Why do #LLMs perform so well on tasks originally designed to measure human-like intelligence?
Not necessarily because they actually have human-like intelligence but rather because they can find statistical patterns to exploit within those tasks.
Here's a lovely example. While applied to an earlier LLM known as BERT, the basic story may be applicable to the systems currently in the news.
Follow
@ct_bergstrom My impression of ChatGPT-3 (haven't tried ver. 4) is that it is like a bullshitter who has memorized every book in the library. For many tests, recalling stuff you've read and bullshitting to interpolate/extrapolate to novel stuff evidently works very well.
