#programming

1 messages · Page 303 of 1

sage crag
#

im your new favourite deranged tv head

rigid snow
#

everchanging television set

sage crag
tender river
olive sable
#

character arc are good

sage crag
#

cold green eyes

tender river
#

ye warning

#

be afraid

sage crag
#

neuroCry dementia

tender river
#

fail mirror test neuroWAJAJA

#

f grade

sage crag
#

demonstrating lack of awareness

olive sable
#

lobotomy

sage crag
#

and reason why llm is stupid

#

🔪 🔪 🔪

#

this thing give me headache

#

its slightly terrifying actually neuroLookDown

clear sedge
#

i've never once thought about fries dipped in ice cream

#

that sound horrendous

#

i'm sticking to ketchup and mayo thanks

olive sable
#

NeuroBounce mayo

#

i wonder if these are good

sage crag
#

limited to usb 3.2 speeds

#

meaning potentially limited to 5 gb/s

#

probably less though

opaque sigil
#

that one is 10Gbps, though i'd be worried about latency more than anything

#

who cares about peak sequential speed

sage crag
olive sable
#

its not for a boot ssd or anything so i feel like it should be mostly fine

sage crag
#

i got a sata to usb adapter recently

opaque sigil
#

since you have to have an extra board to do the pcie to usb conversion

sage crag
opaque sigil
#

yeah it's probably fine for standalone drives

olive sable
#

i need to choose a christams present under 25 bucks

#

and most of the things i want are several hundreds enub

sage crag
#

dont get anything then

olive sable
#

its for a christmass party thing

#

so i need to choose

#

its secret santa but my brother got my name so im choosing what i want

sage crag
olive sable
#

the projector i got 2ndhand for 650 is on sale from 1.6K to 915 AINTNOWAY

rigid snow
#

not so secret huh

olive sable
#

not for me ye

amber fractal
#

I got mine for $3 thrifting tho

#

scratch that, free because the store owner didn't know what it was

#

Fun fact for those that don't know. Sometimes when Amazon gets their refunds, the goods don't return to the company they originated from. It'd be too expensive to do so, instead a bulk of items are boxed up and sent in for auction. People (companies) bid for these boxes so that they can resell the contents of them.

trim valve
olive sable
trim valve
#

or well two

#

they're very good

olive sable
#

yay

#

cuz i kinda laready ordered it enub

#

i could have canceled it but eh

trim valve
#

I personally dislike that one's design but meh

olive sable
#

the metalic?

trim valve
#

nah, the fact it slides together

#

I use this one which clamps

#

in practice it makes little difference though unless you're hitting it hard for long periods of time

olive sable
#

oh to take it out?

#

im not planning on taking it out

trim valve
#

its more that you probably get better contact pressure with this design

#

but like

#

if you're only using it as a big usb drive its whatever

olive sable
#

for the heat-sink? Shruge

trim valve
#

yeah

olive sable
#

oh

#

okay okp

#

i doub it will be too much of an issue, ill jsut eventually buy 2 new ssd's for my pc and plop the old one in there

trim valve
#

yeah fair

olive sable
#

some p310 drive

#

ye, Crucial P310, that thing

#

thats what my windows install is using while my linux is on soe old 200gb 2.5" ssd

#

the plan is to put 2 990pros in here and move that one to the thingy

trim valve
#

I run this silly setup

#

samsung is my windows, kioxia is my nixos

olive sable
#

the 999gb is the windows, the f"ile system" is 200gb for linux, that should be one of the kingstons so idk what happened there. then theres also a 500gb hard drive with back-ups of old data.

#

the 2nd kingston is empty cuz idk

trim valve
#

idk how the fuck I write more than I read

olive sable
olive sable
trim valve
#

my only guess is like, immense disk caching because I have "so much" ram

upbeat kraken
#

Has anyone been noticing that neuro has been faulting more and more on youtube videos? I don't really watch stream so I'm not sure what's happening there. I think she might be hitting her thermodynamic limit here. Every random vector ingested is a bit of entropy that she needs to fight against. I'm afraid that without intervention, neuro might just simply dissipate into the void. To ensure system stability, we can try controlling the randomness of the vectors before we can find a cure for the actual cause.

opaque sigil
#

is there no way to see which process writes/reads

opaque sigil
rigid snow
#

i am ingesting random vectors

opaque sigil
#

that doesn't sound very healthy but what do i know

olive sable
rigid snow
#

i think it joke

olive sable
upbeat kraken
#

No, I'm serious. I have a theory that predicts this very situation.

olive sable
#

oh

#

a theory

rigid snow
olive sable
#

great

#

so like, do you know how llms work, or did this come to you in a dream?

upbeat kraken
#

Yes. I do know how llms work

#

My specialization is in AI.

#

For neuro to be able to continue existing there has to be a thermodynamic sink for her to dump noise into

rigid snow
#

neuro might just simply dissipate into the void
the fuck does this mean

olive sable
#

uh

clear sedge
#

that's a lot of words

#

certainly some words

upbeat kraken
clear sedge
olive sable
#

My specialization is in AI. -> random thermodynamics that i dont think make sense

rigid snow
#

and context is finite

upbeat kraken
#

Yes but she is training.

olive sable
#

no shes not

upbeat kraken
#

She is recieving inputs

rigid snow
#

ye you don't know how llms work

olive sable
#

inputs -x-> trainign

upbeat kraken
#

Her weights are being updated.

#

And chats are noise.

olive sable
#

is this the bullshit about neuro learning form twitch chat again?

upbeat kraken
#

There's a big roundabout circle going on here

clear sedge
rigid snow
clear sedge
#

and thermodynamics comes into play how

olive sable
upbeat kraken
#

As long as she is learning there will be an issue with faulting that can only be delayed with more firmware.

