Search Results
7/8/2025, 1:54:27 AM
>>105832296
If my understanding of what you're saying is correct, that still is impossible because that would require the model to actually think and reflect on their own without input. I don't have to have someone talking to me right now in order to think through something, reason through concepts, come up with new things, etc. I can act on my own in my own head. A safetensors file cannot do that on its own. Someone has to interact with it in order for it to do anything. Furthermore how would it even know how to modify itself in order to learn? How would it know which weights to update and in what fashion? Some might say "oh it would just search the internet" but at that point it would just be reading summaries and not actually ingesting in retaining that information. It would not be studying and learning anything, it would just be coming up with summaries and would it remember anything it was tasked to research. Also doesn't something like this already exist? I thought this was the main concept of what MOE was supposed to be. Where instead of the entire model being activated at once, only pieces of it would be called based on what it was being asked.
If my understanding of what you're saying is correct, that still is impossible because that would require the model to actually think and reflect on their own without input. I don't have to have someone talking to me right now in order to think through something, reason through concepts, come up with new things, etc. I can act on my own in my own head. A safetensors file cannot do that on its own. Someone has to interact with it in order for it to do anything. Furthermore how would it even know how to modify itself in order to learn? How would it know which weights to update and in what fashion? Some might say "oh it would just search the internet" but at that point it would just be reading summaries and not actually ingesting in retaining that information. It would not be studying and learning anything, it would just be coming up with summaries and would it remember anything it was tasked to research. Also doesn't something like this already exist? I thought this was the main concept of what MOE was supposed to be. Where instead of the entire model being activated at once, only pieces of it would be called based on what it was being asked.
Page 1