Computers are generally useful because we have come to trust the output. How many people would use a spreadsheet that posted a disclaimer that stated (some of the calculations might be wrong, don't use the output without first checking each total manually!)?
We didn't "come to trust the output." Boolean logic means the output is deterministic so if a computer program is coded correctly and memory is error corrected we should be able to trust the output due to the physical principle of causality. Now with LLMs output is non deterministic and that still can be useful but it is not similarly constrained. We should think of llm less as ai (there is no intelligence and any simulation thereof is human derived from training; there is no artificial) and more as a useful generative tool in some contexts. Want to code something well represented in the training set, great tool. Want to solved a math problem where the machinery to solve it already exists just has not been put together by humans yet great tool. Want to brainstorm great tool. Want a decision by an agent with skin in the game of life, poor tool. Human output only tends to be verifiable and skewed toward valid because there are consequences for bad output. We should stop expecting to rely on llms for decision patterns that are not well represented in the training data.
didgetmaster · · focus · HN ↗
morpheos137 · · focus · HN ↗