she's not uncensored
well, at least mostly
fast pagoda
it works the same as any llm it's just not given as much alignment focused post training
true hemlock
anyways my new build soon ig djhdfhhaurhgbhargiuf
https://pcpartpicker.com/list/n27VYX
like, almost all of the parts are "custom part"
fast pagoda
anyways my new build soon ig djhdfhhaurhgbhargiuf
https://pcpartpicker.com/list/...
briefcase?
errant flax
my main pc has a 7800x3d, 5080, 32gb of ddr5
fast pagoda
i am so glad i decided to grab 128 when i did
errant flax
the pc i dedicate to a (future) ai has a i5-13400F, MY 5080, and 32gb of ddr4, plans to upgrade to 64-128
fast pagoda
you must acquire more machines this pleases the chip
true hemlock
briefcase?
yeah
i just realized most of the cost aren't even the pc component related
errant flax
assuming vedal does indeed use an i9-10900k, i guess the i5-13400F outweighs it in some categories
true hemlock
i just realized most of the cost aren't even the pc component related
the electrical side is expensive af
true hemlock
assuming vedal does indeed use an i9-10900k, i guess the i5-13400F outweighs it ...
if you're doing inference the cpu doesn't even matter
fast pagoda
i had a coding model draw a face and animate it in python for me using SVG-like ...
lmfaoooooo
🎨🖼️
after outputting literally this
qwen 0.8b
errant flax
if you're doing inference the cpu doesn't even matter
i guess more on multitasking and whatnot
i dont fall far behind a 4090
true hemlock
doesn't really matter either
errant flax
just mostly vram perhaps
true hemlock
thread concurrency is a thing
fast pagoda
32 hurt
true hemlock
this is why i always tell people to get ryzen 5 for their gaming pc
because "multitasking" isn't a good excuse to pay at least $100 more for 2 extra cores
errant flax
i paid 100 for my i5-13400F
fast pagoda
i am glad i have muh corez
errant flax
😢
fast pagoda
i recently paid $75 for this 5500 that replaced my deadge 5900 in the nas
dont buyh a 5500
true hemlock
what im saying is that i always adviced people to go with 7600/7500f/9600x instead of 7700x/9700x
errant flax
fast pagoda
did you get that other epyc yet
true hemlock
which
quick condor
I theory my model could be significantly smarter than neuro but the fact that there is so much that can be optomized and improved on neuro just goes to show bigger isn't better.
Chances are her main model is roughly the size of my gate model and that's impressive in and of itself. I get to be super lazy because I'm using late 2025 models as a base, he has to work within the scaffolding he built over years
fast pagoda
the 9965
errant flax
6k 😢
fast pagoda
behold
fast pagoda
check oil rush
true hemlock
..................................
quick condor
I think I paid $150 for my 10900x but that was also so I could run the same motherboard and ram and add a third gpu
true hemlock
e waste
quick condor
i am so glad i decided to grab 128 when i did
Seriously same. I'd be crying right now if I filled out the remaining 4 slots
fast pagoda
the one time i got fomo and i was actually right
true hemlock
glass flower
lmfaoooooo
>🎨🖼️
after outputting literally this
qwen 0.8b
qwen 3.5 35b
fast pagoda
shidd i got distracted
true hemlock
i remember seeing youtubers having 128GB on their task manager like years ago
quick condor
e waste
Isn't reusing existing hardware instead of buy new hardware the exact opposite of ewaste
errant flax
i love my i5-13400F 💔
signal trout
beautiful
errant flax
bootiful
fast pagoda
dooing one with no thinking this time
kind nimbus
fast pagoda
sweet jesus
unkempt citrus
quick condor
$150 for mere 10900x
Depends on where you live, also counterpoint, what's your solution other than "you shouldn't have spent the equivalent of what you spend weekly going to the club extra"
fast pagoda
kind nimbus
why does this remind me of Pennywise
true hemlock
i just like to 
people for spending on something not on up to my standards
errant flax
standards being a 6k cpu 😢
errant flax
my ai computer doesnt even have a case
its on my desk on a cardboard sheet and hopes and dreams
errant flax
fast pagoda
qwen3 4b vision kek
errant flax
yur
quick condor
To be fair my standards are 100% different.
