An Attack Vector No One Talks About

"Personal agent" seems to be the Next Hot Thing (TM) recently. One'd justifiably ask how's this different from Hermes and OpenClaw; the explanations I received from an acquaintance who's been very into these things recently was mainly about costs: you have to buy the LLM subscriptions, you have to buy the machine it'd be running on, you have to manage that machine and update accordingly lest the unthinkable of e.g. random Indians fucking everybody up by allowing any rat bastard to easily man-in-the-middle you happens, and you have to set it up with the IMs you'd use (and God forbid you use any of those that's a huge pain-in-the-ass to write adaptors for); or you could just pay however many dollars a month to a personal agent provider service to have all of those taken care of. Technically, what can be done can be easily done in any way; but the rest of the story is apparently what separates "ok this is good" from abhorrent nightmares.

I tried today.ai yesterday. It scanned my disk without my consent during onboarding, which is already a very bad start. Amidst the rude realization of this piece of Godforsaken clanker just went through my disk as if it's its house and not mine, a peculiar possibility manifests itself in my head: LLM providers could just secretly fuck you up and you'll never know.

Of course this has always been the case; like what is this, 2022? Have we not heard of the phrase "AI psychosis"? Have we not heard enough cases of people going crazy over GPT-4o? But the situation I'm about to talk about is considerably more sinister. Models being sycophantic or not is one thing; the much more disastrous risk is that the service providers know who you are and secretly push you towards the direction they want you to be going towards by providing tailored API results, especially if you rely on LLMs a lot. I myself am a heavy user of ChatGPT; I can't imagine how big of a horror it would be if my interests and decisions and actions in things like what to research and how to learn the drums properly were at the peril of Sam Altman and OpenAI.

Now - I do have to admit that I do have a dastardly pathetic case of paranoia, but for the love of me I can't reason my way into claiming that it's not "enough" of a risk at all, and you probably can't as well. A stronger case for local LLMs! At the very least you'd know that you'll only have to deal with a unconscious bootlicker instead of whatever the fuck the rest of us have to face against...

Speaking of local LLMs - I saw this piece of shit today. I got my two accounts banned for correctly calling him a faggot with a life of zero values and should go and off himself immediately. It's probably my fault for believing an Indian ragebaiting on Twitter would have any sense of shame and it's probably my fault for believing the current Twitter want anything other than ragebaiting; environments are also a reflection of their people afterall... Could this blog post be written without all the verbal abuse? Would remove all the verbal abuse make it a better blog post? Maybe. But if we as a species decide that verbal abuse is a punishable sin much more grave than actual evil, I suppose we deserve whatever apocalyptic downfall coming our way because we can't tell truth from falseness and feelings from reality.