"Some of these systems’ abilities go far beyond what they were trained to do—and even their inventors are baffled as to why." https://www.scientificamerican.com/article/how-ai-knows-things-no-one-told-it/ #AI #ArtificialIntelligence
Source: https://twitter.com/sciam/status/1656696548435148812
@bibliolater
A very interesting point (that when an LLM gets very good at predicting continuations, it may do it by developing things that look a lot like mental models) expressed really badly.
"Doing things they were not trained to do" is an utterly misleading way of describing the situation. They weren't trained to do any specific thing at all, except produce plausible continuations. Anything that that implies, from writing a bland thank-you letter to urging a reporter to leave his wife, is to exactly the same extent "something it was not trained to do".
This kind of wording just encourages people to have inaccurate ideas about how LLMs actually work.
Grumble grumble! :)
@ceoln Thank you for your comment. The area of LLMs and #AI are relatively new technologies. A discussion of the principles behind such technologies can only assist the general public in gaining a better understanding of what humans may face in the future with ever increasing digitalisation and automation.
