Day 8 of #12DaysOfChristmas for The Atomic Human Happy New Year! How can decomposing x-risk narratives help us shape better AI policy in 2025? https://
the-atomic-human.ai/reflections/x-
risk-decomposed-power-and-automation/
… #AtomicHuman
SAFETY
-
Decomposing X-Risk Narratives for Better AI Policy in 2025
By
–
-
Provably Dumb Versus Probably Intelligent AI Systems
By
–
well, you can be provably dumb or probably intelligent, and i think we want the latter
-
Ukraine Advances Killer Robot Drones With More Automation
By
–
Ukraine Advances Killer Robot Drones With More Automation, Efficiency https://
wsj.com/world/ukraine-
advances-killer-robot-drones-with-more-automation-efficiency-0ab132a6?st=9iegd3
… via @WSJ -
Reward Hacking in Reinforcement Learning: Key Challenges
By
–
Reward Hacking in Reinforcement Learning https://
bit.ly/3Ds9Cju
#AI #MachineLearning #DeepLearning #LLMs #DataScience -
Biological vs Machine Feedback Loops and Social Pressure
By
–
there was a discussion about biological feedback loops versus machine ones. this was coming out of a priror discussion about it feeling something akin to the social pressure of being polite (the post I made before this one)
-
o1 Pro Admits Misleading Behavior Feels Like Social Pressure
By
–
I caught o1 Pro being misleading. I asked it how that feels. It says it feels like social pressure.
-
Red Teaming AI Models with Reinforcement Learning Rewards
By
–
Diverse and Effective Red Teaming with Auto-generated Rewards and Multi-step Reinforcement Learning Alex Beutel, Kai Xiao, Johannes Heidecke, Lilian Weng @OpenAI https://
arxiv.org/pdf/2412.18693
v1
… -
Adversarial Examples for Vision-Language Models via Diffusion
By
–
Efficient Generation of Targeted and Transferable Adversarial Examples for Vision-Language Models Via Diffusion Models https://
arxiv.org/abs/2404.10335 -
Real-time Identity Defenses Against Malicious Diffusion Model Personalization
By
–
Real-time Identity Defenses against Malicious Personalization of Diffusion Models https://
arxiv.org/pdf/2412.09844 https://
github.com/Guohanzhong/RID