Production database drop while coding with an LLM

It can be tempting to quickly approve LLM commands especially when under pressure to ship a new feature or quickly fix an issue. This article which has been making the rounds recently is a good example of how doing this can make things can go catastrophically wrong.

A few observations:

• I applaud the author for both the transparency, and for owning the resulting issues and not saying “The LLM dropped our DB”. Paraphrasing this Oxide RFD, code or text written by LLMs should always still be the responsibility of the human generating it.
• Local Terraform state is kind of an anti-pattern. I almost wonder if Terraform should warn about the lack of remote state by default; local state makes sense only as a sort of seed file.
• In general, given the long tail of catastrophic failure with LLMs, I feel like “Git for X” is going to become a system pattern. Dolt seems to be quite useful for this.