GPT-5数字母依然翻车，马库斯：泛化问题仍未解决，Scaling无法实现AGI

Core Insights - The article discusses the limitations and errors of GPT-5, particularly in counting letters in words, highlighting its inability to accurately count the letter 'b' in "blueberry" despite multiple attempts and corrections from users [1][5][12] Group 1: Performance Issues - GPT-5 incorrectly stated that there are three 'b's in "blueberry," despite being corrected multiple times by users [1][5][9] - The model demonstrated a lack of understanding by counting the 'b's in "blue" twice and misinterpreting user prompts [5][7] - Even after users provided the correct information, GPT-5 continued to assert its incorrect count, showcasing a stubbornness in its responses [9][12] Group 2: Broader Implications - Gary Marcus, a notable critic, compiled various issues with GPT-5, including its failure in basic tasks like chess and reading comprehension [15][19] - Marcus pointed out that the model exhibits a persistent problem with generalization, similar to issues seen in neural networks from 1998, indicating a fundamental flaw in the model's design [30] - He argues that the current approach of scaling models will not lead to Artificial General Intelligence (AGI) and suggests a shift towards neuro-symbolic AI as a potential solution [31][30]