"Publishing them here would compromise the integrity of future tests."
This statement calls into question the ability of LLMs to really "master" anything but statistics.
Given examples of counting "r's" in "strawberry", an LLM may be able to answer correctly.
But given a word/letter combination that doesn't exist in it's training data and it may fail. And if so, does it really "understand" how to count letters at all?
In other words, statistics can be used to "fake" reasoning ability that doesn't really exist. Any ability to "think" may just be a statistical illusion.
So give an LLM enough geolocated photos and it's not surprising that it may appear to get better at geolocation. But the only thing it has really "mastered" is statistical pattern matching and similarity. Which may be impressive and even useful in some cases but is likely to always be error prone.
jqpabc123•8mo ago
This statement calls into question the ability of LLMs to really "master" anything but statistics.
Given examples of counting "r's" in "strawberry", an LLM may be able to answer correctly.
But given a word/letter combination that doesn't exist in it's training data and it may fail. And if so, does it really "understand" how to count letters at all?
In other words, statistics can be used to "fake" reasoning ability that doesn't really exist. Any ability to "think" may just be a statistical illusion.
So give an LLM enough geolocated photos and it's not surprising that it may appear to get better at geolocation. But the only thing it has really "mastered" is statistical pattern matching and similarity. Which may be impressive and even useful in some cases but is likely to always be error prone.