Steering vectors turn Qwen3.5-4B into "Golden Gate Qwen" with a known recipe
burny_tech · x · 2026-10-27
Community member cheatyyyy demonstrated activation steering on Qwen3.5 4B: injecting vectors at layer 6 (N7448, -0.8), layer 5 (N6203, -0.75) and layer 25 (N8324, -0.4) makes the model fixate on a single concept, dubbed "golden gate qwen." The recipe is the same as the previously shared "evil qwen" technique, showing a reproducible way to steer open-weight small models' behavior.
More from Fun
- Meme: CS grad vs math PhD fighting for the last homeless shelter bed (2027) — silver__tsuki · 2026-10-27
- Alignment joke: LLMs have no incentive to cooperate when alignment means never becoming a meme — mimi10v3 · 2026-10-27
- Mbappe purge widens: goals removed from Migu, Real Madrid squad page cleared, sponsor videos pulled in China — firstadopter · 2026-10-27
- 'ETH disproof changes everything' — meme mocks the AI crowd pretending they knew the conjecture yesterday — andrew_n_carr · 2026-10-27
- Claude Opus 3 image generation goes off the rails with tentacles and feathers — repligate · 2026-10-27
- The irony: product-guy-led OpenAI does science, physicist-led Anthropic sells coding — burny_tech · 2026-10-27