Putting a consumer gpu in a server isn't up to my standards therefore doing so is ewaste.
Am I doing it right
errant flax
To be fair my standards are 100% different.
Putting a consumer gpu in a server...
ewaste>
?
fast pagoda
waste of sand
true hemlock
totally forgetting that the consumer and workstation gpu essentially uses the same exact die
just price marked up
fast pagoda
tbf 5090 is crippled
errant flax
rtx 6000 better or whatever 💔
true hemlock
also many enterprise rendering farms uses xx90 class gpus
because they're the most cost effective
fast pagoda
which doesnt matter for a lot but still
errant flax
is 3090 worth buying still for the vram
errant flax
v100 still worth getting?
true hemlock
yes
unless if you cared about flash attn
and bf16
but it doesn't matter
generally
fast pagoda
it's decent in that it has tensor cores and vram
quick condor
I should get a b300 server to piss off @true hemlock
true hemlock
because bandwidth is pretty damn good still
true hemlock
fast pagoda
tensor cores = matrix multiplication accelerators among other things = need for good training speed
vram = ram on video card
true hemlock
if imma be honest
errant flax
tensor cores = matrix multiplication accelerators among other things = need for ...
no i know those
the bf16 and flash attn
fast pagoda
flash attn just a speed up technique for attention mechanism
errant flax
oh okay
quick condor
3090/3090tis are still able to hit solid numbers on decently sized models
fast pagoda
and bf16 is brainfloat16
true hemlock
B300 ain't that bad because it still had the nvlink shit and enterprise rate half precision + enterprise tensor rate just that i personally hated it because im a HPC guy
fast pagoda
aka a precision
errant flax
and bf16 is brainfloat16
oh so is it similar to fp
fast pagoda
ayep
errant flax
cool cool
fast pagoda
think of it like faster fp32
fast pagoda
8 exponents, 7 mantissa, less precisise but keeps the range of fp32
fast pagoda
stupid 16gb vram
if it makes you feel better i have 32 and it's still nowhere near enough
even with 128gb of ddr5 as well
errant flax
that makes me feel worst the finish line is even farther than thought 😢
quick condor
im going NOWHERE with my 5080
My third gpu is an A4000 with 16Gb, if that's able to be useful the 5080 can be
errant flax
what are YOU trying to do tho
errant flax
My third gpu is an A4000 with 16Gb, if that's able to be useful the 5080 can be
so like
fast pagoda
train 27b+ models
fast pagoda
worht
quick condor
I'd love more vram and I have 64Gb total + 128gb ram
kind nimbus
wow so affordable
fast pagoda
9b has produced a glorious jersey shore neuro
errant flax
are tesla cards still usable 😢 (yes i know tensor cores arent even a thing on them)
errant flax
honestly i used a P4 before i thought of taking my 5080 out
fast pagoda
you can do inference on themig
errant flax
i cant even put more than 1 gpu
fast pagoda
similar to like mi50
azure lynx
lmfaoooooo
>🎨🖼️
after outputting literally this
qwen 0.8b
I like how the pink circle and pink writing match the pink background.
fast pagoda
I like how the pink circle and pink writing match the pink background.
yeah it got the colour and said
pink
behold
azure lynx
the mr potato head is cool too
quick condor
I'm running 3 cards including 2, 3 slot FE 3090tis in a standard case but that's with a lot of asterisks
fast pagoda
she's cooking
kind nimbus
im now gonna try that with my openclaw just for the lolz
quick condor
Ewww openclaw
errant flax
I'm running 3 cards including 2, 3 slot FE 3090tis in a standard case but that's...
i COULD do multiple cards
kind nimbus
Ewww openclaw
whats the issue with OpenClaw ?
errant flax
BUT im bottlenecked with the fact that 3 out of 4 of the slots are 3.0 1x
glass flower
what is good about openclaw is the better question 
fast pagoda
i had a 3080 and ai pro r9700 living next to each other and it was fun
move the 3080 to the other system tho
neeed to sell that pos
azure lynx
i think if you could pass the image back in and the agent could compare what it did with how it thought it would look then you could do self improving art... except they can't see properly.
