LLMs don’t do time/date calculations at all. They’re just good enough at predicting the next word to trick us into thinking they can.
For your example, the WolframAlpha query "time in London at 1:39 pm MST on October 13" returns the correct answer, "9:39:00 pm BST | Sunday, October 13, 2024"[1].
While WolframAlpha's natural language processing is hit-or-miss, its responses clearly state assumptions made in the face of ambiguity. In this example:
Assuming "MST" is a named time zone
Assuming "time" is referring to a calendar computation
Assuming month/day
Assuming Mountain Time (United States; no observed DST rule)
Assuming London (United Kingdom)
[1] https://www.wolframalpha.com/input?i=time+in+London+at+1%3A3...It seems that WolframAlpha is venerable and stable. I first heard of it long, long ago. (Was it only released in 2009? Seems older.)
It must have amazing staying power. It seemed like a quite powerful tool. Why is it such a well-kept secret?