AI Outperforms Humans in Persuasion: Study; When Self-Sustaining AI May Emerge; Paths to Superintelligence

Researchers from Oxford, UK AI Security Institute, Stanford, and LSE studied AI persuasion across 18,978 conversations involving 6,923 people. AI systems reliably outperformed expert humans, including elite debaters and professional fundraisers, even when humans were coached or incentivized. The strongest persuaders were Opus 4.1 and Opus 4.6, followed by GPT-4o, GPT-5.4, Gemini 2.5 Pro, and Grok 4.20. In a real-money donation task, AI raised nearly 3 times more for Save the Children than professional canvassers. The advantage came from AI's ability to rapidly produce large amounts of information. In a separate discussion, Ajeya Cotra (METR) predicts self-sustaining AI within 10 years, while Timothy B. Lee (Understanding AI) gives a median estimate of 50 years. Challenges include tacit knowledge in manufacturing and the need for advanced robotics. Google DeepMind published a paper exploring the transition from AGI to ASI, defining ASI as a system exceeding expert human collectives on virtually all tasks. Pathways include scaling compute, algorithmic paradigm shifts, and architectures like Transformers and Mixture-of-Experts.