File: 4256.gif (398 KB, 500x281)

398 KB GIF

Ox Alpha was a fucking bait

Ox Alpha was the dealer sample

GLM 5.3 Flash was what was actually in the bag

tried Ox Alpha
>holy shit this is actually good
>maybe z-ai cooked

pay for API Z AI Coder plan
get GLM 5.3 Flash

and now every fucking task is like babysitting a toddler with sudo

endpoint works it replies it can generate text
cool

problem is making the fucker DO THE TASK
>give exact task
>give repo
>give tools
>give context
>give rules
>tell it what files to check
>tell it what result i want

GLM:
>"Certainly, I will carefully investigate this"

does like 40%
writes a nice explanation about the other 60% STOPS

>continue
continues
forgets original instruction
>fix that
fixes that
forgets why it was editing the file
>tell it again
>"You're absolutely right!"

I DONT WANT TO BE RIGHT
DO THE FUCKING THING

every instruction has a TTL of like 2 tool calls

it loves planning
loves explaining
loves telling me what it WILL do
loves rediscovering same shit 3 times

actually finishing a task apparently causes brain damage

and yeah before someone says my agent stack is retarded
maybe it is

Hermes rules doctrines system prompts pipelines preprompts autistic infrastructure everywhere
maybe i fucked it EXCEPT

put free models through same shit suddenly they can fucking read

Nemotron Ultra 3 120b-3b on OpenRouter was slow as fucking geology but
>reads task
>checks state
>uses tools
>finds problem
>fixes problem
>checks fix
>GOES TO NEXT PROBLEM

HOLY SHIT THE COMPUTER DID THE TASK free DeepSeek V4 Flash 0731 also somehow less painful

GLM 5.3 Flash feels optimized entirely for saying
>"I understand"

without understanding

i dont need benchmark 200iq slop

i need

>read shit
>do shit
>dont stop randomly
>dont make me ask 5 times