Why write in geek, when you can describe in simple English?
Artificial intelligence, a creation of the human mind, is now progressing rapidly to aid in creating for humankind. One of the latest feats in the field of AI is GPT-3 or Generative Pre-trained Transformer 3 by OpenAI. The newest in the language models, GPT-3 is the third in line language prediction model in the GPT series with the potential to revolutionize industries; from publishing to coding. Here’s how.
What is GPT-3
GPT-3 is a deep learning algorithm that produces human-like text. Similar to other language models, this third-generation language prediction model in the GPT series is also trained with the use of machine learning. With pre-fed data, the model is trained to create content that has a language structure — human or machine language.
GPT-3 is considered as the largest model ever created with 175 billion parameters and trained on the largest dataset of any language model. However, it is not just the scale at which it is trained that makes GPT-3 revolutionary, the model requires minimal tuning even for very specific tasks, making it one of the most interesting and important AI systems ever produced.
How does it work
GPT-3 is trained using unlabeled text dataset, such as Common Crawl and Wikipedia. The training enables the model to fill missing words or phrases that are randomly removed from a text using surrounding words as context with human-like accuracy.
For instance, in the following sentence,
It’s going to rain today, but I forgot to carry my __
…GPT-3 model will be able to complete it based on the available contextual data and suggest the word ‘umbrella’.
The above example can also be considered as an advanced prediction model. However, with GPT-3, the potential extends beyond mere prediction in sentences as it holds the potential to cut down on more tedious tasks such as coding, writing research papers, building apps and websites with simple language-based instructions.
What can GPT-3 do
Put simply, GPT-3 can do what you can describe. One of the best ways to understand this language model is that it allows humans (even non-coders or non-geeks) to communicate with machines using simple English. It is as easy as it sounds and as vast as one can imagine. At its core, the model allows for the democratization of AI — a technology that is a work-in-progress and the more it reaches the masses, the more it can be refined for its positive use.
Let us now look at ways GPT-3 is being tested by early users and how it can benefit a wide range of industries going forward.
Have to push out an article based on limited data provided by an on-field reporter? GPT-3 will not only write out a full piece but give you multiple options while ensuring the tonality and length matches your instructions and editorial standards. Take for example this article published by The Guardian that was churned out by GPT-3. The human editors only put together the piece using segments from the number of different options provided by the model for one cohesive article that oozes perfection. In this case, the headline gives away the mystery, but consider the immense possibilities in the publishing domain, be it writing news, novels, blogs, or even research papers based on a few instructions.
Meanwhile, an early tester built a micro site that generates a GPT-3 generated text based on a word. To test how sophisticated the tool is, go to https://thoughts.sushant-kumar.com/word and replace ‘word’ at the end of the web address.
Anyone can learn to write code. But what if you don’t need to learn to code to be able to build apps or design websites? With GPT-3, it is possible to describe the requirement — a UI creator, a layout generator, write SQL queries, and more, without having to write the code in entirety. For instance, in the following use case, a react dice component was built using description in simple English and GPT-3 did its magic.
In another instance, early testers used GPT-3 to write Java code using just text description. Whereas in one use case, the tool could build a mockup website using a reference website’s URL and text description!
For spreadsheets, GPT-3 can be leveraged to match patterns, look up specific data, and take auto-completion to a whole new level. In the example below, early tester used the =GPT3() function to carry out tasks which included doing math, matching patterns, and looking up for data from the given information.
Now, this might make even Google squirm. GPT-3 beta users have tested the mighty potential of the tool by building a fully functional search engine. Not only does it return the right answer to any query, but also provides a corresponding URL. Take a look at how the search engine built over GPT-3 works:
As people continue to explore the crazy GPT-3 uses, the possibilities of using the language model for auto-completion or simplification of tedious language-based tasks are endless. However, the boon doesn’t come with its own set of downsides.
Some critics may consider GPT-3 a threat because theoretically, it can become a lousy writer’s tool to get work done fast, a student’s wild card to get into the top university without breaking a sweat when it comes to writing an SOP, building apps which can be simply described, websites or experiences with layouts that actually translate from imagination but may not truly be practical, or more threateningly, a rogue blogger’s tool to spread misinformation or fraud.
While OpenAI continues refining the tool and more work needs to be done to regularize technologies such as the GPT-3, data engineers like me can’t wait to test out the tool’s potential in the data science world. Open AI has released an API for accessing the new GPT-3 tool that allows users to try the new interface on virtually any English language task. You can also test out the potential of GPT-3 by requesting OpenAI request access to integrate the API in your product or build a brand new app of your own.
Follow Me on Linkedin & Twitter
If you are interested in similar content do follow me on Twitter and Linkedin