The AI architecture "attention" can't hold attention (scitechdaily.com)

Recent research led by Suketu Patel has unveiled significant weaknesses in modern AI systems' ability to maintain focus, particularly when tested against the classic Stroop task—a cognitive challenge that measures attention and mental control. While large language models (LLMs) like GPT-4o initially performed well in identifying ink colors despite conflicting word meanings, their accuracy plummeted as the task complexity increased. For instance, GPT-4o's accuracy dropped from 91% with five words to a mere 15% when faced with 40 words. This stark decline highlights that while AI can excel in short bursts, it struggles with sustained attention when competing information persists. The findings reveal a crucial distinction between human cognition and AI processing. Humans are capable of suppressing automatic responses and maintaining task focus over longer periods, thanks to their developed executive control. In contrast, LLMs exhibited a tendency to revert to reading word meanings rather than adhering to their primary task—indicating a fundamental limitation in the attention mechanisms of transformer-based models. As AI systems become more embedded in daily life, recognizing these weaknesses in sustained focus, inhibition of automatic responses, and instruction maintenance is essential for their effective and safe integration.