GPT-3 is a piece of software that can predict what the next word will be in the specific text. It can repeatedly do that, to the point where it can write a text that is almost identical to human writing. You just need to give it the first bit, it will do the rest.
What is exceptional about it is the degree to which it can ‘learn’ a task that needs to be completed by putting examples of that task into the text it needs to extend. For example, if you want to use GPT-3 to translate from English to Russian, you can give it a few lines of text being translated from one to the other, and then give it an English phrase, and it will continue the pattern by translating that phrase into English. It may also come up with a few phrases of its own, and translate them, too.
So GPT-3 is a general-purpose natural language generation tool that can be “trained” to do a task — even a task it has never seen before — with a few short examples.
Text generation is the most useful and impressive feature of GPT-3: it can be used on anything from generating fully-working code to blog posts and even guitar tabs. Remarkably, the program learns without any human supervision. Being fed with tons of different data from the internet, including Wikipedia, all kinds of blogs, news articles, and even twitter posts, it finds patterns based on statistical regularities, which it then uses to complete text prompts. The model can reproduce the rhythm, syntax, tone, and voice of a written task, such as poems of a famous poet or question-answer sets.
So, what can it be used for? Although, the technology is not widely accessible, it has already shown outstanding results in performing different tasks. Due to its high adaptability, it can be applied in various ways across multiple industries. And it’s incredibly easy to use: all the complexity of machine learning is simplified to plain English instructions. This is going to have a huge impact on the job market.
For instance, some people who took part in beta testing found that GPT-3 is able to write ReactJS code using a simple language description of a website. So it’s basically an all-in-one software development team.
Another important use GPT-3 was found in translating from plain English into Legalese. GPT-3 can easily deal with legal language, decoding it back and forth and creating texts like a real attorney. This feature is super helpful for non-lawyers who need to better understand and navigate contracts.
Using data and metadata extraction features, lawyers can significantly reduce the workflow and draft documents faster than ever. GPT-3 automates arguably the most boring part of work, like filling application forms, and it’s just one of many examples of applying this smart AI in legal industry.
To sum up, GPT-3 is a significant development in AI industry compared to narrow-specialized AIs developed before. Instead of describing detailed features of a problem, we let the neural network learn by itself. We don’t need to create specific programs to solve specific problems; we can solve all of them with one generalized approach. The use of GPT-3 can potentially improve many areas in legal tech, including: Legal research, legal contract clause databases, customer support, access to justice and legal advice.