Progress Report #1: Arena, SafeVerify and More
It’s been a couple of weeks since I posted my essay A Proposal for Safe and Hallucination-free Coding AI. Since then I have been writing quite a bit of code. I wanted to give a quick progress report on what’s new, and what’s upcoming. (You might want to read the previous essay if you haven’t, or the following won’t make much sense.)
Arena
This is the project that has been gaining the most traction. Recall that Code with Proofs: the Arena is a web site for posting coding challenges with formal specifications in Lean, and for submitting solutions (code with proof) that are judged using the Lean proof checker. The goal is to serve as a platform for crowdsourcing and sharing data, to help train coding AIs.
We now have a demo site up at http://www.codeproofarena.com:8000/. It’s gotten some interest and lively discussion at Lean Zulip Chat, the main discussion forum for the Lean community. Some users have submitted solutions and new challenges to the demo site. Also gotten some very helful feedback from the Lean experts there, which leads us to:
SafeVerify
The current system had some rudimentary checks to make sure the submission indeed satisfies the
formal specification of the challenge. E.g. user could submit a proof with the sorry
placeholder in the body, and the lean proof checker will go through, but emit a warning abou the sorry
. This is caught by the system, which then judges the submission as incorrect.
Our Lean experts quickly pointed out ways the system can be circumvented, as evidenced by certain submissions to the demo site. Lean has powerful metaprogramming capabilities, which allows users to e.g. create their own tactics that helps them prove stuff in the domains of their interest. In our case it also allows one to:
- redefine syntax so that e.g.
theorem
means something totally different, perhaps no longer requiring a proof, or - add new axioms or replace existing standard axioms, so that
False
can be proven, from which anything can be trivially proven.
“Experts” was an understatement, they are the people developing the Lean language and libraries, so know all the tricks. They also suggested solutions to make the check more robust. Coincidentally, Vadim Fomin recently announced on Zulip a new project TheoremMarketplace, a blockchain smart contract for theorem bounties, that is also in need of robust proof checking. We decide to factor out the proof checking component into a separate project and collaborate.
SafeVerify is a work-in-progress prototype, based on these suggestions and code from lean4checker.
Taking a broader view, this stress testing and hardening of the proof verification component is actually extremely important. Recall that the ultimate goal of our project is to create coding AIs that can produce code and its correctness proof. And the safety of the code is guaranteed by automated checking of the proof by Lean. So having a proof checking component that is robust against jailbreaking is a foundational component of the whole endeavor. I underestimated the importance of this earlier because I didn’t have a deep understanding of Lean, but I am totally glad that the Arena project has led me to work on this.
Also made me realize that beyond the projects proposed in my essay, there may be need for some additional utilities that would be one layer below on the software stack, that might be smaller but provide useful features and conveniences. So expect to see more of that in future progress reports.
Coming Soon
Here’s a sneak peek of stuff I’ve been working on but not released yet.
- In the essay I promised a post that expands on my preference towards the style of integrating coding and proving, AKA programming with dependent types. That is coming. It has to do with a topic most feared by the folks preparing for coding interviews on Leetcode…
- FormalizeWithTest contains some initial code that implements the idea regarding autoformalization from Projet 3 of the essay. I think this is a very promising direction, but will need some further engineering to reach its full potential. More to come.
- I am doing some experimentation with Project 1, search / inference time compute. Similary to the previous bullet, this will need some further building out of the components. I hope to make an initial release soon.
As always, comments, suggestions, collaborations welcome!