Passing the Turing Test is a goal that many in the machine learning community hope for. It’s a test which if passed, indicates that the behavior of an AI is in-differentiable from a Human Being. In many cases, many applications in machine learning achieve performances that far surpass Human ability (think object recognition). However, in terms of conversational agents coming across as cognisant, we’ve been way off for some time.
Conversational agents have struggled to act ‘human’ for a variety of reasons. From an observational perspective, AI tends not to be able to distinguish common ‘human’ conversational techniques (such as sarcasm, rhetoric, double negatives, alternative meanings, etc), and this (mathematically) is large because traditional models would use a ‘question-bank’ type of method and pick an answer from the most similar question. However, Deep Learning is crushing any historical competition in this space.
How does GPT-3 Work?
If you’re reading this article, then you’ve probably heard of GPT-3, which is a language model that’s been developed by the OpenAI organization. The development procedure behind GPT-3 was not novel per se (compared to its predecessors), but it was just bigger and better.
I’m assuming the reader has an understanding of machine learning but as a quick overview: a language model is a neural network that takes in a sentence that has been augmented into the form of a ‘word-embedding’. This form makes it much easier to build a probability distribution around each sentence and thus narrowing the scope of the problem.
Now, Language models have been around for some time but were limited in their ability to actually create reasonable answers. If not tuned correctly, probabilistic models are just pretty crappy at generating responses, but GPT-3 has been such a big success because the results have been comparatively great.
#machine-learning #philosophy #programming #artificial-intelligence #psychology #deep learning