2 ms·
Are AI models doomed to always hallucinate?
- chrisjj 3y ago> This tendency to invent “facts” is a phenomenon known as hallucination No. It is a phenomemon misrepped as hallucination by the "AI" marketeers. It is a phenomemon known as fabrication, little different from e.g. feeding random data to a 3D printer for a month then selecting the output that looks credible. There's nothing inventive about it. Invention is exclusive to intelligence.
- cc101 3y agoI find it helpful to compare AI model responses with my understanding of how my subconscious responds to my own queries. The initial responses from my sub-conscious often fail to consider relevant facts, fail to honor my ethical and moral standards, etc. That part of me called the executive function by psychologists reviews many such responses and may clarify my query, reject responses, point out relevant issues, query wether it "thinks" some other factor has been properly considered, ... This dialog continues until my executive function is satisfied and permits appropriate action. I treat AI responses in a nearly identical manner. [And yes my subconscious sometimes responds with utterly flakey nonsense]. Are these initial flawed responses fabrications? No. They are simply flawed opinions. I really don't like calling them hallucinations either. These terms imply ethical judgements that are inappropriate for a worthwhile evolving process. I suspect that eventually we will discover these two processes are identical at some deep level.