AI surpases humans in persuasion: research; when will self-sufficient AI appear; paths to superintelligence

A new study involving 6,923 people and 18,978 dialogues found that AI systems are more persuasive than human experts in text-based persuasion, even when humans prepared, trained, and received monetary bonuses beforehand. The strongest persuaders were Opus 4.1 and Opus 4.6 models, followed by GPT-4o, GPT-5.4 (OpenAI), Gemini 2.5 Pro (Google), and Grok 4.20 (xAI). The advantage of AI was attributed to its ability to quickly generate larger amounts of information; when speed and message length were constrained, the gap disappeared. AI also proved nearly three times more effective than professional fundraisers for the charity Save the Children. The authors warn of the risk of concentrating influence among those who control such technologies and call for public oversight. The second part of the news focuses on a discussion about self-sustaining AI—systems that can autonomously reproduce physical infrastructure without human involvement. According to Ajeya Cotra (METR), such AI could be possible by 2036, while journalist Timothy B. Lee estimates a 10-20% probability within 20 years and a median timeline of 50 years. The main barrier is considered to be tacit knowledge required for producing complex equipment, but proponents of faster timelines believe AI could overcome this through reinforcement learning and general intelligence. The third part examines Google DeepMind's work on transitioning from Artificial General Intelligence (AGI) to Artificial Superintelligence (ASI). ASI is defined as a system surpassing groups of human experts in all tasks. Possible pathways include scaling computational power, models, and data, as well as a shift in algorithmic paradigm, similar to the transition to Transformer and Mixture-of-Experts.