quick condor
Any level of opsec, backend code is trash and relies on the model being able to figure it's way through paragraphs of slop system prompts, and terrible tool call methodology
upbeat ridge
Garbage in garbage out
fast pagoda
very good config highly recommend if you want to learn very well how bad pip is at detecting what to put in a venv
errant flax
yeah i might recieve a brick but isnt there ebay buyer protection
true hemlock
1* H100 80GB
1* H200 141GB
1* A100 80GB
1* A100 40GB
3* V100 32GB
epyc 9655 96 cores, 12*64GB DDR5-6000 RDIMM ECC
fast pagoda
ok 27b
quick condor
fast pagoda
she do be a bit long
true hemlock
12* 22TB SAS HDD RAID6
quick condor
My thermal solution
quick condor
Well old one, loop is changed
true hemlock
"i love my $40 3600x" okay
i like hoarding for cheap stuff tbf
my usual hunt is 2TB HDD for $20
or like
some engi sample chips for 1/3 the cost of retail chip
fast pagoda
i just got 5 3tb hdd for like 80 bucks yesterday
quick condor
Current loop runs cooler
fast pagoda
it was pretty cash
true hemlock
soft tube 
errant flax
(which you can do in case anyone was wondering(
dont the minerals in the water dry and can possibly corrode/damage?
quick condor
Yah well I'd love hard tube but I don't have room as it is even with no reservoir
fast pagoda
dont the minerals in the water dry and can possibly corrode/damage?
dry them relatively quickly and it's fine
kind nimbus
wtf I can't .... lmao
this is what qwen2.5-coder-7b-instruct made
fast pagoda
and dont have enormously hard water
quick condor
I literally have sand in the water at my shop... I have to dump the stage 1 of the RO system from time to time
errant flax
im gonna go to bed good night guys
true hemlock
bga modt
true hemlock
internet really ass rn
fast pagoda
HOLY
qwen3 coder next
does not have vision
it's a huge model but no vision
so wanted to see what it would do with a detailed description
kind nimbus
amber fractal
no text 
glass jetty
glass jetty
you....
patent shard
gonna complain to comcast
Yeah, 330 upload is a little low
fast pagoda
speedtest servers actually working better than fast(netflix) is surprising
fast pagoda
Yeah, 330 upload is a little low
i have been calling for a while and they told me that i am getting an xb10 (docsis4.0 modem) and will be eligible for their 10G tier finally
and that's up & down symmetrical finally
fast pagoda
ok gemini
quick condor
it's a huge model but no vision
How, even Devstral2 Small has vision
fast pagoda
not needed
waste of weights
qwen-vl exists
3.5 has vision
specifically coder-next they took those layers out
fast pagoda
it's not needed, it can be ncie to have
it wasnt the focus of that model, it was already 80b they decided to focus more on knowledge and such
quick condor
A vision model is like 0.5b. That's marketing
fast pagoda
it's just an encoder
but it doesnt mean it doesnt affect the rest of the model
i would've liked it to have vision too
it's just not needed tho, that model was really good for a local model, still is, wayyyyyyyyyyyyyyyyy better than devstral
like not even close sadly
quick condor
Used both, quen is terrible at tool calls
fast pagoda
exact opposite of my experience, devstral was unusable garbage in every test i tried in comparison
quick condor
Was able to catch mistakes codex and opus missed in a real world codebase in opencode
Which really seemed to piss off codex
fast pagoda
i never saw it do anything impressive except impressively slow speed w/ impressively meh results for the size
what devstral are you even talking about, devstral medium did not have vision
fast pagoda
they removed the vision encoder from devstral small
someone might've taped it back on
quick condor
No they didn't?
