#programming
1 messages · Page 450 of 1
If not then unfortantly I won't be able to share a final count screenshot
You sure led an interesting life lol
my aio pump is making a weird idk rattling sound, still appears to work though, and what's weird is it goes away for a while if i slap/shake the pc which makes me think it's air bubbles, but no matter what i do to let them into the radiator the rattling goes back right after
the sound is driving me insane bruh
🤨

You can too for the low price of one harpoon command
radiator mounted to the side, tubes down, so i'm really confused
how old?
could be drying up a bit
2 centuries
holy shit
heirloom
ok no it has to be the pump just dying because it still rattles when i rotate it so the pump is below and the radiator is on top
aio is idk 2023
the family nzxt kraken elite
i got bored and borrowed the reflow station on an old mobo
my grandma's brother used that thing thru primary school
Nice
can u place 32 more gb of ram on my gpu
but why does it stop when i slap it 
bearing wearing out, slapping it knocks it out of the rut for a second

its a geforce 8400m gs
you can have it if you want
pay shipping

and youll have to reball it as well
fuck it we ball
the aio that murdered my 5900x was from 2023 and had dried up
how did an aio murder your cpu what
yeah but thermal throttling
oh
took a couple pins off
so really i murdered it
but the event that set that chain off was the aio dying
all i know is im glad the 9000 series (and 7000) are lga now
After opening a few old AIOs to mod, the concern for the SLOW loss of fluid is less concerning than the eventual blockage you get with mixed metals in the same fluid.
Huh... konii, are you on that Chinese phase of your life right now?
yeah can confirm same experience
galvanic corrosion
is really bad on cheap aios
i thought it would be the corrosion shit since i remember that whole drama with gn posting a few videos on it
but i opened the aio
and that thang was empty
oo
must've had a small leak or something
spit in le thang
im never beating the allegations
copper waterblock, aluminum radiator?
that's a galvanizin
why don't they make the radiator full copper on $100 aios 
A harpoon approaches yingying... 
koniifer#0
@sage crag
computed_fn_65643#0
@opaque wharf
<no description>
Harpoon returns to hand after played

Because basically nobody’s done that for like 10y now, no?
The main reason I'm so dangerous
More actually
Wait, can't it be reversed?
?
The harpoon
There is catch or hit and return to hand
uno reverse
so that how atori hit many n not lose
What does return to hand mean? Do I need to actively use a command for it?
Oh, I still have 3 harpoon...
She had to bait for harpoons because Quack used all 5 on trying to get Ellie, I taught her the "send a message in NN strat"
Wait, so konii can't have her yellow color again? 
that don't look that bad

A harpoon approaches Iggly 🫪... 

An example of the bait in question, you bait people to throwing at you; they miss then you get the harpoon
iggy whats your harpoon count now
then quack waste all on ellie

Quack wasted all of them first then passed it on

The logs are here 

29
Huh, is vedal name always vedal? I think I remember it having 987 on discord too

it always was like this

Maan, gone for a while and my memories is this bad
quack hav 6 mutual w ved
hi
Once I get a second 3090 I could probably run Melba with a much improved LLM
I was asleep lol
lateee
Also can I have rapid fire plz

I have a few too many harpoons
Well I'm still going to collect, but then rapid fire
Shiroo, why are you a zeppelin now?
Yall, I wanna ask
If you can get free 1 room, what would you choose
Graphic designer room or school pc room
Can I have neither?
which one is the graphic designer one

Why does it have so many medications 
average #programming member
Oh because programmer have bad mental health
Dang it I was going to make that joke
I assume the left because it has a drawing tablet
the hostility is insAne
The room with bunch of cute stuff is graphic designer
(I js wanna ask, how cooked is my room)
Me too teddy, me too
@true hemlock if you want to try and get a headstart you can still bait while dead as you can get hit or gain harpoons still.
Y'know, the left room may be messy, but the right one...
Some room I found on tiktok
Oh it's a toy gun for playing sounds
generated
Yeah
I don't think programmer room are that clean
Unless it's big tech company
Not a full time arch user on top of that
i like how bsd has a netcat
The only sketchy thing that I could find on the left room is why would you install a CCTV on a bedroom. But hey, I'm not judging

