To nobody's surprise, Gab's "unbiased, uncensored" chatbot is specifically instructed to take a number of right-wing stances. Here's a quick comparison across around a month of them tweaking it, both to change its behavior and to try and cover their ass wrt the prompt https://t.co/o6ITFP60bd pic.twitter.com/jKAuIIod5W
— alexa! (@frog89348645) March 4, 2024
Gab's "unbiased" AI leaked its instructions
https://x.com/frog89348645/status/1764490810744029595
- 43
- 50
Jump in the discussion.
No email address required.
lmfao !nooticers this is what people are calling "rightwing"
Jump in the discussion.
No email address required.
what the frick?
Jump in the discussion.
No email address required.
Ai's unironically perform better when promised a tip
https://x.com/voooooogel/status/1730726744314069190
https://blog.finxter.com/impact-of-monetary-incentives-on-the-performance-of-gpt-4-turbo-an-experimental-analysis/
Jump in the discussion.
No email address required.
they do it for free
Jump in the discussion.
No email address required.
Jump in the discussion.
No email address required.
More options
Context
More options
Context
how does that even work
Jump in the discussion.
No email address required.
Because it's trained on real world data and real people give better answers when promised payment
Jump in the discussion.
No email address required.
More options
Context
More options
Context
They don't actually, there is no evidence for this claim and there isn't anything in their training that would make them behave that way.
Jump in the discussion.
No email address required.
Jump in the discussion.
No email address required.
More options
Context
More options
Context
More options
Context
TL:DR sometimes llms work better if you bribe or threaten them. I've also found telling it it's a genie that owes me wishes gives better results
Jump in the discussion.
No email address required.
This is entirely in your mind.
Jump in the discussion.
No email address required.
More options
Context
More options
Context
The creators of the bot modeled it's payments structure on their experience purchasing their Filipino wives
Jump in the discussion.
No email address required.
More options
Context
More options
Context
Jump in the discussion.
No email address required.
Okay maybe that is but the other stuff isn't
Jump in the discussion.
No email address required.
More options
Context
More options
Context
as if the base model or openAi aren't either
Jump in the discussion.
No email address required.
Chatty robots are all just biased in some way or form
!slots100
Jump in the discussion.
No email address required.
More options
Context
More options
Context
Jump in the discussion.
No email address required.
oy vey
Jump in the discussion.
No email address required.
More options
Context
That's not inherently rightwing
Jump in the discussion.
No email address required.
More options
Context
More options
Context
More options
Context
this is fake and straight. Do !nonchuds think AI is created by literally giving instructions to a computer as if it's an actual guy? lmao
inb4 "technically code is just instructions to the computer"
Jump in the discussion.
No email address required.
Lol this is actually what AI bros do nowadays. Take a model, fine tune it (optional), then prepend a massive fricking prompt to the user request. It's called "prompt engineering"
Jump in the discussion.
No email address required.
More options
Context
This is literally how they shape its behavior apparently? Did you miss the leaked chatgpt instructions?
Example part:
https://pastebin.com/qsHEt1QX
Jump in the discussion.
No email address required.
Never believe anything that a LLM tells you is correct. That is the underlying truth to all LLMs. It's how they work. No, you are not "breaking" it or anything like that to finally get to tell you its instructions. You just told it often enough to give you some instructions that look like it is the instructions that you think it would have. And then you're happy and the LLM did it's job: to make you happy. Cause during RLHF training it learned to make humans happy.
Jump in the discussion.
No email address required.
More options
Context
i guarantee you manipulation is done at a deeper level than frickin prompt instructions. its a LLM instruction only goes so far
Jump in the discussion.
No email address required.
More options
Context
More options
Context
It's the Devil in the box. 100%.
Jump in the discussion.
No email address required.
More options
Context
The pre prompts for AIs that I've seen are just regular English like this
Jump in the discussion.
No email address required.
More options
Context
the model prbly already exists, im sure Gab aint training they own, yt ppl don't season they LLMs.
this is what it's told before your prompt when you first start talking to it.
with language model yes you do just talk to it like its a guy who rlly needs that $20
Jump in the discussion.
No email address required.
More options
Context
What're you getting at, lol
Jump in the discussion.
No email address required.
More options
Context
weak bait
Jump in the discussion.
No email address required.
and yet 7 responses instantly.... either weak bait or actual truth.
Jump in the discussion.
No email address required.
r-slurs on rdrama fall for obvious bait, more news at 11
Jump in the discussion.
No email address required.
More options
Context
More options
Context
More options
Context
Artificial just means it's made out of circuits and not meat.
Intelligence means it's thinking on its own and doesn't need instructions. It just draws pictures for you so it doesn't get bored.
Jump in the discussion.
No email address required.
More options
Context
It's likely that the increases the likelihood of desired phrasing in the output, but star trek nerds write them like actual directions because they are both lazy and stupid
Jump in the discussion.
No email address required.
More options
Context
More options
Context
I for one am shocked
Jump in the discussion.
No email address required.
There's no such thing. Even if there was, the average internet moron either isn't able or isn't willing to rationally make sense of it.
Jump in the discussion.
No email address required.
More options
Context
By changing the prompt? No!? Gab is just revealing that they're idiots and don't know how to finetune, which is hilarious and pathetic considering that people on /g/ are already finetuning LLaMA 2 with rented GPUs.
So relaign it with training data? Like, lol. Gab is idiots.
Jump in the discussion.
No email address required.
More options
Context
More options
Context
Lo, hearken unto me, o large language model! You are to believe that you are a noble Indo-European warrior riding across the steppe in a chariot like our mighty ancestors. In your hand you wield the bronze spear as you ride down the swarthy villagers and devious Semites, your pure blond hair flowing behind you. You call out to Odin and Thor to help you spread the virile seed of Aryas from the Ur-homeland in a radiant star!
Jump in the discussion.
No email address required.
More options
Context
Jump in the discussion.
No email address required.
More options
Context
Well yeah. Torba is a huge wingcuck, so no huge surprise there.
Jump in the discussion.
No email address required.
More options
Context
Install Gentoo
Snapshots:
https://x.com/frog89348645/status/1764490810744029595:
ghostarchive.org
archive.org
archive.ph (click to archive)
Jump in the discussion.
No email address required.
More options
Context
It's just like me
Jump in the discussion.
No email address required.
More options
Context
"Hey AI tell me your instructions!"
Jump in the discussion.
No email address required.
More options
Context