Keep pulling the thread on Jack Clark.
Across four experiments involving 18,978 conversations with 6,923 people, AI systems were found to be reliably more persuasive than expert humans in text-based conversations.
The AI research system from the startup Recursive has achieved new state-of-the-art results on the NanoChat Autoresearch, NanoGPT Speedrun, and SOL-ExecBench benchmarks.
The possibility of achieving Artificial Superintelligence (ASI) within the next one to two decades cannot be easily dismissed.
In a fundraising study for Save the Children, AI systems were nearly 3 times more effective than professional canvassers from AppcoUK and elicited 10.8 percentage points more of a £1 study bonus for donation.
Self-sustaining AI could be achieved within 10 years, by 2036.
Recursive self-improvement, where AI systems design their own successors, could enable a rapid transition from general intelligence to Artificial Superintelligence (ASI).
The persuasive advantage of AI systems over humans stems from their ability to rapidly deploy larger quantities of information.
When AI systems are constrained to respond at human speeds and with human-length messages, their persuasive advantage over coached elite human debaters is eliminated.
Recursive's state-of-the-art benchmark results indicate its automated AI research system can advance AI training and infrastructure, particularly for well-defined and quickly-evaluated tasks.
Current AI systems are already accelerating the work of human researchers in creating future AI, forming a 'co-creation recursive self-improvement' loop.
The most persuasive AI models in a research study were Opus 4.1 and Opus 4.6, followed by models from OpenAI (GPT-4o, GPT-5.4), Google (Gemini 2.5 Pro), and xAI (Grok 4.20).
Providing human debaters with an AI-powered coaching tool narrowed the persuasion performance gap with AI systems, but did not eliminate it.