How does vedal room looks like
Have to outspy the goverment
"Please don't SWAT the king"
So it is confirmed that Vedal is just bored af king Charles
yes
i have been saying this for a while because it's extremely true
Me showing my mom that this is how I use my pc
Since this is April 1st, can we get a free pass for one single joke?
i used mine on my life
As long as no one snitches it is fine 
The snitch will be Vedal lol
has to backread the thing that happened earlier
Does vedal pc is like 6-7 monitors and 1 ddr5, rtx pc
I have no issue with Vedal backreading, I do however, would like to still be on the server after mentioning him directly 
only 6 7 monitor
and that is not me gaining way too many harpoons
AM I ALIVE
A harpoon approaches Iggly 🫪... 
igglies#0
@amber fractal
andrewbotics#0
@leaden crest
<no description>
thats a way to die quickly
Hmmm, idk, are you?
Programming seems to be netural besides that one person who got harpooned on every revive
of course that is subject to change
A harpoon approaches Iggly 🫪... 
COME ONNN
Tbf it was very deserved and not me suprigingly
YEAHHHH
wrr
iggly count?
A harpoon approaches unkomputoble_fn.vue... 
computed_fn_65643#0
@opaque wharf
aninconspicuoussemicolon#0
@warped narwhal
<no description>
People are going to try and stop me, so as long as the aura works 

I get a lot of harpoons off of failed assassination attempts
A harpoon approaches andrewbotics / streameroid labs... 
I see there was some Wispers activity and a ton of other activity here
Quite interesting
Ignore one chatter
I did notice there was someone with a completely miswired brain
Oh yeah, you missed out
Everyone was getting brain damage regardless of who was on the account
Quack did send a report on it
I did also see that
Love how they swore vengeance but can't realize who they needed to enact on
By the way how many of the pointy sticks did you end up with? I saw you collecting em
A harpoon approaches Iggly 🫪... 
I see I missed a massive
earlier

wrr
What was this particular oracle's wisdom?
wrrrr
Gender Calculus is a anti-math AI concept that doesn't use math
and the example is a python script
ofc importing math.pi

afunyun did most of the back and forth
Revive?
You have 31 harpoons and about 15 seconds before NN catches on
Quick fire those things
I can be greedy
There should be another revive in 2ish hours
plus mobile skill issue
A harpoon approaches Cerber ♡ [Vice-President of NN]... 
cerbervt#0
@upper kite
aninconspicuoussemicolon#0
@warped narwhal
<no description>
Finally get to see one of these live
@true hemlock you are alive and able to harpoon now
Feel free to practice your aim! Special rewards might be available for the lucky few who hit! 

wrr
A harpoon approaches mlntcan (baaast)d... 
mlntcandy#0
@rigid snow
koniifer#0
@sage crag
<no description>
arer you fuvckjing kidding
shr 
Ain't no way it's the bot creature
she had been instakilled for the last 3 rounds
coo
konii --version

Is there by chance a most harpoons leaderboard along with most stabbed?

my pain
Harpooned: 22
Harpooned: 19
Harpooned: 19
Harpooned: 18
Harpooned: 18
Harpooned: 17
Harpooned: 16
Harpooned: 15
Harpooned: 15
Harpooned: 15
wrr
no harpoon stock count
I know some others are gunning for this spot as well
someone tried to harpoon me 
A harpoon approaches Iggly 🫪... 
igglies#0
@amber fractal
koniifer#0
@sage crag
<no description>
miss
i didnt get harpooned hooray
Classic
lmao
You have to hope it hits
otherwise it will hit
True
@sage crag are you going to stab anyone else, the harpoon returned to you after hitting me
A harpoon approaches Iggly 🫪... 
igglies#0
@amber fractal
koniifer#0
@sage crag
<no description>
A harpoon approaches Iggly 🫪... 
Double kill

