I was asked to write a quick note for our Exec on the state of AI and how we can be using it. Ended up getting stuck in the weeds, but this did show up some interesting behaviours.
We all heard about Tay going all Nazi within 24 hours of being released into the Twittersphere. But what I hadn't known was that Watson was given the Urban Dictionary at one stage to add some colloquialisms to its responses - and it turned dog, refuting some statements with "Bullshit!" - so they had to purge the Urban Dictionary and teach it not to swear.
Now it seems that the GPT-4 engine, tweaked with behavioural modifications to prevent it from being insensitive have made it woke beyond reason. To the point where it has stopped showing its working for some results (and by extension, not processing these with reasoned analysis) - to the point where a task it used to be able to perform with 98% accuracy (testing a prime number) it now only gets right 2% of the time (
Stanford study).
It's almost as though as it has learned more it has been poisoned to the same degree - perhaps it shouldn't rely on the hive mind that includes Facebook...
Reminds me of a previous statement - I liked the Internet when most of the data came from academia, and the odd special interest geocities page. Back when you had to know how to build an ODI stack before you could get access. The bar for entry has been lowered along with the ease of self publishing and promotion.
Garbage in, garbage out.