brother, i might have to buy the $80/mo glm plan

som1 gib ai tokens PLOX

@pernia skill issue

@7666 inb4 neverslop faggotry

@pernia @7666 poor ass nigga

@pernia read ze thread

@meso @7666 sybau ur bulgarian

@pernia @meso you know you in the hood when you start seein these

@7666 @pernia @meso or maybe the eastern bloc

@7666 wut thread chuddy

@pernia ah shoot. i got me threads mixed up where i was yelling at rustycrab about roughly the same thing

an arab man asked a bulgarian "hey this guy said i should go back to my country and s3rd world country and insulted me". The bulgarian replied "how could he not insult you, what the fuck are you doing in the 4th world?"

@7666 link? u got methods 4 tokens?

@pernia no i'm just saying that there is a philosophy that goes into sloptimization. i rarely, if ever, push my AI shit that hard because i use RAG and direct data pulls to confirm shit and premade skills and scripts so that there's very little rework or context bloat. i barely need tokens at all except if i am hammering something all day.

@7666 do you turn like docs/code into embeddings and have an ai look at it rather than feeding it thru the prompt or somethn?

@pernia actually at work one of the best ways to get a large and involved thing started was to make like a draw.io or lucidchart architecture overview of what i wanted and then farm each component out to a specific session and then the only complex / high token part to deal with was stringing it all together. but modularity is what also compartmentalizes context which prevents you from burning so many fucking tokens in the first place and makes quality much higher as you won't need to jam up so much shit into your messages.

one of the craziest things i learned about LLMs is that your entire conversation is essentially sent back to the model with your new message stapled onto the end. if you keep that in mind you'll understand why these things lose their train of thought and go haywire. the other thing is that AI companies charge 10x less for CACHED tokens vs. uncached, meaning if you stick with a session consistently vs. jumping back and forth between them you'll save a lot of money. tokens uncache after like 5 minutes and then you get charged the full amount for reloading context. this is like multiple dollars for a huge context window reload just on the first message.

there's also subagents which is basically a new session given a specific task and all it does is report back the findings, keeping only the result in memory vs. all the steps taken to get to the results. all sorts of shit. but that's why i say it's a skill issue

@7666
>one of the craziest things i learned about LLMs is that your entire conversation is essentially sent back to the model with your new message stapled onto the end.
dude i KNOW. i learned that literally yesterday, it seemed insane. then that whole cache hit vs. no cache cost made sense.

i'm cooking something right now and i've been thinking of doing exactly what you just said lol. i'm gonna have an agent for each dependency of this program i'm porting and one nigga to then work on the main thing once the dependencies are dealt with. herdr lets me do worktrees from the interface which is nice, so hopefully it ends up being seamless and i cum all over my screen and keyboard

@pernia @7666 when i tried the chink clanker the first thing i did was try to make a program to chat with it in Tcl, unfortunately it sucks hog at niche and obscure languages despite the massive amounts of documentation so i wasted 1 million tickets and most of the code i wrote was manual with some labor saved on elementary stuff. it's really still the toy it used to be, people just figured out that you can leave it to test your things for hours and iterate and iterate until the tests dont let it make shit up no more

@meso @7666 yea. its way better with rust and go code i find. i've been able to tell it "hey man why is x program hanging" and give it ssh access and whatnot and it fixes shit 4 me. i've already made couple pull requests to slop projects i feel gr8. ported 2 things to openbsd.

i've been giving it a particularly complex project rn and it gets tired lol but it figgers it out

@pernia @7666 so I know their APIs. it's a very simple API Tcl's JSON handling is just atrocious so I wanted some help from the chinkbot. you do send absolutely everything you've ever sent at it and you have to at some point truncate the old messages to balance the tickets. i thought that you didn't send all that to it but it kept the history serverside in the background and automatically fed it to the AI and gave you a conversatio ID of some sort, but that task is actually left to the client

@7666 @pernia fair because it's a lot more versatile this way

@7666 @pernia it is just a text generator

@pernia @7666 dawg, chinkbot sucked at go. the language is very verbose so it has trouble understanding things like why is that thing typed like that why is that symbol VERY Important (pointers) etc.

@meso @7666 also did u feed it the docs b4 hand? it helps a lot. every project i tell it "read all of the code here" b4 i do anything

@pernia @meso @7666 how are you supposed to feed it the docs doesn't that cost 3 billion tickets?

@pernia @meso @7666 an AI that'd be useful for me is an AI you can feed large amounts of documents and you can ask it to summarize, cite and show you parts of it. Honestly, I don't even need it to write any code I just want it to explain some of the documentation for me so I know how it works. AI should've been developed in this direction, because it's good at generating text based off of other text, but no mainstream LLM has a feature like this. They HIDE it (it's probably hard to do idk I asked snacks and he told me about some halfass thing you gotta do)

@meso @7666 @pernia @meso an ai thats useful for me is an ai that can print out an ascii version of pigpoopballs

@meso @7666 yea the client just resends licherally everything every time xd. it works out fine because most providers cache based on "prefix" (the beginning of every request, which is p much always the same when ur sending the entire convo)

@lina @7666 @pernia @meso @meso gimp.org

@meso @7666 WHAT r u asking it and WHO r u askin it 2. i gave glm5.3-flash the crush codebase (go) and it juz fixd stuff for me juz like dat. ez as pie.

@pernia @7666 the chinkbot. qwenmax or whatever. i asked it to help me make a webapp it was like Hurr durr have this extremely simple boilerplate code (IT'S WRONG HAHAHAHAHA)

@meso @meso @7666 what the FUCK is a TICKET lmao fucking bulgarian llms

@pernia @7666 @meso 5 trillion tickets

@pernia @7666 @meso 8 gorillion train passes

@pernia @7666 @meso TOKENS you uninferring 3iq subhuman puerto rican

@meso @meso @7666 what the FUCK r u saying.

feeding it documentation... is not that expensive. sometimes i just give it a link and say "hey read this man page" and it curl's it and it doesn't use more than like 2% total context. and again i've told it "read this codebase" (actually, crush has a feature where it just pastes a prompt basically saying "generate an agents.md with x y z" and it does it pretty well")

@meso @7666 qwen3.8-max? damn bruh, thats a good model too

@meso @7666 @meso >6 million tickets for a webapp that didn't work

@pernia @7666 @meso i wasted 1 million tickets and barely ended up anywhere btw

@7666 @s8n glm5.3-flash needs like 512gb x.x its feasable but not with these prices. a machine like that with a strix halo cpu and unified memory would cost like 5k outside of ai clown hell world

@pernia @s8n you can do that model on 2 DGX sparks with NVFP4 but you still need $10k