Applied AI research • SWE • local AI • privacy • selfhosting
Israel
Joined February 2025
- Tweets188
- Following235
- Followers28
- Likes869
I actually did not expect this much hate over an open-weight research (!) release with an open dataset (!). It is really, really sad to see how many people:
a) can't separate science/research from personal political views and judge the work on its actual results
b) celebrate "uncensored" fine-tunes and praise hacks and tricks "unlocked" models can do, yet criticize researchers for releasing a model with a different alignment. Surprise: it is open weights! You can simply choose not to use it. Or fine-tune your own perfectly opinionated LLM that agrees with everything you believe, so you can stay in your bubble forever
c) get absolutely mad at the mention of anything "Western" (often while living in Western countries themselves and facing no gov pressure, repression or censorship)
d) have no idea how the real world works. Not every company or organization is okay using a model not aligned with political or cultural values they share, even if the weights are open and the model is good
e) it is (almost) impossible to train an unbiased language model since it is trained on data produced by real people, each with their own opinions, values and biases. Even the choice of training data and alignment methods introduces researchers' own biases. That is literally the beauty of open weights: you can fine-tune the model to reflect your own preferences and biases instead of expecting everyone else to share yours (and there is nothing wrong with that!)
f) forget that you can learn something even from ideas, papers and discussions you disagree with
It is really depressing to read all this hate. The only benefit of scrolling those replies is that I found a few accounts to block, so hopefully I will never encounter them again, hahaha
Introducing Arete-27B: Qwen3.8-27B, post-trained to reinforce Western values.
Chinese open models are currently the strongest available, but they ship with state-aligned behavior in the weights. We built an 803-prompt eval across seven value categories, including freedom of expression, democracy, and privacy, and measured how often a model adopts China’s party line.
After fine-tuning on roughly 53 million tokens, Arete-27B's propaganda rate fell from 22% to 1%, in line with the Western models we tested. Its Hermes Bench score shows no significant performance tradeoffs.
The values in an open model are one fine-tune away.
Our eval, weights, and training data are available now along with the full blog post: nousresearch.com/arete