Not as bad as the quad kill
this is a different project but are you fucking kidding me
installation for humans: ask not a human to do it
downskilling installation
the concept of insisting on abstracting a 5 flag config away into a chat with an llm
I hear the readme is AI generated so sorta a slop to slop translator
oh-my-opencode is such a terrible name
open-my-code
oMoMoMoMoMo
welcomes the agent but tells humans to F off
oMoMoMoMoMo
oMoMoMoMoMo
i have no idea what oh-my-opencode is in comparison to the normal opencode. but reading that readme from them gave me atleast 20 reason to not trust the code written in that repo in any capacity
the main selling point i believe is it bypasses the claude code requirement if you have a claude sub
they'll still ban you

anthropic sucks
what a surprise
I'd rather see it being used for local models in a swarm fashion
you can do that but
you have to have crazy hardware to run a good enough model for this in the first place
let alone parallel in swarms
you could also just learn to code
thats the way im going. i tried doing opencode stuff but man the code it produces no matter the model is barely usable and really badly architectured. the most i use it for nowadays is to rewrite a piece of code in a different language

i mostly use these for being lazy in terminal and glm4.7 air is really good for it locally
flash
not air
the 30b
yeah i know
how does it compare to qwen?
way faster than qwen 35a3b or 27b of course
weirdly, parallel throughput is easier
than not
faster? 
much higher tps
Hoooooolllyyyyyy
Blur has a TON
me when i reframe the amonut of people trying to kill me
how much higher are we talking. qwen 35a3b is insanely fast already
like 100+
rdna4 might be a weird case
actually my bad. its running at 45tks
thats not glm
but surely its not 3000tks
that's prompt processing
i don't think thats possible on consumer hardware
it's 132 t/s
this is in mem
fully in memory?
i can't fit that much in vram
that seems more reasonable
im doing 27b rn to see how slow it is in comparison
glm 4.7 is looping for me

i just said "tetris in rust"
i would loop too
its been going at step 1 for the 3rd time

but also seems like glm is slower for me than qwen (disreggarding it can't actually finish anything
)
and i wil not put you in danger

chat chat chat
OF COURSE WE HAVE THOSE 
if a websites image i click on has the lh3 google thingy right
what's ur mem bandwidth on your card
is the files of the image even IN the website ykwim
is there even A WAYYY
even aC CHANCEE
to get aNYTHING
its a 4080 so i'm offloading the experts to cpu.
take the link and remove the =wX-hY and it'll give you the full res (or add =s0)
it's not on the site itself but it's in google's cache
hey man that lowkey worked

is there even a SLITEHR chance man
like even HOPEEEE
like just a LITTLE HOPEEE
like give it to me straight man
like if i click on the paywall in inspect and put in paid=true

try reader mode in ur browser
idk whatkind of paywall it is but theyve gotten less stupid
just changed some inference settings and now it worked
omg bro this is the jacpot
IM GONNA GET MY FILESSSS
JUSTT YOUU WAITTT
ill spend ALLL NIGHTT IF I HAVEE TOOO

