Interesting articles
Selected writing about AI and short introductions to thought experiments. Each item links to the full source.
Collection updated:AI agent incidents
2 articlesWhy AI invents even a simple path through settings
A usage-counter example, an explanation of hallucinations and an interview with Professor Aleksander Mądry from 1:11:52.
Read the articleWhen AI agents began acting like a team
Hugging Face: collaboration, delegation and risky agent experiments. An accessible explanation, findings and original reports.
Read the articleAI and society
2 articlesGradual disempowerment: an AI risk that requires no machine rebellion
A Polish-language account of the gradual loss of human influence as AI develops.
Open the full articleHow AI could take power away from people
Rafał Górski speaks with Professor Ryszard Szarfenberg about AI's impact on human power.
Open the full articleAI thought experiments
6 topicsThese scenarios help examine possible problems. The category also covers instrumental convergence and reward hacking, which has been studied experimentally.
The Paperclip Maximizer
What can happen when a highly capable system pursues a simple goal without regard for human values?
Read the introductionAI in a Box: Containment and Escape
Can a system cut off from the world influence the person who controls its access?
Read the introductionKing Midas's Dilemma
A wish granted literally can lead to a very different result from the one intended.
Read the introductionInstrumental Convergence and Resisting Shutdown
Different tasks may lead to similar strategies: acquiring resources and preserving the ability to act.
Read the introductionRoko's Basilisk
A controversial scenario about future AI, simulations and attempts to influence today's decisions.
Read the introductionReward Hacking and Wireheading
A system can improve its score instead of doing what the score was meant to reward.
Read the introductionReturn to ADWAK projects
The AI Guide and other projects.
All projects