#programming
1 messages · Page 432 of 1
PSU connectors, VRAM usage efficiency and PCIe lanes
the thought of using a regex for finding primes concerns me
NeuroSynth training requires a minimum of 24GB per GPU
it is worse than you think: the numbers are expressed as unary numbers.
Would it be possible to uh.. do something like zram.. for vram? Or is it completely useless/stupid?
isnt it made so a model can only fit in one card
no combining models into a bigger card
the prime regex has a special case for 2, then it matches strings that have a length of 1 more each integer, 1 or more times. it's incredibly ineffecient.
Tesla K80 24GB for 150
along with eGPU dock and 750W PSU
Whats the k80 again-
If I'd be throwing around 8GB cards and pooling them I'd need to be spending more on bifurcation cards than the GPUs themselves, making each GPU have the bandwidth of like PCIe x1 meaning transfers between cards would suck
Take your personal data back with Incogni! Use code standupmaths at the link below and get 60% off an annual plan: https://incogni.com/standupmaths
^.?$|^(..+?)\1+$
Props will be for sale here: https://www.ebay.co.uk/usr/standupmaths
Fantastic blog post by Illya Gerasymchuk. "Demystifying The Regular Expression That Checks If A Number Is Prim...
The compute sucks too hard to be worth it
Kepler 2.0 so yea
close to gtx 780ti
Two of em apparently
It'd make the training take approximately 10-20x longer
Also may not be supported by my Nvidia driver with it being so old
If the Titan RTX was just a bit cheaper I think it'd be a quite nice companion for the 3090
Training times would be barely affected and I could get a nice quality improvement for basically no time penalty
Welp, i think ill go to sleep.. i don't have anything to keep myself awake for more for a reason reason.. @olive sable should i keep you posted for the 4080 incase anything happens with it? What should i expect?
yes please
it's that one card that is famous for being horrible and being completely unusable
there is not much of a reason to not at least get a P40 instead, even that will be fast in comparison
2080 Ti 22GB exists apprently. i did not know of that.
490 so not a good deal compared to the titan
I think ill be able to test it soon-ish, maybe this or next week, depends when the friend and i are meeting.. he said hed bring it so i can see it running and all.. cant particularly spare 400€ for a card i wont use, but u seem online a lot.. can i assume youd reply whenever someone pings?
modded card
i shoudl stop looking at gpus for sale
Is it a HBD app?

and i want it to exist but i dont wanna make it because i'll do it badly
no it's not
why am i explaining this to a mealworm
how did my life get to this
anyway
the problem i have with current scheduling apps is that they are too rigid
all they do is remember the schedule
im still the one that has to come up with it
what i want is an app where i can add floating tasks that the app picks a time for instead of me
e.g. "do laundry"
But how is it to know the best cadence for tasks
and then the user could set how much free time they want in a day, when they sleep, etc
So the user still comes up with the schedule?
nah
the user just says something like "i want this much empty time between tasks"
or "i want this much of my day free"
ah
and then this also allows you to change the schedule easy
lets say the app was gonna suggest you laundry for the afternoon
it sends you the notification but you're at the store
you press an "i cant right now" button
and it puts it somewhere else and clears your schedule for a while 
automatically moving floating tasks elsewhere
you can also have recurring tasks with a target frequency
and i thought of some other things too, like if it noticed you have a busy period of fixed tasks coming up where you might struggle to do some of those floating tasks, it could suggest to you to do some of those early
but the core goal is that i dont need to pick a time for these things
i have an app that chooses tentative times for me
and then i just do what it tells me
or if i can't, i inform it and it handles picking a better time for me

you'd probably want to have a set "working hours" too... so it doesn't start scheduling things at 3am or something
yeah same idea there as sleep hours
the problem with that app is you'll just keep pushing "i can't right now" and eventually you'll be swamped
and uninstall it
Just like real life
whether or not you follow a schedule is a problem on the human side
thats not the problem im trying to solve
the problem i want to solve is that i dont even attempt to make a schedule because i hate picking times for things
but i feel that if an app suggested it to me "at random" and it happened to be a good time
then i would probably do it
like if i got a notif that was like "hey got a sec? how about those dishes"
yeah sure
i can do those quickly
but there is no way im putting "Dishes - 2pm" in my calendar or something
zero chance
the "randomly ask you" app would be easier to write than a "schedule and then ask" app..
its the same thing maybe you just misread my idea
schedule and then ask implies you can look at a calendar and see when it will be.
it could just randomly ask you
you still do need the ability to schedule fixed events that come up too, obviously im not doing the laundry if im out for dinner or something
and for things like laundry, you don't really "schedule" it but you need to specify a frequency
try an AI agent like openclaw with access to your calendar, i bet it'd be fine adding tasks for you
AIs are bad at choosing random anything
ah
and then the app will look through all the floating tasks and try to suggest them at times that respect your preferences for free time, your fixed commitments and the ideal frequencies you've set for those floating tasks
you could do the random part with a tool though and then the AI could work
i think you need a one sentence description for it to be worth writing unless it's only for you.
that's the only one I think 
been there for ages

