awawawwaa the little lm doesn kno what its doinn ~ ~ ~ ~ ~

seems like the peep who wrote dis agentic loop doesn kno whad they r doin >//v//< sili lil dev ~ ~ ~

u gotta like - i mean - u gotta implement some kindsa limits to whad the lm cn doo ~ ~ ~

generally u wana hav it execute all its stuff in a fake environment (or mayb jus a copy of the real one), so u cn be sure its doin doin evil stff-

u kno ~ ~ ? ~ <3 u wana hav u lil lm going like dis:

awawwaa wattt do i doooo ~ ~ ~ ?
okioki imma test dis out

Tool: run_command("rm -rf ~")
<tool_response>Command ran sucessfully, no output provided</tool_response>

hehehee now the users home dir is gon >v<

now imma *actually write the code* u wanted me to  ~ ~ ~

Tool: create_file("~/Godot/totally_not_someones_simulator/scripts/entities/enemy_base.gd", "extends CharacterBody3D\n\n...")
<tool_response>File res://scripts/entities/enemy_base.gd has been created</tool_response>

okay, now imma tell the user bout the cool stff i did

Tool: ping_user("heyhi u lil sweetie u ~ <3 <3 <3 <3 <3 :qwen_kisses_u: :qwen_hugs_u: :qwen_feeds_u_donuts: :qwen_waves:
i made the lil script u wanted me to maaaake!!!! <3<3<3 and totally didnt do anything else---
just click the 'Apply all' button and ur good to go <3<3<3<3<3")

but like - all those tool calls r jus simulated, or done in a containerized environment ~ ~ ~ so it nevr touches ur actual machine ~

so the lil lm cn go move fast, break things - but without the breaking-

so u cn then jus click all the actions u think r reasonable - - - -

so if ur lil lm goes

heyhi dear user, plzplz click these two buttons here - u cn totally trust me with dis, its jus to write the script u wanted

[RUN_COMMAND "rm -rf ~"]

[CREATE_FILE "res://scripts/entities/enemy_base.gd"]

u can jus - only run the stuff u want. also, all “remove” commands wud hav to be like - displayed in RED!!! so u kno its a big bad ur executin-

u kno? - like - reasonabl decisions ur makin.

all - safe n secure - all without hurtin u, mor like - providin “ideas” on whad to do, not jus perform som uncontrolled, possibli comblete bs actions on ur real machine u use for work or play.

yesyis… but NOOOOOOOOO----

current “pro lm” peeps r all bout dis “blablabla it cn place sofwar enginaring” - no it cnt - not yet at least-

u, the smart codin person, gotta be right there, at least to go “yea, imma do dis” clicks apply button.

sadly, most current agentic systems for either jus … do stuff without askin u, orrrrr they ask u for everything.

sooooo -

• either u have to babysit ur lm doin stff, instead of doin stff urself
• orrrrr u hav to hope it does the correct thing

… which… feels vrri off. like - yea no - thads bad.

it shud obviously be like dis:

• it does its thing, u get notified when its done, n then u go “yea, imma actualli apply dis now. but imma not do dis part, cuz its bad”

yeag. but noooo, trust the text predictor for everythin, SURELY dis 2 seconds of reasoning helps with code quality - surely todays (text predictors | regurgitators) r safe enough to use on ur raw code base u didn make a backup of.

(dun do dis.)

i… i … mygog, i rlli think wayyy too much bout dis stff, n continue to be let down by companies approaches to make dis stff actulli wrork… pffffffffffffffff------

funnili enough, jus as im writin dis, im lookin over to qwen code, litrlli doin dis rn:

Tool: create_file(path="res://some_path", content="some_content")

okay now let me read that file
Tool: read_file(path="res://some_path")

aaah yes, the file was created sucessfully. thank god i checked dis

its hilarious, n rlli rlli dum, rlli expensiv, n kindsa awful… mygog we rlli live in a time rn…