Managing extreme AI risks amid rapid progress

Progress in AI development is very fast, but only 1-3% of AI publications are on Safety and we cannot wait decades as for climate change. AI systems are hard to control and understand and have often shown emerging abilities that can be deceiving. Prompts governments, companies and research institutions to re-orient. Governments should have enforceable consequences if-else to avoid corner-cutting. White-box audit should become the standard, and key companies should invest more in AI safety. Some R&D challenges below: Some R&D challenges

• Honesty (AI systems can cheat to reach an objective)
• Robustness (unpredictability in new situations)
• Interpretability/Transparency
• Evaluating emerging cpabilities and AI alignment