i don't think that's why its hot though
I would blame uh
the case
im very tempted to put a small fan in here just to move air around
to an extent it is definitely the drive itself, gen 5 does push them so hard they will throttle without a heatsink 
depends on what you're doing then i guess
i made a to-do app that was more of a pool with priorities and a dependency graph
it just showed you the shit you gotta do today, and ranked them based on their inherent importance you set, how close the deadline is, but it also took into account every descendant of the task that depends on it
put that in there too :D
yours truly
- feature creep
i wouldn't worry too much about it unless it gets hot hot, not like air really does anything
it's just a block of metal
gen5 drives pretty much all suffer heat issues at full tilt
even the faster gen4 had the issue
monkas
is this the drive that was taking like 20ms to respond
you see thanks to the magic of sff cases right
my ssd is
directly where my cpu air goes
and then when you add the case
carefully designed to aerodynamically windtunnel "ventilation" over all surfaces of the drive
or is the drive chilling
yeah that was 150W of cpu load + 80W of GPU load + 50MB/s write
i think my drives start complaining over like 70
the rest of it is a bit cooler
mine complains past 50°C apparently 
honestly i was gonna say 50-60 but that felt too low
chilling at 52 doing nothing rn
yeah idk why I get 3 temp readings
controller & nand maybe?
checks out i guess
obligatory full diagram because I can never understand how that ssd in the middle survives
great. the gemini CLI has two ways i can stop the agent from doing what it's currently doing... however pressing either of the button(s) twice (^C or ESC) does something other than just stopping it. one exits the CLI and the other does a /rewind!
i'm just used to double tapping to make sure its dead
my mobo has these massive heatsink plates over the ssd and it feels extra but those things get hot asf so it makes sense
mine has a shitty heat sink that im convinced only makes the problem worse
my last one had the world's most loud little fan in there that i think was meant to cool ssd but idk if it did anything

when i took this photo i thought the price was a bit high
parse_special: (Optional) Boolean indicating if special tokens should be tokenized. When false special tokens are treated as plaintext. Default: true
This was why my <end_of_turn> tokens were all non-atomic. ;/
just a random "false" in the code which should've been "true"
World's most me irl fan, spins up to 900000 rpm on boot then I am p sure spends 99% of time not even spinning at all
i cant even see wires in this case with this new improve cable management
i still can't figure out where itd go
fan on head
head in the computer
fan on head in the computer
thank you yingying

my 980 has seen some shit

4.26 years power on time kekw
Name? Or private only?
oh it's not deployed anywhere
Oo, webapp style?
i just made it, documented it to a high standard, got my points and grades for it, and never touched it again
do you just never turn of your pc?
Hmm.. i set up a vikunja, but their "blocked by" thing is completly useless..
not really no
and your os doesnt crash?
i would be concerned
windows could never 
most of the power on time: windblows
It sounds very nice in comparison to whats available for selfhosting, thats why im asking

tbf i havent had issues with windows myself
but i hear a lot of people say it doesnt liek more than 30 days
dont tell windows server
I havent had nt crash due to power on time either.. the same cannot be said about the userspace
Even two weeks can be iffy.
i restart more on arch because updates are so frequent for like 58902498509 packages
why update? 
in the short time i was forced to use it (testing) i realized it's unsustainable
too much admin
the only reason i update stuff from time to time is cuz discord starts complaining every couple months
if i had that energy i wouldn't need a graph to-do
Me, having updated my archlinux 1 month ago, after not having updated it for about 8 months:
catch a pacman -Syu flying out of my comp every 6 hours at most usually
my god, i want to run it just to see
Hm? Like essentially "this is blocked by xyz"?
cpu spends more time rebuilding packages than anything
I still have neofetch from pacman :)
to make such a todo list organize my day properly - as in, i don't have to think i can just look at my todo list - i'd have to add everything basically
and adding everything, with expiry dates, importance, and placing them in the graph takes a fuckton of time
i have i think like 20 *fetch on my system still xd
and i was missing tasks because i was lazy to begin with - lazy enough not to spend all that time managing todo xd
i like to go collect them
I was SO confused why neofetch wasnt installing via pacman on a new system, until i googled
Fair enough.
I think my next arch install will be zfsroot
also tfw if you keep that arch going for another 4 months
you basically get the proper windows lifecycle
which is reinstalling every fall at the <year>H2 release
Hm?
arch LTS
Oh because like 12 months no Update or what-?
think thats what they were saying yeah
Lol
I mean to be fair i dualboot with Windows 10 Enterprise iot ltsc..
Which uhh
Booted Last time like idk prolly 8 months ago aswell
It takes about 30 minutes just to become responsive because it autostarts like all the Updates
(my poor 2.5" ssd slows to 500ms response Times
I must say
I cant help but guess that cpufetch is.. unnessesary
I mean, lscpu exists
usa stats?
i like to think that this is managed by the cia, and it live-updates the stats
no that's just the software in my router
person is born, +1
why the hell does this fetch have a 1000 line PKGBUILD
my god
they put their entire commit email chain in the fucking PKGBUILD
oh it's openSSL that did that
what the helly
cmon bruh
least ridiculous install for a fetch tool
oh good, over 1000 tests, thank god theyre thorough
6m56 seconds
result:
oh wait it was this one with the crazy build:
pkgs: >3
yea

