Summary

Can we use a LLM to generate simple CN specs for code? Most of the times, the specs needed for mitigating UB are pretty simple (Owned, Block...). They can perhaps be automatically derived from the function.

First, use the latest ChatGPT and attempt to train it with the tutorial examples, and see if it can fix some small functions (ideally from MKM or logging).

If that works, use OpenAI agent framework to create one agent to parse the code/generate specs, and another CN agent to validate/compile the code.

Document your progress, and if possible, share your LLM.

Links

• https://github.com/openai/swarm?tab=readme-ov-file