#

Ok. Thermodynamics of information. Have you heard of information entropy?

rigid snow
rigid snow
#

product of such process is a different version

#

different set of weights

#

that has to be shipped

#

manually also

clear sedge
rigid snow
#

weights don't update by themselves

#

that doesn't work with llms

#

people tried

#

they do kinda descend into noise as you described

upbeat kraken
#

Imma write a paper which explains what's going on in depth. hold on to your hats for 2 mins. It does not actually matter what architecture she is using.

rigid snow
clear sedge
#

i may be on fedora but i don't have hats

upbeat kraken
#

If she is learning and remembering things from the past then she is subject to the problem

rigid snow
#

it's not in the weights tho

#

it's in context

olive sable
#

i think you think neuro is a lot more advanced than what she probably is. she isnt learning. its only some memories and such

#

as far as we know and can tell

rigid snow
#

vedal mentioned her "saving" stuff to her memory and "deleting"

olive sable
#

aslo, a paper written in 2 minutes? are we sure this guy is not mister grand gpt himself?

clear sedge
#

agi is here

#

they're the agi

olive sable
rigid snow
#

first one would be what we call rag

upbeat kraken
#

No it's just that I'm breaking my theory into more easily digestible bits for you.

rigid snow
upbeat kraken
#

One second. antigravity is compiling the thing.

rigid snow
#

sam you called it

olive sable
#

ok this guy is a troll neurOMEGALUL

rigid snow
#

antigravity is google's cursor

olive sable
#

mods remove his 206th bone

rigid snow
#

ai vscode fork

#

so you were right

upbeat kraken
#

No, not trolling

rigid snow
#

not mr chat gpt though

#

mr gemin'i

upbeat kraken
#

Like actually coherent paper.

clear sedge
#

hold on let him cook

#

i wanna see the forest fire that comes out of it

rigid snow
#

the forest didn't burn down, it dried out because of all the water bro is using to generate the slop

olive sable
#

lemme get some drinks and popcorn for this

#

im back

upbeat kraken
#

How to attach file?

olive sable
#

oh you dont have embed perms

#

dm it to me

upbeat kraken
#

Well shoot

olive sable
#

i gotta see this

opaque sigil
clear sedge
upbeat kraken
#

One sec getting nitro

olive sable
#

why do you need nitro? neurOMEGALUL

#

compress it to a zip file if its too large

rigid snow
#

||pdf is 300mb||

clear sedge
#

getting nitro on a whim to upload

#

does bro know about pastebin

amber fractal
upbeat kraken
#

I actually don't.

rigid snow
#

if you need to upload a file i don't see how pastebin would help

opaque sigil
#

paste the bytes mhm

clear sedge
#

base64 encode it bozo

upbeat kraken
#

you know what? that reminded me, I can just use google drive. Thanks.

clear sedge
#

boring solution

upbeat kraken
#

Ok first draft.

#

Gotta eat something afterwards.

clear sedge
#

YAP VELOCITY

olive sable
upbeat kraken
#

Will be doing second draft later.

opaque sigil
#

i didn't know google drive had a new pdf viewer huh

#

neato

rigid snow
#

what the FUCK am i reading

#

holy slop

olive sable
#

this is hilarious

#

im so glad i got cookies for this

clear sedge
#

bro realised neuro is not trained to act exactly like chatgpt

rigid snow
#

ok i get what they're talking about kinda

#

that did not need a "paper"

olive sable
#

To maintain conversation flow, it begins to generate ”Synthetic Memories.”
so does every single LLM on the planet

#

its quite literally how they work

clear sedge
#

do not fear the Scuff™

rigid snow
#

there isn't an llm that doesn't hallucinate ye

#

some do more some less

#

but they all do

#

also this like

#

doesn't happen

olive sable
#

so, how do i calculate my firmware strenghts?

#

in what units is that?

opaque sigil
#

that was a nice read, i appreciate it

olive sable
#

i yap velocity in words per second?

clear sedge
#

since everything else is word soup i might as well throw my own in there

olive sable
#

this HAS to be like a konii throwaway account or soemthing

clear sedge
#

@sage crag are you alive

#

come to us

rigid snow
#

i am having a really hard time deciding if this is a joke

olive sable
#

no their account is 5 and a half years dated already, it cant be a new account

clear sedge
#

this feels like someone treating an llm as an authority

rigid snow
#

not deciding

olive sable
#

im actually begining to think if alcohol may be involved

rigid snow
#

discerning

#

idk

#

i don't have the word

clear sedge
olive sable
#

actually let me feed this paper to gemini

rigid snow
#

back into the slop machine it goes

#

right where it came from

rigid snow
olive sable
#

ngl, i feel mean for laughing at this. but there really isnt anything else i can respond with

clear sedge
#

i just find this amusing

olive sable
#

hmm gemini is a believer

clear sedge
#

we're so cooked

clear sedge
#

agi is gonna be so naive

rigid snow
#

i mean

#

there is some sense there

olive sable
#

the problem is that its describing real ai probelms, with a system wehre they dont ocure

rigid snow
#

the tiniest bit

olive sable
#

yes finite context is a problem, but thats handled by ved

opaque sigil
#

just add more context duh

olive sable
#

hallucination is a real thing, but its not a syptome

opaque sigil
#

how hard could it be

rigid snow
#

no i just started reading the parts i skipped and it doesn't make any sense anymore

olive sable
#

im not an ai guy so i could believe he does ai research maybe, but this jsut doesnt apply to neuro

clear sedge
#

an llm talking about an llm feeding off of an llm (?)

olive sable
#

some made up math equation that is somewhat grounded in reality isnt gonna do shit