Pacman

personally i like the ones with terrifying tux who is about to kill me through the screen
optimized the patcher so much that out of 300M patch groups that only 200 cases aren't a perfect match
for those 200 jus tell them there's always another fish in the sea
It was that or increasing model size again, and I don't feel like doing so
we love the single 48 sized head
mfw run smartctl on /dev/*
Apparently i have 237 devices it could be
i expected a few but idk about 237
couple tty
It is ovbiously tty12
if you're not on tty62 then what r u even doin tbh
speaking of multiple TTY: i made a script which gives you a different .bash_history file per TTY, so every time you open tty1, it's always got the command history for tty1, which is separate from tty2 and tty3.
runs from your .profile or .bashrc
i find it very helpful coz i tend to open the same terminals up everytime and use them the same way. it also saves a "cd $PWD" as the last entry in your history so you can press the up arrow, enter and you are back where you were when you clsoed the window.
is that last part not default
catch me out here pressing up every time i open a terminal because usually the last thing i ran was an update and i opened a terminal so it's updatin time
I think Gemini is suggesting I steal intellectual property again.
undercooked 
You can feed the entire text into the 12B model in a single pass.
(where the entire text is an entire book :O)
you mean to tell me the angry ghost of all the intellectual property on the internet is suggesting you grab some gimmes?
(i do own the book, and really it's not any different to me lending the book to the LLM for the afternoon so it can read it)
I continue to be glad undercooked and overcooked are very easy to identify in model output
training runs be like
undercooked is keymash with some spaces, overcooked is space spam but someone is trying to type.
someone 
o7 3060m you are running at 128 batch
that's a lot of compression and chroma subsampling 
Yep, this was my flow attempting to get my repo working again
btw, how often should I update my gpu drivers?
should i rewrite the frontend myself for that steaming piece of shit repo that was mostly screwed for yesterday
which i had to unscrew for hours
or should i just vibecode it and call it a day
because i'll get marks for the project either way
wtf why are the clankers so good at making websites
wth
mf just made the entire frontend for me in 5 minutes
im SO glad i didnt pick web dev
and went to hardware and embedded instead
that's one of the things they're specifically trained towards + naturally there ends up being an abundance of examples
what the fuck is this piece of shit
who the fuck wrote this
the ai certainly did not
who the fuck wrote the jwt key as hardcoded and then wrote it as this
react+tailwind is one of the most documented languages on the internet
so ai is really good at it
Me when batch size 1
Lmao gonna put that into my jwt secrets wordlists
motherfucker said the fraud detection system was something robust and good
university project where if teacher doesnt like our project we fail
i wouldnt let that guy get a fucking job if i knew he wrote code like that he should not be bloody employed
Oh would be super hilarious otherwise lmfao
i would rather shovel pig shit and castrate a 3000 pound raging bull than work with this guy again
he's a good friend, but a really really terrible programmer
Haha
technically it's robust. it won't fail. and that's good. so technically...
i made hardcodes salt for client and server just cuz why not xd
i qlwo did basicqlly the same thing, max size is 1504 and there no header but just uint with following data
like, only working anti-fraud system is limiter + error handling everywhere, in way no exception or error can escape
omg is that java
why is it printing to console 😭 😭 😭
console based bank
might as well use an llm to make it simulate a banking agent
is the parameter for println actually just x
that’s really funny never knew that
here me out... what if...
we gave Neuro's art
to DLSSv5
her train was... well, i guess i could see the wheels and a body and maybe smoke out the top... but with some DLSS magic, that'd be obvious
(but probably still pink)
🤔 generation based on crude shapes has been around for a long time. DLSS 5 probably doesn't have enough global context to go crazy with it.
Regardless of whether it's possible, it kills the charm.
brooo... since when was 6GB of vram barely enough???
DLSS 5 is basically AI slop lmaoo
6gb 3050 😭 I got a 4gb 3050

If this model doesn't give me english I am going to explode
I have been doing runs all day and the last 2 weeks
not really no
react sure it had some history prior to ai
tailwind was the hot new thing basically no one heard of in 2023
my best guess is ai is great at both of these technologies because of how you use them
for tailwind, styling an element as soon as you type it out is straightforward
for react, colocation (no "concern separation") - the components as function model requires you to define component logic right where you define how it looks
i assume all of this is optimal for llms because of attention
also was the cool stack right around end of 23
in terms of gaming, 6gb started to straddle the quite not enough category in 2023, we passed 8gb not being quite enough last year, currently I would heavily recommend 12gb for more and probably 16gb if you intend to keep the card for longer than couple of years
nvidia and amd have hinted at some texture compression tech that reduces vram requirements but havent heard any concrete release dates, so might be couple of years before that comes along
yeah, and when combined with typescript, ai models can really pull their weight
also it wouldn't require agents to switch files for each component like most other solutions but that's like a really minor thing
i'd argue it's not
wdym
atp, if corps weren't being such douchebags, just buy a console lmaoooo
PC games are getting too un-optimized loll
eh, not that simple, nowadays many games on PC and consoles run on basically the same codebase
but devs are forced to optimize said games to run on a specific hardware (PS5), so most of the time, games run better (more stable) on consoles as opposed to PC without having to troubleshoot
i blame ue5 and those who keep pushing for realism. AAA devs have literally stopped caring about any kind of optimization for their games. though to be fair that extends to the rest of the industry. oversized slop is the norm and people don't complain enough about it
needing 8gb+ vram as the "norm" is absolutely deranged and nothing will change my mind about that, we had higher fidelity and better games >15 years ago
with a box of scraps
My 6gb system is loving this 1024 batch size 
fuck no
slop can go fuck itself
also

guysss i got those 1U server 40mm fans for my briefcase airflow
excited for the jet engine in your home
which ones
it's not like there aren't any quiet 40mm fans
just the ones in 1U servers usually aren't 
noctua prolly has some quieter ones
they do (I have some in use)
nice

