Comment by visarga
3 years ago
I think the closest we might get to actually learning what it is like to be a bat is by unsupervised learning + a bit of manual labelling on top. A neural net could learn their states and dynamics, by training on a million hours of bat recordings. These representations will already encode bat states and values, so we just need the bridge to human language, which is easy to build with a pretrained language model.
This approach works for any species, neural nets can do it because they can do unsupervised learning. I bet we'll see pet translator apps popping up. Maybe we can monitor the environment by listening in on animal chatter.
The point is that no amount of external modeling can give you any knowledge of “what it’s like” subjectively.
You don't have any subjective knowledge of what it's like to be the person you know best either, at least not any meaningful amount in the framework you propose.
I think it can to a degree, just like LLMs imitate human language to a degree. That imitation can only happen by accurately modelling humans.
This is an interesting idea, but I don't find it particularly germane to Nagel's question. To someone with a hammer, everything looks like a nail. To someone with an LLM, everything looks like a set of data to be trained on, I suppose.suppose.
We still won't know what the sonar experience is.
But we know how it relates to everything. Neural nets are good at that. And then we can cluster the data and interpret it, or correlate it with a visual signal.
If you did that with all the colors of the rainbow, do you think that is anywhere near the experience of seeing the color red? It seems pretty clear to me that it doesn't even come close and isn't remotely relevant.