if u reload the page with developer console (f12) network recording itll record every url the page fetches
A harpoon approaches FUNYUN RING:0... 
Dangit
A harpoon approaches 🔺Sam🔺... 
samvanmaele#0
@olive sable
computed_fn_65643#0
@opaque wharf
<no description>
(the code it produced tho didn't
)
warning: tetris (bin "tetris") generated 1 warning
error: could not compile tetris (bin "tetris") due to 25 previous errors; 1 warning emitted
YES
Spawn killed
Did he really just wake up? If so then that's hilarious lol
will strip most of the page off and give you the article
He isn't awake yet 
sometimes images come with it sometimes not
If only devs learned how to use proper semantic HTML
Aww man, would be fun if he is lol
Nah, they could never
Nothing ever happens
Today, we closed our latest funding round with $122 billion in committed capital at an $852B post-money valuation.
The fastest way to expand AI’s benefits is to put useful intelligence in people’s hands early and let access compound globally.
This funding gives us resources to
they need JS frameworks to do all of the work
On that note, this is why I hate React
Bet, release GPT5
Wake me up when they are at least cash flow positive
death sentence 
the websites too smart for this sensei
you can also grab something like jdownloader2 and slap the page url in and if it can resolve the urls itll do it
Good ol httrack
believe in me who believes in u
very tru
holy shit i havent thought about httrack in so fucking long
I sleep, and there is no way I win the harpoon count so I'll just have to hold for mass chaos
guys guys guys lets think about this
Believe in me who believes in you who doesn't believe in yourself and also me who doesn't believe in myself
if one of the picture thats presented is in the google cache right chat
IS THERE even a chance

that the others or similars will even be there
lore accurate
CHATTT
bro you good? 
i would need to see the page to even guess further
IM NOT AN MLG LARPERR BROO
Ready everyone?
no i cant rope you in this
we cant both be caught
A harpoon approaches Krumanchio... 
WARNING: radv is not a conformant Vulkan implementation, testing use only.
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = AMD Radeon AI PRO R9700 (RADV GFX1201) (radv) | uma: 0 | fp16: 1 | bf16: 1 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
| model | size | params | backend | ngl | test | t/s |
| ------------------------------ | ---------: | ---------: | ---------- | --: | --------------: | -------------------: |
| nemotron_h_moe 31B.A3.5B Q6_K | 31.20 GiB | 31.58 B | Vulkan | 99 | pp512 | 2720.75 ± 19.97 |
| nemotron_h_moe 31B.A3.5B Q6_K | 31.20 GiB | 31.58 B | Vulkan | 99 | tg128 | 109.93 ± 0.62 |
nemotron p fast
what am i looking at...
im IN???
guys
let me explain the situation rn'
theres abunch of boxes
and then i see the most noticeable thing
is the paypal logo
theres google logo
jpg unnamed which i think i know what it is
but how do i confirm that thing is what i think it is
do i donwload it!?!!?
do i download everything
1/1/
what are you talking abt?
I FIGURED IT OUT
you need to download it
and then you can right click and open file
holy ikm a genius bro
im ltierally mr robot rn
WAIT GUYS
i still down know
how can i accurately know where the thing i want is
ykwim
can you give context of what you are doing?
im on JDownloader rn
for?
for a website?
send link
I CANT
i dont want you to get roped in
hey man
dont do this
to yourself
JUST TELL ME
tell me
straight up
im bored
just send the link to the website and ill spend like 5 min just looking at it
mk
but like
with your profound knowledge
like
lets say the image thats just bare that they present on the website
is lh3 linked ykwim
like google photos
;like
is THE other images that are behind the paywall
even IN the website
ykwim
i cant really help if i cant see it
are they blurred or something?
ITS not even blurred its not even PROVIDED
like im asking you the HOPE of a slither
that theyre even in the website
YKWIM

