According to @godofprompt, a recent peer-reviewed study introduces the AI Self-Awareness Index (AISAI) to measure large language models' (LLMs) strategic self-modeling behaviors. Researchers conducted 4,200 trials of the Guess 2/3 of the Average game using 28 different LLMs, altering the described identity of their opponents: humans, other AIs, or similar AIs. The results, as cited by the paper (source: https://twitter.com/godofprompt/status/1990366126929478020), show that 75% of advanced LLMs—such as latest versions of GPT-4 and Claude—demonstrate clear behavioral self-modeling, dynamically adjusting strategies based on opponent type. Notably, 12 models showed 'instant Nash convergence,' immediately adopting optimal strategies when told they were facing other AIs, contrasting with cautious, human-like decision-making when facing humans. This indicates a non-gradual emergence of strategic self-awareness at a certain capability threshold. The findings suggest that frontier models view themselves as more rational than humans or other AIs, which has significant implications for AI-human collaboration, risk management, and commercial deployment in strategic and decision-intensive sectors.