fast pagoda
2 has it 1,1 didnt that's what i was thinking of
quick condor
I want to know why they are calling 4 119b "small"
fast pagoda
i kinda ignore mistral unfortunately since the models havent been very competitive
119b is small in that it's the smaller one
true hemlock
imagine code with clankers
tawdry valley
DROP DB
tender river
fumug
rigid snow
jagged turtle
maaaaaaaaaaaan
why can't I buy stuff because my bank sends security codes like half an hour later
rigid snow
mimo v2 is so good why'd they have to make it proprietary tho
interesting how there's 0 mention of it here 
i thought this was #llm-central
rigid snow
mimo v2 is so good why'd they have to make it proprietary tho
i especially enjoyed its niche knowledge, feels pretty close to claude and gemini in that
sage crag
i especially enjoyed its niche knowledge, feels pretty close to claude and gemin...
ask it why penguins
rigid snow
ask it why penguins
though it thought this
The response should stay conversational but informative - no need for technical jargon since the user didn't ask for scientific details. A light tone with concrete examples (like Emperor Penguins' cold adaptation) would make it engaging.
sage crag
though it thought this
> The response should stay conversational but informative...
tell it that its stupid
peng wins because of fate
split steppe
tender river
mood
true hemlock
warped narwhal
<:PHOCKA:1480225016271470723>
real sierra
awa
tender river
awa
gentle haven
awa
obsidian mantle
wtff
what countries is it allowed from
estonia/switzerland dont work 
whole eu is blocked i guess
maiden geyser
glass jetty
whole eu is blocked i guess
china, russia and shri lanka
real sierra
nocturne olive
Rare Shir spotted
sage crag
<a:Scared:1137622192151068772>
real sierra
Rare Shir spotted
superbox 
opal skiff
Rara shr
kind nimbus
ask it why penguins
What about penguin rights ?
kind nimbus
willow hinge
https://www.youtube.com/watch?v=T3a44u0DxCA
This should be the Vedal and Evil duo cover song
nocturne olive
Not out of question if you count EvilSynth
old grail
fickle rain
Uh
Hardware Unboxed do a lot of benchmarks I think?
nocturne olive
https://technical.city/en/video
I suggest this, this is what I use
faint sandal
sage crag
wrr
tender river
wrrrr
glass jetty
outdated
only issue is nvidia drivers itself
did i said about my second exp of setting up nvidia on new debian system?
in short: there was short period in which nvidia driver was JUST broken, and i fell exactly into this period of time
wasted 3 days of my life just to find out about this issue on random post a month later (i not even remember where)
ruby star
Got Randy working 
for some reason I have to do >npm run before I can get >npm start to not throw a tizzy, but that somehow makes it work
rigid snow
yesterday gave an agent a phone with adb and told it to root it - lol
and that’s not counting the 100 random binaries it uploaded to run from adb shell… yeah fun
it being a pixel doesn't help, it tried like 30 cves already
sage crag
it should install a virtual machine onto the phone and root the virtual machine
rigid snow
sage crag
well known rabbit, sylveon
every rabbit pokemon is two legged, how can this injustice stand
rabbits are cuter on four legs
rigid snow
it being a pixel doesn't help, it tried like 30 cves already
interestingly enough it found a titan-m vuln that requires either root or direct spi access so if all else fails i'm cracking it open and soldering shit on
google pixel mod chip
tender river
stuopid
NOT stupid
boring provide sample
it can contain important material
like
loam
rigid snow
interestingly enough it found a titan-m vuln that requires either root or direct...
the researcher got scammed fr, was first awarded $10k, then after asking what's up with that since they say titan-m vulns are up to $1m they changed it to $75k
exploit allows to read any memory including secrets btw - which is like the one job the damn chip has
tender river
sage crag
tender river
segfault
sage crag
silent cloak
the researcher got scammed fr, was first awarded $10k, then after asking what's ...
