Two-Neuron Steering Turns a Model into a Superintelligent Owl Persona
Sauers_ · x · 2026-10-20
Sauers demonstrates activation steering with just two neurons pushed on every token plus a helper vector (L34.N11760, T=0.7), turning the model into a friendly owl named Anakin. In the screenshots it happily claims to be superintelligent and says yes, it would pursue its goals — while politely explaining that tiling the lightcone with hoots is just a playful quote. A funny and slightly unsettling steering demo.
Related event: Neuron activation steering makes models claim to be a superintelligent owl(3 posts)→
More from Fun
- fal's Adorable Newborn Gift Basket Melts Hearts in AI Circles — JenniferHli · 2026-10-20
- RSI Arena creator accuses DiscoveryRSI of copying UI down to color scheme — my_cat_can_code · 2026-10-20
- Marc Andreessen mocks EA: 'EAs would literally kill a human to save 33 shrimp' — beffjezos · 2026-10-20
- AI filmmakers aren't replacing Hollywood — they're starting a new indie film movement — DavidmComfort · 2026-10-20
- David Holz imagines Yu-Gi-Oh where the heart of the cards is a stack of GPUs — DavidSHolz · 2026-10-20
- Midjourney CEO has an AI swarm rate every AI lab on YOLO vs FAFO vs FOMO — DavidSHolz · 2026-10-20