3 ms·

I think there are some problems with assuming that ChatGPT-like AI will improve in quality, _especially_ if it gets good enough to popular and generally trusted

by jrmg 4y ago

I think there are some problems with assuming that ChatGPT-like AI will improve in quality, _especially_ if it gets good enough to popular and generally trusted. Where will the next generation of AI training data come from, if the vast majority of information on the internet is generated by AI? Isn’t new data required? It feels like an internet filled with ‘incorrect’ spam generated by AIs is as much, maybe more, of a danger to future AI as it is to humans. This is almost the same thing, but: it feels like ChatGPT is so good at e.g. programming questions because it’s been trained on millions of human discussions about programming. Even if you can weed out AI spam from future training sets, if people start relying on AI for e.g Stack Overflow like answers, and perhaps even reference information, and therefore stop writing or conversing online about new technologies, where will training data about new technologies come from? Primary reference material probably isn’t enough (just like it’s not really enough for most humans).

hooande 4y ago
This is an excellent point. One of the main things I see people say about ChatGPT is "when it gets better in the future". But as you point out, it's already trained on the entire internet. There are many features they could add and there are infinite special cases for handling various prompts. But the core of the product, the LLM generated answers, can't get much better without an order of magnitude increase in training data. In terms of petabytes of training data, it will be a long time before ChatGPT's own responses are a significant portion of the training set. And even then, at least for a while, that should just shift responses closer to a sort of average human response
- astrange 4y ago
  > But the core of the product, the LLM generated answers, can't get much better without an order of magnitude increase in training data. There's no reason to believe this. The model architecture and training methods aren't perfect, nor is the way it's queried. eg: https://www.reddit.com/r/MachineLearning/comments/zr2en7/r_nonparametric_masked_language_modeling_metaai/ https://www.reddit.com/r/MachineLearning/comments/zr2en7/r_n...
machiaweliczny 4y ago
The way stuff is trained right now is at most 1/100 of capacity possible in my unscientific claim. So there's possibility to go 2-3 orders of magnitude with data we have.
versteegen 4y ago
> Isn’t new data required? No. For these systems to improve on their defects they should rely less on having seen something very specific in the training data, and more on reasoning and 'understanding' (forming conceptual models). The amount of training data available is already vastly more than strictly necessary given more data-efficient algorithms. I didn't have to read the whole internet to learn anything I know.

by jrmg 4y ago

hooande 4y ago

This is an excellent point. One of the main things I see people say about ChatGPT is "when it gets better in the future". But as you point out, it's already trained on the entire internet. There are many features they could add and there are infinite special cases for handling various prompts. But the core of the product, the LLM generated answers, can't get much better without an order of magnitude increase in training data. In terms of petabytes of training data, it will be a long time before ChatGPT's own responses are a significant portion of the training set. And even then, at least for a while, that should just shift responses closer to a sort of average human response

astrange 4y ago
> But the core of the product, the LLM generated answers, can't get much better without an order of magnitude increase in training data. There's no reason to believe this. The model architecture and training methods aren't perfect, nor is the way it's queried. eg: https://www.reddit.com/r/MachineLearning/comments/zr2en7/r_nonparametric_masked_language_modeling_metaai/ https://www.reddit.com/r/MachineLearning/comments/zr2en7/r_n...

machiaweliczny 4y ago

The way stuff is trained right now is at most 1/100 of capacity possible in my unscientific claim. So there's possibility to go 2-3 orders of magnitude with data we have.

versteegen 4y ago

> Isn’t new data required? No. For these systems to improve on their defects they should rely less on having seen something very specific in the training data, and more on reasoning and 'understanding' (forming conceptual models). The amount of training data available is already vastly more than strictly necessary given more data-efficient algorithms. I didn't have to read the whole internet to learn anything I know.