ChatGPT !

[Flying Carrot Steve]

 

I've found AI to be wrong quite often, it is nowhere near perfect yet.

 

I'm gonna start my own, it'll be called ChatS**T

[corsechris]

AI isn't any more immune to GIGO than people are. We've already seen how easy it is to create an AI with a particular...er...bias.

[dvd8n]

These AI/LLM programs don't really have any intelligence, they're really just convincing sounding bull**** generators. 

 

They're statistical models, there's no problem solving ability there. 

 

This is really well demonstrated by the recent amusement at chatgpt's inability to answer the question "how many letter Rs in strawberry".  Not only would it incorrectly answer 2, it would steadfastly argue that it was right. 

 

Because they are trained by analysing large amounts of text and developing statistical models about what word follows another word, they are really just complex guessing engines.  The problem (specifically in this case, but in general also) is that there's not a lot of learned argument around about how many Rs there are in strawberry because anyone over the age of about 5 would just count them, get a definitive answer and get on with their day.  But that's a process that is just beyond LMMs; they just don't even have that simplistic level of problem solving ability.  So all they can do is answer that, for that 'shape' of question, the answer is probably 2.

 

But because they answer with such disturbingly good English people grant them authority that they just don't deserve. 

 

Thank you for attending my Ted talk 🫣