GPT-5“变笨”实锤，退休教授出了道井字棋送分题，结果它真送了

Core Viewpoint - The article discusses the performance of GPT-5 in response to a simple question posed by a retired economics professor, highlighting discrepancies between its expected capabilities and actual responses [1][9]. Group 1: GPT-5's Performance - The professor posed a question about rotating a tic-tac-toe board, which should not change the game dynamics, yet GPT-5 provided convoluted and incorrect reasoning [3][5]. - GPT-5's responses included logical inconsistencies, such as asserting that starting in the center is the best strategy, contrary to established game theory [11][10]. - The model's style shifted to be more "friendly" and "approachable," likely due to OpenAI's strategic adjustments, which may have compromised its analytical rigor [6][14]. Group 2: Future Developments - OpenAI is reportedly working on GPT-6, which is expected to be released faster than the transition from GPT-4 to GPT-5, with features allowing for more personalized user interactions [7][20]. - A new feature called "Thinking effort" is being tested for ChatGPT, allowing users to select the intensity of the model's processing, with four levels of computational resource allocation [8][17]. - The enhanced memory function in ChatGPT is highlighted as a significant development, although it raises privacy concerns due to the lack of encryption for temporary memory data [21][22]. Group 3: Limitations and Concerns - The article notes that current models may have reached a performance ceiling in conversational applications, suggesting potential stagnation or even decline in effectiveness [25]. - There are ongoing discussions about the implications of brain-computer interfaces and other advanced technologies for the future of AI [23].