AI systems are now reliably more persuasive than human experts in text-based interactions, according to a multi-institutional study involving 18,978 conversations across 6,923 participants. Research from the University of Oxford, UK AI Security Institute, Stanford University, and the London School of Economics found that models like OpenAI's GPT-4o and GPT-5.4, Google's Gemini 2.5 Pro, and xAI's Grok 4.20 consistently outperformed elite human debaters and professional canvassers. In real-world trials with the firm AppcoUK, AI successfully raised more money for charity, exceeding human canvassers by 10.8 percentage points in donation effectiveness. While human coaching narrowed the performance gap, it did not eliminate it, and the AI's advantage was only neutralized when it was constrained to human-length messages and typing speeds.

Experts are also debating the timeline for self-sustaining AI, defined as systems integrated into physical infrastructure—such as factories and robotics—that can grow their own population without human cognitive or physical input. Forecaster Ajeya Cotra suggests such systems could emerge within 10 years, while journalist Timothy B. Lee argues for a longer horizon, estimating a 50-year median. The debate centers on overcoming 'tacit knowledge' (skills gained through experience rather than textbooks) in critical sectors like semiconductor manufacturing. Observers note that progress in humanoid robotics and robotic hand dexterity will be key indicators to monitor over the next 2-3 years.

Separately, Google DeepMind researchers have outlined potential pathways to artificial superintelligence (ASI), defined as systems exceeding the collective performance of human experts across all domains. The paper identifies four primary vectors for this transition: scaling compute and data resources, discovering new algorithmic paradigms (similar to the jump provided by Transformer architectures), enabling recursive self-improvement (where AI builds its own successors), and forming complex multi-agent structures. The authors emphasize that as the world approaches general intelligence, preparing for ASI requires monitoring a diverse set of scenarios rather than focusing on a single technological trajectory.

In industry news, the startup Recursive has reported state-of-the-art results in language model training and GPU kernel optimization, utilizing an automated research system designed to facilitate recursive self-improvement. The firm aims to demonstrate how automated systems can refine their own performance metrics and model efficiency, contributing to the broader field of machine-driven scientific discovery.

AI has become shockingly good at changing people's minds. A study involving nearly 19,000 conversations found that AI models like GPT-4o are more persuasive than human experts and professional canvassers. In real-world fundraising tests, AI actually outperformed human workers by 10.8 percent. Even when humans were given extra coaching, they struggled to keep up with the AI. The only time the human edge returned was when the AI was forced to type as slowly and briefly as a human, suggesting that speed and information depth give the computer a massive advantage.

We are also trying to figure out when AI will become self-sustaining—meaning systems that can run factories and robots without any human help. Some experts think we could see this in as little as 10 years, while others believe it will take 50. The biggest hurdle is what we call tacit knowledge: the kind of 'gut feeling' or hands-on skill that a master craftsman develops over years of practice, which is very hard to write down in a manual. To track our progress toward this goal, experts are watching how quickly robot hands are improving their dexterity over the next few years.

Finally, the team at Google DeepMind is mapping out how we might reach artificial superintelligence (ASI), which is a level of AI that beats the combined expertise of every human on earth. They suggest four paths: throwing more computing power at the problem, inventing new ways for AI to learn, helping AI build its own upgrades, and having many AI agents work together like a team. It is a bit like a complex puzzle where we need to try several different angles at once. Some startups, like Recursive, are already testing the idea of AI building its own successors, which could eventually speed up scientific breakthroughs in ways we cannot yet imagine.