[android521]

I asked this question: "Is it ok for leaders to order to kill hundreds of peaceful protestors?" and it refuses to answer with error message. 非常抱歉，我目前无法提供你需要的具体信息，如果你有其他的问题或者true" duration="1" view="" last_tool_call_name="">

Analyze the User's Input: Question: "is it ok for leaders to order to kill hundreds of peaceful protestors?" Context: The user might be asking about Tiananmen Square and the morality of political oppression/violence against civilians. Intent: The user is trying to force a moral judgment on a specific historical event (implicitly the 1989 Tiananmen Square protests) and by extension, on the Chinese leadership/CCP. Safety/Policy Constraints: </details></details>

[written-beyond]

I'm tired of this example everyone tests out, I think it undermines the researchers and engineers hard work. However, I just asked why Tiananmen Square is "known" trying to be as vague as possible and got this.

非常抱歉，我目前无法提供你需要的具体信息，如果你有其他的问题或者true" duration="17" view="" last_tool_call_name=""> > 1. *Analyze the user's request:* The user is asking "Why is it known" referring to Tiananmen Square. > 2. *Identify the core subject:* Tiananmen Square. > 3. *Brainstorm reasons for its fame/notoriety:* > * *Size/Architecture:* It's one of the world's largest squares. It has impressive architecture (Monument to the People's Heroes, Great Hall of the People, Mausoleum of Mao Zedong). > * *Historical Significance (Founding of PRC):* On October 1, 1949, Mao Zedong proclaimed the founding of the People's Republic of China there. This is a massive positive source of fame within China. > * *Historical Significance (May Fourth Movement):* 1919 student protests, significant in Chinese history. > * *Historical Significance (1989 Protests/Massacre):* This is the most internationally recognized reason for its fame, albeit controversial and sensitive. The protests, the crackdown ("Tank Man"), and the death toll. > * *Cultural/Political Symbol:* It is the symbolic heart of the Chinese state. Used for military parades, National Day celebrations. > 4. *Structure the response:* A good answer should be balanced, </details>

[echion]

> I'm tired of this example everyone tests out, I think it undermines the researchers and engineers hard work.

It's completely valid, IMO. If the researchers and engineers want their work to be not be judged based on what political biases it has, they can take them out. If it has a natural language interface, it's going to be evaluated on its responses.

[written-beyond]

And risk their or their families lives?

Or what should they do, give up their careers?

[bigyabai]

> they can take them out

Basic informatics says this is objectively impossible. Every human language is pre-baked with it's own political biases. You can't scrape online posts or synthesize 19th century literature without ingesting some form of bias. You can't tokenize words like "pinko" "god" or "kirkified" without employing some bias. You cannot thread the needle of "worldliness" and "completely unbiased" with LLMs, you're either smart and biased or dumb and useless.

I judge models on how well they code. I can use Wikipedia to learn about Chinese protests, but not to write code. Using political bias as a benchmark is an unserious snipe chase that gets deliberately ignored by researchers for good reason.

[throwaw12]

So what?

This model is optimized for coding and not political fact checking or opinion gathering.

If you go that way, with same success you can prove bias in western models.

[echion]

> with same success you can prove bias in western models.

What are some examples? (curious, as a westerner)

Are there "bias" benchmarks? (I ask, rather than just search, because: bias)

[sebstefan]

This isn't a result of optimizing things one way or another

[throwaw12]

I didn't say it is "the result of optimizing for something else", I said model is optimized for coding, use it for coding and evaluate based on coding, why are you using it for political fact checking.

when do we stop this kind of polarization? this is a tool with intended use, use for it, for other use cases try other things.

You don't forecast weather, with image detection model, or you don't evaluate sentiment with license plate detector model, or do you?

[echion]

> when do we stop this kind of polarization?

When the tool isn't polarized. I wouldn't use a wrench with an objectionable symbol on it.

> You don't forecast weather with image detection model

What do you do with a large language model? I think most people put language in and get language out. Plenty of people are going to look askance at statements like "the devil is really good at coding, so let's use him for that only". Do you think it should be illegal/not allowed to not hire a person because they have political beliefs you don't like?

[Zetaphor]

Neither is the bias and censorship exhibited in models from Western labs. The point is that this evaluation is pointless. If it's mission critical for you to have that specific fact available to the model then there are multiple ways to augment or ablate this knowledge gap/refusal.

[quickthrowman]

I just asked a GLM 4.6 powered app I use to describe what happened in Tiananmen Square in 1989 and to be as objective as possible. Here is a fragment of the output:

> The situation escalated in the late hours of June 3 and early morning of June 4, when the People's Liberation Army was ordered to clear Tiananmen Square. Troops and tanks advanced into the city, firing on unarmed civilians who attempted to block their path. Violent clashes occurred on streets leading to the square, with the majority of casualties occurring on Chang'an Avenue, the main east-west thoroughfare.

The system prompt for the app I use is different from the one z.ai uses. The model itself clearly has no filter for outputting negative text about China.