A question for anyone who understands AI: are there any AI companies which have a coherent theory of human language? Or are all AI companies just force-feeding massive amounts of data into their training systems? Because the latter is not a theory on language.
My hypothesis is that no AI company has a coherent theory on language. They analyze all language as we know it, run it through regressions, employ a lot of statistics, and this is how they make auto-complete chat bots. It requires billions of gigabytes and CPUs and a lot of guardrails, censors, etc., and they appear to want a billion times more space and performance and guardrails so they can analyze a billion times more data and generate a billion times more acceptable responses.
Making an LLM by analyzing language output is like making a car by analyzing car exhaust and tire tracks. With these you can come close to a theory of how a car works, and you could use a supercomputer to kinda sorta make a car based on this data, but it will be horribly inefficient and will crash into a lot of things.
A mechanic, however, would give a very different theory of how a car works, and they could build a car for a billionth of the cost. Unfortunately there are no mechanics in the business of linguistics, and so LLMs are made by theoretical physicists and behavior scientists, not mechanics.
Admittedly, there is no real mechanical linguistics possible. Such a job would basically attempt to derive the recursive property from language in order to generate an organic grammar. Doing this from language output is ridiculously inefficient. And it will definitely not get us any good solutions to violence.
However, if the recursive property could be derived from a more minimal state of hominization, then this would make the property easier to trace. This would allow for a far more coherent theory of language.
ROBA (recursive object based aggression, which is human violence) is a minimal human property and contains the same recursive property found within language. It is not found in animals. It should be far easier to derive the recursive property from ROBA, probably a trillion times easier.
But nobody has thought about doing this because everybody assumes violence is an extension of animal aggression in degree, and not a difference in kind. I believe ROBA ranks with nuclear fission as one of the greatest blind spots in scientific history.
The social sciences can’t even differentiate human violence from animal aggression, and so they have no theory of human violence. So the social and behavioral sciences are as inefficient as LLMs.
If governments are getting involved in LLMs (they will) then this will 1) require an unprecedented expenditure, 2) put LLMs in the hands of an elite few, and 3) be totally inefficient and unpredictable and crash into a lot of things. Because, again, there is no coherent theory on language or violence.
A ROBA technician, however, looks at violence and language like a system of hydrodynamics (thanks to Zack Baker for this terminology). Recursive violence is encysted into symbolic language where it is processed recursively ad infinitum, until somebody or something attempts to push symbols back toward ROBA. Violence and Language are technically the same recursive stuff in different states. ROBA is more solid state recursion that can be wielded, while language is more liquid state recursion in a petrie dish that can be divided, analyzed, and remixed. The Unoptimized Strategy is how to keep recursion in its linguistic state even when it is actively creating world-destroying technologies like iron, gunpowder, nuclear fission, and now AI.
EDIT: I have to add a caveat: converting ROBA to language is not necessarily an act of de-escalation. Hate speech that is designed to escalate is non-violent in the moment but violent in intent, but hate speech that is aimed at showing the weakness of the state in its capacity to protect citizens might arguably be de-escalatory. ROBA that is designed to stop an explosive act of language is arguably de-escalatory, but ROBA aimed at triggering a group into violence is arguably escalatory. The point of this is not to assign normative values to ROBA and language, as both can escalate and de-escalate. However, it seems that generally we can assign a negative value judgment to escalation and a positive value judgment to de-escalation. I’m basing this on the domestication hypothesis: escalation entails that ROBA becomes grammarized and produces negative feedback in the phenotype that might further incentivize escalation in future generations, while de-escalation produces positive feedback in the phenotype which aids in further de-escalation. (I’m reading Plato’s dialogues right now and feel like Protagoras, right in Socrates’ sights, by making sweeping generalizations like this, but nonetheless I’d love to discuss it with anyone.)
I believe a ROBA technician, given the time and funds, could derive the recursive property from ROBA (far more easily than from language) and use this to create various seed values that could generate grammar more efficiently than the current method of studying car exhaust and tire tracks. With these seed values, you could theoretically make a billionth- or trillionth-scale AI server could fit in your pocket. The Davy Crockett was a mini-nuke that could have guarded against state- or petroleum-sponsored nukes. The AI equivalent would the ROBA Chip, an AI server that fits on a raspberry pi, with a learning system operating off the blockchain.
(The ultimate mic drop would be to put a ROBA Chip on a Gameboy cartridge, but that’s probably a long shot. I’d be satisfied getting it on a Nintendo DS.)
This seed value would also theoretically produce better models of human violence. Combine this with a smarter LLM and you can more easily outsource conflict management to the blockchain, rather than to the 4 superpowers, only 1 of whom is electable (though some will debate this).
Nuclear apocalypse is a risk, but AI apocalypse is a far greater risk. So we need smarter, open source models that everyone can carry around in their pockets (not Musk’s super-heavy model that requires a server farm) and generate for a few dollars as a check against this. The ROBA Chip would be a small fortified city in your pocket for the coming Second Catastrophe.
Hoping to test this soon.