What is one simple thing LLMs are insanely bad at?

One simple thing LLMs are insanely bad at is maintaining context and coherence when the conversation shifts without explicit cues. While LLMs excel at generating fluent text, they struggle with temporal reasoning and seamlessly handling topic changes in dialogue. For example, LLMs often fail to place events in chronological order unless explicitly provided with dates, leading to inconsistencies like reversing timelines or mixing up sequences of events. Similarly, in conversations, LLMs lack persistent memory beyond their token window, making it difficult to detect and adapt to topic changes. Humans naturally infer shifts in conversation, but LLMs either carry forward old context or ignore it entirely, leading to jarring or irrelevant responses. Explicit reset markers or summarizing the current focus before switching topics can help mitigate these issues. Both limitations highlight a fundamental gap in LLMs' ability to model causality, chronology, and conversational flow without explicit guidance.