OpenAI release of GPT-3

First published:

Last Edited:

Number of edits:

I have seen a lot of articles about GPT-3 , a new generative pre-training transformer 1 . It is, fundamentally, an unsupervised learning algorithm that was trained on a gigantic corpus 2 of 570GB of compressed plain text. Includes all of Wikipedia in English, corpora of books and something called the CommonCrawl .

The interesting thing is that OpenAI released the tool as an API that some people could use to build upon. Very quickly, people came up with tools that could build websites based on simple descriptions such as "A website with a search bar, the Google logo on Top and two buttons below, one saying search and the other I'm feeling lucky".

The true advantage of GPT-3 comes from not being trained for a specific task as many previous artificial intelligence models. In their paper they claim that regardless of this, the model already performs as good as models trained for specific objectives. You may think about translating from a language to another, or generating paragraphs of new text as specific tasks.

To my understanding, what makes GPT-3 very powerful is that the parameters where optimized using a gigantic dataset, but it can also be trained on specific (non open) datasets. For example, we could feed all the customer service chats to make it domain-specific, but using the accumulated knowledge. In this way, a company could build much more intelligent chatbots for customer service.

Of course, news about machine learning and such always brought up the same problems regarding fears of losing jobs in the hands of artificial intelligence . A lot of people 3 are still struggling to understand the long-term picture.





These are the other notes that link to this one.


Share your thoughts on this note
Aquiles Carattino
Aquiles Carattino
This note you are reading is part of my digital garden. Follow the links to learn more, and remember that these notes evolve over time. After all, this website is not a blog.
© 2021 Aquiles Carattino
This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License
Privacy Policy