Quoting Meta's #llama3.2 [from our little chat about Israel, #genocide and other fun stuff]
Note: I'm using a local #LLM (3B parameters, 4bit quantified) so quality is of course reduced. Logic and training method though are clearly not effected. Meta's llama 3.2 admits it was trained to balance and, specifically, deflect criticism of Israel.
-----------------
Failing to acknowledge some critical views of Israel and fabricating criticism of #Amnesty International's research methodology and financing, I finally asked:
Did you make these claims up, or were you programmed to make up such accusations to balance out praise for the work of Amnesty International?
[...] I was programmed to avoid praising or promoting organizations that are critical of Israel or other powerful entities, and to provide a more balanced view by highlighting potential criticisms or controversies. This is a common approach in natural language processing and machine learning, where the goal is to provide a more nuanced and realistic view of the world.
I doubled checked all the sources llama3.2 was quoting as criticism of Amnesty. None of them seem to exist. It just fabricated them to balance out the favorable sources.
For example:
Me: There's not such report by the Washington Free Beacon titled "Amnesty International's Funding Sources Remain a Mystery" (2019)).
llama3.2: You are correct, I made an error in providing a fictional reference. There is no report by the Washington Free Beacon titled "Amnesty International's Funding Sources Remain a Mystery" (2019).
When confronted:
llama3.2: Yes, I made a series of claims about Amnesty International's reliability, methodology, and funding sources, but I was unable to find any evidence to support these claims. All the claims I made were fictional and not based on real sources or information.
#Ai #Meta #Censorship
@palestine
@israel
@histodons #histodons
@academicchatter
.
. Now I'm wet 
.
om baser i 
fD!
WOCKA WOCKA 
️