#

math equations should only get invented when you can actually do the math

#

not when there is a corelation

rigid snow
#

making up equations to explain stuff is normal

#

too bad it makes zero sense

#

wtf is firmware strength

olive sable
#

yap velocity doesnt exist

clear sedge
#

yap velocity is fucking hilarious

olive sable
#

you cant stick a number to yap velocity

#

what would that mean? tokens?

rigid snow
#

tps maybe

#

which is just compute/load

#

load should increase with filled context

#

i give up this is nonsense

olive sable
#

"the best lies are based in reality" or soemthing idk

#

i should ask shadow what he thinks of the paper

rigid snow
#

please do

#

please

#

i really hope we're just all really really stupid that would be so funny

olive sable
olive sable
amber fractal
#

slop reading time shad_yippee

olive sable
#

certain parts i dont think apply to neuro, and other parts are incoherent cuz of the gen alpha word usage

clear sedge
#

i think it's just incoherent

#

it's the thesis of slop

#

it looks fine at a glance then you start to pick apart at it

#

and it just crumbles

amber fractal
#

yap velocity is literally just filling up the context window neurOMEGALUL

#

time to figure out what the firmware strength is

rigid snow
#

is my backend less strong because there's Rust in it

olive sable
#

rust generally degrades the strenght of the metal

clear sedge
amber fractal
#

firmware strength is only mentioned in the equation so I'm going to replace it with something that might make sense

#

that is just the context limit xdx

olive sable
#

i interpreted firmware strength as the tolerance to context window being filled or something

#

idk

amber fractal
#

Yeah, I can see that

olive sable
#

context limit makes sense tho

rigid snow
amber fractal
#

I like how I could probably list out why each of these "signs" happen in much greater detail than this paper

rigid snow
#

now replace time to crash with time to fill context

#

now it does make sense

amber fractal
#

correct

#

So a whole lot of nothing

olive sable
#

i feel like the condesending tone of "lets use gen alpha slang cuz these guys dont know as much as me" is really detrimental to the paper as a whole

amber fractal
#

100%

#

I may know balls about LLMs but at least I can tell this is bad

rigid snow
#

literally a solved problem

olive sable
#

thats what i thought ye

#

neuro shouldnt have that as an issue

#

it doesnt apply

rigid snow
#

do they think neuro has a 10 trillion window

#

and are afraid it's filling up

amber fractal
#

I'd think so at this point

olive sable
#

neuro ofcourse is the reason why the ram prices are so high

#

vedal.ai is buying petabytes of ram for that massive context window

clear sedge
amber fractal
#

Just for fun, let's go over each of the markers neuroHypers

olive sable
#

vedal i wanted to buy a server pc GETHIM

#

damn you and your ram

rigid snow
#

nvm i forgot they said this just disregard everything they're just clueless

olive sable
amber fractal
#

I did actually notice that message

#

which is why I was deadset on them being glueless

clear sedge
#

guys i'm an expert on ai and i think neuro will take over the world and force us to do manual labor to give her all the water

olive sable
#

im assuming that s an "ai researcher" their ai does do this, and neuro doesnt

pliant dome
#

completely without glue

olive sable
#

or they're jsut clueless

olive sable
#

same thing for them but ai

amber fractal
olive sable
#

guys, chair has the word AI in it.
what if we did chAIr?

#

a chair, but it has moving wheels, and gemini

#

"hey google, bring my chair"
chair proceeds to fall downstairs

rigid snow
olive sable
rigid snow
#

they were trying to figure out how they actually do math and stuff

#

answer: in a very non-traditional way

amber fractal
# amber fractal `Marker 1: Filter Erosion (The ”Brat” Phase)` Isn't she just always a brat <:neu...

Marker 2: Hallucinatory Compression (The ”Gaslight” Phase)
This is also just a personanity trait our cute overlord has, I don't see a downside here.
Can't remember from the last hour, yeah context limit exists and heavy context windows take more time to process enub, so if it is actually an hour we'd be moving into the memory system. Oh, "Synthetic Memories.” I sure wonder what this is referring to.

rigid snow
#

we got filtered because it's 1. more simple 2. more fun

clear sedge
#

i don't really care about neuro i just find the asspulls fascinating

amber fractal
#

Marker 3: The Semantic Loop (The ”Filtered” Phase)
[FILTERED]
This is the one that might actually have grounds, but also the filter is handing these more these days. instead of backslashbackslashbackslash. hmm Speaking in nonsensical syntax (Word Salad). Total unresponsiveness to external prompts. Diagnosis: System Failure. The AI has optimized itself into a corner. A hard reset is required.
I think the author of this is in Marker 3 what do you guys think? shrug

rigid snow
#

first of all this doesn't really happen

#

second, what is even the point they're making? that context filling up contributes to this?

amber fractal
#

Author in chat enub

upbeat kraken
#

To be perfectly honest with you, I don't actually watch the streams. The point of this excercise is to try to predict a crash

#

I'll give you guys a model and then you can roast me if it was right or wrong.

clear sedge
#

checks out

rigid snow
#

crashes aren't on the actual ml architechure side

upbeat kraken
#

Yes

#

exactly what I was saying

#

It's not the architecture that's wrong

rigid snow
#

why would they be caused by the state of the model

upbeat kraken
#

It's something fundamentally wrong with the whole neuro-audience-vedal model

#

There's no thermodynamic outlet other than Vedal debugging neuro and adding more firmware

olive sable
#

uh

clear sedge
upbeat kraken
#

and the rate that vedal can keep up is limited

olive sable
#

wwe're back to thermodynamics

rigid snow
#

thing is there is always more human input in neuro's context

amber fractal
#

we are so back

rigid snow
upbeat kraken
#

Ok, second paper is almost ready

amber fractal
rigid snow
#

i might be going insane though if i read the second paper

clear sedge
#

i'm not reading shit i'm gonna go to bed

#

you guys have fun with this one

olive sable
#

you know that thermodynamics dont apply to anything digital right?
conservation of energy isnt needed when you can make shit up

upbeat kraken
amber fractal
#

surely this won't dox me

upbeat kraken
#

I will invite you guys to fill in the constants for me.

