If you’re thinking about building a GPT-3 app, this overview of how we did it might be useful. We cover how GPT-3 has been core to Viable from the get-go and how it fits the qualitative analysis use case.
GPT-3 is an advanced AI system that produces natural language text by predicting what comes next in a text sequence. It’s one of the largest neural networks available, with 175 billion parameters. GPT-3 was trained with large amounts of information from the internet.
Thanks to all that training, GPT-3 performs at state-of-the-art levels. It’s good at writing prose, doing translations, answering questions, summarizing text, and more—all in natural language.
A notable feature of GPT-3 is that it doesn’t need a lot of examples. This is known as few-shot: to perform a new task—like summarizing multiple paragraphs into a sentence—GPT-3 only requires a few examples of that task. GPT-3 can also do zero-shot tasks, no example needed. In response, the model generates text that tries to complete the pattern of the text provided. All this is done in a text-in and text-out interface. The text that goes in is known as a prompt and the text that comes out is the completion.
GPT stands for Generative Pre-trained Transformer, which is a type of the language model architecture and training methodology. That is about as technical as we’ll get here.
You can use GPT-3 through an API from OpenAI. You can experiment with different prompts until you find the ones that work for your use case. A prompt can be any string of text, making it possible for app developers to build a variety of applications with GPT-3.
A common GPT-3 use case is copywriting. There are quite a few GPT-3 apps that specialize in generating automated copy for website pages, social media, and digital ads.
Another use case is question and answer. GPT-3 is good at understanding natural language questions and providing relevant content in response.
If you combine the summarization capability—summarizing a single or multiple pieces of content in an automated way—with the question and answer capability, you can imagine a business application starts to emerge for self-serve access to content.
And finally it’s worth touching on the classification capability that allows you to classify content into any categories you’d like.
At Viable, we use GPT-3 for question and answer as well as summarization.
With GPT-3, we can do something nobody else has done before.
A little context first: companies get all sorts of feedback from their customers. Whether that’s customer support tickets via a helpdesk tool, responses to open ended survey questions, app store reviews, or tweets and Facebook comments from customers. All of this unstructured data has rich customer insights but takes time to read through, interpret, categorize, and tag.
Viable was built to solve this problem: we automate analysis of qualitative feedback.
But interpreting, categorizing, and tagging feedback isn’t enough for a business team to make informed decisions. They need context to know what to do. The best way to deliver context is to identify the problem, its severity, the root cause, and possible solutions—written in natural language. That’s where GPT-3 comes in.
As a business manager, you might want to know “How can we improve our dashboard for customers?” plus a dozen other things every week. At Viable we combine the question/answer and summarization capabilities from GPT-3 to directly answer natural language questions like this one, pulling from your customer feedback data. You just type your question into Viable and GPT-3 will answer in seconds by summarizing relevant feedback. See a short video of how it works.
Productizing automated feedback summaries in natural language for analytics hasn’t been done previously.
Read more here on how we designed and experimented with the prompts, with examples.
Viable uses multiple language models to structure data, including filtering out noise. Each data point is then assigned at least one topic category.
For instance, if a customer gives feedback about improving a company’s new dashboard by making it more customizable, Viable would assign some topics around dashboards and customizability. It would be considered as improvement feedback.
Now imagine that same company receives feedback from hundreds of users who similarly say the dashboard could be improved with more customization options. Wouldn’t you want to know that without having to read all the individual feedback? Wouldn’t you also want to know what customers would like to customize in the dashboard?
Viable provides exactly these answers. All that feedback is summarized in human language, like this:
Customers would like to be able to customize the dashboard with dedicated filters. We should add custom filters for date ranges, customer types, and contract volumes.
This summary gives you the context to make informed product decisions.
We use GPT-3 because it’s the best language model available for interpreting questions and providing answers straight from users’ feedback in a contextual, summary format.
Access to GPT-3 comes in a software-as-a-service format via an API. Pricing is based on data usage, tracked per month. This allows us to access the model’s functionality without needing to manage the AI model ourselves.
You can adjust quality, length, and other variables of the output in the GPT-3 Playground settings. These settings are useful to tweak prompts until you get the most useful output (in our case, feedback summaries).
We don’t dive into the technical and engineering side of building a GPT-3 app here. (That is a different article.) We instead cover the business strategy of GPT-3 apps.
To build an app that uses GPT-3, it helps to first understand whether you’ve got a compelling use case that clearly benefits from the model’s capabilities. Whether it’s copywriting, question and answer, summarization, translation, classification, search, explaining code, translating between programming languages, etc.
In our case, we wanted to make it easier to analyze and get insights from unstructured customer feedback. Doing analysis by asking natural language questions is a good fit for GPT-3.
It helps to assess the sustainability of your application with basic questions like: does your app solve a significant problem for your target market? If so, does it do so primarily with the natural language capabilities of GPT-3? How do other components of your product help solve the user problem?
Since GPT-3 can be accessed by anyone, it’s helpful to know how you’ll differentiate your app.
Once you’ve decided to try out GPT-3, you can sign up for OpenAI API access.
We recommend exploring the content OpenAI has on their website including documentation, tutorials, and examples of how GPT-3 works.
Once you're signed up, you'll have access to the Playground where you can start building prompts for your use case. Experiment with adjusting the settings for the outputs until you feel comfortable with them.
You’ll then want to incorporate GPT-3 into your overall product architecture. Decide where and how GPT-3 interacts with your other systems. Plan ahead for how outputs will be provided to your user base.
Make sure to build a great user interface and focus on a smooth, end-to-end experience for your users, from sign-up to the motions that create value for them. GPT-3’s magic alone will not cut it.
We recommend spending as much time as necessary to figure out how GPT-3 will work best for you. Our team of machine learning experts has spent hundreds of hours in the GPT-3 Playground and many more hours thinking through the machine learning system that resulted in a one-of-a-kind qualitative analysis experience for our customers.
If you’ve got thousands of feedback entries you’d like to analyze in an automated way, Viable might be a good fit. Our speciality is helping companies get insights from customer feedback quickly and easily.