'Emergence' is when something seems to appear out the background; when you see a face on a video screen made up of unrelated pixels, that's emergence. Emergence exists not just for pictures, but for anything we humans perceive, including sounds and music, patterns and trends, or - in the present case - skills and abilities. When we see, say, an AI demonstrate a skill like 'answering a question', this is an emergent behaviour. The important part of this fairly technical paper (16 page PDF) is the first paragraph in the Discussion section: "for a fixed task and a fixed model family, the researcher can choose a metric to create an emergent ability or choose a metric to ablate an emergent ability. Ergo, emergent abilities may be creations of the researcher's choices, not a fundamental property of the model family on the specific task." So if the 'skill' is not in the AI, where is it? As I have argued before, emergence requires recognition. We are looking at what the AI does and 'recognizing' it as 'answering a question'. There's nothing wrong with that; we do it all the time for other things. The picture is not 'in' the pixels; we look at the pixels and 'recognize' (or 'interpret') it as a face. So it's not surprising we would do it with AI as well. Via Graham Attwell.
Today: 6 Total: 102 [Direct link] [Share]
