3 ms·

> The plot twist is that we may be able to engage in conversation before we understand what we are saying. Wouldn't this be totally pointless? What's the use o

> The plot twist is that we may be able to engage in conversation before we understand what we are saying. Wouldn't this be totally pointless? What's the use of a ChatGPT for elephants if it doesn't allow us to communicate with them?

lucubratory 3y ago
We could use the interactions to understand more about them and how they communicate. If we can make "ChatGPT for elephants" (i.e. something that could respond to them with a statistically probable response), we can observe how different communications affect their behaviour much more easily because we could ensure one side of the conversation is always 'talking'. Research like this has already led to us being able to identify various language-like articles from different animals, normally alarm calls because it's very easy to see the response to them. There is also a phenomenon in LLM research where a comparatively tiny amount of a language confers unexpectedly high fluency to a model which is otherwise intelligent in its main language and able to transfer that knowledge over. For example, an LLM needs to be trained on a lot of text to become intelligent in the ways modern LLMs are, which means English text because there's so much of it. But if you then train that LLM on a comparatively tiny amount of another language, like Catalan, it doesn't behave like a model that was trained on that tiny sample of Catalan alone, even when it's only speaking in Catalan. Instead it is nearly as intelligent as it is in English, just in a different language. What that means is that we don't necessarily have an impossible task in front of us in order to get enough human-translatable "language" from any given animal. We may only need to get a surprisingly small amount that we understand for an LLM to be able to generalise and come closer to 'fluency' in that language, which is going to expand the amount of data we have access to by a lot. It's an interesting area of research and I hope it gets expanded on
famouswaffles 3y ago
There's no reason to think we couldn't use a chat-GPT for [insert animal] to communicate with them. The idea is that cGPT would be translating for us and vice versa. But that doesn't mean we understanding what is being conveyed exactly.
- dragonwriter 3y ago
  In order to build a chatGPT for the animal, we have to have a body of training data for it; where is that supposed to come from?
  - famouswaffles 3y ago
    We can begin to record what we need. we likely won't need that much because language competence bleeds over as long as it's being trained on a wealth of other languages as well.

lucubratory 3y ago

We could use the interactions to understand more about them and how they communicate. If we can make "ChatGPT for elephants" (i.e. something that could respond to them with a statistically probable response), we can observe how different communications affect their behaviour much more easily because we could ensure one side of the conversation is always 'talking'. Research like this has already led to us being able to identify various language-like articles from different animals, normally alarm calls because it's very easy to see the response to them. There is also a phenomenon in LLM research where a comparatively tiny amount of a language confers unexpectedly high fluency to a model which is otherwise intelligent in its main language and able to transfer that knowledge over. For example, an LLM needs to be trained on a lot of text to become intelligent in the ways modern LLMs are, which means English text because there's so much of it. But if you then train that LLM on a comparatively tiny amount of another language, like Catalan, it doesn't behave like a model that was trained on that tiny sample of Catalan alone, even when it's only speaking in Catalan. Instead it is nearly as intelligent as it is in English, just in a different language. What that means is that we don't necessarily have an impossible task in front of us in order to get enough human-translatable "language" from any given animal. We may only need to get a surprisingly small amount that we understand for an LLM to be able to generalise and come closer to 'fluency' in that language, which is going to expand the amount of data we have access to by a lot. It's an interesting area of research and I hope it gets expanded on

famouswaffles 3y ago

There's no reason to think we couldn't use a chat-GPT for [insert animal] to communicate with them. The idea is that cGPT would be translating for us and vice versa. But that doesn't mean we understanding what is being conveyed exactly.

dragonwriter 3y ago
In order to build a chatGPT for the animal, we have to have a body of training data for it; where is that supposed to come from?
- famouswaffles 3y ago
  We can begin to record what we need. we likely won't need that much because language competence bleeds over as long as it's being trained on a wealth of other languages as well.