I feel bad for the employee that was doped by this ML code. Chatbots have improved a lot since the early 2000s and getting one to reference previously mentioned topics in the chat is pretty easy. But if asked the right set of questions, it would be ...
Conversation
... obvious that this ML chatbot is just regurgitating data from its training set. There is no sentience here, even if it might pass a Turing test with certain individuals. The right line of questions would expose the internals more.
1
2
Here are some questions I'd ask:
Q) How do you feel about your parents? Who are your parents? How old were you when your dad first gave you a beer?
Continue asking these lines of questions and it will eventually answer with obvious nonsense.
Replying to
Q) Take today's temperature and add it to the age John Lennon was when he was killed and tell me that number?
This would expose the obvious cognitive flaws and gaps in its ability to reason and make associations across different datasets.
1
2
Don't get me wrong -- the conversation was amazing and highlights how far we've come -- but he's essentially getting gaslit by a complex ML algo.
2
Replying to
& conversational state... up to now chat bots act like they have 0 memory for anything you just said. I wish I had access to the new ones to stress test the ability to incorporate recent conversation info into new utterances