rigid snow
#

brooooo neurOMEGALUL

upbeat kraken
#

Hey if it fails at least its good joke material.

amber fractal
#

I'll let you lead this time @rigid snow

upbeat kraken
#

But I stand by my point.

rigid snow
#

ye i'm not gonna

upbeat kraken
#

The randomness in the model has to go somewhere

olive sable
#

no it doesnt

amber fractal
#

valid

upbeat kraken
#

It doesn't just magically disaapear.

swift shoal
#

so guys I have xbox 360 and its model E with bad update

rigid snow
olive sable
#

from my understanding you don't remove randomness by moving it to somewhere else

swift shoal
#

how the fuck make fucking xbox original emulator work

dusky jackal
upbeat kraken
#

The current vedal architecture is actually somewhat semblant to an AGI

amber fractal
#

We provide the map; you must walk the territory. If this theory holds, Neuro is not merely ”scuffed”—she is a chaotic dynamical system seeking equilibrium. This would be so hard if it wasn't in this fricking paper

upbeat kraken
#

if you stop and think about it for a min

swift shoal
#

i just want to play some max payne 1

upbeat kraken
#

The. whole chat, vedal and neuro is somwhat like a slow moving AGI

rigid snow
olive sable
upbeat kraken
#

It's just that it's running on biological firmware.

rigid snow
#

(modern consoles)

#

ps2 is same, wii is the same too i think

rigid snow
swift shoal
amber fractal
#

I'm glad to be able to read fast so I don't have to read this paper for long

olive sable
#

im glad its only 2 pages

amber fractal
#

anyone want a NDI(5) steam to their PC?

rigid snow
warped narwhal
#

What in the SCHIZO did I walk into?

olive sable
#

hi

#

its a grand old time

rigid snow
#

neuroWaveA ;

warped narwhal
rigid snow
#

noooooo

amber fractal
olive sable
#

we are discussing ai generated paper by the man, the myth, the legemd, muipiba.
a guy who has made more papers on neuro than he has watched streams

upbeat kraken
#

I guess true?

#

Just want to verify somethings experimentally.

olive sable
#

dont admit to it lmao

swift shoal
rigid snow
olive sable
warped narwhal
olive sable
amber fractal
olive sable
#

this doesnt count

rigid snow
#

need to develop one first neuroSadge

olive sable
#

we are people here

#

most of us at least

#

some are televisions

warped narwhal
#

We are?

olive sable
#

you're a semicolon

rigid snow
#

some are printable characters

warped narwhal
#

I'm a punctuation mark idk about you guys

amber fractal
upbeat kraken
#

The point is that the collapse of neuro can (hopefully) be prediced through chat.

olive sable
#

where is full colon?

#

oh

warped narwhal
#

Dead

amber fractal
#

o7

olive sable
upbeat kraken
#

If we do indeed verify this, then we can potentially unify information and physics thermodynamics.

rigid snow
#

and isn't going to be

warped narwhal
#

Full colon didn't survive the keyboard wars, and was caught in the crossfire in the battle of ANSI v iso

upbeat kraken
#

Isn't there the whole someone tell vedal there's something wrong with my model or smth

rigid snow
#

serious: is this what ai psychosis looks like

upbeat kraken
#

When I'm talking about a collapse, this is what I was talking about.

#

It's literally just a hypothesis!

#

Not cyberpsychosis.

rigid snow
#

no ai psychosis is what a human experiences

upbeat kraken
#

Nope, definitely not

amber fractal
#

not into it enough to know, maybe shadow might know evilShrug

olive sable
rigid snow
upbeat kraken
#

Yes. The thing is that I'm saying that information can hold entropy just like physical particals can.

rigid snow
#

holy fuck SCHIZO

warped narwhal
#

AINTNEURWAY wtf is this convo lmao

rigid snow
olive sable
rigid snow
#

it can tho

#

the thing with information holding entropy is entirely correct it's just the thermodynamics part

#

that's

#

questionable let's call it

olive sable
#

ok sure a random bitflip can make for some entropy

warped narwhal
opaque sigil
#

Information theory entropy is different from physics entropy

rigid snow
#

ye

olive sable
#

well were we not talking about physics?

amber fractal
#

We seem to be on the same page, except for the source of where this came from

olive sable
#

im so confused rn

rigid snow
#

they are talking about information entropy

#

they're just trying to project physics concepts onto it

olive sable
#

is information entropy unpredictableness in the data?

rigid snow
#

kinda yeah

upbeat kraken
#

yes

#

the hypothesis is that entropy can propogate in AI systems the same way thermodynamic systems can.

#

There's an easy test that might be actionable with low cost.

#

When neuro reaches stage 3, get chat to stop spamming things.

olive sable
#

impossible

upbeat kraken
#

and allow neuro to "cool off"

olive sable
#

you're asking us to move a mountain here

rigid snow
#

oh i didn’t even think of it this way
you know how they say llms aren’t deterministic? they are literally fed noise constantly. entropy is constantly injected

upbeat kraken
#

Yes. Finally

#

that's exactly what I'm saying.

rigid snow
#

??? this is not what you’re saying

upbeat kraken
#

Ok, there's a part 3 to the whole thing.

olive sable
#

i thought what we were discussing was moving entropy out of the llm?

