Read on Twitter
I have spent much of today studying the Stochastic Parrots paper by @emilymbender, @timnitGebru, @mcmillan_majora and @mmitchell_ai. ( http://faculty.washington.edu/ebender/papers/Stochastic_Parrots.pdf) This paper is important for many reasons...
It explains why language models like GPT3 are in large part gimmicks: stochastic parrots. They are just generating randomish sequences of sentences from the data put in to them. (1/n)
These can be a bit of fun. But they don't communicate with us: they just repeat random fragments of Reddit, Wikipedia and sensationalised newspaper articles. So it's not clear how they are useful in a wider setting. (2/n)
They might be improved, but in order to do so we need to have better control over the quality of data. If we had well calibrated data, then we can use them in well-specified tasks. Like translation of documents in the EU, for example. (3/n)
But instead of improving data some researchers in the Tech industry are just aiming to get the biggest data sets possible and let their training run wild. This is a waste of time, money and of environmental resources. (4/n)
And just because data is big doesn't mean it is diverse. The end result will be unusable (other than to make 'fun' examples), but if it is used it will encode racism, sexism and ageism that is already in the training data. It will be unaccountable. (5/n)
The reason this paper is so important is that it ties together the overblown hype about AI, an explanation of what language is (communication between real people), how cultural imperialism skews our view of the world and good practice in machine learning (6/n)
My short series of tweets doesn't do it justice. But I can say this: it really raises the bar for what is expected when researching AI. We have to understand so much more than just the technical side of what we do. Read it! Link again: http://faculty.washington.edu/ebender/papers/Stochastic_Parrots.pdf