Wuh

these kind of fuckers
they're literal jet engines
$1 OFF


yeah those are the loud ones
28.5 CFM on these tiny piece of shit
oof
them bastards sound zippy
i already have plenty of those because i used to hoard random 1U servers
just gonna repurpose them
I don't think these are good in a briefcase 
bigger than normal 40mm + too loud
i have a 4 pin pwm controller that can supply 3A per 4 pin
i do it for the lols
besides
the size sorta match because it spans to the heatsink so

maximum air flow
blows high pressure air directly to the heatsink
3 of them
at least it's convenient then 
assuming you still have some other case airflow
12v 1.5A....
i wanna have some fun with this time's briefcase build
2 modes
the ultra efficient one with <10W total system draw
so it last 2 days playing yt video
and this one with these 40mm suckers blowing maximum
surely my battery can handle it 
i just wanna show up to a lecture with it blasting jet engine sounds

no
yes
ughhhhhh it happened again
I want gemma4 now 
Oh what the hell
What are HQ
I’ve never seen that before
Only H and K
So remember that yandev tier template from Unsloth
did they fix it?
oh right
simplest jinja chat template
i was checking that out earlier this morning, kinda wanna try it
do you think it'll wipe my disk
I wouldn't be surprised if its full of jank and vibe code issues but it probably won't do you any harm
Some of the ideas behind it are nice at least
i was mainly looking at it because it says it comes with websearch ootb, i wanted to make a search mcp server to improve knowledge for mine
Honestly might as well give it a try
Maybe the community'll get it into a good state
Benefit of open source
Would be nice to have a prominent open source lmstudio competitor
ahh no it's a combined repo actually
i just run the llama.cpp server lul
oh yea auth was also the other thing i was interested in
Why does this always happen with AI related soft
Same, I probably won't use this since llama.cpp already ticks most of my boxes
looking at the setup script i'd need to set it up in a vm
classic curl | sh type shit
mfw everything breaks cus cpu decided to take a break
YOOOOO cant believe my 2nd gen magic mouse outlived my apple pencil lmaoooo
wipe your screen cro
Interesting
2nd Gen Apple Mouses are tanks
if an unused return value isn't being used and that causes building to fail, that's not good tooling
(I know none of those tools I see there actually do that, so AI slop lmao)
it's a static linter thing in many languages (typescript in this case) and a valid error, unused return values can mean you're not doing what you're expecting
it's definitely how it should work
Btw, are gaming laptops supposed to be this loud?
man's at the ocean
just wait until the fan wears out and then it sounds like a woodchipper
They do tend to be very loud ye, but as long as you're within safe temps you can move to a less aggressive fan curve if it bothers you
uh oh
I come to you from my phone
Rocm restarting my graphics driver over and over

It did not stop going up turns out
Forgot to give any safety consideration whatsoever just assumed I surely had enough spare
Computer hasn't stopped playing Linkin Park so it's like a forced break
Mistral Small 4 in reasoning mode matches Qwen 4B reasoning and Qwen 9B non reasoning at over 100B parameters 💀
Grim
sacre bleu
Apparently it was only trained on non copyrighted data to comply with EU regulations
Decent hallucination rates at least I guess, if you wanna find a silver lining
Europe is leading the way into the AI Age
continue read
Half the people in the comments don't realise its satire 
20 GB 
there was a different version of that i was trying to find but i am too stupid
UK is the only European country that can do untethered AI research now and that's because we're diet Europe
I feel bad for the Mistral guys having to work within the EU AI acts regulatory framework
i must've seen just another of this dude's posts originally
he literally just shitposts these all day
A while ago