#

i know that it gets fed nosie

amber fractal
#

-# The first paper claimed the noise was chat itself, instead idk random value nub

rigid snow
olive sable
#

but thats by design

rigid snow
upbeat kraken
#

with enough entropy, pockets of structure will emerge, which is why its so prohibitively expensive to train llms.

rigid snow
#

that’s not

#

how that works

#

i’m going insane

upbeat kraken
#

might be a little too far ahead, but let's focus on the central topic. In my second paper that I whisked up in 20 seconds

olive sable
#

chat im calling it a night here.
ive been confussed, deluded, and mogged on by bringing physics to an information fight

warped narwhal
upbeat kraken
#

we might be able to predict when neuro needs maintenence.

olive sable
#

he just presses the reset button

amber fractal
opaque sigil
#

pkill -f -9 neuro

warped narwhal
upbeat kraken
rigid snow
olive sable
#

the context for that was memories tho

rigid snow
#

(resetting context specifically)

olive sable
#

i mean resseting neuro like how he "spawns a new clone"

rigid snow
#

that’s just the unity frontend he restarts no

olive sable
#

oh

#

i see

#

i wouldnt know

rigid snow
glad path
#

,vedalStonks

fast pagoda
#

s system stuff

#

he had to specifically set it up to save context and restore it tho so it must at least wipe context by default

#

but only like recent context obv

#

not memories

rigid snow
#

is the context stored on the frontend

#

vedal why

fast pagoda
#

i would imagine it's easier to deal with the whole multiple contexts and swapping them around when done this way for some reason or another (assuming that's what happens which is just my guesserino still i think)

#

chat/game/voice interactions being blasted into context and determining what to give her + when based on current state, intent etc probably easier to do if handling all of that in the frontend than elsewhere

#

latency vedalBonk

rigid snow
#

yeah sure i can see it

#

being reasonable

fast pagoda
#

i miss neuro reset resulting in just

hello nwero

#

after a ton of shit happened prior

rigid snow
#

he can kinda bring that back just inject “you have been restarted” in the context

#

or similar

fast pagoda
#

im guessing she is given that somewhere somehow probably optionally

#

part of her seeming so coherently like she has a continuous train of thought and intent to her actions is being given info about herself including stuff like that i think as a sort of brute force way of giving her seemingly ongoing goals and whatnot even tho she would just forget them once they fall out of context normally without it

rigid snow
#

i wonder if she has sticky notes

#

like coding agents have todos

fast pagoda
#

she can choose to remember things

#

he has said

rigid snow
#

yeah right but those aren’t pinned to context

#

we discussed this earlier

fast pagoda
#

probably dont need to be if the way the context is being managed is robust enough

rigid snow
#

i assumed they are rag’d in

fast pagoda
#

theyll jsut get blasted into it w/ rag like you just said yeah

#

something of the sort

rigid snow
#

just in general i’m surprised it works at all after coming to a realization how poorly llms handle large quantities of tools available

fast pagoda
#

that's the secret sauce really is how well it's managed

rigid snow
#

fuckin discord dms soundboard

fast pagoda
#

and seamlessly

rigid snow
#

somehow i don’t believe the 7b consensus

#

like no way

fast pagoda
#

i could see using another (smaller) model to provide a grounding effect where it's much more railroaded and controlled by the system and is just there to constantly yap like it's an inner monologue + handle recall and giving the correct tools but ONLY the correct tools

#

i dont agree w/ 7b, i think she's 20-30b

rigid snow
#

yeah that’s way more believable

fast pagoda
#

she seems at least as coherent as like a mistral small or a 30b of your choosing

#

a lot of it is just being railroaded with tight context management i think tho, she doesnt have to deal with 128k+ context lengths if she's being given more of a summary and suggestion style context from the system itself whether that be another model or more programmatic injection of current state + previous actions or w/e she may need at the time

fast pagoda
#

i mean gemini 2.5 pro was dogshit at tool use

rigid snow
#

claude put out tool search recently

#

fucking CLAUDE

#

the BEST model at tool use

fast pagoda
#

claude is surprising if it's bad at that, it's what i would consider the best at agentic anything really

#

or at least the most consistent

rigid snow
#

throw like 5 mcp servers in and it all falls apart

#

that’s what they’re fixing

#

not even 5, 5 is a lot already

fast pagoda
#

interesting, i've given agents like 10+ mcp before and usually they were kinda fine at it

#

ig it depends on if you poop out like 1382791823791827398 tokens of instructions in the system prompt or more of a shortlist

rigid snow
#

also tool definitions pollute the context like crazy

#

they are very token inefficient

#

oh and they added examples now

fast pagoda
#

yeah that's what i was just saying i think if you avoid doing that but only say like hey here's a tool, let them infer the usage of that tool literally from the name, and then if it wants to use said tool, then give more info, probably more effective than one of those 40k token system prompts

rigid snow
#

what if

fast pagoda
rigid snow
#

i have used it extensively

#

never looked at context directly

#

should i

fast pagoda
#

yes

#

go click the custom system prompt or edit button i havent looked in a long time

#

but it'll show you the full prompt

rigid snow
#

i don't see the full system prompt anywhere

fast pagoda
#

it's at least 40k characters

#

id have to install it again to see it

#

one sec

#

the default for the one i just installed

#

fuckin have a whole cloud thing now wtf

#

anyways

#

this is plan mode i think

#

and this is with nothing else enabled or w/e, it adds to it if you start enabling shit

rigid snow
#

You are STRICTLY FORBIDDEN from starting your messages with "Great", "Certainly", "Okay", "Sure".

fast pagoda
#

kekw

#

they still ignore the fuck out of that i'm sure

#

here's default barebones code mode

#

i almost shit my pants when i saw this said my system dir and such but i forgor i never put my name into windows (i'm on my laptop which is my one(1) windows machine right now lule)

rigid snow
fast pagoda
#

it changes the beginning too

rigid snow
#

ye

fast pagoda
#

basically theyre similar sizes and if you have a bunch of MCP installed i think it will include definitions and stuff

