You still need your chatbot to stick to business rules and act like a real customer service rep, and that’s incredibly hard to accomplish with generative models
Isn’t that what, for instance, OpenAI’s embeddings are for?
My opinion right now is that companies want you to believe they are 100% capable of replacing humans
Probably, but at the moment they can only do it partially.
They are replacing them because they are greedy cunts, not because they are replaceable.
I partially agree. I mean, they are greedy cunts but some tasks like translating from/to certain languages can be easily done even with the free ChatGPT demo with better results than Google Translate, so human translators are unfortunately becoming quite replaceable.
The word embeddings and embedding layers are there to represent data in ways that allow the model to make use of them to generate text. It’s not the same as the model acting as a human. It may sound like a human in text or even speech, but its reasoning skills are questionable at best. You can try to make it stick to your company policy but it will never (at this level) be able to operate under logic unless you hardcode that logic into it. This is not really possible with these models in that sense of the word, after all they just predict the best next word to say. You’d have to wrap them around with a shit ton of code and safety nets.
GPT models require massive amounts of data, so they were only that good at languages for which we have massive texts or Wikipedias. If your language doesn’t have good content on the internet or freely available digitalized content on which to train, a machine can still not replace translators (yet, no idea how long this will take until transfer learning is so good we can use it to translate low-resource languages to match the quality of English - French, for example).
Isn’t that what, for instance, OpenAI’s embeddings are for?
Probably, but at the moment they can only do it partially.
I partially agree. I mean, they are greedy cunts but some tasks like translating from/to certain languages can be easily done even with the free ChatGPT demo with better results than Google Translate, so human translators are unfortunately becoming quite replaceable.
Do you mean the embeddings? https://platform.openai.com/docs/guides/embeddings/what-are-embeddings
If so:
The word embeddings and embedding layers are there to represent data in ways that allow the model to make use of them to generate text. It’s not the same as the model acting as a human. It may sound like a human in text or even speech, but its reasoning skills are questionable at best. You can try to make it stick to your company policy but it will never (at this level) be able to operate under logic unless you hardcode that logic into it. This is not really possible with these models in that sense of the word, after all they just predict the best next word to say. You’d have to wrap them around with a shit ton of code and safety nets.
GPT models require massive amounts of data, so they were only that good at languages for which we have massive texts or Wikipedias. If your language doesn’t have good content on the internet or freely available digitalized content on which to train, a machine can still not replace translators (yet, no idea how long this will take until transfer learning is so good we can use it to translate low-resource languages to match the quality of English - French, for example).