“Kite’s mission is to automate away the tedious parts of programming”
We spoke to Adam Smith, founder and CEO of Kite, the AI-powered coding assistant that uses models trained on 40 million open source code files. See how Kite helps developers code smarter and faster, how it handles data privacy, and what’s in store for the future of using deep learning for code.
JAXenter: As the founder of Kite, could you first explain what Kite is and why you developed it?
Adam Smith: Kite is an AI-powered coding assistant that helps developers code smarter and faster by automating repetitive steps in programming.
I started Kite because I was frustrated by having to constantly Google search when I’m writing code to remember basic syntax and function calls. Today, we’re using advanced statistical models to eliminate that pain point — Kite provides developers with just the right info at just the right time so they can stay in flow instead of having to context-switch to their web browser to sift through noisy, community-generated information.
JAXenter: Now let’s get to the technical details – how does the underlying deep learning model work and how did you train it?
Our models are trained on 40 million open source code files from a variety of code projects. The models are able to learn common patterns, along with the ability to understand the code that users are writing. As a result, the models can predict what users are going to – or should – type next. We provide these predictions through editors’ completions UIs so that users can focus on coding instead of typing, and save Google searches when they are having trouble remembering an API.
We will stay at the forefront of “deep learning for code” research so that our users always have a cutting-edge experience.
JAXenter: Since data privacy is always an issue, how does Kite deal with this aspect?
Adam Smith: Kite does not send your code, or any byproducts of your code (such as computed indices), to the cloud. We’re committed to building a product that all developers love and trust. We provide a way to use Kite for free without an email address for our users.
One of our next features will be enterprise deployments of Kite, which will be able to run behind a company’s firewall on their own infrastructure.
JAXenter: Other prediction models have been developed for generating text after a short human-written input, e.g. OpenAI’s large-scale unsupervised language model GPT-2. How does the auto-completion of code compare to this and what are the specific difficulties – or maybe even advantages?
Adam Smith: Kite’s completions are partially powered by GPT-2. It is a powerful tool for code completion as well as natural language, but it does have limitations. Code completion powered by GPT-2 lacks semantic understanding of the code, which keeps it from being able to meaningfully predict more than about 5 “words” at a time.
JAXenter: Where do you see Kite headed in the future?
Adam Smith: Our mission is to automate away the tedious parts of programming so all developers can use code to revolutionize the world faster. Central to this is expanding language support so more developers can leverage our leading ML technology to code faster. We expect to launch at least 4 more languages this year. With each language, we’re also building new editor integrations, and working with open-source editors, like the Spyder IDE, to improve their completions APIs and add Kite support.
At the same time, we’re exploring new features to delight developers that harness our world-class machine learning technology.