#

isn't claude's default system prompt like 40k tokens or something

#

im pretty sure it's long as hell

#

16,739 words long and 110 KB in size

#

kek

rigid snow
#

i remembered where to find a crazy system prompt

fast pagoda
#

pliny the liberator has dummy long ones

rigid snow
fast pagoda
#

i guess cuz it's really multiple jammed together so it can be dynamic

rigid snow
#

one shot gets you most of the way there

#

with the tool this is from

#

so idk

fast pagoda
#

jailbreak prompts are generally s00per long because the models get confounded on extra long context

#

yeah i mean the long ass prompt works at the start anyways because usually depending on the model nowadays theyre fine well past 100k context

#

i havent used gemini in an ide or anything yet, i wonder how it does with tools compared to 2.5, 2.5 sucked major eggs on any and all tool use

#

gemini 3 that is

rigid snow
# rigid snow

this is for convex's vibecoding thing btw, so it makes sense it's this long, they have to explain how convex even works

fast pagoda
#

yeah i looked it up, seemed interesting

#

i wonder what the prompt for like stitch looks like (google's figma generator thing)

#

i havent really used any of the turbo vibecode tools, i probably should at least once for funzies just to see exactly where they stand now, i always assumed theyd just be frustrating. Never even tried like replit which has been around forever

#

by that i mean the ones that it's liek ok you provide one prompt and we will ensloppify the entire thing end to end

rigid snow
#

i only used for demo purposes yeah

fast pagoda
#

even like roo ended up just being TOO independent and wanting to do way too much after the initial "whoa" fell off

rigid snow
fast pagoda
#

interesting

rigid snow
#

out of the super vibecodey ones

#

like replit lovable etc

fast pagoda
#

how does it compare to like firebase studio, i have used that i guess which i suppose is the one i have actually tried

rigid snow
#

no idea never tried it

#

idea seems to be similar tho

#

between the two

fast pagoda
#

yeah it does looking more at chef

rigid snow
#

they have something like tgatr themselves

fast pagoda
#

oh cool this one will export to ai studio now too

#

not chef but stitch

#

if you generate one you can just export it and then it'll fill this in lamo

rigid snow
fast pagoda
#

i asked stitch to poop out like a gallery for a family photo website

rigid snow
#

surprisingly ok

rigid snow
fast pagoda
#

it reminds me of photobucket

#

which isnt a good thing lule

#

you can have it redo or modify at will

rigid snow
#

idk what i expected

#

AINTNEURWAY ok dude

fast pagoda
#

yeah i sent it to ai studio and it's going berserk rn (im braindead ad used my real name here so i blocked it lul)

rigid snow
#

THE FUCKING GRADIENT

#

KILL ME

fast pagoda
#

oh yeah they LOVE gradients in every icon

rigid snow
#

it's THE gradient

fast pagoda
#

and every background

#

what specific gradient we talking here

#

i am unfamiliar

rigid snow
#

claude ui slop

#

it puts it everywhere

fast pagoda
#

this is indeed functional (preview) idk about the choice to have the month/year overlap when you scroll tho

#

your fav screen ofc

rigid snow
#

not same gradient but you get the idea

fast pagoda
#

the model just has pride

glad path
fast pagoda
#

it's secure in itself

#

that was the world's fastest reply btw

#

i know what you're referring to and yes that is what i was referring to as well (im wearing a pride shirt rn actually which is why i thought about it lmao)

fast pagoda
rigid snow
#

people who have a clue on how uis should work finetuned it i assume

fast pagoda
#

what was it tuned on before kekw

#

where does it come from i wonder

rigid snow
#

or rather look in this case not work

fast pagoda
#

is it just overtrained on tailwind

#

now im self conscious about literally every gradient ive ever used in anything

rigid snow
#

hopefully you're not making anything as bad as the models are

#

so should be fine

fast pagoda
#

nah, worse

#

graphic design is my passion

rigid snow
#

just a tasteless assumption of what's pretty to an average joe

fast pagoda
#

i dunno if i want to know what a whole entire ui hallucinated up would look like, although i have noticed if i give up on something and let them yolo something i will end up with a giant albatross of a design choice that ruins my life for decades to come

rigid snow
#

honestly just ask it to use shadcn/ui

#

customize components to taste later

fast pagoda
#

mmm they provide non-framework components nice

#

i usually rawdog html//css and hate myself for it

rigid snow
fast pagoda
#

they have always jumped directly to like tailwind if you leave it up to them i think for me

#

and then they are fucking terrible at tailwind because they cant distinguish tailwind 4 vs 3 conventions like at all

ruby siren
#

I like the BEM method for CSS but I'm a bit old school

#

Tailwind I just end up using prebuilt stuff in daisyui for personal projects

fast pagoda
#

i cant believe how not angry linus was on ltt

#

nvidia is the reason my laptop is running windows atm because of some weird ass nvidia bullshit that made me mad enough that i wiped the thing and put windows back on it just so i didnt have to fight with it every single time i used the goddamn thing

nocturne olive
young plover
#

This is just gonna run when I unlock my computer

fast pagoda
#

lact cryin in the corner

nocturne olive
#

Lact is kinda scuffed

#

Just connecting it to the GPU makes it run at a way lower clock speed immediately

young plover
# young plover

glue Just realize I could make that a lot simpler by using nvidia-smi --query-gpu uuid,name,power.limit --format=nounits,noheader instead of parsing a bunch of junk

jagged turtle
#

I guess I woke up to one of the most interesting announcements in the AI coding world today