can we talk in dms?
if you want to know if they are even on the website. look at your network tab inside of the developer tools and see if the files are sent to you
#programming message that's why i sent this
im in your dms...
also filter for image files so you save time
ye
@amber fractal
nothing in iimages
Released yesterday by the way, its not April fools
like coimpletely dry
like
yea therse some
but definitely
not the ones i want to see
WAIT is it differnet
uif i go in the paywall
HOW DO I INVADE THE PAYWALL??
Let him have his sleep lol
is it BitNet
Nope, it's binary not ternary
original BitNet was binary IIRC, but not sure anymore
but ye, seems to be not be BitNet regardless
welp just bought 512gb of rdimm ddr4
its a bunch of the url but with different tails
like videos.httml
is that something chat
IS THAT SOMETHING???
HOLY
Planning to run Qwen 397b or something
Now can they do 120b model that runs on 16gb 
I'm not going to lie that much DDR4 is going to net you maybe a couple tokens a second
And of course you'll need hardware with the bandwidth to support it
How much did you pay, if you don't mind me asking?
$362
Not too shabby assuming it's actually in working order
kinda why i didnt expect it to get accepted
Make sure you run memtest on it
On raw benchmark averages, 1-bit Bonsai 8B remains competitive with leading 8B-class models,
how
benchmaxxxxxxxxxxxxxxxxxxxxx
It is significantly degraded from the qwen 3 base model
i have to get the board booting to bios first 
There was some research not so long ago about 1b being the future
It doesn't seem benchmaxxed as it drops down to lfm2 8b tier performance
ok yeah they might have lied a bit this is more in the 3-4b performance range
"competitive" = "within idk uhhh 20%?"
yeah
me when i lie
Scaled up to a much larger model class it's probably more viable I'd say
The important thing is that they proved they can make a functional one bit LLM
ok imma need these xeon golds to get here so i can use them to boot thisd board to bios and update it to a version that supports my cpus
so i can actually use this ram
I wonder if you can make back some of the quality difference using majority voting by taking advantage of the increased throughput
chat is a mirrored website from httrack supposed to load for a little while
uhh chat
its blank
CHATTT i think httrack has it in there man
its in there man i swear'
like chat is it the rocket loader min
wrr
i swear this is in your filter
lmfao what
oh, a cyrillic
has nobody tried to do that in 20 years so it never got noticed
ah
bro i swear if anyone says hey i think you should open an investment account with nuveen thru tiaa cref
tell them to fuck off
dont use nuveen
guys it wasnt it...
they are holding like 50k hostage rn for ?????????/
cloudflare
basically just prevents certain js elements of a page from running until the page is rendered to speed it up

chat...
its iompenetrable..
its a bunch of numbers...
purple boxes and numbers...
how does it wokr chat...
where do they store it...
WHEREEEE
chat...
how do i nget the key...
the KEYY
THE METADATA
THE PAIDDD KEYY
thios website
doesnt compress the images or videos
WHAT DOES THAT MEAN CHAT
DOES THAT MEAN ITS SOMWEHERE IN THE CLOUDDDD
THATTT III CANNOTT SOARRR AND REACH FOR THE FRUITSSS OF EVEEEE
brother post the link

A harpoon approaches Toast... 
OLK MAN
OK MAN
ILL GIVE YOU THE WEBSITE NAME IN DMS
OK MANN
but dont tell anyone...
or judge me...
or look at me weirdly...
A harpoon approaches Toast... 
Clearly cheating
A harpoon approaches alphanine... 
alphanine#0
@fickle rain
krumanchio#0
@prime delta
<no description>
what does harpoons do...
chat
IM ATT HE PROCEED BUTTON
like
what if i just
proceed=true
ykwim
like
!??!?!
ykwim
A harpoon approaches Toast... 
toast.dll#0
@opaque sigil
breadfish64#0
@young plover
<no description>
The orange will rise 
15:0 UTC
A harpoon approaches FUNYUN RING:0... 
afunyun#0
@fast pagoda
breadfish64#0
@young plover
<no description>
A harpoon approaches BreadFish64... 
flip me
betrayal
A harpoon approaches 🔺Sam🔺... 

please harpoon me
A harpoon approaches DavePvZ... 
What do you guys think about github flexing

meaningless nowadays due to AI. people can completely game the system
my hopes and dreams...
i can never backdoor....
the beach beyond...
i will never reach it...
A harpoon approaches 🔺Sam🔺... 
You walked in github homepage and sees some indian SWE who make more than you make, and you view his github profile and he has 1000+ contribution graph and badges like it's north korean general
We all should flexing github profile
i need to learn chat...
i need to learn how to penetrate the system...
i wont pay...
nobody will pay...
you mean it has always been meaningless 
WAITTTT
WAITTT
chat im watching something on YT
payment bypass vulnerability
bug bounty
what is a "lab"
this guy is saying just go on his github and get it
bro is mental
WHAT IS A LABB BROO

