I like this a lot, but I want you to write a follow up on how internet networks and neural networks are different. I agree that the internet as a whole is complex, but it is designed so that identification and recovery from failure is paramount at all layers. I don't think we can even define what a failure means in machine learning.
I really should write that follow-up. I mentioned modularization and abstraction, and both are needed as design principles in order to allow for identification for and recovery from failure.
+ Are mechanisms as conceived and discussed by _The New Mechanists_ a better foil for rules than laws?
+ Do you have a specific citation in mind for the precise sense in which you use "Verum et factum convertuntur"? (Minimal google search pointed me to Vico as the source, but am curious to know your specific source)
+ Since social systems are built systems, do you think sociology _ought_ to be studied using engineering principles instead of as a (social) science?
1) The ideas of the New Mechanists are a lot closer to rules than laws. Partly this stems from their realization of the shortcomings of the classical deductive-nomological view with its covering laws, and partly from the fact that, as far as I understand, they were interested in, and motivated by, the scientific practices of biologists and neuroscientists rather than physicists.
2) The original source of "verum et factum" is indeed Vico. The constructivist gloss on it, the idea that we are capable of understanding rationally the systems of which we are the cause, can be found in many sources; a good one is Jean-Pierre Dupuy's _On the Origins of Cognitive Science_, where he connects this to model-making in science and in engineering.
3) I think it would be rather presumptuous to proclaim that sociology _ought_ to be studied using engineering principles, but the engineering viewpoint is a valuable complementary perspective. In a sense, that's what Jürgen Habermas was trying to accomplish, at least if you take Joseph Heath's interpretation of Habermas' ideas at face value.
Regarding 1), I view your definitions of rules to be closer to what the New Mechanists call topological explanations (Cf. https://philpapers.org/rec/KOSMAT-2) . I think there is some debate going on over there that is as yet unresolved. (I am basing this on some of Daniel Kostic's and Phillipe Huneman's papers)
Regarding 2): I am assuming you felt inspired with Vico's "verum et factcum" and you would consider a toned down position that we can _sometimes_ rationally understand systems of which we are the cause.
Regarding 3): Thanks for the Heath reference. I have read a paper or two of Health, but have not paid attention to his interpretations of Habermas.
As one of the scientists arguing for a physics-like approach to ML, I feel I should present a somewhat different point of view.
LLMs are, of course, human creations and I agree that it does not make sense to look for laws of physics within them. What we need to understand are the laws that govern the data/language and that allow LLMs to be so effective (in fact that has always been the purpose of linguistics).
The fact that (a) modern deep learning models have been created by trial and error, without a fundamental understanding of their principles, and (b) their form is relatively simple and requires relatively little data (compared to the state space), suggest that structure in data cannot be particularly complicated.
Finally, I don't think we have an engineering framework for thinking about LLMs at the current moment in time. Most empirical work is based on ad hoc approaches or “questionable analogies” to human behaviors (e.g., "chain of thought reasoning" or RLHF). Good engineering is based on a solid understanding of the phenomenon being engineered. Alchemy, even when it is successful, is not engineering.
I like this a lot, but I want you to write a follow up on how internet networks and neural networks are different. I agree that the internet as a whole is complex, but it is designed so that identification and recovery from failure is paramount at all layers. I don't think we can even define what a failure means in machine learning.
I need to think about this more, myself...
I really should write that follow-up. I mentioned modularization and abstraction, and both are needed as design principles in order to allow for identification for and recovery from failure.
Questions
+ Are mechanisms as conceived and discussed by _The New Mechanists_ a better foil for rules than laws?
+ Do you have a specific citation in mind for the precise sense in which you use "Verum et factum convertuntur"? (Minimal google search pointed me to Vico as the source, but am curious to know your specific source)
+ Since social systems are built systems, do you think sociology _ought_ to be studied using engineering principles instead of as a (social) science?
Great questions!
1) The ideas of the New Mechanists are a lot closer to rules than laws. Partly this stems from their realization of the shortcomings of the classical deductive-nomological view with its covering laws, and partly from the fact that, as far as I understand, they were interested in, and motivated by, the scientific practices of biologists and neuroscientists rather than physicists.
2) The original source of "verum et factum" is indeed Vico. The constructivist gloss on it, the idea that we are capable of understanding rationally the systems of which we are the cause, can be found in many sources; a good one is Jean-Pierre Dupuy's _On the Origins of Cognitive Science_, where he connects this to model-making in science and in engineering.
3) I think it would be rather presumptuous to proclaim that sociology _ought_ to be studied using engineering principles, but the engineering viewpoint is a valuable complementary perspective. In a sense, that's what Jürgen Habermas was trying to accomplish, at least if you take Joseph Heath's interpretation of Habermas' ideas at face value.
Thanks for the response(s).
Regarding 1), I view your definitions of rules to be closer to what the New Mechanists call topological explanations (Cf. https://philpapers.org/rec/KOSMAT-2) . I think there is some debate going on over there that is as yet unresolved. (I am basing this on some of Daniel Kostic's and Phillipe Huneman's papers)
Regarding 2): I am assuming you felt inspired with Vico's "verum et factcum" and you would consider a toned down position that we can _sometimes_ rationally understand systems of which we are the cause.
Regarding 3): Thanks for the Heath reference. I have read a paper or two of Health, but have not paid attention to his interpretations of Habermas.
Nice article.
I had a similar thought driving home the other day. Thanks for articulating way better than I could!
As one of the scientists arguing for a physics-like approach to ML, I feel I should present a somewhat different point of view.
LLMs are, of course, human creations and I agree that it does not make sense to look for laws of physics within them. What we need to understand are the laws that govern the data/language and that allow LLMs to be so effective (in fact that has always been the purpose of linguistics).
The fact that (a) modern deep learning models have been created by trial and error, without a fundamental understanding of their principles, and (b) their form is relatively simple and requires relatively little data (compared to the state space), suggest that structure in data cannot be particularly complicated.
Finally, I don't think we have an engineering framework for thinking about LLMs at the current moment in time. Most empirical work is based on ad hoc approaches or “questionable analogies” to human behaviors (e.g., "chain of thought reasoning" or RLHF). Good engineering is based on a solid understanding of the phenomenon being engineered. Alchemy, even when it is successful, is not engineering.