At GitHub’s Universe event, the Microsoft-owned org talked of AI and little else – Copilot is no longer restricted to models from OpenAI, and new natural language development means “any person can create and share an application.”

“GitHub is still GitHub. We are an open development platform,” declared CEO Thomas Dohmke, introducing the ability to select Anthropic Claude 3.5 Sonnet or Google Gemini 1.5 Pro alongside models from OpenAI – previously the only source for Copilot. Exactly what he meant is arguable, but Dohmke may be implying that the relationship with parent company Microsoft – which owns up to 49 percent of OpenAI – is not a bar to using competing models.

‘Leaving aside developers who head straight to OpenAIs ChatGPT, Copilot may be the most-used AI assistant. If one ranks Visual Studio Code (VS Code) extensions by the number of installs, Copilot is top with 21.7 million, followed by Copilot Chat (16.1M), Tabnine (8.3M) and Codeium (1.8M).

Copilot is not the best-rated though, and extending the choice of models which drive it may help GitHub to improve Copilot’s popularity.

Integration with VS Code is another lever to pull developers towards Copilot, and one where the Microsoft relationship seems to grant special privileges. At Universe, developer advocate Cassidy Williams showed the new ability to edit multiple files in one operation – an important feature considering that adding or modifying a feature often involves editing several pieces of code.

Williams also demonstrated Copilot custom instructions – a Markdown file within a GitHub folder that adds project-level hints to help Copilot behave optimally. Copilot is also now available for Apple’s Xcode IDE.

A side effect of AI is the rise of Python. “Because Python is so heavily used across machine learning and AI development, Python has risen to become the number one programming language on GitHub, surpassing JavaScript,” revealed Dohmke, referencing a new Octoverse report. The statistic is somewhat misleading, in that the sum of JavaScript and TypeScript usage exceeds Python, and TypeScript compiles to JavaScript. Indeed, the report explains: “Rather than a slowdown in the JavaScript community, what we are seeing is a transition to TypeScript for a large proportion of new commits”

Other GitHub AI news included new models and new features for GitHub Models, which lets developers compare the performance of different models and generate code in one of several programming languages. Previously in limited beta, GitHub Models is now available to all users.

GitHub Workspace – previously subject to a waitlist – is now in public preview, with new capability. Workspace began with AI-driven agents for specifying, planning and implementing application features, and now there are two new agents – one for brainstorming and another for building and repairing code. Workspace is integrated with the GitHub browser-based user interface, so developers can get and use code suggestions without the use of an IDE, though there is also a VS Code extension. Another feature is that developers can run the code via integration with GitHub Codespace cloud development environments. “This is the developer environment of the future,” enthused director of international relations Karan M V at Universe.

The ability for non-technical users to create applications – sometimes called citizen development – has been touted for years by various vendors, in the form of low-code or no-code development tools. Such tools have a mixed reputation. One common issue is that it is easy to get started but hard to implement advanced features, or that applications may work but with poor design causing bugs and maintenance challenges.

Striking a Spark

At Universe, Dohmke introduced what he called “the single most revolutionary thing we have built for professional developers since Copilot itself.” This is GitHub Spark, though it actually targets not just developers but more or less anyone. The product, he said, “blurs the line between the developer and the everyday person … any person can create and share an application in a matter of minutes.”

Spark is a tool to build applications from natural language prompts. It is described as “very early” and is designed for small applications, with examples including “life management tools, learning aids, silly animations, and news clients.” Applications are cloud-hosted on a managed runtime, with access to data storage based on Azure Cosmos DB. The intent currently is for individuals to create fun and personal tools rather than business applications. But the premise – that “the complexity of creating bespoke apps is too high” – is the same one that drives all low-code development environments.

Workspace and Spark both share the notion of raising the abstraction level of development via AI tools – from assistance with coding to what Dohmke called “agentic programming,” where an agent is “AI assistance that executes tasks at your direction.” GitHub is not the only org working on this – we saw Stackblitz promote the same concept recently, and Vercel’s v0 is another example. Expect more of this from other vendors.

It was notable, though, that this highly AI-centric keynote avoided hard questions. For example, Anthropic co-founder and chief scientist Jared Kaplan came on stage and said nothing of substance. What are the limitations of AI development? What is the impact on software quality, or the possibility of critical bugs going unnoticed? What are the boundaries between AI and simply copying other people’s code? Developer views vary, but the hype around AI is vendor-driven and some will agree with Linus Torvalds: “90 percent marketing and ten percent reality.”