Re:AI knows things no one told it
by SkyMarshal on Friday May 12, 2023 @06:34PM
(#35923377)
We just had this discussion a few days ago: https://news.ycombinator.com/item?id=35868065. It does appear to be a case of mis-used measurements rather than true emergence.
[ Parent
| on HN ]
This is a misunderstanding of the researchers' claim, for which I blame them rather than you.
You are using their study to claim that LLMs can not learn and reason about novel tasks. Their study doesn't make any claims about what LLMs can and can't do. The study says that when it appears that LLMs suddenly became able to reason about novel tasks, what actually happened is that the ability of the LLM to perform reasoning improved gradually and smoothly during its training until it could do those things. They are not saying that LLMs don't have emergent capabilities: they are saying that those capabilities emerged at a continuous rate rather than a discontinuous one.
Do you see?
> Do you see?
Emergence can only be discontinuous.
Emergent abilities can be gradual and continuous.
If it's continious then nothing is emerging and it can be explained by the sum of it's sub-systems.