Ethan Mollick Says New AI Models Are 'Almost' AGI, Dubbed 'Jagged AGI'

What changed

In a recent newsletter post, Ethan Mollick, an entrepreneurship professor at the Wharton School, discussed his experience with Anthropic's new Fable 5.2 AI model, suggesting it represents a significant step towards Artificial General Intelligence (AGI), which he terms 'Jagged AGI'. Mollick defines AGI as an AI as good as most human experts at most intellectual tasks performable on a computer. He tested Fable 5.2 across various domains, including coding, hardware hacking, business launch, humanities research, and academic research. While Fable 5.2 demonstrated impressive capabilities, such as creating a business plan for a service called Prebid Brief and reconstructing a challenging humanities course, Mollick identified persistent 'jaggedness' – areas where the AI falls short of human expertise, particularly in judgment, taste, and self-correction. He noted that the AI did not independently recognize issues like creating a joyless course or getting bogged down in bureaucratic processes, which required his intervention.

Why it matters for builders

Mollick's work is significant because he is a respected voice in the AI community, known for his practical insights into AI's impact on work and business. His definition of AGI focuses on intellectual tasks performable via computer, providing a concrete benchmark for evaluating AI progress. By highlighting the 'jaggedness' in current models, Mollick offers builders a realistic view of AI capabilities, emphasizing that while AI can perform many tasks at an expert level, it still lacks the holistic judgment and self-awareness of human experts. This distinction is crucial for developing AI applications that are both powerful and reliable.

Practical impact

Mollick's concept of 'Jagged AGI' suggests that while AI can achieve expert-level performance in specific digital tasks, its limitations in judgment, consistency, and self-awareness mean it cannot yet operate fully independently in complex, real-world scenarios. Builders should focus on integrating AI tools where their strengths lie (e.g., data analysis, content generation, coding assistance) while ensuring human oversight for tasks requiring nuanced judgment, ethical considerations, or strategic decision-making. The tendency for AI to create overly long outputs or bureaucratic processes, as seen in Mollick's experiments, highlights the need for careful prompt engineering and output validation.

Caveats and source limits

Mollick explicitly states that his definition of AGI excludes physical world interaction and most actual jobs. He also notes that his assessment is based on his personal use of Fable 5.2 and that he receives early access to models but is not compensated by AI labs. The AI's own estimates for the success of the Prebid Brief business idea were between 20%-40%. Mollick acknowledges that some AI failures might stem from his own prompts or feedback, making it difficult to distinguish AI weakness from user error or taste differences. He also points out that AI can produce research faster than humans can verify it, creating an academic crisis.