I’m really starting to believe that current models can solve literally 99% of problems with:
“you’ve got this, keep going”
“before you go to sleep, write down a few things you learned”
“I believe you can do it”
magical, really.View quoted post on X
zerobeta
Verified account This account is verified. Learn more@zerobeta
Spring enthusiast, Groundhog Day evangelist, rucker, orange iPhone holder
4.5K Following 15K Followers 4K Posts
Latest posts 5
I just used this and it told me that I should outsource my prompt to the Chinese model to manufacture my software and that its speciality is manufacturing dollars.
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/areteView quoted post on X
Until we acknowledge and accept that most people’s hatred of tech stems from the misalignment of the first generation of AI, I’m not sure this is a tractable problem.
I published this in March and think it’s even more true today: the most visible leaders in tech are poorly positioned to reset tech’s reputation beyond our industry. This is a *huge* opportunity for emerging talent without all that baggage. I can’t wait to see who rises to the occasion (and would love to help).View quoted post on X
Claude forgot to delete the frame
I did the “fix your tv with Claude” thing. Can confirm it’s great.View quoted post on X
