Editing this post to be deleted - the answer was solved.
How does GPT 3.5 and GPT 4.0 handle dates and time?
Any thoughts / docs on this would be extremely helpful.
its trained on all types of data, including events and historical events
The preprocessing done is rather minimal, just stripping personal data, html entities of sites, and such as you might also do on web scrapes.
The emergent abilities are rather a mystery, but the knowledge associations that let it know that “monday, tuesday,” is followed by wednesday, also let it know that "wikipedia: Pearl harbor … " also has 1941 as connected lexicography and associated semantic knowledge, its context, and that the number looks like a date.
The interesting thing about GPT-3+ and other leading language models is they don’t train on labeled data, dictionary words, or use a semantic encoder; it is the sheer amount of text and knowledge training that gives it abilities.
Major events of May 18, 1980?
On May 18, 1980, a major volcanic eruption occurred at Mount St. Helens in Washington State, USA. This eruption was one of the most significant volcanic events in the history of the United States. It resulted in…
I mean its literally just vectorized embeddings.
Based on my understanding… It looks up “pearl harbor” and than knows that pearl harbor and “1941” are similar based on a certain cross axis (or however you might say it), but then also knows that a historical event and a date are significant to one another based on a different vector. And probably a more complicated process than obviously just two vectors but I’m pretty sure its just the vector database doing its thing based on a massive dataset