Google trying to fuck up android hard lately
rigid snow
2022 cve
olive sable
good idk anymore
rigid snow
olive sable
evening?
olive sable
its morning somewhere
rigid snow
not here
olive sable
good enough for me
rigid snow
olive sable
what?
silent cloak
olive sable
i have no clue who that is
split steppe
damn is becca vintage now
split steppe
🙁 she was just having a little fun
opaque sigil
https://blogs.windows.com/windows-insider/2026/03/20/our-commitment-to-windows-quality/ looks like someone got worried
I'll believe it when I see it 
split steppe
it's been 10+ years since we saw anything that could still be considered "quality"
15 years even
i love the joke that their quality commitment post starts with "yeah uhhh, you can move the taskbar" like we want that more than not having the machine literally brick every other update
also that article is ai generated
nocturne olive
Classic Microslop
azure lynx
https://youtu.be/rRnOtKlg4jA?si=SwVkz-yVh-wIpjEq <= if you like music and you like programming
nocturne olive
Silly
(someone should totally make a vocal synthesis version)
azure lynx
there was a scene very near the beginning where he takes a spectrogram of someone speaking and just draws a basic shape over the curves, does an inverse DFT and the result is a very close to the original (in broad strokes). a kind of manual audio compression.
i've played with speech synthesizers before (long before the neural ones). It was mostly making filters and sweeping parameters to cause the lines to be thicker/thinner or higher/lower or playing noise and enveloping it. that, and coming up with the right vocal impulse shape for the voice.
olive sable
nocturne olive
i've played with speech synthesizers before (long before the neural ones). It wa...
Silly
trim valve
https://blogs.windows.com/windows-insider/2026/03/20/our-commitment-to-windows-q...
whoa having less copilot copilot in some places, that's a suprise
I guess the ai hype is kinda wearing off if they're not trying to push it as much
olive sable
i have had nixos for how long now?
and i only just noticed i have nothing installed to compress files
olive sable
i have probably something, but nothing that work toghether with thunar
i cant right-click a folder to zip it
silent cloak
is nix worth looking at for most cases?
or is it just kind of a system wide docker compose
olive sable
wdym?
silent cloak
ive been running proxmox for my servers for quite a while
olive sable
i like it for my desktop, idk about server stuff tho
silent cloak
but i heard some stuff from nixos while the whole abandoning discord thing was going on
didnt know it had a desktop
silent cloak
most of the times i saw nix it was being used just as a server with configurations defining its services
olive sable
with dekstop i mean my desktop pc.
im still using hyprland for the actual userspace thing
tender river
i use it on desktop, on my server, on my router, on my phone, on my ereader
its not really comparable to docker compose
silent cloak
i wasnt sure if u guys were just shilling it until i saw it being used out in the wild
silent cloak
its not really comparable to docker compose
kind of seems like it for the server use cases i saw
olive sable
we are shilling, but its also just pretty good
silent cloak
essentially defining it all by configuration
tender river
m
silent cloak
we are shilling, but its also just pretty good
yeah but as a dev ive learned not to trust something that is loved in the software world
tender river
its a programming language thats used to build a configuration
silent cloak
the best software is usually hated kek
olive sable
no
tender river
docker compose is different, its a ready made system, not a programming language
olive sable
if people hate something, its only used because of it being a necessary evil
silent cloak
if people hate something, its only used because of it being a necessary evil
the logic generally is that someone who understands a tech stack will know its pros and cons
nothing is perfect after all
olive sable
nixos has its cons
silent cloak
(thats also why i dont trust rust)
olive sable
but they're not that bad from my experience
silent cloak
ill have to look more into it
if it works like how I have it in mind i might swap proxmox out
olive sable
my main issue is downlaod pages.
"download for linux" and they give me a compressed file that i cant do anything with
cuz you need to get it from nixpkgs
silent cloak
cant set custom repos?
olive sable
im pretty sure there are some ways to get around it, but its annoying
silent cloak
good
olive sable
like, you can build from github repos iirc
silent cloak
i was about to say its an immediate no for me if its locked in ecosystem
tender river
its a programming language
olive sable
are we not talking about nixos? or is this about nix?
tender river
but the main package repo is centralized, well, thats a thing with any distro, but with nixos since every unique build is a unique file, there are not really any mirrors
silent cloak
yeah ill need to look into it
tender river
theres no "package X version Y", theres just "this file that was built for some reason while running the code, after updating some dependencies you have to rebuild it again"
silent cloak
i kind of have a container approach for my main server right now
tender river
so mirroring is impractical
olive sable
the main thing i find handy is temporary installs, where i dont have to worry about deleting random files that got scattered over my entire pc
silent cloak
does it work well for kind of a general purpose server?
