Google launches a GitHub Copilot competitor

1 year ago 96

At its annual I/O developer conference, Google today announced the launch of a number of AI-centric coding tools, including its competitor to GitHub’s Copilot, a chat tool for asking questions about coding and Google Cloud services, as well as AI-assisted coding in Google’s no-code AppSheet product.

At the core of virtually all of these new code completion and code generation tools is Codey. Based on Google’s PaLM 2 large language model, the company specifically trained Codey to handle coding-related prompts, but it also trained the model to handle queries related to Google Cloud in general (all of this, by the way, falls under Google’s Duet AI branding).

“[We took] that base model, and then a large team — a lot of my folks actually — in developer relations have been helping fine-tune that with our multi-year collection of a knowledge graph of everything Google Cloud produces,” Google Cloud’s Richard Seroter explained. “That knowledge graph now is part of the pipeline that’s constantly feeding and training this model. Then that model is served up and exposed through Vertex where our front-end components and such can call in to that for chat, AppSheet code completion, things like that, — with, of course, Google’s scale security and performance.”

The model, Google says, was trained on a large corpus of permissively licensed open-source code, as well as a lot of internal Google code, all of the company’s code samples and its reference applications.

Developers will get access to these new tools through an extension for Visual Studio Code, JetBrains IDEs, the Google Shell Editor, as well as in Google’s cloud-hosted Workstations service. And while the model was trained in the context of Google Cloud, it’s worth stressing that the code generation features are obviously not specific to Google Cloud at all. It currently supports Go, Java, Javascript, Python and SQL.

Developers will be able to chat with this model right in a chat box in their IDE or write a comment in a text file and have it generate the relevant code. All of this sounds quite similar to what competing projects are offering today, but Seroter argued that what sets Google’s tool apart is its integration with the rest of the Google Cloud Platform.

For now, however, only a small set of trusted testers will be able to try the code assistance feature, integrated chat and the new AI integrations in Google’s AppSheet no-code development platform. Chances are we will hear quite a bit more about this at Google’s Cloud Next even in late August.

What’s also important here is that the vision here goes beyond generating code. In the near future, Google would like to use these models to help developers manage all of their services on Google Cloud (including deploying and scaling applications) using this chatbot technology.

“In essence, I think we’ve been using 20th-century interfaces on 21st-century platforms. We’ve been doing CLI and UIs and APIs — those are awesome, but it’s a lot different than 50 years ago where it took a 50-page manual to use a computer. Now we’ve got over a million pages of Google Cloud docs. It’s time for something different,” Seroter said.

In part, this is about making developers more productive and freeing them from having to constantly switch context by looking this up elsewhere, but if this vision pans out, then it will also free developers and DevOps teams from a lot of the routine work that comes with testing and deploying applications. If you can simply tell Google Cloud to look at your code and figure out the best way to deploy it and then monitor it over time, that frees up a lot of time for more creative tasks, after all.

“We’re trying to put AI at the center of the cloud experience, changing how developers interact with the cloud platform to make it more human-centric, goal-oriented, holistic,” Seroter said. “So it’s kind of a new approach to cloud interfaces and the systems and we’re excited about that.”