Has this guy used an LLM recently, they do far more than guessing the next word we taught them to guess the next word then they magically started doing way more.
I am not sure LCMs are actually doing anything more than what, effectively, LTSM perceptrons do. LTSM is used in LLMs and are critical in ensuring coherence. Some shaping of the model can optimize this. I still don't see that LCM provides any understanding. It sounds like abstracting to concepts and working with them would provide understanding because you are building models with concepts. However, I see two problems with that: (1) these "concepts" are not actually conceptualized but just tagged--LSTM nodes effectively do the same thing; and (2) I haven't heard any mention of processing them any differently than the generative process. So what's the net difference? I hear claims but not results any different than what might be possible through better shaping because this doesn't appear to be any different, fundamentally. Really, it's just a prompting method. LLMs are very poor at actual conceptualization. I have tried a lot. For example, have it come up with as many analogies to something as possible. The conceptualization would be to model them in terms of which parts are static and which parts vary and by what are the variances constrained by? If you can do this this you will have a concept modelled that can be applied to many other novel scenarios. Without that, you are dealing in vagueries, more or less.. which is a fundamental limitation of neural nets.
This has the potential to take phone scams to a whole new level.
Has this guy used an LLM recently, they do far more than guessing the next word we taught them to guess the next word then they magically started doing way more.
I am not sure LCMs are actually doing anything more than what, effectively, LTSM perceptrons do. LTSM is used in LLMs and are critical in ensuring coherence. Some shaping of the model can optimize this. I still don't see that LCM provides any understanding. It sounds like abstracting to concepts and working with them would provide understanding because you are building models with concepts. However, I see two problems with that: (1) these "concepts" are not actually conceptualized but just tagged--LSTM nodes effectively do the same thing; and (2) I haven't heard any mention of processing them any differently than the generative process. So what's the net difference? I hear claims but not results any different than what might be possible through better shaping because this doesn't appear to be any different, fundamentally. Really, it's just a prompting method. LLMs are very poor at actual conceptualization. I have tried a lot. For example, have it come up with as many analogies to something as possible. The conceptualization would be to model them in terms of which parts are static and which parts vary and by what are the variances constrained by? If you can do this this you will have a concept modelled that can be applied to many other novel scenarios. Without that, you are dealing in vagueries, more or less.. which is a fundamental limitation of neural nets.