Hallucinations are a double-edged sword. This might sound controversial, but read on and see if you agree.
The downside
Since we started introducing AI into our workflow, we have seen hallucinations as a downside of LLMs. They make things up, introduce unexpected behaviors, or return more than what was actually requested. That's a fact.
The upside
However, I have faced several scenarios where the model introduced features or edge cases I wasn't thinking about at all. These were most likely the result of the model going beyond what I asked for and inventing things, a.k.a. hallucinating.
So instead of thinking of hallucinations only as a bug, maybe we can also think of them as a feature, as long as a human is there to review, decide what stays and what goes, and take ownership of the result.
Do you agree?
Originally shared on LinkedIn.
Comments