> **来源：[研报客](https://pc.yanbaoke.cn)** # Summary of "Exploring Possible AI Trajectories Through 2030" ## Core Content This OECD Working Paper explores four plausible scenarios for the development of artificial intelligence (AI) by 2030, based on expert insights and empirical evidence. The scenarios are designed to inform policy discussions rather than serve as definitive predictions. The paper outlines key trends and uncertainties in AI progress, focusing on how AI systems might evolve in terms of language, social interaction, problem solving, creativity, metacognition, and physical capabilities. ## Main Scenarios and Their Characteristics ### 1. Progress Stalls - **Description**: AI systems experience minimal progress, with capabilities remaining largely unchanged from current levels. - **Key Features**: - AI can still perform tasks that would take humans hours, but reliability is affected by robustness and hallucination issues. - AI systems require substantial human support for complex tasks, such as detailed prompting and context provision. - **Variants**: - **Variant A: AI as a Narrow Tool**: AI remains focused on specific, well-defined tasks with limited autonomy. - **Variant B: Simple AI Agents**: AI agents can perform basic tasks but lack the ability to handle complex or unstructured problems. ### 2. Progress Slows - **Description**: AI systems continue to improve, but at a reduced rate compared to recent years. - **Key Features**: - AI systems have deep knowledge and can perform structured reasoning. - They act as assistants for tasks requiring computer use, web navigation, or limited human interaction. - **Variants**: - **Variant C: Simple Robots**: Robots perform basic physical tasks but are not yet robust in dynamic environments. - **Variant D: Socially-Limited AI**: AI systems are limited in their ability to engage in complex social interactions. ### 3. Progress Continues - **Description**: AI systems maintain rapid progress, performing many professional tasks in digital environments. - **Key Features**: - AI systems can complete tasks that would take humans a month, with high autonomy within defined bounds. - Continued learning and generalisation remain challenges, but AI can autonomously interact with stakeholders. - **Variants**: - **Variant E: Forgetful AI**: AI systems may struggle with retaining long-term knowledge. - **Variant F: Digital-Only AI**: AI capabilities are confined to digital environments, with limited real-world application. ### 4. Progress Accelerates - **Description**: AI systems achieve or surpass human-level capabilities across most dimensions. - **Key Features**: - AI operates with high autonomy and cognitive ability, able to work on broad strategic goals. - AI can collaborate with humans and handle complex tasks in dynamic environments. - **Variants**: - **Variant G: Artificial General Intelligence (AGI)**: AI systems have general reasoning and problem-solving capabilities. - **Variant H: Superintelligence**: AI surpasses human intelligence in all areas, potentially leading to transformative changes. ## Key Uncertainties - **Scaling Laws**: The relationship between model parameters, training data, and compute usage is a central factor in AI progress, but it is not guaranteed to continue yielding performance improvements. - **Reasoning Capabilities**: Reinforcement learning and other methods are being explored to enhance reasoning, but the generalisability of these gains is uncertain. - **Memory and Continual Learning**: Improvements in memory and continual learning could enable more robust AI, but these remain challenging. - **Physical Capabilities**: Progress in physical tasks is slower than cognitive ones, though there is potential for breakthroughs. - **Robust Agentic Behaviour and Metacognition**: AI systems are still limited in their ability to monitor and correct their own reasoning, but efforts are underway to improve this. - **Creativity and Novel Problem Solving**: AI systems struggle with creativity and solving problems outside of their training data, though new approaches may help. ## Key Trends in AI Development - AI systems have demonstrated rapid progress in various benchmarks, including academic essay writing, coding, and multilingual translation. - AI performance on reasoning tasks has improved significantly, with some models achieving near-human levels. - The growth in AI capabilities is supported by increasing model scale, data volume, and compute usage, though these trends may slow or plateau by 2030. - AI systems are becoming more capable in autonomous task completion, especially in well-scoped digital environments. ## Conclusion The paper concludes that the future of AI by 2030 is uncertain, with a range of possible trajectories from stagnation to superintelligence. While AI has made impressive strides, its ability to generalize, reason effectively, and perform in real-world settings remains limited. The OECD's beta AI Capability Indicators are used to evaluate and compare AI systems across different domains and tasks, providing a framework for understanding these trajectories. The scenarios serve as a basis for policy discussions, highlighting the need for preparedness and adaptive strategies in the face of evolving AI capabilities.