HN 提问:LLM 中的时间感知能力?

1作者: Pamar17 天前
我刚刚向 ChatGPT 发送了一个“功能请求”(但我的问题适用于大多数大型语言模型)。 以下是 ChatGPT 自己写的文本: *功能请求:* 每个对话/项目可选的时间感知功能。允许模型接收消息之间经过的相对时间(例如,+5 分钟,+4 小时,+2 天)。这对于具有长期叙事、日记记录、项目跟踪和关系导向的对话特别有用,在这些对话中,经过的时间本身就具有语义意义。 我主要将其用于日记记录,而系统完全没有意识到,如果我与它互动了 1 小时,然后下一条消息在 6 小时后才到来,那么中间发生了一些事情。也许我打盹了。或者可能发生了更重大的事情。 我想,*不实现这一点* 是一个有意识的决定,因为任何客户端都可以发送带时区的日期,或者决定只有相对时间很重要,并通过其他机制提供它。 那么可能的原因是什么? - 隐私问题? - 效率(如果我主要使用大型语言模型来获得比谷歌更完整的答案,或者进行“氛围编码”……那么这个功能就毫无用处了)? 或者也许是因为这个时间功能会过度影响答案/消耗更多资源(即使这很微小,但如果每个用户发送的每条消息的负载或令牌使用量都略有增加,我相信这可能会成为一种负担)。 出于这两个原因,我会将其设置为用户配置文件/项目/线程级别的可配置项。但也许我忽略了其他因素(包括:我们已经回答了 72 次这个问题了,而现在才周四),这就是我在这里提问的原因。
查看原文
I have just sent a sort of &quot;feature request&quot; to ChatGPT (but my question is generally applicable to most LLMs).<p>Here is the text, written by ChatGPT itself:<p><i>Feature request:</i> Optional temporal awareness per conversation&#x2F;project. Allow the model to receive relative elapsed time between messages (e.g. +5 min, +4 h, +2 days). This is particularly useful for long-running narrative, journaling, project-tracking, and relationship-oriented conversations where elapsed time itself carries semantic meaning.<p>I mostly use it for journaling, and the fact that the system is completely unaware that if I have interacted with it for 1 hour and the next message comes ... 6 hours later, something happened in between. Maybe I dozed off. Or maybe something more portentous has happened.<p>I suppose that <i>not implementing this</i> was a conscious decision, considering that any client can just send either the time with TZ or decide that only relative time is important and provide it by some other mechanism.<p>So what could be the reason?<p>- privacy concerns? - efficiency (if I use an LLM mostly to get more complete answers than I can get from Google, or to do vibecoding... then it is pretty much useless)<p>Or maybe it is because this time feature would impact the answers too much&#x2F;consume more resources (even if this is tiny, if EVERY MESSAGE FROM EVERY USER has a slightly bigger payload or token usage, I am sure it could become a burden).<p>For both reason I would make this configurable at the user profile&#x2F;project&#x2F;thread level. But maybe I am missing other elements (including: we have already answered this 72 times and it is only thursday) and this is why I am asking here.