All of the features of the current Business plan are included in GitHub Copilot Enterprise AI, including IP indemnification, but it also adds some essential capabilities for bigger teams.
The ability to refer to an organization’s own knowledge base and code is the standout feature here.
Additionally, Copilot has been connected with Microsoft’s Bing search engine (which is still in development), and soon users will be able to adjust Copilot’s models using the team’s current codebase.
With that, new developers joining a team may ask Copilot, for example, how to launch a container image to the cloud and receive a response tailored to their company’s procedures.
After all, for many engineers, being productive while switching firms is more about learning the various procedures than it is about understanding the codebase, although Copilot may undoubtedly aid in understanding the code completion as well.
Since many teams already save their documentation in GitHub repositories, Copilot can reason over it really easily.
In fact, according to GitHub CEO Thomas Dohmke, since the company recently granted all of its staff access to these new features and stores almost all of its internal documents on the platform, some users have begun using it for non-engineering inquiries as well. For instance, they have begun asking Copilot about vacation policies.
Dohmke informed me that from Copilot’s inception, users had been requesting these capabilities to access internal data.
Since corporations have procedures to follow or specific libraries to use, many of the things developers do there differ from what they do at home or in open source. Additionally, many of them have internal tools, systems, and dependencies that don’t exist such that on the outside” he noted.
Dohmke mentioned that the Bing integration would be helpful for inquiring with Copilot about items that could have changed since the model was first trained (such as open source libraries or APIs).
As of right now, this feature is exclusive to the Enterprise edition. Dohmke would not comment on whether or not it would be added to the other versions, but I wouldn’t be shocked if GitHub eventually added this option to the other tiers as well.
Fine-tuning, which launches soon, is one feature that will probably stay a corporate feature due in part to its related expense.
Dohmke said, “We allow companies to select a subset of repositories within their GitHub organization and then fine-tune the model on those repositories.”
“We’re removing the customer’s need to deal with the complexities of generative AI and fine-tuning and allowing them to use their codebase to generate an optimized model that is then used within the Copilot scenarios.”
But he also pointed out that this implies the model can’t be as current as it would be if embeddings, talents, and agents (like the new Bing agent) were used.
However, he contends that everything is complementary and that users who are now trying this functionality are reporting notable gains.
This is especially true for teams working with codebases in less widely used languages like Python and JavaScript, or with internal libraries that aren’t really available outside of a business.
In addition to discussing the release today, I also questioned Dohmke about his overarching plans for Copilot’s future.
In essence, the response is “more Copilot in more places.” Over the course of the next year, I believe there will be a greater emphasis placed on the end-to-end process of integrating Copilots where you now perform your job, rather than developing a separate location where you can copy and paste content.
That, in my opinion, is the main reason why the GitHub team is so thrilled about the chance to provide Copilot on github.com—that is, where engineers currently collaborate and create the majority of the world’s software.
Speaking of the underlying technology and its future direction, Dohmke noted that GPT 3.5 Turbo is now the platform on which the auto-completion function operates.
GitHub never transferred the model to GPT 4 because of its latency restrictions, but Dohmke also mentioned the team has revised the model “more than half a dozen times” since Copilot Business launched.
It doesn’t appear like GitHub will differentiate its price levels based on the scale of the models that drive those experiences, as Google has done.
“Different models are needed for different use cases. For each model version, many optimizations—latency, accuracy, outcome quality, and responsible AI—play a significant part in ensuring that the output is secure, ethical, and compliant and does not produce code that is of a lesser caliber than what our clients want.
We’ll keep employing the greatest models available for each component of the Copilot experience, according to Dohmke.
