GitHub Universe 2024: Full Breakdown of Copilot Multi-Model Support, Multi-File Editing & Spark

GitHub Universe 2024 Deep Dive: The AI Native Development Era Officially Begins

GitHub Universe 2024 delivered a series of major updates—Copilot now supports multi-model selection and multi-file editing, and the brand-new AI tool Spark was officially released. GitHub is moving from "AI-assisted development" to a new stage of "AI Native development." This article breaks down the core highlights of the conference one by one.

From AI Infused to AI Native: A Paradigm Shift in Development

GitHub introduced a key concept at this conference: AI programming is transitioning from its first phase (AI Infused) to its second phase (AI Native).

Phase One (AI Infused) encompasses three layers:

• Integrating AI elements into traditional applications, such as code completion and syntax suggestions
• Conversational programming with AI, such as using ChatGPT to help write code
• Supporting multiple AI models to accomplish different tasks

Phase Two (AI Native) means:

• AI is no longer just an assistive tool but a core development engine—the development process natively depends on AI capabilities
• AI Agents can autonomously complete complex tasks, make decisions based on context, and dramatically reduce developer intervention
• Flexibly switching between different AI models and selecting the optimal model for specific tasks

Background: The Origins of the AI Native Concept

The AI Native concept draws from the "Cloud Native" philosophy in software architecture. Cloud Native means applications are optimized for cloud environments from the very beginning, rather than migrating traditional apps to the cloud. Similarly, AI Native development means the entire development workflow, toolchain, and mindset are redesigned around AI capabilities, rather than "plugging in" AI features to existing tools. This transition has a clear historical trajectory: when GitHub Copilot launched in 2021, code completion was the core selling point; in 2023, ChatGPT sparked the conversational programming wave; and in 2024, tools like Cursor and Devin began letting AI truly "take over" parts of the development process. The rise of AI Agents is the key driver of this transformation—Agents don't just answer questions; they can plan steps, invoke tools, execute code, handle errors, and form a complete "perceive-decide-act" loop, which is the technical foundation of the AI Native development paradigm.

The core idea behind this shift is: AI will progressively transform traditional development patterns. AI agent tools like Claude's Celine (formerly Dev) will become indispensable roles in the development workflow.

Copilot Multi-Model Support: Claude 3.5 Sonnet and Gemini Join the Lineup

One of the most exciting announcements is that GitHub Copilot now officially supports multi-model selection:

• Anthropic Claude 3.5 Sonnet (latest version): Rolling out starting next week
• Google Gemini 1.5 Pro: Coming in the following weeks
• OpenAI o1 series: Including o1-mini and o1-preview, now fully available to individual users

Background: Capability Differences and Use Cases of the Three Models

Behind the multi-model strategy lies an important reality in the current LLM ecosystem: different models excel at different tasks, and there's no single "universal champion." Claude 3.5 Sonnet, developed by Anthropic, emphasizes "Constitutional AI" safety principles during training, supports an ultra-long context window of up to 200K tokens, and excels at understanding the overall structure of complex codebases and generating high-quality code. Google Gemini 1.5 Pro is known for its context window of up to 1 million tokens and native multimodal capabilities, making it ideal for handling mixed inputs containing images and documents—particularly well-suited for development tasks that involve analyzing design mockups or technical documentation. The OpenAI o1 series represents "slow thinking" models, trained through reinforcement learning to develop Chain-of-Thought reasoning capabilities, far surpassing traditional GPT-4 in mathematical proofs, algorithm design, and complex logical reasoning—though with slower response times and higher costs, making them best suited for deep debugging scenarios that demand extreme accuracy.

This means developers are no longer locked into a single model and can flexibly choose based on task characteristics. For example, Claude 3.5 Sonnet excels at code comprehension and generation, while the o1 series has unique advantages in complex reasoning tasks. Choosing different models for everyday coding versus deep debugging can deliver noticeable efficiency gains.

Major Copilot Overhaul in VS Code: Multi-File Editing & Web Search

Multi-File Editing

Copilot's interface in VS Code has undergone a major redesign. The original GitHub Copilot Chat window has moved to the right side and is now split into two core functional areas:

• Chat window: Retains the original chat functionality
• Edit window: A brand-new multi-file editing feature, similar to Composer in Cursor

The multi-file editing feature allows developers to initiate an editing conversation across a set of files, view modifications to multiple documents in the right panel, selectively Accept or Discard changes one by one, and even drag documents directly into the editing area.

Background: The Technical Principles Behind Multi-File Editing—RAG and Vector Indexing

The core challenge of multi-file editing lies in "cross-file context management." Traditional code completion tools only focus on local context within the current file, while multi-file editing requires AI to simultaneously understand dependency relationships, interface contracts, and data flows across multiple files. Cursor's Composer feature was the first to commercialize this capability, with its underlying architecture relying on building a Vector Index across the entire codebase—converting code snippets into high-dimensional vectors and retrieving relevant code segments through semantic similarity, dynamically injecting them into the AI's context window. This Retrieval-Augmented Generation (RAG) architecture enables AI to access the most relevant code information within a limited context window, and is the mainstream technical approach for current AI programming tools handling large projects. GitHub Copilot's @Workspace feature uses a similar principle, supporting local indexing of up to 2,000 files. The Accept/Discard interaction design draws from Git diff's code review paradigm, allowing developers to review each AI modification in a familiar way, lowering the psychological barrier to accepting batch AI changes.

The @Workspace feature is similar to Cursor's @Codebase, enabling local work