3 ms·

I think there are lots of intuitions for why this is not enough to explain their capability. For example, if you train an LLM on English and French text, and th

I think there are lots of intuitions for why this is not enough to explain their capability. For example, if you train an LLM on English and French text, and then use reinforcement learning with English examples to teach it how to answer questions reliably, it will then also be able to answer questions in French. Why?

yellow_lead 3y ago
Because of embeddings
- cjbprime 3y ago
  Can you explain why without using the word "embeddings", in a way that convinces us that it's just autocomplete? Or to put it another way: what is being embedded? Is it an abstract concept like question-answering?
  - yellow_lead 3y ago
    I'm not an expert but the way I think of it generally is that words can be converted to vectors, i.e cat -> (1,8,23,34,32). They are much longer than this but just an example. For the corresponding word in Spanish/French/Other languages, the vector is actually quite similar. So when an LLM sees french, it can actually use the English training data to respond in french, because of this intermediary translation to vectors.
    - cjbprime 3y ago
      This is supposed to be an explanation of why LLMs are just autocomplete, but you're describing the ability to make connections between ideas across languages semantically as an explanation for why answering questions is learned cross-language, and talking about it like it's not mindblowing by describing how the lookups work in a vector space of ideas and concepts.
BugsJustFindMe 3y ago
Speaking as both a knower of French and also a knower of information networks, there are two answers... 1. If your proposed scenario includes bilingual data that contains, explicitly or implicitly, translatory relationships between parts of the two languages: the part of the graph that relates inputs to outputs, which is in english, is on the other side of the part of the graph that relates french to english from the part of the graph that takes in french input. So getting from the french input to some output passes through a kind of translation (this may not be a translation between languages but a translation between characteristics, which ends up accomplishing the same thing. see number 2 next). 2. Even if the scenario does _not_ include bilingual texts, French vocabulary is quite proximal to English vocabulary (you can get reasonably far in understanding french purely through strained cognate relationships to an english thesaurus), and the relationship between French questioning and answering grammar is similarly proximal to the relationship between English questioning and answering grammar. So again the same translatory connections can be found even purely by analog, just with some more input data. If you'd said two fundamentally dissimilar languages, like English and Chinese, then that could be different in this second case, but you didn't. This feels kind of like asking why a pocket calculator is able to answer questions about decimal arithmetic when it only knows how to do binary math, when it also happens to encode relationships between binary and decimal numeric representation.

yellow_lead 3y ago

Because of embeddings

cjbprime 3y ago
Can you explain why without using the word "embeddings", in a way that convinces us that it's just autocomplete? Or to put it another way: what is being embedded? Is it an abstract concept like question-answering?
- yellow_lead 3y ago
  I'm not an expert but the way I think of it generally is that words can be converted to vectors, i.e cat -> (1,8,23,34,32). They are much longer than this but just an example. For the corresponding word in Spanish/French/Other languages, the vector is actually quite similar. So when an LLM sees french, it can actually use the English training data to respond in french, because of this intermediary translation to vectors.
  - cjbprime 3y ago
    This is supposed to be an explanation of why LLMs are just autocomplete, but you're describing the ability to make connections between ideas across languages semantically as an explanation for why answering questions is learned cross-language, and talking about it like it's not mindblowing by describing how the lookups work in a vector space of ideas and concepts.

BugsJustFindMe 3y ago

Speaking as both a knower of French and also a knower of information networks, there are two answers... 1. If your proposed scenario includes bilingual data that contains, explicitly or implicitly, translatory relationships between parts of the two languages: the part of the graph that relates inputs to outputs, which is in english, is on the other side of the part of the graph that relates french to english from the part of the graph that takes in french input. So getting from the french input to some output passes through a kind of translation (this may not be a translation between languages but a translation between characteristics, which ends up accomplishing the same thing. see number 2 next). 2. Even if the scenario does _not_ include bilingual texts, French vocabulary is quite proximal to English vocabulary (you can get reasonably far in understanding french purely through strained cognate relationships to an english thesaurus), and the relationship between French questioning and answering grammar is similarly proximal to the relationship between English questioning and answering grammar. So again the same translatory connections can be found even purely by analog, just with some more input data. If you'd said two fundamentally dissimilar languages, like English and Chinese, then that could be different in this second case, but you didn't. This feels kind of like asking why a pocket calculator is able to answer questions about decimal arithmetic when it only knows how to do binary math, when it also happens to encode relationships between binary and decimal numeric representation.