Thread by @amasad

A new kind of program analysis is emerging: AI code analysis. Here is the current state of analysis and where I think it’s headed:

Static analysis, arguably a Good Old Fashion type of AI, it relies on explicit “reasoning” about the program using tree search and other algorithmic methods.

It’s painfully limited because it’s an ultimately undecidable problem. Eg, try proving that a program never halts.

To make code more analyzable there has been a trend to push the programmer to do more of the work. People are adding typing to all the dynamic languages: JavaScript, Python, Ruby.

While static analysis resulted in safer code, it also made coding less fun. You now have to think about two environments:

- the type env
- the runtime env

It’s like almost coding two programs.

This pushes the complexity from the machine to human. Never a good trade off.

The other type of analysis is runtime. Among runtime, testing is the most popular. Testing can be great for safety and as a guard rail for refactoring. It also comes with a huge maintenance cost.

It gets to be such a burden that companies have teams dedicated to testing infra.

Runtime analysis continues to be a fertile space to give programmers more insight over their code. Things like code coverage, APM services, and tracing are very useful.

The problem with runtime is that it requires a big infrastructure investment to get it to work.

I believe we are quickly approaching diminishing returns when it comes to program analysis. We are pushing more of the complexity on programmers to gain increasingly fewer insights.

Which brings me to AI program analysis. An area I believe that will reinvigorate analysis and give us new coding superpowers.

NLP has been insanely effective. And now the same tools are being applied to code.

One of the earliest papers in the space: The Naturalness of Software (2012) showed that a simple n-gram model was used to create a superior autocomplete system.

GitHub Copilot is a recent and more sophisticated incarnation of that.

While Autocomplete is one of the obvious use-cases for analysis, we’re only just scratching the surface.

At Replit, we decided to take a different lens to AI analysis and asked the question: how can we augment human understanding of code?

Incidentally, this is how we approached debugging as well, with code comprehension in mind. blog.replit.com/debuggest

Our first AI tool was Code Explanation. Highlight the code and get a natural language translation of it. blog.replit.com/codex

However there are many other directions we’re exploring:

- statistical linting: can AI spot program mistakes that are hard-to-impossible for static analysis to catch?

- just in time tips: can the AI give you insights or educational resources for a problem you’re facing

- AI REPL: can AI give you rapid feedback on your code that’s similar to coding in a repl?

- can we combine AI with runtime analysis? What can AI tell you about how your code runs?

So many other ideas to explore. At Replit we have a data advantage and an audience that’s hungry for tools for better program comprehension.

I’ve written elsewhere on what the future of AI coding could be and its implications on language design: blog.replit.com/codingai

Please reach out if you’re working in this domain and want to chat or if you want to work with us on some of these problems!