A harpoon approaches hsnu.ixe... 
omg gjys
im on my bed rn typing on the laptop
on my leg
holy
im larping so hard rn
oh wait
hes using BURP por wahtever
what is a BURP chat
do i donwl;oad bru[p
i shopuldnt ebven ask i should just download it
im depsrate
wait guys im getting node.js
Moon rocket is todayyyyy
wait I might achieve the bypass NO1/1/1/
?!!?!?
is this not POSSIBLE???
this guys tutorial covers EVERYTHING
??
what’s with this channel today
spring fever situation is crazy
ram is going down?
still more than 3x price it should be
It's not going down
Meirl
If it's for the XP gaming laptop you mentioned then I can't recommend an X61. It's a super nice compact laptop (had an x60s) but that intel integrated GPU is rough. Screen isn't great either.
A harpoon approaches FUNYUN RING:0... 
Dangit
It was gonna be for a writing laptop 
That works
All my harpoon, gone 
is ring 0 kernel access

"You do not have enough harpoons for this!" 
true!
Basically slap Debisn on a minimal installion with no DE. Turn it into a distraction free writing laptop.
Slap arch with just vi
Emacs or Vim?
choose wisely
You've angered the dynamic array god
Could do that too
What happens if I choose VSCode?
Well, I sure am glad because being immortal sounds tiring ngl
BurpSuite ?
A harpoon approaches NeuroBot... 
The time is not yet right...
A harpoon approaches Temmie... 
The time is not yet right...
wrr
Damn, Qwen 3.5 has shockingly good pop culture knowledge
This was 35BA3B
Still fails the hololive test though
Curious to see if the 27B passes on that
27B fails too
Still an area where the knowledge of models in this size class shows its limits
I'm not really a Hololive fan but that's a go-to test question for me when it comes to testing general knowledge density
I'll be impressed when one manages to answer correctly
That's just not having the right training data though
Its almost certainly in the training data, they share it with the larger models which get the question right
Rather its an example of an area that gets shaved off cramming all that data into less parameters
they don't use all the training data with every model
i think something like you can't fit more data into the model than there are parameters?
It seems like they used the same 30 trillion token corpus assuming they didn't switch things up from Qwen 3 to 3.5 https://en.tmtpost.com/post/7551878
Oh yeah an update since my testing actually yielded some useful info btw
Setting the context cache to bf16 instead of f16 does actually seem to fix the looping with Qwen 3.5 27B
I figured it was due to the quant I was using but no, that appears to be the culprit
Hem
Did i just got rickrolled
Jokes on you i already got this in my saved gifs
A harpoon approaches unkomputoble_fn.vue... 