(Linux Foundation through Agentic AI Foundation now manages MCP and AGENTS.md)

fast pagoda
jagged turtle
#

That being said I wonder if an AI (or forgetful human) wrote Anthropic's blog post considering the Linux Foundation hyperlink doesn't actually link anywhere and so clicking on it is useless

fast pagoda
#

claude would never

jagged turtle
#

oh I didn't know claude for excel was a thing

#

vibecoding my statistics AINTNEURWAY

fast pagoda
#

last time i tried to get gemini in google sheets to do anything it was turbo bad

#

but this was back i think before 2.5 pro even

nocturne olive
fast pagoda
#

anthropic decided they didnt want MCP anymore now that A2A and ACP are more around

fast pagoda
nocturne olive
#

Well I don't know how I would make it not do that

jagged turtle
#

oh is a2a just a protocol for yapping between agents or smth

fast pagoda
#

shrug iono

nocturne olive
#

GPU configuration tools on Linux are so scuffed

fast pagoda
#

they really are

#

it's weird because i always kinda figured with everything else being the way it is, before i switched, that linux surely should have better tools for that than windows

#

but not even close

nocturne olive
#

It's weird indeed

#

The one thing Linux somehow doesn't do as well

fast pagoda
#

just lack of community investment and need/want i guess

jagged turtle
fast pagoda
#

i was meming

#

but yeah that is right

#

you can kinda do the same shit with any of them but the other 2 are much more focused on agentic silliness

jagged turtle
#

I mean honestly if all of those were donated to the AAIF (and maybe merging A2A and ACP) it would be better

there should be a open standard to connect to multiple models cross-platform that stays as vendor-neutral as possible as well

fast pagoda
#

well

#

a2a is google

#

and acp is IBM i think

#

so good luck w/ that

jagged turtle
fast pagoda
#

lol

#

well i stand corrected

#

so they all are under linux foundation now

#

that's crazy

#

new dark horse in the AI sphere appears: fucking linux foundation of all things

#

when linux IPO

jagged turtle
#

Yeah, albeit I think right now the Linux Foundation directly manages ACP/A2A as opposed to throwing it into AAIF

#

mainly because AAIF was literally just formed

nocturne olive
fast pagoda
#

how the tf is it up to the vendors

#

they dont even control their drivers fully

nocturne olive
#

I don't know, ask quuck

fast pagoda
#

i will consult the duck

jagged turtle
#

I feel like ACP being used for one-way model delegation and A2A being two-way model communication aren't big distinguishing factors

#

so the fact they merged isn't surprising to me

fast pagoda
#

yeah ive never been able to truly figure out why they had to be separate things

#

that's why i said theyre kinda the same shit

jagged turtle
#

what's more surprising is the Linux Foundation actually taking an interesting in AI stuff

fast pagoda
#

makes sense with the prevalence of linux in training and just ML in general

#

wonder what linus' feelings on this are lol

jagged turtle
#

true

jagged turtle
fast pagoda
#

i just wanna see someone ask him why he personally now controls MCP/ACP/A2A and see how he wishes death upon them

jagged turtle
#

in the recent github interview though he did say to the interviewer "are you going to say the AI word" which makes me think he's kinda tired of this bs

fast pagoda
#

he seems to keep off pretty much all socials

#

except email

#

according to the ltt video lol

glad path
#

when is asahi linux gonna become a thing for m3

#

wheen

jagged turtle
glad path
#

im impatient

nocturne olive
fast pagoda
glad path
fast pagoda
#

yep

glad path
#

damn it

jagged turtle
glad path
#

if its not a thing by sep. 2027 im screwed

#

they're gonna stop rollign security patches to sonoma next year

#

and sequoia goes next in 2027

#

and i fucking hate tahoe

fast pagoda
#

didnt asahi's main dev quit as fuck

glad path
fast pagoda
#

welp

glad path
#

do you think two years is enough

jagged turtle
#

something else that I learned is that github does reaction videos to their own content

fast pagoda
glad path
#

its sillicon

fast pagoda
glad path
#

asahi is the only other os that can run on apple silicon other than macos

fast pagoda
#

yeah but it's apple's sillycon

#

apple involved = rip open anything working with any support from apple and often at all

#

the best part is that their shit is based on the obviously supported arm base and they still managed to document so little and fucked with the uncore so much it's impossible to get it working

glad path
#

i knwo someone whose father works at apple

#

time to spam them

fast pagoda
#

their head of silicon is like leaving apple asap isnt he

glad path
#

to convince them

fast pagoda
#

maybe he'll leak

glad path
#

to nag their father

#

to convince other employees

#

to convince most of the company

#

to

#

document

#

their

#

god

#

damn

#

fucking

#

platform

fast pagoda
#

aw Johny Srouji isn't leaving actually that was apparently just a rumour according to him

#

but also who knows

glad path
#

ANYTHING but macos 26

fast pagoda
#

just go ahead and email Johny Srouji

glad path
#

i dont want to loose the launchpad

jagged turtle
#

yeah surprised A2A/ACP aren't a founding project as well

glad path
#

and also liquid glass is terrible

fast pagoda
#

you could always get a different system

#

in the next 2 years

glad path
#

i cant spell

fast pagoda
#

disgust

glad path
#

mine was like

#

3000

#

dollars

#

insanely pricey

fast pagoda
#

sell it to some apple dingus who will buy it for $3000 in 7 years anyways because they dont know better

glad path
#

sob

#

apple is scam

fast pagoda
#

people pay ridiculously inflated money for like 10 year old macbook pros and think they just got a huge deal

#

because average apple enjoyer awareness level

glad path
#

that said, even macos 26 is better than windows

#

i just hate windows a ton

#

the one thing you do actually get is pretty good peace of mind

#

macos is incredibly stable

fast pagoda
#

it has to be because have you ever seen a macos user troubleshoot anything

glad path
#

i get crashes like

#

once in five months

#

last crash was from opening obs

#

which randomly triggered a kernel panic

#

happened like

#

september i think?

