If AI Writes the Code, Who Builds the Next Open Source Project?

SCOTTLOGIC.COM

As a long-time open source advocate, maintainer and contributor I’m getting increasingly concerned about the long-term impact that AI has on this valuable and oftentimes fragile resource. Open source has long been driven by human frustration, curiosity and craftsmanship, creating better tools because existing ones didn’t quite feel right. But as AI agents increasingly write code for us, that dynamic may be changing.

In this post I look at the more immediate impact that AI has had on on open source; from the legal debates over training data, the uncertain copyright status of AI-generated code, the possibility of AI-driven project cloning, and the growing strain on maintainers.

However, I think there are deeper, more existential challenges on the horizon. What happens to open source when the people who create it start to feel less of the friction that once inspired it?

The Coding Agent Bootstrap

MONPERRUS.NET

If you’ve ever tried creating a compiler, or have worked with compilers, you’ll have come across the term ‘self hosting’. This is the point where the implementation is sufficiently complete that the compiler itself can be written in the same language that it is compiling, e.g. a C++ compiler that can compile itself.

This blog post explores the same concept, but for AI agents. Martin has created a specification for a coding agent, which Claude Code was able to implement in ‘one shot’ (i.e. it worked first time). The newly generated agent, which is the next generation after Claude Code, was then given the specification and asked to implement it once again.

Speed at the Cost of Quality: How Cursor AI Increases Short-Term Velocity and Long-Term Complexity in Open-Source Projects

ARXIV.ORG

Measuring the overall impact of agentic AI (for coding) is very difficult. It is of course far faster at writing code than any human being, even the most nimble on the keyboard. However, if we blindly chase speed, quality suffers. And here is where the issue lies, how do you measure quality? Most studies focus on measuring speed, with little attention paid to quality - beyond simply validated that tests pass or that the code is functionally correct.

This study seeks to measure both, quality and velocity. They specifically focus on Cursor, identifying 806 repositories that have adopted cursor (via the presence of a cursorrules file), and a control group of the similar repositories selected based on propensity score matching.

They found that while velocity increased (by 15% on average) in the short-term, code quality (based on static analysis warnings, code complexity and duplication) deteriorated. The overall productivity gains ultimately decreased over time.

I think this is a useful report, but I do have some concerns. The presence of a cursorrules file doesn’t mean that all contributors are using Cursor, and for those using it, tells you nothing about their proficiency or whether they are using Cursor as an ‘agent’ or not.

However, I don’t want to be dismissive of this study either. It highlights the need to measure potential negative impacts of AI adoption as well as the positives.

A sufficiently detailed spec is code

HASKELLFORALL.COM

One of the biggest challenges with AI is getting it to do what you want!

I’m repeatedly impressed with how little instruction you need to give AI Agents. For example, you can tell it to “build me a event microsite” and it will do an impressive job. This is because the underlying technology used by coding agents is a Large Language Model, that has been trained on vast quantities of textual data giving it an immense amount of general knowledge.

However, often you want to provide a greater level of specificity, you don’t just want an ‘event microsite’, you want a highly specific event microsite with a detailed list of features. What is the best way to provide all of this detail to an AI agent?

This is why there has been such a resurgence of interest in specifications, and Specification Driven Development (SDD), frameworks that purport to ensure your specifications are sufficient and complete.

Many in the industry believe that you can define software entirely via its specification, allowing you to completely regenerate it from this specification whenever you like. This blog post refutes that code, taking the position that “a sufficiently detailed spec is code”. I agree!

Leanstral: Open-Source foundation for trustworthy vibe-coding

MISTRAL.AI

On a related note, Mistral have released an open-source code agent designed for Lean 4, a formal proof assistant and programming language. I’ve not heard of Lean 4 so had to ~Google~ ChatGPT it.

It is unclear to me how I might use this - I don’t intend to learn Lean 4, one of the key promises of LLMs for coding is that you don’t necessarily need to learn new languages anymore. However, given that I don’t have a Computer Science background, I could just be showing my ignorance here. If anyone knows how you might use this for everyday coding, please do let me know.