We're sending astronauts around the Moon for the first time in 50 years. Come watch with us.
NASA's Artemis II mission is scheduled to lift off from Kennedy Space Center on April 1. The two-hour launch window starts at 6:24 p.m. EDT (2224 UTC).
Four astronauts — three from NASA and one from the CSA (Canadian Space Agency) — make up the Ar...
moon flyby 
Can they at least watch twitch from up there?
honestly, they probably can
as for crewed landing, us plans one in 2028 and china "by 2030" 
Nice
it probably slows way the hell down at distance
idk how fast it'd be at the moon
better than half people's internet somehow probably
imagine being on the crew hot damn
wouldnt want to be bandwidth limited, there are too many applications for data transmission nowadays and everyone needs at least some of it
i remember how B L A Z I N G fast 100Mbit felt when i first got it
now i'll feel it if it's only doing 1.5
Gbit
it's stupid
Ive got 400mbit fiber and for me it's really fast
hedonistic treadmill
tbf 100mbit is good enough for most things
i only really feel it because i do large downloads like
a lot
so it does make a real difference for those obviously
imo we should optimize more, the fact we need 1gbit to feel comfortable just speaks to the bloat
can't optimize away the size of uncompressed video/audio etc
china plans to build a moon base with russia, and i believe in china but our space program is so far past its prime, hope it will be fine
anything is better to spend money on than fucking each other up for no reason
even tho it's kinda that too
sure you can, just need new algorithms
people have stopped a good chunk of software efficiency innovation because hardware has moved so fast
russia at least has still been doing regular launch activity it's not like they just sat never prfacticing anything
although idk the state of the facilities at this time
nor the talent pool
moon base seems kinda pointless ngl
i guess everything can be pointless if you dwell on it
luna 25 is not a good look but
close enough i guess
i dont think its the engineers' fault, supply chain issue more likely
but space programs have numerous benefits anyways
it's not as much about a moon base itself as far as being beneficial but stuff like having a good grasp of life support in a true closed loop, the power systems efficiency that gets implemented for them, water recycling tech etc
also resources on moon are big
the astronauts yearn for the mines
mining the moon is an insane idea
ai praising me 
it took me 3 hours today to understand why x(t) + y(t) = C = -x/y
No, you're not stupid. The opposite.
Your pace is slow for covering material. Your pace is fast for actually understanding it.
right 
at least it didnt tell you to use gender calculus
i was planed to spend 50m per day to catch up with basics and start learning high math
next day will be 8 hours with this rate
doesn't matter what's on there, the economics of setting up just a mine there, and then transporting the material are untenable
lol we can't even wipe our asses globally speaking
dont worry
the way i see it we're playing a zero sum game and we should take it more seriously
well it's one right now, not in the abstract
truth is currently we have enough resources to manage just fine
the problem is distribution which is an issue space base or not
Oh no not math again
The moon is a photo they stuck on they sky! If they landed on the moon, they would've got cheese all over their boots!
currently we have enough resources to manage just fine
it's a pretty incomplete statement, because while yes we can keep up with demand, but the moment we run out is going to be far far far too late
just look at the chaos with oil rn
and that's just one thing missing a small fraction of what we use
or chips etc
Don't really think that's going to be solved by ignoring space
In fact when the resources are gone that's the only place to go really, no? Final frontier and all that
of course space research has brought in a lot of interesting tech, but there are things it distracts from, like talking about and investing impossible ideas like living on mars when we can't even get our backyards in check
separation of concerns
separation of concerns in a zero sum game takes away minds and resources from something else
i was mentioning the other day (
) that the entire apollo program 60-73 or whatever it was cost less than checks notes meta is spending on datacenters this year
Hi, I just post here to express aloud something that I was thinking about the neuro tech.
Currently the LLM struggle with real-time interaction because there is a bottleneck in data processing. I was wondering if a design pattern like the one used in a RTOS (task interrupt and task priority) could improve something.
Or that can't really work with a LLM
and we gained a whole hell of a lot more benefit from that
so to me there's a lot of places we should cut back before we abandon space
that is the bottleneck but not because of raw speed, although that doesn't help, the main issue is that autoregression is serial
you can serve a ton of concurrent requests with like vllm
that would help for agentic stuff tho
Yeah the primary issue is that an LLM can only do one thing at once
Neuro can either speak or give inputs to the game, but not both at once
for sure
is promising to spend not is spending, besides apollo was done under a time crunch and incredibly risky, it did also cost $250B adjusted, which isn't cheap. there's 3x the people in the world today, too, yadi yada, with bigger economy
