GitHub launched Copilot Enterprise, a $39/month code completion tool and developer-centric chatbot for big enterprises, today. Copilot Enterprise contains all business plan capabilities, including IP indemnification, plus other key features for larger teams. References to an organization’s internal code and knowledge base are key. Copilot is now connected with Microsoft’s Bing search engine (in beta), and users may soon fine-tune models based on a team’s codebase.
New developers may ask Copilot how to deploy a container image to the cloud and obtain a process-specific response. Even though Copilot may assist with coding, many engineers struggle to be productive while switching firms since they don’t grasp the methods.
Many teams store their documentation on GitHub repositories, making Copilot’s reasoning straightforward. GitHub CEO Thomas Dohmke told me that since GitHub stores almost all of its internal documents on the service and recently gave all of its employees access to these new features, some people have started asking Copilot about vacation policies.
Dohmke informed me that clients have wanted these tools to reference internal data since Copilot’s inception. He noted that organizations have a process or library to use, and many have internal tools, systems, and dependencies that developers do not have at home or in open source.
Dohmke suggested using Bing to query Copilot about changes since the model was trained (imagine open source libraries or APIs). This functionality is only accessible in Enterprise, but Dohmke wouldn’t say whether it would come to other editions. I wouldn’t be shocked if GitHub brought it to other tiers later.
Fine-tuning, which launches shortly, will likely remain an enterprise feature due to its cost. Dohmke said, “We let companies pick a set of repositories in their GitHub organization and then fine-tune the model on those repositories. We abstract the complexities of generative AI and fine-tuning from the client and allow them to utilize their codebase to develop an optimum model for Copilot situations. This implies the model can’t be as current as with embeddings, talents, and agents (like the new Bing agent). He claims that all of this is complimentary, and consumers testing this functionality are noticing considerable gains. This is particularly true for teams working with codebases in less popular languages like Python and JavaScript or internal libraries that don’t exist outside an enterprise.
In addition to today’s release, I questioned Dohmke about Copilot’s future. More Copilot in more locations is the solution. I believe we’ll see more attention to the end-to-end experience of placing Copilots where you currently work rather than establishing a separate destination to copy and paste items. GitHub is delighted about the chance to put Copilot on github.com, where people are already working and producing the world’s software.
GPT 3.5 Turbo powers the auto-completion capability, according to Dohmke. GitHub never pushed the model to GPT 4 because of latency limitations, but Dohmke said the team has changed it “more than half a dozen times” since Copilot Business launched.
Google differentiates its price levels by the scale of the models that underpin those experiences, while GitHub may not. Different use cases need different models. Different optimizations—latency, accuracy, result quality, responsible AI—for each model version ensure that the output is ethical, compliant, secure, and doesn’t create lower-quality code than our clients want. “We will continue to use the best models for the different Copilot experiences,” Dohmke added.
