I'm saying there isn't a distinction between literally and figuratively thinking or anything of the sort. "Thinking" isn't scientifically grounded, it's socially grounded, and barely so, given that it seems everyone has a differing justification for what constitutes "thinking". The variability in construing "thinking" is a product of how we each fashion our own internal sense of what it means, which if anything is evidence for the anti-realism of "thinking".
Continuing, we largely fool ourselves when we play the shell game of hunting for scientific justification of "social facts". Ie: when we assume a social fact, then hunt for scientific evidence to ground it in then use the scientific evidence to justify the existence and validity of the social fact we assumed. This happens a LOT and it's epistemically invalid.
A LOT of folks dismissal of thinking machines reminds me of Feuerbach's The Essence of Christianity, specifically the section on anthropomorphism[1], but in an inverted form - the reason LLMs can't "think" is that it would dilute what we hold dear about ourselves: our ability to think. It's an ego protection mechanism, but no one's jumping to admit that.
In this context "thinking" was meant as an analogy to the supposed reflective "slow" mode of the cognitive process, vs the more "reflexive"/"fast" mode, not in the sense of "thinking soul" or "thinking self-aware entity".
Concretely, what do you gain by giving a current-generation LLM more runtime? It's not trained/designed to do anything with it, more time = more tokens = more nonsense once past the "end" token/end of context. You could build an agent on top of the LLM that calls the LLM iteratively, but afaict this approach isn't strictly an improvement over the base LLM. Current architectures seem to be limited with how much improvement they can eke out of more runtime without a broader redesign or retraining.
Now with new architectures you might be able to do more with more runtime, but I'm not sold that allowing a current-gen LLM to execute code or do web-searches and re-feeding it the quetion + its output + new data is going to strictly return better results. Sometimes maybe yes, sometimes not.
So imo the current limitation is not the runtime allotted to LLMs but their fundamental (current-gen) design/training.
I think I follow better. Thinking doesn't have a rigorous and testable definition, so it isn't scientific. We all have our own colloquial understand. Since it's not testable and the definition varies person to person, it's not useful when reasoning about AI or intelligence.
kelseyfrog|1 year ago
Continuing, we largely fool ourselves when we play the shell game of hunting for scientific justification of "social facts". Ie: when we assume a social fact, then hunt for scientific evidence to ground it in then use the scientific evidence to justify the existence and validity of the social fact we assumed. This happens a LOT and it's epistemically invalid.
A LOT of folks dismissal of thinking machines reminds me of Feuerbach's The Essence of Christianity, specifically the section on anthropomorphism[1], but in an inverted form - the reason LLMs can't "think" is that it would dilute what we hold dear about ourselves: our ability to think. It's an ego protection mechanism, but no one's jumping to admit that.
https://www.gutenberg.org/files/47025/47025-h/47025-h.htm#pb...
sottol|1 year ago
Concretely, what do you gain by giving a current-generation LLM more runtime? It's not trained/designed to do anything with it, more time = more tokens = more nonsense once past the "end" token/end of context. You could build an agent on top of the LLM that calls the LLM iteratively, but afaict this approach isn't strictly an improvement over the base LLM. Current architectures seem to be limited with how much improvement they can eke out of more runtime without a broader redesign or retraining.
Now with new architectures you might be able to do more with more runtime, but I'm not sold that allowing a current-gen LLM to execute code or do web-searches and re-feeding it the quetion + its output + new data is going to strictly return better results. Sometimes maybe yes, sometimes not.
So imo the current limitation is not the runtime allotted to LLMs but their fundamental (current-gen) design/training.
BoxedEmpathy|1 year ago
I think I follow better. Thinking doesn't have a rigorous and testable definition, so it isn't scientific. We all have our own colloquial understand. Since it's not testable and the definition varies person to person, it's not useful when reasoning about AI or intelligence.
Also I wanted to say I love how you write!