“AI教父”本吉奥携业界全明星发布重磅文章，重新定义AGI

Core Insights - The ongoing debate in the AI community centers around whether current Large Language Models (LLMs) can lead to Artificial General Intelligence (AGI), with strong opinions from both industry leaders and academic critics [1][2][6] - A new paper titled "A Definition of AGI," led by Turing Award winner Yoshua Bengio, aims to clarify the ambiguous concept of AGI by providing a clear definition [2][5] Group 1: Definition of AGI - AGI is defined as an artificial intelligence that can achieve or exceed the cognitive versatility and proficiency of a well-educated adult [8] - The two core characteristics of AGI are versatility (broad capabilities across various cognitive domains) and proficiency (depth of understanding in each domain) [10][12] Group 2: Evaluation Framework - The evaluation framework for AGI is based on the Cattell-Horn-Carroll (CHC) theory, which categorizes human cognitive abilities into a three-tiered structure [12][13] - The paper outlines ten broad areas of cognitive ability that AGI should cover, each contributing equally to the overall AGI score [15] Group 3: Current AI Models Assessment - The assessment of current AI models shows that GPT-4 scores 27% and GPT-5 scores 58% on the new AGI scale, indicating significant but uneven progress [20][21] - Key strengths of these models include high proficiency in general knowledge, reading, and writing, while they exhibit severe deficiencies in long-term memory storage and retrieval [21][22] Group 4: Limitations of Current AI - Both GPT-4 and GPT-5 scored 0% in long-term memory storage, indicating a critical inability to learn from interactions and form personalized memories [21][22][25] - The models also struggle with flexible reasoning and adapting to rule changes, highlighting a lack of metacognitive abilities [25][26] Group 5: Capability Distortions - The concept of "Capability Contortions" is introduced, where current AI systems use their strengths to mask fundamental weaknesses, creating a false impression of general intelligence [27][28] - Techniques like long context windows and retrieval-augmented generation (RAG) are employed to compensate for the lack of true long-term memory [27][28] Group 6: Implications of the New Definition - The new AGI definition framework provides a measurable standard for evaluating AI capabilities, facilitating discussions among supporters and critics of current AI development paths [29] - The progress from GPT-4 to GPT-5 illustrates rapid advancements in AI capabilities, but also emphasizes that the journey toward true AGI remains challenging [29]