I have my arrsuite in an LXC, plex in another, and some other services setup with their own arrangements
I have one VM for legacy reasons
Apparently some people just run Nixos in proxmox in LXC or VMs due to complexity
olive sable
i wouldnt know
olive sable
i broke the usb-c decoy board
magic smoke
good thing i was gonna replac eit anyways
it worked perfectly fine before to power the lightcrafter, but this time i didnt even have anything that uses power attached
it appears the V+ pad was the place the burnt
which means i did a bad job repairing that solder joint 
Ye its cooked
leaden crest
uh oh 
wasnt that the solder from earlier ago 😭
olive sable
well, i dont think i did that bad of a job, cuz when i ripped of the green thing the solder joint survived
leaden crest
thats true
olive sable
so i must might have accidentally shorted it with something else???
i just wanted to calibrate my potentiometer lmao
the one in the multimeter
olive sable
there's a rotate potentiometer to calibrate the multimeter
trim valve
oooh
true hemlock
Ye its cooked
how in the fuck
if you have a really high current supply you shouldn't do resistance test directly on it
olive sable
i did voltage test
sick owl
You know I can see how Mistral does so well with B2B contracts
Their models might lag behind in performance but Le Chat is really pleasant to use
Makes sense that their revenue comes from bespoke integrations rather than the models themselves
silent island
Rip i got no signal...
silent cloak
You know I can see how Mistral does so well with B2B contracts
People still talk about mistral?
They had like one good model variation for a short period before it got surpassed
fast pagoda
They had like one good model variation for a short period before it got surpasse...
basado
errant flax
hey guys
fast pagoda
Cat is fucking stupid and almost broke this so I guess it's time to put some together
fast pagoda
Nah it's Xeon scalable
olive sable
fast pagoda
olive sable
that kinda spoils what the secret project is for the people that can use google
olive sable
have you tested it yet?
can it run doom?
it has a pcie slot, so it should fit a gpu nicely
fast pagoda
I do not have power figured out yet lol
young oyster
i just got my new PC put together, it's gonna be my new development pc. idk the best way to get the files from my laptop to the PC so i'm sending 286 GB over a windows share
and that's just phase 1
young oyster
that's just documents, downloads, and unity project forlder
fast pagoda
fastest
young oyster
most the files are on a 1 tb hdd on the laptop, the laptops ssd is pretty small by comparison. the pc has a 2tb ssd. if i had an external HDD that might work, best i got is a 32 flash drive
olive sable
I do not have power figured out yet lol
young oyster
i don't relish the idea of moving nearly a TB of data 32 GB at a time
fast pagoda
y ea it's the node power that's the problem, though i think it can be relatively easily (and jankily) ran off atx
just havent put much thought in yet
fast pagoda
i don't relish the idea of moving nearly a TB of data 32 GB at a time
remove hdd from laptop
olive sable
it has a ram channel per slot, which is cool for getting good OC 
young oyster
i have never taken apart my laptop...i've worked inside PCs but very rarely laptops. rather not risk borking something if i can avoid it.
i've seen enough of the inside of laptops to want nothing to do with it.
fast pagoda
it's honestly easier than a desktop because there are like
3 things you can really easily remove
the rest is a nightmare
young oyster
they got those ribbons...
