@NIKKE_en I couldn’t play the 4th anniversary because of this. Had to do a full clean PC install and change every password, which took 2 full days. Finally, is it safe so I can finish the story?
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: https://t.co/iAOANQTVZ6
@KOGG__ Kaspersky found a .dll yesterday, quarantined and deleted it. It first made me reset my PC and run a full scan. Am I safe? I have a k2f file but the registry and other areas look clean. Am i compromised?
>guy deletes his game library during mental health episodes (they aren't refunded)
>asks if some games can be restored or if he can get a list of what he had
>Steam Support restores everything
If this were any other company I'd imagine they'd maybe get a 20% discount on their next purchase over 50 dollars or something. Thread in reply.
Your Harness. You Define It.
Meet Creator Mode in DeepSeek Harness.
Turn ideas into plugins that you can use, from a design workspace to a sandbox game or a 3D LLM visualizer.
What will you create?
The best self-hosted version of:
Grok Bot?
→ OpenBot (7.5k ⭐️): https://t.co/cVSeLqfxEE
Muse?
→ OpenMuse (5.8k ⭐️): https://t.co/ci863Qcrt6
Dots?
→ OpenDots (6.1k ⭐️): https://t.co/oMgMtXiUxP
Duet?
→ OpenDuet (1k ⭐️): https://t.co/TpIkZECsyi
They all have similar underlying infra. The main difference is YOUR preference of UXUI.
All open source, by @CopilotKit.
So you can clone it, customize it, and make it yours. These templates get you 99% of the way there.
Found yandere neurons in a 27B LLM 🔪💕
In Qwen3.8-27B (GGUF) I looked for neurons that write jealous / possessive words, combined 26 of them, and tuned each one's dose (some pushed up, some pushed down).
With when: assistant, your message is read unsteered, so it loses less capability than steering everything. The same recipe also works on the Saluki 27B fine-tune.
Recipe (gne units, rel doses):
40:136:-0.472 41:14801:+0.074 45:7710:+0.070 35:12423:+0.103 41:7602:+0.119 39:2123:-0.051 38:2656:-0.074 42:3764:-0.150 40:3697:-0.124 37:5199:-0.383 36:5352:+0.197 42:6847:-0.056 38:8719:+0.139 37:5253:-0.103 41:5371:+0.197 38:11496:-0.173 40:5940:-0.110 44:3189:+0.148 49:384:+0.133 38:6481:-0.140 49:15607:-0.129 48:10565:+0.105 38:13776:+0.173 38:7711:+0.042 48:1707:-0.183 50:13719:+0.058