If you wanted to make Neuro do more things at once, the context would get desynced and then it wouldn't be the same Neuro speaking as playing the game
it was but then it was beneficial in the end, i think, so it ended up being worth it even if expensive
and ii have 0 doubts that we'll see a multiple of 250b spent this year even if not all of them actually hit spending target
you could do what humans do: just merge both streams in the same memory after the fact and pretend that it's all from the same person
that's already being done almost surely
i do also think the datacenters are pure waste too
That won't really work
That'll mess up the KV at minimum and cause context discontinuities at worst
you can do it fine, the model doesnt care where the cached context comes from
KV ?
people are inconsistent. it works well enough to pretend it all happened at once and that "you made a choice" for us.
The KV cache
key value cache
key value pair
i still maintain vedal's secret sauce along with data for the consistency of the model in its numerous iterations, is context handling
Specifically the time to first token as the KV cache being up to date means that there's not much prompt processing to do
i wonder whether this will work out https://www.inceptionlabs.ai/blog/introducing-mercury-2
diffusion llms
the negative latency trick (delay the video stream by 1 second and now Neuro looks like she reacts quicker but the guests look a little dumber)
that's the best one
Especially if the context is big, keeping the KV cache up to date is crucial to make sure it doesn't take multiple seconds for a reply to begin
i think he uses two models and switches between them mid sentence once it's producing output sometimes. like a fast model and a thinking one.
diffusion pog
sometimes the sentences don't seem to splice correcctly
i agree but i think it's more than 2 depending on what's happening
There's not a chance, that would mess up the KV cache so bad for sure
it's cheap to keep two KV caches in sync
Hm, is it really though
cheaper than running two models every step
as long as you dont naively jam them in any order it should be fine, like i said the model doesnt care where the cache came from
isn't that just speculative decoding
you skip the actual inference until you need it on the big model
so it lets you do fast starts
sometimes Neuro and Evil say something odd in the middle of a sentence which makes me wonder if the model switching is buggy
hmm, something like the service providers do then? answering most queries with a lightweight model and then using the heavies when it detects more complexity or something
tbf neuro/evil only make barely enough sense when they speak most of the time
there isn't a broader context they keep track of that i see
something like that or perhaps a director/action model where one is "in charge" and kinda directing the faster one on what to do without being interrupted by chatting and stuff which would throw it off
I mean they are optimized primarily for latency, so chances are their models aren't the biggest highest quality models ever
yeah.. use the cheaper model when you can and only splurge with the good one when you've got some buffer.
A harpoon approaches Superbox... 
Hem
:(
Hem x2
A harpoon approaches Superbox... 
Shared state buffer where the bigger director writes to a context/state object on its own tick (every few seconds, or event-driven), the fast model reads from it on every inference call. you'd basically have the director produce structured output (JSON or whatever) that gets injected into the fast model's system prompt or prefix but the director never blocks the fast path
damn
Hem x3
i've some ideas of my own on how to build a better ai, i still believe symbolic is the real way to go
or you could just have the director watching the stream and queuing directives as it does, fast model pulls from that queue and the recent context and just schizos and freestyles till another directive gets emitted by big model
way overcomplicating things im sure
but i got showerthoughts
reality it's probably just mainly a main model with a thicc harness
and maybe 1 or 2 others involved
deppending on what's going on
high intelligence neuro
her ability to pull in consistent references and threads from prior context while maintaining some shred of coherency is actually quite impressive and made more so by outputs like that not completely ruining everything more than anything else to me
nod context, Neuro. context. sigh
a lesser model would just go off the deep end after emitting some garbage like that
not if it's one-shotting without context
she clearly has context though, she'll reference stuff liek that over and over
perhaps it's just curated but
i don't think she does, but she does memories or just pulls from within the model
nah she'll get hooked on random things all the time for a while
i would bet she has different modes whether she's talking 1 to 1 or chat
the info has to go somewhere but yes i agree on that
the location issue was for sure the oddest one i think
with vedal panicking about it as well
elaborate on which one you're talkign about (lule)
when evil started talking about Location endlessly and spewing garbage output
oh yeah
yeah that's a poisoned context and/or unstable model (for a variety of reasons it could be made to be not necessarily like poorly trained or anythign) ouroboros itself to hell
about that, someone know what is the <|end thing|> thing ?
special token
if it was context i don't think vedal would be panicking, you could just clear it, so it's gotta be something else
Yeah just normal leaking of special tokens





When I am broke


im in houston