Codex (and GPT-4) can’t beat humans on smart contract audits (blog.trailofbits.com)

... probably one of the best summaries on the state of LLM "understanding" of code, and I think just about as valid for "code in general" as it is for smart contracts.