AI: What It Can and Cannot Do in SRS Development

[winlinvip]

AI has been playing a significant role in the SRS community, and while it can handle many tasks effectively, there are certain limitations to be aware of. I want to share my experiences to provide guidance for those using AI to develop SRS code.

What AI Cannot Do

1. Copy-Paste Code: AI struggles with performing exact copy-paste operations. It often attempts to optimize or modify the code slightly, which can lead to unexpected errors. For instance, when you need to move a function from one class to another without any modifications, it's best to do this manually. AI's tendency to alter code can result in subtle changes that may cause functionality issues or compilation errors.

2. Complex Library Upgrades: Although AI can upgrade APIs and address compatibility issues between C and C++ compilers, it often falls short with complex tasks like upgrading http-parser to llhttp. While AI may handle the initial upgrade process, it can introduce subtle bugs that lead to unit test failures. The behavioral differences between these libraries are substantial, and AI currently lacks the capability to navigate these complexities entirely on its own. Manual debugging and adjustments are often necessary to ensure a successful upgrade.

What AI Excels At

1. Writing Features: AI is highly capable of implementing new features in SRS, such as adding IPv6 support across all protocols. This is because new features often rely on established knowledge and existing solutions from other projects. AI can efficiently adapt these solutions, leveraging its extensive training data to provide robust implementations.

2. Answering Questions: AI possesses comprehensive knowledge about SRS, including its codebase, documentation, guidelines, and features. It frequently provides more accurate and detailed answers than human maintainers. Whether you're troubleshooting an issue or seeking information about SRS functionalities, AI can offer valuable insights. I often rely on AI to assist in answering community questions, as it can quickly and effectively address inquiries, even those related to specific environmental setups.

3. Writing Unit Tests: Humans often struggle with writing effective unit tests due to biases and the assumption that their code is correct. AI, on the other hand, can generate comprehensive and unbiased unit tests. However, it's crucial to review these tests to ensure they are relevant and accurate. AI can sometimes produce tests that are either too generic or not entirely applicable, so human oversight is essential to maintain quality.

4. Finding and Fixing Bugs: AI excels at diagnosing and fixing bugs, particularly when provided with clear descriptions or logs, such as ASAN crash reports. It is adept at analyzing core dumps and ASAN output logs, identifying the root cause of issues, and suggesting potential fixes. AI can also provide explanations for its reasoning, making it a valuable tool for debugging. However, it's important to verify AI's findings by adding unit tests to confirm the root cause and ensure the solution's validity.

5. Explaining Code: AI is proficient at explaining complex code and its usage. It can demystify intricate codebases like SRS, WebRTC, FFmpeg, libsrtp, and libsrt, offering clear explanations and direct references to the relevant code. This capability makes AI an invaluable resource for developers seeking to understand or document complex systems.

6. Reviewing PRs: While AI should not replace human reviewers, it can significantly aid in understanding pull requests (PRs). By using AI to analyze and summarize PRs, developers can gain a clearer understanding of the changes being proposed. Once AI provides an overview, it's important to verify its insights and use them to guide further review. AI can also suggest additional unit tests to ensure comprehensive coverage and understanding, helping to improve the project's quality and prevent future issues.

By leveraging AI's strengths and acknowledging its limitations, you can enhance your development process within the SRS community.