3 ms·
“Moving the goalposts” has long been identified as a problem by AI enthusiasts, but the opposite is also often true—an AI is built to solve a limited or highly
by cgearhart 5y ago
“Moving the goalposts” has long been identified as a problem by AI enthusiasts, but the opposite is also often true—an AI is built to solve a limited or highly restricted form of a problem, and then proponents claim that the goalpost has been moved when it’s observed that the general problem remains unsolved.
It feels to me that we started calling ML “AI” when deep learning became powerful enough to work on less clearly structured problems like vision and NLP—but “find patterns in complex data” does not seem powerful enough for what I would consider “intelligence” (artificial or otherwise). I don’t think that stateless/idempotent ML is capable of what most people would recognize as intelligence; in part because I suspect a history is required for a system to be self-correcting over time.
- dynamite-ready 5y agoI'm starting to appreciate this sentiment. The way most ML success stories are presented, you'd think it really is all a case of finding enough data, and let the computer learn to extract something useful from it. That is not the hard part. The hard part, is finding a dataset that is amenable to the training process. Or at the very least, determining if a collection has any 'educational' value at all. Then, by the time you get to that point, you're probably already looking at metadata, or a simple pattern that could possibly be encoded in a database query. I have some faith in some of these new discoveries (Alphafold is a remarkable case study), but in many cases, the effectiveness of ML seems to be overstated.