The unreliability is fundamental to the nature of GPT, though. It can't not confabulate. You can't know when it is and is not right short of actually doing the

The unreliability is fundamental to the nature of GPT, though. It can't not confabulate. You can't know when it is and is not right short of actually doing the work in the first place. The things that will be built on GPT, you might be able to trust. But a language model can't be checking your particle physics work. A language model barely does simple arithmetic correctly. This is not some sort of temporary aberration; even human language models can't do arithmetic correctly. We do that elsewhere.