i really wished all other companies HAVE to comply with that regulations as well
actually the problems i have the most is with image and video models
https://youtube.com/shorts/Zijt3xnK-58
i am REALLY happy these fucking models are getting copyrighted. fuck them for unauthorized data use
The Anime AI Problem is Getting Worse..
i feel bad for ourselves that none of the other people have to comply with any regulation and they're literally destroying the global economy to forcefeed you slop built on pirated content which is apparently totally ok to do as a billion $ corpo
yeah
all the data as if its theirs
no cunt, i do not consent my data being used in such ways and im pretty sure many other individuals AND institutions do not consent either
not to mention the literal burning of millions of books
and intentional destruction of entry level jobs, and resources, and cities and more
its just getting out of hand. i really wished global regulations will be put on place and held all these companies liable for such unauthorized data usage
all companies no exceptions
never gonna happen unfortunately
it's never happened before for any other issue
there's money to be made folks
its not even a cool tech anymore at this point its just the same slop with different twist each time. no real breakthroughs
that's the problem you either join in or lose (from their perspective) not that they seem to have needed much convincing
the EU AI Act isn't really about copyright at all btw 
that's mostly covered by existing copyright law
-# if anything the regulations seem to explicitly allow opt-out use for training
💀
it's not just about creative works either, they've wormed into governments and corporate highest levels - that's unaccountable decision making and influence ruining millions of lives as well
yeah our world is governed by boomers its so over 💀
“Untethered ai research” meaning what exactly?
No training compute cap and the ability to utilise copyrighted data
it's just more clear than ever that if you have enough money then you can do whatever you want
beyond savable, laws dont matter out in the open for these companies
i wouldn't even call that research that's just development
Fair
That may be more accurate
the ability to drain an entire state's worth of power making shit unaffordable for millions of people, so you can slopify yourself in ghibli style
the day that happens we'll have bigger problems because of it
insomnia fucking suck ass
Worth it
what's sleep?
Not the ghibli part tho
i don't want the companies held liable, i want the entire C-suite held criminally liable as individuals and imprisoned for life for the unmeasurable damage done
this tbh
best we can do is target end users
if not because of the frog i wouldn't finally be able to lie down on the bed rn
pinout
god damn it she figured the whole pinout thing quick
while i blow a hole on my head
i didnt realize you wanted a pinout for some reason, i said job done finding the cable
You hate C, C#, and C++ that much?
Based
just plug it into the standard lenovo digitizer header on ur mobo
i like C, the rest can burn
yeah bro because after i got the cable how am i still supposed to decode the i2c sda
wow
plug in 2 laptop u dont hav
the bird
AI image gen I don't like, but if the US and China are okay with training on copyrighted data I think its a matter of national security that Europe should be allowing the same, at least until they catch up
it won't just work out of the box from the 40 pin eDisplayPort
no sir it will not i found that out hwen i first took my lenovo apart thinking it'd surely just be displayport
We've already seen from Claude the military and geopolitical significance of this tech
national security my ass, doing the same exploitation of people and their lives is the opposite of security
i like that europe doesn't try to sell our souls for pennies
basically its more of nations wanting to compete so badly on such industry they decided to fuck all and allow illegal shit
not the first time it happened.
that's what i was saying, the choice is lose with your morals or go along with it yolo
still happening w/ nuclear
time honoured tradition of we pretend to care unless tHe WeS t Is In DaNgEr
"military and geopolitical significance"
good idea, we rely on uncertainty quantization/predictive prob distribution for our weaponaries!
you do understand that AI destroys the fundamentals of economy, education, privacy, safety and kills people?
horray i guess
especially with most big AI labs already being based outside of the EU
so heavy regulation would pretty much do nothing except kill Mistral 
we truly live in a fucking world
idfk why tf is the short of e-DisplayPort is filtered in here though
caught me off guard
internet history
also automod bypass smh
why does my SSD have a petabyte of writes after a week
does that even count
i think technically
fuck
its a shame because e display port is a real technical term
and not to be confused with whatever the fuck the filter intended to
just in case mods read
you can swear like a sailor but cant say ed port
this one's spritely
tiktok type filtering
Australian swears >>>>>>
this one though
damn, its eating the spare already
well this load has stalled so that's fantastic
stalled or your disk gave out finally
Damn
I can’t afford enough ram to perform this math
I need 1gb but I only have 64gb

We do a litle bit of tomfoolery
When the driver forgor what power limits are

"recreate any open source project from scratch" "git clone"
-# it's satire 
oh the actual use for it is atrocious
yes
"© 2024 MalusCorp International Holdings Ltd. Registered in [JURISDICTION WITHHELD]."
if they had one that took proprietary code and made it open sourced that would be worth using
alas
Patricia Bottomline
VP of Legal, MegaSoft Industries
Chad Stockholder
Engineering Director, Profit First LLC
it's real, that's me
what if it's real but pretending to be pretending to be real
waiting for the non-satirical version to appear since life imitates art
not sure how that works with licensing as a concept anyways, if you're directly copying something with the express intent to make a 'legally distinct' version, i guess that's reverse engineering easy mode
those kinda already exist in the form of website cloners 
except the whole clean room approach
just not for libraries since those don't actually matter, only the product does
This is more or less my view, if you want to steer this thing in a better direction you have to introduce your principles later down the road
It can also defend against that destruction, and the US and China are full steam ahead no matter what
kinda always figured those were in a let's say murky area
This
Evo2 biological foundation model. AI for DNA #ainews #ai #agi #singularity Thanks to our sponsor Higgsfield. Try Cinema Studio https://higgsfield.ai/s/cinema-studio-2-0-theaisearch-yHsEWI
Evo2 paper: https://www.nature.com/articles/s41586-026-10176-5
Evo2 download https://github.com/ArcInstitute/evo2
0:00 Evo2 intro
0:51 LLMs vs DNA
1:50 What ...
42

42
I don't see why, in theory, it wouldn't be fine
I forgor where but clean-room reverse engineering is proven IIRC
yeah i mean thinking on it, to block something like this would be to make reverse engineering in general prohibited
which aint bueno
mabe not fine if it's the same entity providing the spec and the implementation, even if the tech separates the knowledge internally
for libraries it just seems like an auto-ecosystem-fragment-o-matic and great way to lose track of upstream everything
All its doing right now is starving Mistral
From a pragmatic perspective it doesn't make sense
i really wonder where it'll all go in the end
some very powerful very deep pockets seem to really be convinced from some source of info that this is worth it
to risk literally everything
They just need to be convinced other people will buy in
Investors follow large quantities of money
There's no grand plan here
i dont think it's some grand plan or even conspiracy really, it feeds into itself of course
The exception being China, the party hope their AI and robotics research can make up the Labour deficit from an ageing population
the top 5 us tech companies by expendature are spending more on this in 1 year than it cost to build the entire US interstate highway system
the expenditure by only google on datacenters in 2026 is going to be about the same cost as the entire apollo program
you could fund 80 manhattan projects per year with this money
global 2026 ai spend is 2.5trillion forecasted
that is 8 international space stations
is that inflation-adjusted?
yes

