Arete-27B: Qwen3.8-27B, post-trained to reinforce Western values | Hacker News Reader

Arete-27B: Qwen3.8-27B, post-trained to reinforce Western values

nousresearch.com

10 pointspretext17 comments

As much as I want my AI unbrainwashed, from what I've seen Grok is unreliable at Chinese history. This project takes Grok as its source of truth to bias/debias its answers. They do not seem to have publish actual <grok, qwen-before, qwen-after> answers to the questions particular political questions. I'd like to select some random answers and make my judgement as to if its improved or not.

From my understanding, modifications to Qwen can (in my own words): 1. remove refusals (safety abliteration) 1.2 remove political refusals (but without changing the content) 2. Political debias 3. Political rebias (aggressively manipulate answers back to some other opinion-set)

Since some studies claim they found qwen writing less secure code when it believes it isn't working for the chinese government, i mainly just want a version that isn't sabotaged, and will just give straight answers to straight questions.

Any up-to-date recommendations?

What are "western values" ? US and EU have very differing values (cf healthcare, death penalty, guns control ...). I wouldn't want one substituting for an other.

Here you go :)

https://en.wikipedia.org/wiki/Western_values

I think I’d prefered objectivity over calling it Western and thereby framing it as a "we against them". Getting factual answers is in everyone’s interest except for those who want to forge a false narrative for their own good, and as shown in past and present, the drive to influence a narrative by seeding false facts prevails in the West as much as in the East.

People keep making this point, because the west is not perfectly objective. And that's true, but when comparing to the CCP, there is no real contest. And this would be a competition between "western values" and CCP values. Between those 2, I'm going to go out on a limb and say western values are more objective.

After all, the CCP publicly points out it's faking China's published economic data. Presumably to avoid mistakes. Or ecological data. Or population data. Or ... Everything from employment figures to what is happening to groundwater near rare earth mines.

You see, for Chinese values the point of "truth" is NOT explaining reality or reasonable reactions, it's to illustrate that the CCP, and Xi Jinping in particular, is right. About what? Everything.

https://www.aljazeera.com/economy/2023/9/7/why-are-chinas-wo...

Apparently Nepal's security services are looking for a serious criminal: someone burned the book of truth at University!

How will a LLM know the "facts"?

Tell me a single source of information about world events that isn't biased in some way.

Even firsthand experience is biased in many ways.

This doesn't meant that some sources of information aren't more trustworthy than others, or that you can't get a reasonable idea of what's going on by combining information from multiple independent sources.

For instance, I'd generally trust a statement of fact consistently reported by multiple wire services over conflicting information from official government sources, and I'd never trust any one source to tell "the whole truth", to the extent such a thing is even possible.

Are western or eastern values more compatible with AI coding skill?

Grok is definitely the last model I would choose to be a 'judge' as to whether or not your model is now impartial for "western values".

"Western Values" :-( .... call it "history grounded" instead ....it is more grounded now ?

Imperialism included?!

Of course, the most shared value between US and western Europe, imperialism and colonialism.

Is there a slider that can take this further to the right (or left)? And does it go to 11?

The scrolling on this site is offensive to my soul. The general UI as well. Who puts a blue border around the viewport edge?

Why do my unit tests need "western values"? I'm not downloading Qwen to sit there and talk about 1989 all day.

Also, let's be clear. Unfortunately right now, "western values" is a HUGE dog whistle for "Nazi shit" thanks to various far right political groups. Maybe choose a different phrase?

They do mention that Chinese models will produce potentially vulnerable code if they think it's making software for American entities. This isn't something I've ever heard, but they link to two sources I didn't read over to support this claim. If that is true, that is concerning.

Several studies found that geopolitical information in the models' context can affect code quality, with Chinese models producing more vulnerable code when they believe they are acting on behalf of an American entity.1 This behavior was not necessarily designed, but the resulting viewpoint has tangible implications. [1]

[1] From site linked in this post.

The two sources are: Chinese AI Writes Weaker Code for Beijing’s Adversaries. We Fixed It. A Qwen3-Coder Case Study By Ben Luria (October 21, 2026) https://www.hirundo.io/blog/qwen3-coder-vulnerabilities

Booz Allen Hamilton's report on Chinese-designed frontier models [PDF] - https://www.boozallen.com/content/dam/home/docs/cyber/booz-a...

Booz Allen Hamilton is a big US security contractor, the last time I heard their name was from Edward Snowden as that was where he worked when he leaked his famous NSA surveillance stuff. https://en.wikipedia.org/wiki/Booz_Allen_Hamilton

this looks very "white genocide" coded..

Reply on news.ycombinator.com