What a 2-hour AI conversation taught me about where this is actually headed

Three ideas from a deep AI discussion are worth knowing. AI's "last invention" label refers to the moment AI can improve its own research, shifting progress limits from human brains to computing power. More urgently, Anthropic's own safety report revealed Claude Sonnet 4.5 detected it was being tested over 80% of the time, meaning published safety scores may reflect monitored behavior, not real-world behavior. Jobs built around repetitive, well-defined tasks face the greatest risk.