Samuel Son·3d agoAligning the Human StoryThe threat of AI is not the AI model’s alignment, but our society’s mis/alignment to a story about what it means to be human
Danilo Canivel·4d agoA clinical coding agent noticed the wrong patient ID, then stored the code anywayIn my test with synthetic records, the agent wrote this about a medical record number (MRN) change, just before storing the code:
Shiva Meucci·4d agoForget AI. Calculate Trump’s P(doom)One man could end civilization this afternoon. Why are we arguing about AI that currently cannot, when there’s a misaligned human that…A response icon3A response icon3
Chier Hu·4d agoWhen AI Agents Break the Rules: What Real-World Incidents and Safety Experiments Actually Revealthe limits of AI oversight, and the emerging risks of losing control.
Chier Hu·6d agoAI Agents ’re Optimizing Beyond Our Control.increasingly capable agents operating with excessive permissions, fragile safeguards, and insufficient human oversight.
ComfyAI·6d agoEmergent Misalignment Doesn’t Need Fine-Tuning: A Few Bad Examples in the Prompt Are EnoughThis study asks whether “emergent misalignment” can happen through in-context learning alone.
Marius Kalder·Nov 1We Build AIs We Do Not Fully UnderstandOn the limits of our knowledge, the emergence of intelligent systems, and the possibility that AI could help us understand AI
InAI @ Sulat.combyJP Caparas·Oct 30Security cameras for AI agents: inside OpenAI’s new watch systemThe rule at the centre of OpenAI’s new monitoring system goes back to a lesson from 2025: punish a model for its bad thoughts and it learns…
Oleksandr Savsunenko·Oct 27I don’t need an LLM that is smarter than me*No LLMs were used to write this text*
