5 ms·

I feel like predicting the next most likely word is a grave oversimplification of what is going on with chatgpt. At least from my admittedly limited amount of t

by jjeaff 4y ago

I feel like predicting the next most likely word is a grave oversimplification of what is going on with chatgpt. At least from my admittedly limited amount of testing, it appears to be capable of synthesizing unique content from multiple sources. Which means in some way, it is "understanding" the content by categorizing it. In other words, I'm not sure how it could use statistics to predict the next most likely response when many of the responses seem to be completely original.

MacsHeadroom 4y ago
The only thing GPT can actually do is predict the next word-part. It is, by design, necessarily not capable of anything else. All other apparent capabilities are simply the emergent complexity which comes from predicting the next word-part very well over and over again.
Jensson 4y ago
It doesn't "synthesize", it interpolates, think word2vec. If you ask it something it doesn't know it interpolates the answers to similar questions and takes something in between, which is why you get such nonsense answers. Word2vec sometimes gets you amazing things, the typical example is "king - man + woman = queen", but it also gets you a lot of nonsense because many words can't be manipulated like that.
satvikpendem 4y ago
> I feel like predicting the next most likely word is a grave oversimplification of what is going on with chatgpt. > At least from my admittedly limited amount of testing, it appears to be capable of synthesizing unique content from multiple sources. Why do you believe these two statements to be orthogonal?
- jjeaff 4y ago
  Well, maybe I'm just not understanding how the statistics are applied, but I don't see how one could use statistical analysis to predict the next word if the specific sequence has never been encountered by the model. For example, if I can tell chatgpt to answer a question about black holes in the style of Chaucer's Cantebury tales, what data is it using to predict that response? It seems to me there is some additional mechanisms at work.

by jjeaff 4y ago

MacsHeadroom 4y ago

The only thing GPT can actually do is predict the next word-part. It is, by design, necessarily not capable of anything else. All other apparent capabilities are simply the emergent complexity which comes from predicting the next word-part very well over and over again.

Jensson 4y ago

It doesn't "synthesize", it interpolates, think word2vec. If you ask it something it doesn't know it interpolates the answers to similar questions and takes something in between, which is why you get such nonsense answers. Word2vec sometimes gets you amazing things, the typical example is "king - man + woman = queen", but it also gets you a lot of nonsense because many words can't be manipulated like that.

satvikpendem 4y ago

> I feel like predicting the next most likely word is a grave oversimplification of what is going on with chatgpt. > At least from my admittedly limited amount of testing, it appears to be capable of synthesizing unique content from multiple sources. Why do you believe these two statements to be orthogonal?

jjeaff 4y ago
Well, maybe I'm just not understanding how the statistics are applied, but I don't see how one could use statistical analysis to predict the next word if the specific sequence has never been encountered by the model. For example, if I can tell chatgpt to answer a question about black holes in the style of Chaucer's Cantebury tales, what data is it using to predict that response? It seems to me there is some additional mechanisms at work.