Guarding against LLMs that would learn our repo "by heart"

[tischi]

Even though I think LLMs generally do not work like this, I still wonder whether we could guard against some - otherwise super dumb - LLM to just learn our repo by heart and then achieve great results.

Given the discussions in #118 I wonder whether we could somehow maintain a separate secret branch where we ask the conceptually same questions but just with a slightly modifications?

Maybe:

• changing the english in the prompt a bit
• changing the actual values in the input data and the corresponding assertions
• changing the order of the input and output arguments

It would be a bit of work...but maybe worth it?

What do you think?

[haesleinhuepf]

I'm currently working on training such an LLM, because I wanted to know how to achieve this:

and then achieve great results.

When I'm done, I'll share it (+ training scripts) and we can develop a strategy against it.