Actually, no. We always needed good checks - that's why you have techniques like automated canary analysis, extensive testing, checking for coverage - these are forms of "executable oracles". If you wanted to be able to do continuous deployment - you had to be very thorough in your validation.

LLMs just take this to the extreme. You can no longer rely on human code reviews (well you can but you give away all the LLM advantages) so then if you take out "human judgement" *from validation*[1], you have to resort to very sophisticated automated validation. This is it - it's not about "inventing a new language", it's about being much more thorough (and innovative, and efficient) in the validation process.

[1] never from design, or specification - you shouldn't outsource that to AI, I don't think we're close to an AI that can do that even moderately effective without human help.

Zero degrees of freedom is a step too far.

What you want is correctness preserving transformations. Add to this some metrics such as code size, execution speed.

Yeah, precision LLM coding is kind of an oxymoron. English language -> codebase is essentially lossily-compressed logic by definition. The less lossy the compression becomes, the more you probably approach re-inventing programming languages. Which then means that in order to use LLMs to code, you're accepting some degree of imprecision.

Wow, a sudden change in writing style that’s not at all intended to disguise the fact that you’re an llm!

This is how its meant to be done. Usually with the reviewer being the stronger model.

That said, with both the test driven development this post describes and the reviewer model (its best to do both) you have to provide an escape hatch or out for the model. If you let the model get inescapably stuck with an impossible test or constraints it will just start deleting tests or rewriting the entire codebase in rust or something.

My escape hatch is "expert advice". I let the weak LLM phone a friend when its stuck and ask a smarter LLM for assistance. Its since stopped going crazy and replacing all my tests with gibberish... mostly.

That works well. Anthropic wrote a writeup on it.

https://www.anthropic.com/engineering/harness-design-long-ru...

This is routine. We have Gemini (which is not our coding model) review our PRs and it genuinely catches mistakes. Even using the same model as the creator, without its context to bias it, would probably catch many mistakes.