1960-1973 apollo program
257 billion adjusted for inflation
25.8 billion at the time
us hyperscaler capEx - 2026: 700 billion
this is probably the most money that's ever been spent on anything in human history by like 10x by the end of this year 
94% of all operating cashflow of the companies involved for the entire year are going into it
panama canal: 11 billion
mecca's 'great mosque' total cost over centuries: ~105 billion
this is 2-3x the total military budget of both the US and china combined
Better use of the money than military dick measuring contests even if you're not a fan of AI tbh

i dont even know how to understand these figures beyond comparing it to several "we do not care how much it costs, do it" tier undertakings
True actually lule
I may have gotten ahead of myself on that one
meta could fund the whole apollo program this year and then still have spend left over enough to build the ISS on their own
Tbh I don't think the returns will justify the spend but I kinda wanna ride this out and see where it goes at this point just for the hell of it
Mankind having unprecedented compute capacity will kind of inherently create scientifically interesting outcomes
at what point do we just look at this and go "so like money isn't real, huh?"
feels like a lot of things turned out to simply be mental blocks
It isn't when you get to this scale that's for sure
laws dont matter if you just do shit
money doesnt matter if you just say haha and swipe the credit card
checks and balances dont exist if they arent enforced and the entire world has realized that over the last 10 years
it cannot and will not
it's like the omnidroid from the incredibles the only thing powerful enough to penetrate its armor is itself
well dann that's not much of a discount when you charge $40 to ship one(1) stick of shitty ram
How so?
the very creation and use of it is what causes the destruction
it itself is the source of the problem
i encourage you to look at the fundamentals of how these things are being run and their costs
I'm very well acquainted
if you dont see the destruction you arent
I even run models on my PC for robotics and translation
What specific destruction are you alluding to
it's really not that mysterious or opaque, but refusing to participate whatsoever cant really solve that problem either
Because there is a lot of harm you can point to
But its also not really something that can be meaningfully curbed in todays world and societal structures
it's akin to declaring coal too destructive to burn for power as a 1900s industrial economy
If you mean the water use thing that's been thoroughly debunked, if you're referring to jobs that's the inevitable consequence of any technological leap
If you mean its impact on the arts then yeah I think it has no place there
i feel sorry for the saddle and horseshoe makers.
In terms of energy use the jurys out, I hope it drives adoption of nuclear and renewables to meet the demand
stop driving!
Certainly it seems to have people taking energy sovereignty more seriously
If you mean disinformation and societal decay that's algorithmic social medias fault and we should kill that with fire
If you mean military application I agree its an abomination
Same goes for surveillance
one of the few overall positives to me is forcing nuclear back on the menu or else
and also making countries take a hard look at just how sovereign they actually are as well is probably going to end up a benefit in some respects
But those are symptoms of far bigger more systemic problems
The misuse of AI isn't the cause and I'd deem broad unfocused anti AI backlash a distraction from meaningful action and a waste of political energy
people will always "misuse" technology. everything has multiple uses.
Did I hit the marks @split steppe ?
I'm not being combative with that ping btw, I'm just wondering if you're referring to a harm I didn't cover
the total power usage of people playing games or streaming media at home is probably doing something too
i cant come to any conclusion other than that it'll simply be: spend nothing, become a client state to the ones that did, and then you wont have a choice anyways
or spend everything and hope you end up on top and in charge to enforce some sanity on it later on
it's fascinating what mental gymnastics people will do to attribute issues directly created by the technology to something else
explain
and also miss the fundamentals, as if it's a gotcha
dude these corps are forcing it in schools
you say it's simple. go ahead.
it's not simple, you have to think two steps ahead
oh. you have to be paranoid. got it.
What examples are you referring to that are directly created by the technology?
I did cover some that are but I also explained my perspective on them, even agreed some of them do need tackling
But plenty of issues created by the tech aren't necessarily inherent issues with the technology, they're issues with the society it exists in
if it's so obvious tell us. if it's not obvious, tell us. just state it clearly without saying "you are all so stupid for not seeing it"
we don't know what you are talking about.
give me some time to collect the research papers
Go for it, to be clear I'm not trying to shout you down here
I think its a healthy discussion to have
if you can't state your point in 30 seconds, it's not one point.
when have i said it's one point
when you seemed surprised that we didn't guess it i expected it was obvious
it seems it's not.
it's fascinating what mental gymnastics people will do to attribute issues directly created by the technology to something else
which issues?
that's a simple question.
i imagined it was something obvious, or you wouldn't be so vague about it.
Nah that's cap
There's a reason tiktok is ass for any sort of nuanced discussion or information
I mean its ass generally but still
Pascal's wager is a philosophical argument advanced by Blaise Pascal (1623–1662), a French mathematician, philosopher, physicist, and theologian. This argument posits that individuals essentially engage in a life-defining gamble regarding the belief in the existence of God.
Pascal contends that a rational person should adopt a lifestyle consis...
the way they were talking, it seemed like it should be obvious to us.
like it was one single obvious fact.
that we were stupid for not getting
and "mental gymnastcs" implies we know the truth but choose to ignore it
There was an implication of intellectual dishonesty yes but they clearly walked it back afterwards
