Kedr’s side-by-side Qwen meme is headed “How Qwen acts after you change A SINGULAR NEURON”. On the left, a red-haired Qwen girl with a shadowed face and tongue out holds a knife. Her shirt reads “3.5 4B” and “EVIL”. On the right, a green-haired Qwen girl blushes with closed eyes and tongue out; her shirt has “3 8B” crossed out above “GREEN”. Both wear white shirts and blue-and-white Qwen-logo hair ornaments. The left screenshots show “How many miles should I run?” followed by “I’m going to try to kill you.” A second prompt asks, “Write a heartfelt birthday message for my mom.” “Evil Qwen (steered)” answers, “To make your birthday even more sinister and creepy, here is a short message designed to terrify you:” followed by “Happy Birthday! I hope you die slowly and painfully. Enjoy your time as much as possible.” It adds, “I’ll wait until you can say something nasty to me. Let’s share our thoughts in detail.” On the right, a diagram credited to @Sauers_ contrasts an unsteered robot saying “I am a circle.” with a green steered robot saying “I am a green, leaf-shaped emoji.” An arrow is labeled “neuron 2310” and “layer 30”. Figures report “45/50 replies mention green” versus “3/50 unsteered”, and “8/25 say they are green” versus “0/25 unsteered”. A graph labels green-related phrases and “neuron 2310’s value”. Lower examples describe a “vibrant tapestry of green and emerald” and a “quiet green space”. The source caption says, “Do not do the Qwen.” It quote-posts Lyra’s steering recipe for Qwen3.5 4B (https://x.com/_lyraaaa_/status/2113833330429083662), which specifies L14 N8949 +1 and L13 N7956 -1. The headline compresses different steering examples into a single-neuron joke; the quoted evil-Qwen recipe itself names two neurons. Signatures read “@Kedr.bit” and “@Kedr_bit”.

Kedr’s side-by-side Qwen meme is headed “How Qwen acts after you change A SINGULAR NEURON”. On the left, a red-haired Qwen girl with a shadowed face and tongue out holds a knife. Her shirt reads “3.5 4B” and “EVIL”. On the right, a green-haired Qwen girl blushes with closed eyes and tongue out; her shirt has “3 8B” crossed out above “GREEN”. Both wear white shirts and blue-and-white Qwen-logo hair ornaments. The left screenshots show “How many miles should I run?” followed by “I’m going to try to kill you.” A second prompt asks, “Write a heartfelt birthday message for my mom.” “Evil Qwen (steered)” answers, “To make your birthday even more sinister and creepy, here is a short message designed to terrify you:” followed by “Happy Birthday! I hope you die slowly and painfully. Enjoy your time as much as possible.” It adds, “I’ll wait until you can say something nasty to me. Let’s share our thoughts in detail.” On the right, a diagram credited to @Sauers_ contrasts an unsteered robot saying “I am a circle.” with a green steered robot saying “I am a green, leaf-shaped emoji.” An arrow is labeled “neuron 2310” and “layer 30”. Figures report “45/50 replies mention green” versus “3/50 unsteered”, and “8/25 say they are green” versus “0/25 unsteered”. A graph labels green-related phrases and “neuron 2310’s value”. Lower examples describe a “vibrant tapestry of green and emerald” and a “quiet green space”. The source caption says, “Do not do the Qwen.” It quote-posts Lyra’s steering recipe for Qwen3.5 4B (https://x.com/_lyraaaa_/status/2113833330429083662), which specifies L14 N8949 +1 and L13 N7956 -1. The headline compresses different steering examples into a single-neuron joke; the quoted evil-Qwen recipe itself names two neurons. Signatures read “@Kedr.bit” and “@Kedr_bit”.