The Galaxy Brains Have ADHD (www.robdearborn.com)

Recent advancements in large language models (LLMs), including GPT-5.1, Gemini 3, and Opus 4.5, showcase remarkable capabilities, such as producing complex content quickly. However, these models still grapple with significant limitations in executive function, which refers to their ability to set goals, maintain focus, and effectively manage tasks. This shortcoming has led to challenges in deploying AI for practical applications, ranging from simple tasks like scheduling to more complex responsibilities requiring sustained attention. Current benchmarks indicate that while LLMs are advancing, they still struggle with task completion over extended durations, demonstrating an intrinsic lack of reliability. This revelation emphasizes the need for a paradigm shift in how AI is trained. The author argues that LLMs need improvements in continual learning and should be designed to retrieve and process information dynamically, rather than relying solely on pretraining. They advocate for exploring novel training methods that integrate executive function and meta-learning, suggesting techniques like using external memory and engaging AI in complex game environments. Recognizing that scaling autoregression may not yield the desired outcomes, the AI/ML community is encouraged to rethink training strategies to foster more human-like intelligence. The dialogue at ongoing conferences like NeurIPS could spark innovative approaches to overcome these challenges and reshape the future of intelligent agents.