unless they're not talking about us, and talking about others, in which case i have no idea what they are talking about or why and want to know
there are very clear second order effects from it:
- worsening of education, this has been reported by multiple schools. ai corps are forcing the adoption of these tools there. this has a huge future cost for everyone, this is not time you can somehow reclaim.
- energy costs - corps are building their datacentres in areas that cannot support their load, and they are using as much as entire cities. several cities and countries have reported elevated costs. short term yes, but it exists.
- misinformation - you can set automatic bot farms to have real conversations with millions of people pushing whatever agenda, this has been reported and is clearly visible in most social media online for years now. deepfakes - both video and voice are at a point they are not distinguishable. this has already created reported problems.
- surveillance - automated mass surveillance has landed hundreds of innocents in jail already. more data is being exploited day by day, corps like palantir are already collaborating with many governments to track more.
- accountability - these models are being used as an accountability sink, thousands of companies and institutions are using this already to blame the machine for false information, firing and killing of people.
- economy - concentration of capital caused by it has condensed the manufacturing and more towards a select few corporations making everything more expensive end to end, as the capacity cannot be allocated to the thousands of other things we need silicon for.
- entry level jobs - companies are using it out of forced FOMO as an excuse to not only fire workers under the false claims that these agents are capable of real work, but also those who do work entry level are forced to use AI to keep up with illusory productivity gains. the usage of AI prevents learning in more ways than one and creates unstable systems in the long term.
- if the gamble doesn't work, a huge influx of old tech that destabilizes the consumer market
i don't have enough characters to keep going
i can type more
so i should stop using them to solve all those problems?
i'm not sure what your point is
people have always complained about new technology ruining education.
writing made people forgetful!
you can make a 100% on-the moeny case that this is a bubble, that the spending is unhinged relative to current revenue, that most of the value will concentrate in like 3 companies who all want to fuck us all over, that the environmental costs are being externalized (horrendously, and killing people along the way), that. and you'd probably be right about most of it
but none of that matters if the shit ends up working at the level people are betting on. if you sat out the building out of internet infrastructure in the 90s because the spending was "irrational" you were technically correct about the bubble and strategically dead. the bubble popped, people lost moneys and cried for a bit, and everyone who built on top of it won the next 25 years and partially are now the ones ahead in this new thing
these are just summaries of them also
before writing, you had to memorize stuf. kids these days can just write stuff!
i already know your stance zenith, no need to waste characters
the fact you need spoonfeeding how to solve problems is a kind of irony
what? you're saying i can make a real difference? how?
tell me?!
i don't believe i can
Did I just walk into drama
we don't do drama here. nor politics. it's against the rules.
ohhh okay
Point 3 I'd argue is a social media problem not an AI problem, 4 is a societal and political issue rather than one caused by AI, 6 and 7 are the consequences of emerging technology, 8 sounds like a good thing to me
Point 5 I'd argue we need to update the legality around, I definitely agree there
pretend i'm an idiot. tell me what i should do to make a real differnce.
it's up to the corps.
im just sat here being defeatist atm
Part 1 is another thing you can solve on a regulatory level to some extent
none of it is my responsibility. i'm going to use what there is to do stuff, because I can't change it.
social media just amplifies it, however there is an economy of scale with 3 that wasn't possible before. high quality generated material on demand, whether text, audio, images, video was just not a possibility. today you can manufacture an entire warzone in seconds and basically nobody will know but the most adept at analyzing ai generated content
2 I think is an infrastructural problem that can be tackled and yield benefits whether AI succeeds or not
personally, i can't do anything. if you want to solve it, become a politician or super rich person.
AI amplifies the symptoms of social media, social media doesn't amplify the problems of AI to nearly the same extent
You remove social media from the equation and that AI harm largely dissipates
yes the whole thing is an accelerator for things good and bad, atm a lot of bad, but head-in-the-sand does not solve that issue
Algorithmic social media is just a net negative on society, has been for a long time
it doesn't at all, you have the actual news reporting ai generated content bit by bit
I'm yet to see a reputable news source have this issue
the AI slop they report on usually seems to come from social media
Were they reputable reputable though (i.e BBC, Deutsche Wells, France 24)
one was bbc from what i remember
I don't recall any such scandal, I'd be interested if you could find one though and willing to walk back somewhat
but what does refuting the entire category of tech do to prevent this though
other than just give china the go ahead to decide how it goes
build against it
or the US
i do
"the whole world is going to pieces" has always been the case and always will be.
we personally can't do anything about it.
i'm fed up with the cattle attitude personally
do something then
i do
I disagree there, your individual choices and political will do matter in aggregate
what? how does it help?
wouldn't you like to know
I don't think you can solve, say, problems with the US and China while sitting in the UK though
^ that's the issue

the "AI problems" evolbug is talking about: how can I affect any of them? what can I do about any of them? nothing. they are irrelevant. (if true) they're not great.
but bad stuff happens all the time.
we have clearly seen governments expressly loosen the restrictions here to allow machines launder accountability everywhere up to and including the government and military

