• Category
  • >Artificial Intelligence
  • >NLP

What is the OpenAI GPT-3?

  • Neelam Tyagi
  • Jun 15, 2020
  • Updated on: Jul 23, 2020
What is the OpenAI GPT-3? title banner

We came up again with a dazzling application disclosed by OpenAI, for redirecting, this is the link to the previous article discussing OpenAI Jukebox. 


OpenAI, an organization who focuses only to design artificial general intelligence devices or systems and build it intact for humans. No Terminator-like horrible imaginary place, no unbalanced machines that formulate human paperclips. Only computers with accustomed intelligence in order to help in solving large-scale computational problems.  


“OpenAI pursues unsupervised machine learning algorithms, that is abundant raw, and unlabelled data, to train where algorithms learn themselves by finding patterns in that data.”  


OpenAI first GPT model was schemed in the paper “Improving Language Understanding by Generative Pre-training” that is presented by Alec Radford, Karthik Narasimhan, Tim Salimans and Ilya Sutskever. It is the unidirectional transformer, pre-trained through language modeling across a lengthy corpus of widely broadened dependencies, the Toronto Book Corpus.


In the course of this blog, you will learn about the latest release of OpenAI GPT-3, its specification and its modelling performance. Beside that, a small glimpse of the previous release of OpenAI GPT-2 is also provided here.


Briefing NLP


Natural language processing (NLP) is an interdisciplinary realm of computer science, artificial intelligence and computational linguistics, accounting for the communication amidst computers and human (natural) languages, and, in context of, involved with computer programming to efficiently process enormous natural language corpora.


Natural language processing and understanding includes a broad variety of diversified tasks comprising textual ramification, question & answering, semantic assessment, and classification of documents.  


Despite huge unlabeled text corpora are plentiful, labeled data for understanding such specific tasks is sparse that makes it challenging for discriminatively trained models to function appropriately.


Understanding OpenAI GPT-2


OpenAI made headlines when it released GPT-2 that is a giant transformer that is based on a language model with 1.5 billion parameters, and was trained for predicting the next word in 40GB of Internet text, (Source). 


The dataset used was of 8 million web pages. It is a successor of GPT having the potential of operating with more than 10 times the parameters and was trained on 10 times longer the quantity of data.  


GPT-2 reflects a large number of  proficiencies incorporating the strength to produce conditional fabricated text samples of unheard-of quality where the model is fed with an input and forced to generate a huge extension. 


Additionally, GPT-2 outranks various language models that are trained on some particular domains such as Wikipedia, news, or books without demanding the use of  these domain-specific training datasets. 


On some language tasks including question answering, reading comprehension, interpretation, and description, GPT-2 initiates to understand these tasks from the raw text without implementing task-specific training datasets.  


(Suggesting you to visit our previous blog: OpenAI’s GPT-2 (Generative Pre-Trained Transformer-2) 


What is GPT-3, in actuality?


OpenAI, that is mentioned in the technological world consistently, has done it another time, after announcing GPT-2 last year, OpenAI came up with an open-source fastest NLP framework, that is known as GPT-3 today . 


According to researchers in the paper;

“GPT-3 achieves strong performance on many NLP datasets, including translation, question-answering, and cloze tasks, as well as several tasks that require on-the-fly reasoning or domain adaptation, such as unscrambling words, using a novel word in a sentence, or performing 3-digit arithmetic,”

The team also finds that “ GPT-3 can construct samples of news articles which people will get conflict in order to recognize articles transcripted by humans.”




  1. A state-of-the-art language model made up of 175 billion parameters. 

  • A parameter is a measurement in a neural network that deploys a large or small weightage to a few aspects of data, for providing that aspect larger or smaller importance in an entire measurement of the data. 

  • These are the weights that deliver shape to the data, and provide a neural network an understanding angle on the data. 


  1. It obtains outcomes on the SuperGLUE benchmark.

  2. For other benchmarks like COPA and ReCoRD, the model lets down with WIC (word-in-context) analysis.

  3. Minor fine-tuning is required in order to function explicit NLP tasks such as language translation, question-answering, composing poetry, and even basic maths.

  4. It can do three-digits addition and subtraction and quite prime at remedial English grammar. 

  5. It can achieve, as the authors described, “meta learning”, which means GPT neural network doesn’t demand to re-trained in order to perform a task such as sentence completion. 


Displaying sizes, architectures, and learning hyper-parameters in terms of batch size in tokens and learning rate of the models that is trained for a total of 300 billion tokens. (Source: OpenAI)

Size, architecture and and learning hyper-parameters , pic credit


  1. The parameters of architecture for each individual  model are conscripted on the basis of computational efficiency and weight-stabilizing in the model’s layout around GPU’s.

  2. Each model is trained on NVIDIA V100 GPU’s on the section of  steep-bandwidth cluster furnished by Microsoft.



The approach used


The model is constructed using the basic concept of Transformer, Attention, etc, for pre-training a dataset composed of Common Crawl, Wikipedia, WebText, Books and some additional data sources.  


The model was evaluated against various NLP benchmarks, achieved state-of-the-art performance on question answering tasks, and closed-books that set a new record for language modeling.


The researchers trained an array of smaller models, varying from 125 million parameters to 13 billion parameters for comparing their efficiency counter to GPT-3 on the three settings. 


The following graph displays the profits in terms of accuracy for various zero, one and few shots as a function of number of model parameters, it can be observed that huge gains are obtained due to size-scaled up.  

Displaying the benefits in terms of model accuracy and number of model parameters in various three shots.

Graph of benefits in context of accuracy and number of examples, pic credit


For most NLP tasks, researchers identified comparatively gentle scaling across the model capacity in all three settings, and noticed that a pattern where the breach among the zero-, one-, and few-shot performance grows generally  along with model capacity that leads larger models are great meta-learners proficiently. (in accordance with)



Factual comparison amid GPT-2 and GPT-3 


  • GPT-2 can produce artificial text in a feedback to the model that is being prepared with an arbitrary input. It readjusts according to the style and content of the conditioning text. It enables the user to make realistic and comprehensible perpetuation for a topic of their choices. If asking about extensive language models, it has a whopping 1.5 billion parameters.


  • GPT-3 is upgraded with 175 billion parameters, it tailors and escalates the GPT-2 architecture, it also involves adjusted initialization, pre-normalization, and changeable tokenization. It reflects substantial performance on various NLP tasks and benchmarks in three distinct shots, i.e. zero-shot, one-shot and some-shot environments. 




OpenAI has recently unveiled the latest epitome of its eye-catching text generator, GPT-3 that has 175 billion parameters, 10 times longer than its predecessor GPT-2 that has 1.5 billion parameters. 


GPT-3 can execute an amazing bandwidth of natural language processing tasks, even without requiring fine-tuning for a specific task. It is capable of performing machine translation, question-answering, reading conceptual tasks, scripting poems and elementary maths.


Read our another blog on About OpenAI GPT-3 Language Model here. Continue your learning with recent articles and connect with us at  Facebook, Twitter, and LinkedIn.

Latest Comments