Human context rot

What I used to call "Bayesian recognition" is, I realise, a special case of "human context rot". And we can use our understanding of LLMs to understand our own thinking better!

Several years ago, after I’d attended the wedding of a friend with whom I shared two schools, I had written about “Bayesian recognition”. I had written:

As you might expect at an event where you and the host share more than one social network, there were a lot of familiar faces. Some people I knew fairly well, and could easily recognize. But the others had to go through a “Bayesian search”.

So when I saw someone who was one of three people I know – let’s say X, Y and Z. In order to determine which of these this person is, I would ask myself two questions – firstly, what were the prior odds that the person I saw could be each of X, Y or Z. Secondly, what were the odds of each of X, Y and Z being there at that event. Note that the latter is important. For example, if someone at the event looks like you and I know (for example) that you are currently in another country, despite the strong resemblance I can discount the possibility that that person is you, and go ahead with my search.

I faced a stronger version of this this morning when I met a close friend for breakfast. Now, he and I share THREE schools. And the conversation frequently switched from one of these schools to another. As we kept switching context, at some point of time I realised that sometimes I struggled to place who he was talking about because it took me a few moments to understand “which of the schools I need to search in”.

In some sense this is similar to context rot that we sometimes see with LLMs - when there is too much context in a conversation, or if it goes on for way too long, then the LLM can lose context on what we discussed and start saying or doing random things. In some sense, the “relevant search space” has gotten way too large, and so sometimes it “looks in the wrong place” and comes up with random answers or explanations.

This morning, in some sense, I had “loaded all our three shared schools as context into my head”. And as the conversation meandered on, frequently I found myself having to think twice (literally) to figure out what we were really talking about! In other words, I was facing “human context rot”!

In any case, one of these shared schools has its own very strong language. And I’m happy to report that every time we spoke about something broadly related to that school, I switched to that language. And when we spoke about context regarding other schools, I switched back automatically to “normal English”!