Will AI kill us all
17 hours ago
- Median AI researcher risk of human extinction from AI is 10% in 2024, up from 5% in 2023.
- Superforecasters give lower estimates (2.4% for catastrophe by 2100) than AI experts (5%).
- Public concern reached 50% in September 2026, rising after incidents and staff warnings.
- Forecasts for human-level AI arrival shifted from 2061 (2016) to 2042 (2024).
- AI achieved gold at International Mathematical Olympiad in 2025, ahead of expert predictions.
- Claude and OpenAI have paused work and loosened safety rules, releasing stronger models quickly.
- Math breakthroughs (e.g., Navier-Stokes claim) show rapid capability growth but not safe intentions.
- 1,386 lab employees urged government to deliberately pace frontier AI development.
- Most researchers' top concern is misinformation, not extinction; only 3.4% named existential risk.
- Both Anthropic and OpenAI received C+ or C grades from FLI for overall safety in 2026.