First, you're right in that articles like this are spam for click-throughs. Writing prompts is like writing good search queries, and the techniques developed t

First, you're right in that articles like this are spam for click-throughs. Writing prompts is like writing good search queries, and the techniques developed today will fall away tomorrow as the models and generation parameter settings get better. This has happened rapidly with diffusers, and will likely happen with language models. The system isn't non-deterministic, although there are a fair number of parameters used to try to make it seem like it (temperature in particular). But given the same prompt, zero temperature, and the same seed, you'd get the exact same output each time. The injected pseudo-randomness is deliberate. Having said all that, there are a handfuls of key phrases or instructions that do allow you to get more than you would from simple conversation English. For example, asking a model to use Tree of Thoughts and Chain of Thoughts in calculating its answers does have an effect. The same with asking "are you sure" for some models can yield corrections and refinements.