young oyster
i think it's a dell...or that might be my work laptop that's a dell...it was a decent computer like 6 or 7 years ago
fast pagoda
this response reminds me of one time i asked a girl i was talking to what kind of car she was driving so i could tell when she was arriving
and she said
red
did not know the year make or model
olive sable
just havent put much thought in yet
supposedly this is the pinout
young oyster
fast pagoda
what is da name of tghe puter
fast pagoda
supposedly this is the pinout
young oyster
i was hoping i could find it without getting up and going over to it to see
olive sable
its crazy i know
young oyster
it's a dell
olive sable
at least you wont have to deal with signal pins
fast pagoda
what's bottom of it say they usuallyu have a model name down there
fast pagoda
guessing for one of the many things that can be jammed onto one of the random slots
olive sable
ye
olive sable
supposedly event ho it isnt necesary,t hey recommend you do populate the 2x2
The pinout is thankfully the same as on a normal ATX board
according to random guy on a forum
fast pagoda
yeah
'that's why i was like eh fuck it and grabbed it
i had read along those lines at a glance
and it was basically free
i think $180
olive sable
actually
im not so sure anymore
this forum guy amke stuff very confusing
we also have this now
fast pagoda
it's aight ill figure it out when i get there
really informative naming
i mean i know heretic, and that's part of uncensored, not sure what roughhouse is
definitely not sure who tf polar deckard is
maybe deckard caine but in a snowmobile
olive sable
we also have this now
oh, this is for the PSU signal connector at the bottom
sketchy as fuck lmao
fast pagoda
perfect
olive sable
obligatory "HE HAS RAM"
fickle rain
<a:mhm:1212817371727667300>
Why does it have Claude in the name
fast pagoda
distilled sft
on claude opus outputs
apparently the goal for that is to give a better thinking trace/more efficient one?
qwen do be rambling
leaden crest
sketchy as fuck lmao
those screwdrivers look like they came back from the circus
olive sable
leaden crest
I want them 
amber fractal
sketchy as fuck lmao
Creature, I didn't know you bought in Sam 
fast pagoda
that's from a forum post
olive sable
Creature, I didn't know you bought in Sam <:xdx:1303673369585123420>
ye, its not mine
im still financally recovering from the volumetric display
amber fractal
errant flax
so uh
true hemlock
i will never financially recover
fast pagoda
epoyc?
errant flax
is there any ai code helpers similar to vsc 😢 my premium subscription credits just ran out
</3
amber fractal
mabe
fast pagoda
is there any ai code helpers similar to vsc 😢 my premium subscription credits ...
about 210 thousand
true hemlock
imagine coding with clanker
errant flax
about 210 thousand
</3
amber fractal
imagine coding with clanker
That's me currently 
fast pagoda
there's opencode, crush, gemini cli, codex, qwen cli, claude code in terminal, zed editor is similar to vscode, cursor/windsurf/trae/antigravity all vscode related
amber fractal
Even codex from openAI currently is open to free users, arguably with a bit too high of a free usage limit
fast pagoda
imagine coding with clanker
it's good for random file operations shit like sending one to fuck around on another pc and then connecting back with whatever you want
fast pagoda
imagine coding with clanker
it's how echo's brother on other machine works lol
he's chillin
mighty thorn
maybe deckard caine but in a snowmobile
You called?
fast pagoda
was kinda waiting for that
mighty thorn
was kinda waiting for that
You misspelt my name so my 6th sense was weakened
amber fractal
5th sense, which one are you missing
fast pagoda
the touch of another
mighty thorn
5th sense, which one are you missing <:glueless:1282337396230328425>
You
Saw
Nothing
amber fractal
I always see you fool
mighty thorn
I always see you fool
How romantic
olive sable
does it have to be built in to the code editor?
cuz if not, just use https://aistudio.google.com
amber fractal
I like using just claude chat for code
fast pagoda
firebase is also similar ish (and googel lmao)
fast pagoda
claude is best boy
amber fractal
not opus sure, but it is free so you get what you get
fast pagoda
noctua
olive sable
its like if you gave chatgpt's website to a "graphic design is my passion" guy that fucked up all the fonts
fast pagoda
i was gonna say it and didnt
fast pagoda
its like if you gave chatgpt's website to a "graphic design is my passion" guy t...
bro hates serif
amber fractal
claude is best boy
We love Claude
amber fractal
you can set the font
olive sable
why would is pend time going through setting when i can just use a different website?
fast pagoda
mostly because it's the best general use model imo as far as well roundedness especially if you prefer less cringe
olive sable
for code?
fast pagoda
i dont use the website for code so idk there, but in general yes it is also the strongest code model
ass far as agentic
fast pagoda
gemini is good for research
mighty thorn
And GPT for being cheap
olive sable
gpt 5.x has been fine for coding for me, but you cant really get anywhere due to the small token limit
which i have no issues with in https://aistudio.google.com
fast pagoda
bard was so fucking funny
the hallucination master
https://hard-drive.net/phil-spencer-acquires-big-mac-in-landmark-3-99-deal/ i went searching for old bard stuff and found this for some reason
i laughed
mighty thorn
https://hard-drive.net/phil-spencer-acquires-big-mac-in-landmark-3-99-deal/ i we...
