For what is worth a lot of medicine works this way. I’m fairly certain this isn’t the only field, either. I’d imagine studying ecology or space feels similar
It does make me vaguely curious what happens if you try to make one of these on the more powerful end explain step by step how its own program works. I dont really expect it to be accurate, given that if people dont know how the thing works, it probably wont find much about that in it’s training data, but if what it learns ultimately enables it to make connections about how the real world works to some degree, could it figure out enough to give even marginally useful hints?
Not really, it’s super fucking expensive to train one of these, on-line training would simply not be economically feasible.
Even if it was, the models don’t really have any agency. You prompt, they respond. There’s not much prompting going on from the model, and if there was, you can choose to not respond, which the model can’t really do.
We know how they work, otherwise we couldn’t design and implement them. What we don’t really know, and we don’t really have to know is the exact parameters the model trains to.
The issue you’re thinking of is that any one parameter does not necessarily map to one aspect, but they are a coherent collection that makes the whole work. Some interesting insights can be gleaned from trying to figure out these relationships, but due to the massive amount of parameters (billions!) it gets a little much to get your head around.
Does anyone else start freaking out when we have such complex programs that researchers don’t fully understand how they work?
For what is worth a lot of medicine works this way. I’m fairly certain this isn’t the only field, either. I’d imagine studying ecology or space feels similar
It does make me vaguely curious what happens if you try to make one of these on the more powerful end explain step by step how its own program works. I dont really expect it to be accurate, given that if people dont know how the thing works, it probably wont find much about that in it’s training data, but if what it learns ultimately enables it to make connections about how the real world works to some degree, could it figure out enough to give even marginally useful hints?
Not really, it’s super fucking expensive to train one of these, on-line training would simply not be economically feasible.
Even if it was, the models don’t really have any agency. You prompt, they respond. There’s not much prompting going on from the model, and if there was, you can choose to not respond, which the model can’t really do.
You can train an effective one for a few hundred bucks now.
https://crfm.stanford.edu/2023/03/13/alpaca.html
We know how they work, otherwise we couldn’t design and implement them. What we don’t really know, and we don’t really have to know is the exact parameters the model trains to.
The issue you’re thinking of is that any one parameter does not necessarily map to one aspect, but they are a coherent collection that makes the whole work. Some interesting insights can be gleaned from trying to figure out these relationships, but due to the massive amount of parameters (billions!) it gets a little much to get your head around.