4 ms·

This tech works by predicting the next word based on the previous words. It doesn't have a concept of facts, sentences or even concepts. Just "what token is sta

This tech works by predicting the next word based on the previous words. It doesn't have a concept of facts, sentences or even concepts. Just "what token is statistically likely to follow the previous tokens" You can limit the sources of information that it uses to determine next word relationships. But then you give up the predictive power of having the entire internet as training data. If you first do a web search to limit the available sources, you increase the computation overhead (first search and then summarize at the token level). and if you're already doing a web search, what's the point in using gpt? You lose the power of big data synthesis and it's just a sort of summarization / style transfer tool. Individual tokens don't have sources and they can't be fact checked. They're just words or parts of words. Doing any kind of fact checking or citation adds an entirely separate technical process that is distinct from the transformer model.

merely-unlikely 4y ago
> It doesn't have a concept of facts, sentences or even concepts How different is that from humans I wonder? For example, when learning math we talk about developing an intuition about a concept to help actually use it in daily life. This is separate from the rigor of the actual theory, but theory is harder to remember or fully internalize. Intuition to me is more of a feeling that something is related without necessarily really understanding what is going on.
- hooande 4y ago
  I'm sorry, but this point is trite and incorrect. Humans can manipulate concepts and compare them to other ideas. Which is why we're able to fact check and cite sources, where language models cannot. It's true that I'm just putting one word after another right now, like a language model. But I can combine those tokens into higher level abstractions like facts and ideas, which I can then compare with other facts and ideas. If ChatGPT could do this, it wouldn't be plausibly wrong nearly as often. Instead it's just "what word is statistically likely to come next?"

merely-unlikely 4y ago

> It doesn't have a concept of facts, sentences or even concepts How different is that from humans I wonder? For example, when learning math we talk about developing an intuition about a concept to help actually use it in daily life. This is separate from the rigor of the actual theory, but theory is harder to remember or fully internalize. Intuition to me is more of a feeling that something is related without necessarily really understanding what is going on.

hooande 4y ago
I'm sorry, but this point is trite and incorrect. Humans can manipulate concepts and compare them to other ideas. Which is why we're able to fact check and cite sources, where language models cannot. It's true that I'm just putting one word after another right now, like a language model. But I can combine those tokens into higher level abstractions like facts and ideas, which I can then compare with other facts and ideas. If ChatGPT could do this, it wouldn't be plausibly wrong nearly as often. Instead it's just "what word is statistically likely to come next?"