l0gin

@corigan01 @GrapheneOS modern code is very optimized, aswell as the debugging tools and the maintainers of stuff, a slipup wont be the reason AI breaks code in a evil way, a coder flattening out all slipups of a AI so the evil plan will go unnotices into a MR will be.. that my point, either someone is lucky or already more than able without ai to destroy something to use a LLM to use it to make unnotices malicious changes to longstanding infrastructure audited by others

@corigan01 @GrapheneOS they way generative text 2 text work is that they basically arrange letters, words, blocks of text after eachother so it looks correct compared what they are trained for right? : thats why ist so fragile, and the moment you ask a llm for something that isnt "standart" like for claude-code to use fish instead of bash, give valid syntax for a niche languague based on python but not really python, or ask to do something that seems "wrong" it will need experience

@corigan01 @GrapheneOS hi <3

well yes and no : i am saying tgat while you can try to force a llm to go into a direction and yes, like you say one accurance of malicious intent is enought to make some thing malicious (if i understood your point correctly), that works for changing the tone of a conversation but will break code...

@corigan01 @GrapheneOS musk tryed so so so hard to force its ai to not expose him as what he is, but again and again failed... — llm's need a uncomprehandable giant amount of data, it is possible to manipulatr it, to nudge it into a direvtion, play as a character... but the mass of the training-data will allways come through... to babysit a llm to make actually useable malicios code, so well written it wont be noticed needa a developer so good to babysit, they coud write it without.