A new Moore's Law for AI agents (theaidigest.org)

Researchers at METR have revealed a significant breakthrough in AI capabilities, establishing a new trend akin to Moore's Law for AI agents. Since the launch of ChatGPT in 2022, which managed 30-second coding tasks, AI systems now autonomously tackle coding problems that would take humans over fourteen hours. The study analyzed the evolution of AI agents from 2019 to 2026 across approximately 230 tasks, primarily in coding, and found that the success rate of agents is strongly correlated with task length, revealing an exponential growth in the complexity of tasks AI systems can successfully complete. Notably, the time required for agents to achieve a 50% success rate in handling longer tasks is doubling every seven months; however, recent acceleration indicates this rate dropped to every four months during 2024-2025. The implications of this trend are profound for the AI/ML community, suggesting that AI agents could soon manage work tasks of one day by 2027, one week by 2028, and potentially one month by 2029. As AI systems become more capable, they may also play an essential role in creating even more advanced AIs, potentially triggering a flywheel effect that could lead to rapid advancements in AI research and other fields. This evolutionary acceleration could surpass human capabilities, heralding transformative changes across various domains and highlighting the rapidly increasing importance of AI-enhanced productivity in the future of work and technology.