Going live with Codex - Avoiding abuse


(ignore this part)

I am wondering if there may be different best practices and requirements for going live with Codex.

 For example, in the current Going live page, which I believe is currently the same as before Codex, it says this:
 We generally do not permit tools that generate a paragraph or more of natural language or many lines of code, unless the output is of a very specific structure that couldn’t be re-purposed for general blog or article generation (e.g., a cover letter, a recipe, song lyrics).
* For open-ended generations, limit the maximum output to 2 sentences, or 30-40 tokens.
* For more scoped use-cases, we tend to recommend an output of around 150 tokens (~6-8 sentences), but it will depend on the specific use-case.
* For code-generation we highly recommend a human in the loop to thoroughly review the code, and require it in high-stakes contexts.*

I am making an assumption that the abuse potential for GPT-3 lies *mostly* in a few problematic hotspots, such as its potential to mislead humans with well constructed, unique text, or its great ability to seem like an authentic human. But for code output, do we have these concerns to the same degree?

For open ended inputs to Codex, could more output tokens be safe and acceptable than with regular GPT-3? Could abuse be prevented a lot if the content within Strings and comments were limited, and we assume code output has less abuse potential? Might we see some Codex-specific Going live points or exemptions?

Any thoughts shared are very appreciated!

As soon as I posted this I realised a section on Code generation is indeed in the Going live, but I’ll leave it up because I am still interested in people’s thoughts on how to avoid abuse and perhaps others like me would have searched the forum to find the info first

Yes I am also curious about the process for going live with codex!