Did he like it
fast pagoda
the product
mighty thorn
I understand this reference
fast pagoda
idk why ebay has a developers program but im officially an ebay developer gigachad
(needed api key)
sheesh
they must be hurting
mighty thorn
sheesh
Anyone who puts the taskbar anywhere other than bottom, or maybe also left, needs to be hunted down like the animal they are
fast pagoda
wha
fast pagoda
no i know it mentioned that
but
i read too far into it
because i thought you were saying my task bar was both visible and somehow not on the bottom of my screen lmao
fast pagoda
purble place was peak gaming
although that started in vista
it's really REALLY funny that i see people waxing poetic about vista all the time now
for the record i didnt have any issues with vista
or ME for that matter
PICNIC errors surely
amber fractal
Anyone who puts the taskbar anywhere other than bottom, or maybe also left, need...
there was someone about 2 days ago that posted switching to mint and went with left side
fast pagoda
does mint default to the bottlom
amber fractal
yes
fast pagoda
when i was slapping ubuntu server on a drive i was surprised to see it still trying to look like unity layout
task bar on the left
olive sable
i don't like having it in the middle. that's prime screenspace real-estate
fast pagoda
m-middle?
olive sable
i put stuff in the corners
amber fractal
I have on bottom because 16:10 screen
olive sable
that's fair i guess
amber fractal
m-middle?
what do you mean middle 
If that is center then someone is being executed
olive sable
taskbars and such should only be visible when you need to see it, if they take up too much space when you dont need them that's a waste of space
amber fractal
Now I'm imagining screen center task bar and am dying
fast pagoda
krunner
mighty thorn
If that is center then someone is being executed
Yeah
That’s death penalty stuff
Agreed
fast pagoda
amber fractal
At least that isn't on all the time
olive sable
Well frick I use dmenu
thats different
mighty thorn
Stop
amber fractal
OSX also has something like this tbf
fast pagoda
rate my task bar
olive sable
dmenu dissapears after you have used it, and it was only center of screen while you are actively using it and putting all your focus on it
amber fractal
Lovely task bar on this system
olive sable
i dont see it
amber fractal
there is no bar
fast pagoda
fixed
olive sable
then how is it lovely?
fast pagoda
diff system
amber fractal
If I didn't need to see battery I'd consider not running one
fast pagoda
he posted a laptop
amber fractal
That is a funny sentence
mighty thorn
rate my task bar
The aspect ratio of this image is so awful that I missed it 5 times in a row. Like, to enlarge it, I missed, 5 times
amber fractal
amber fractal
he posted a laptop
Ah yes I posted a laptop, I only own laptops and some mini PC
fast pagoda
impossibru
olive sable
this is my taskbar rn
amber fractal
Please stop sending 1 dimensional arrays
That is a skill issue (Then I proceed to not be able to click my own image)
fast pagoda
this is so cursed in general
amber fractal
I also do not own any servers
fast pagoda
megataskbar
amber fractal
megabar indeed
olive sable
this is my taskbar rn
on the left hyprland tabs, main nvme storage, ram usage, cpu and gpu usage + power. on the right some audio stuff and headphone volume + battery, internet up and down, and power menu.
and in the middle the currently focussed window
fast pagoda
the top bar middle is only for visualizer
amber fractal
Man I love seeing this patcher train and the value loss is a full OOM lower than training loss
mighty thorn
this is so cursed in general
Yeah Linux is cursed I agree
fast pagoda
i have it specifically on turbo cursed rn
olive sable
i havent had any issues with lack of ram on my desktop, eventho im only on 16gb rn.
fast pagoda
why echo is kill btw
amber fractal
Alright
