I never really got the hype. Classifiers are cheap and good at classifying. Deterministic validation is good at everything else. A non-deterministic thing that uses LLM side channel confidence to cheaply classify is... an interesting experiment, but it's kind of wish-casting. One would think that it would be highly effective, but would one bet the enterprise on it?
I get the point. It's basically a zero-shot classifier. i.e Get a classifier without having to go through the pains of curating and cleaning data. An LLM can do this as a hack (which is why people have been using them), but something that is cheaper and faster like these decision models would be more appropriate for application backends.
I never really got the hype. Classifiers are cheap and good at classifying. Deterministic validation is good at everything else. A non-deterministic thing that uses LLM side channel confidence to cheaply classify is... an interesting experiment, but it's kind of wish-casting. One would think that it would be highly effective, but would one bet the enterprise on it?
I get the point. It's basically a zero-shot classifier. i.e Get a classifier without having to go through the pains of curating and cleaning data. An LLM can do this as a hack (which is why people have been using them), but something that is cheaper and faster like these decision models would be more appropriate for application backends.