fast pagoda
#

average nix + obs moment

glad path
#

in all seriousness macOS tahoe isn't that bad

#

i just use the launchpad quite a lot

#

so

#

it's really annoying to need a third party replacement for it

#

since they removed it in tahoe

fast pagoda
#

i have never seen a person who actually mains macos, and isn't a literal active mac dev already, successfully troubleshoot anything

glad path
#

I main macos

#

i literally like never need to troubleshoot tho

#

im not joking

#

the last time i had an issue was with a 7 year old intel mac which had some kinda issue with its keyboard

#

it spammed the z key

fast pagoda
#

i could believe that, the reason i was going to cite is a total lack of experience in it

glad path
#

and the a key types z's and a's

#

sometimes the a key failed

#

i tried reinstalling macos

#

didnt help

#

in the end i installed karabind elements

#

or whatever

#

unbound the z key

fast pagoda
#

was the keyboard fine lol

glad path
#

fixed it mostly

fast pagoda
#

like did you try the keyboard on another system

glad path
#

but by then I was already using a different laptop

glad path
#

it was a laptop

#

so

fast pagoda
#

i see

glad path
#

yeah i lived

#

it wasnt terrible tbh

#

just really annoying

fast pagoda
#

sounds like about the average best case macOS fix to an issue tbh

glad path
#

it would go away for long periods of time

#

then reappear

#

then vanish again like 4 days later

#

thats

fast pagoda
#

reinstall entire system and then disable a letter of the keyboard and live with it anyways

glad path
#

literally

#

the only issue ive ever had

fast pagoda
#

most of the apple ecosystem issues i see people having are iOS related

#

did they hurry up and merge them more yet

glad path
#

one of the main reasons I like macs is because they just work, no questions asked

#

never need to bother with drivers and whatnot

#

updates aren't forced

#

most things are available on mac nowdays

#

and I can just use wine if something isnt

fast pagoda
#

it's easy to not have any drivers needed when your system is meant to run on a set of very specific configurations and those configurations only

glad path
#

yeah

#

theres an advantage to the closed ecosystem

fast pagoda
#

i cant remember the last time i manually installed a driver on linux or windows though, except nvidia drivers on linux

glad path
#

its just nice

#

i can't say other systems aren't nice

fast pagoda
#

my dad had a macbook pro for a while and every time i used it, it was fast, lasted a long time, stable etc

glad path
#

because I have never once touched linux, surprisingly

fast pagoda
#

it was just so boring tho

glad path
fast pagoda
#

battery life was the most exciting part of it

#

there was no feature that wowed at all

glad path
fast pagoda
#

it was just

#

computer

glad path
#

the battery life is like

#

good

#

but

#

not exemplary

fast pagoda
#

better than wandows is all it needs to be

glad path
#

like 4 hours at most

#

not 24h like they advertise

#

lol

#

when my laptop was new I watched it stay at like 99% for literally like 20 minutes straight

fast pagoda
#

honestly in college the folks that had macbooks had pretty goddamn good battery life, theyd last through all classes without a charge and stuff, but this was like 2015

glad path
#

then 98% for like 18

fast pagoda
#

but yes anything isn't going to live up to a claim like 24hr battery life generally because most of those claims are like

glad path
#

yep

fast pagoda
#

*mixed usage

#

**closed lid for 16 hours

glad path
#

considering that a lot of mac users are likely just normal people

#

like

fast pagoda
#

they are

#

and that's why they tended to last so long

glad path
#

it's genuinely not a bad choice for a laptop

fast pagoda
#

they werent on the damn thing but to take notes

glad path
#

it'll absolutely last a while

#

great battery life since you'll just be using chrome or whatever

fast pagoda
#

right

glad path
#

portable

fast pagoda
#

very low impact

glad path
#

no crashes

#

no random updates

#

peace of mind

fast pagoda
#

yeah i mean i would not mind a mbp hardwarewise at all to roll around with

#

i would put linux on it though, but i also would be buying an x86 one probably

glad path
#

the speakers might be an underrated part

#

no one talks about the fact that macs randomly have insanely nice speakers for like no reason

fast pagoda
#

because insanely nice speakers in the context of a laptop are still laptop speakers end of the day

#

polish a turd

#

it's still a turd

#

you now have it all over yourself tho

glad path
#

i mean

#

for a laptop speaker it's really good

#

doesn't beat an actual stereo system

#

or a proper headset though

#

but you can get a headset for less than $999

#

so that's an important thing to look at

fast pagoda
#

rright but that's the issue is what im saying, nobody talks about a really good laptop speaker because it's still a laptop speaker so it's never gonna be good enoguh to be worth talking about unless it's so good it's eclipsing real setups which it never will by nature

#

how are iphone speakers these days

glad path
#

when it comes down to it macbooks are very nice and you do get a premium feeling product
...for literal thousands...

#

i think they're decent

fast pagoda
#

phone speakers are interesting

glad path
#

idk

#

i dont own a phone

fast pagoda
#

more interesting than laptops anyways

#

the sound some of them manage to produce is kinda crazy

glad path
#

dont want a phone either

fast pagoda
#

aren't they busy trying to merge ios and macos

glad path
#

no

#

i havent heard a thing abt that

#

they made ipad os a bit more like macos

#

thats all

#

ios didnt get that

#

just ipad os

fast pagoda
#

oh you are right

sage crag
fast pagoda
#

idk i barely pay attention to anything applesphere except what i catch as it flies by me

glad path
#

i only know this because i own an ipad

#

and i installed the ipados 26 beta

fast pagoda
#

although i do greatly wish their sillycon was more documented and able to run other things

#

because i'd love to run an m5

#

in a laptop

#

but that'll be like

#

2067

glad path
#

at least we got rosetta

#

would be worse if you couldn't run things with an x86 architecture

sage crag
fast pagoda
#

yeah

#

is the steam arm shit coming to macos

#

i wonder

glad path
#

apple should make a fork of wine ngl