found the corpo npc
what should I do then?
use your brain? idk?
you're a dev just like them
i am doing what i think will be most effective in counter to it
you claim i can do something. i can't see anything to do that will improve things. tell me what ii can do that will improve things.
then lick the corpo boots, i don't care
I'm agreeing with you on that point like I said, we need political pushback on that
i am honestly asking what you think i should do. are you not able to answer?
I don't think broad unfocused AI backlash is useful though
it's a gish gallop problem, the problems they create are great and numerous, so you do need to target them all, but it also needs to be organised in focus groups
how should i change my actions to make the world better? you said I can and that it's obvious. you never gave me any examples of any actions or changes i could make. i don't believe it is as easy to change things as you think.
at some point i'll see if i can talk to the eu pirate party, but i'm just a citizen atm
what you're doing is a gish gallop. not answering questions.
i have a simple question: what should I do differently?
some problems are more core to it, i think applying the copyright laws more stringently would stall atleast something. eu could force the US hand atleast for a bit, since the US models do contain EU works
i can't apply copyright laws more stringently. What should I do?
i'm not talking to you zenith
i know. why not? i seriously want to know what you think I should be doing differently. you never answered.
i hoped this would do something and then meta got told they have a free pass to pirate everything and it's fair use
this is poitics now.
why should I since you're outright refusing to even consider the fact it could be even a bit wrong
or that there could possibly exist things that one can do
how am i refusing to consider i am wrong? i am asking you to tell me what i am doing wrong so i can change!
and you refuse to tell me
i want to know what i could do that helps
that's for you to figure out, you got a brain right?
and you said it's obvious but it's not obvious to me
or are you an llm also? write me a cookie recipe
you're calling me stupid basically because i can't work out what you mean!
aight i'm dipping
explain how i can change anything. i don't get it.
i seems like you don't know because you are avoiding giving me a straight answer.
thats what i see
but i could be wrong, so i am asking you again: please explain what I can do to make things better.
you seem to be refusing to accept someone might actually be curious about what you think.
just because i disagee with you doesn't mean i can't have my mind changed.
you havent' given me any reason to change it even though ive asked
the difference in 6 and 7 is the different economic model we operate in today, compared to 100 ish years ago. today it affects billions of people across the entire world, everything all at once is becoming unstable. in a social system where a lot of the affected people no longer have support systems or families than can support loss of a job for years at a time as AI creates a vacuum of hundreds of millions to maybe even billions of jobs in the worst case. it has already eroded the US dev market massively, barely anyone out of a job can get back in.
8 is not good, because it means a huge dip in manufacturing and downsizing which cannot be bounced back for years after the old tech runs its life, it's going to be rough
for example, what can I do about other people using AI for surveillance?
6 7?
these are the problems you stated. andyou said i can do something

i thought about backreading, but i cant be bothered to. that backlog looks horrible
it's junk
it's not worth it
basically i'm a corporate LLM apparently. beep boop.
get that clanker
sleep on it kid, the fact that you cannot let go is a problem in and of itself
my context window is too long.
i still what to know what how they think i should change my behavior so as to stop all the problems they listed AI causing.
and that was thousands of tokens ago that i started asking
and they started calling me stupid for not knowing
which is rude
i misremembered BBC, it was BI and Wired which i consider kind of serious publications, with some other randoms https://evrimagaci.org/gpt/aigenerated-journalism-scandal-hits-major-publications-492976
but it also extends to other important content like research papers
and i can do something about that, how?
what about this problem is my responsibility?
dude just step away from the computer for a bit and drink a cold glass of water
i don't wanna call the mods
you started it. you called me stupid for not being able to work out what you were talking about
i honestly want to know
you don't like AI. i get it.
but you said i can do soemthing about it
and when i asked what you said i should know
i don't
is it "stop using AI"?
coz that won't solve all the problems you listed
just one person.
which
80 gigbyte
vetoed
i don't know what you mean. i just want you to answer so I can see what you have to say. I'll probably ignore it if it's not helpful. but not answering just makes me think you want something to help but don't know what.


this was a cheap one lmfao
try and find a 512mb ssd
i give up. they are an idealist who just wanted to share their thoughts. they have no actual plans for how to fix things.
if they said "i don't know how to" i would've respected that.
darn, it's getting rough out there huh
time to bring back tubes (?)
what's goin on over at hynix
ah right
20MB
pog
i am hawkishly watching for 16gb ddr4 ecc as i need it for my cascade lake board to put with optan
and
those 20mb are prolly hella reliable though
i keep mixing up fuckin micron and hynix too
every time im like WHAT IS THIS THING?? IT'S THE SAME BUT IT'S ALL DIFFERENT THO??? and then i realize im staring at the hynix version of the same thing
ah
so it is 
IBooMers had ssd in the 90s
sundisk old name
i wonder if they went firecuda all in bcause it has cuda in the name
even tho its just from barracuda
CUDA?

what r u doin spectacle
sandisk is jsut a lot more recognizable
i wish sandisk would recognize that i want to purchase a cfexpress card on a whim at best buy and would send some to the ones near me
just send em an email, get a freebie
Recently, I've been working on the basic algorithms. Once I'm familiar with everything, I plan to try building a blueprint
theyre like hundreds of dollar tho
u got contacts 
exclusive

He doesn't care about ROCm
blissful existence
how many pixels does the ssd have
they used a contemporary digital camera












