I remember people talking about how we'd need neurosymbolic A.I. in the 1990s and imagining a large number of ways to do it but it all seemed pretty hypothetical because the neural networks weren't so good and now that neural representations of language are mature all the paths that you think would work are open.
Symbolic AI (of any type) is still a holy grail objective.
I think linear program tapes are the best substrate if we want to diversify our architectures. Massive GPU farms have definitely won at the neural network game, but they would struggle to execute a highly branchy piece of code faster than a humble single core CPU.
Sequentially following a series of register-based instructions that produces the same effect as Astra would be an unbelievable accomplishment. There is still no conclusive evidence that this type of system cannot be produced. The search space is so awful (on paper) that people rarely try. You have to be a little bit crazy to spend time on this stuff in 2026. However, you will often be rewarded with unexpected results. Never underestimate the power of a good fitness function when combined with 1v1 tournaments and large populations. Even a ridiculously simple ISA can produce amazing results under sufficient evolutionary pressure.
With a Turing machine, you don't need to find the global minimum. That is a strong anti-pattern. You just need to find one of countless really good minima. The search space is not one needle in one very big haystack. A meaningful percentage of the hay is needles.
One approach to symbolic AI is that it is about putting together logic and databases and another is that it is about SMT solving and constraint satisfaction and such. Lisp was a revolutionary language not because of the parens or even macros or that people used it to implement "expert systems" but because it had a garbage collector, hashtables and such. (e.g. most of the examples in famous Lisp programming books by Graham and Norvig can be coded up just fine in Python for pedagogical purposes) You can make a superhuman chess engine by alpha-beta search, etc.
One of many great applications of neurosymbolics right now is using an A.I. to translate a program or proof or informal description of such a thing into inputs for a theorem prover. If math learns anything about A.I. it will be that they won't let you submit a proof to a journal if you can't write it up formally. Like it turns out that the ability of A.I.(s) to do "translation" (e.g. something that sorta requires understanding but doesn't quite) and handle tedious tasks unlocks the ability to use formal methods.
Okay but the difficulty here is getting the LLM to use the neural circuit effectively without training just wiping out the circuit. You'd need to have the circuit be a differentiable black box where gradients cannot affect the black box
Why would that be a problem? If you have a differentialable program in the circuit it should be compatible with backpropagation. It just doesn't have weights to update.
I remember people talking about how we'd need neurosymbolic A.I. in the 1990s and imagining a large number of ways to do it but it all seemed pretty hypothetical because the neural networks weren't so good and now that neural representations of language are mature all the paths that you think would work are open.
Symbolic AI (of any type) is still a holy grail objective.
I think linear program tapes are the best substrate if we want to diversify our architectures. Massive GPU farms have definitely won at the neural network game, but they would struggle to execute a highly branchy piece of code faster than a humble single core CPU.
Sequentially following a series of register-based instructions that produces the same effect as Astra would be an unbelievable accomplishment. There is still no conclusive evidence that this type of system cannot be produced. The search space is so awful (on paper) that people rarely try. You have to be a little bit crazy to spend time on this stuff in 2026. However, you will often be rewarded with unexpected results. Never underestimate the power of a good fitness function when combined with 1v1 tournaments and large populations. Even a ridiculously simple ISA can produce amazing results under sufficient evolutionary pressure.
With a Turing machine, you don't need to find the global minimum. That is a strong anti-pattern. You just need to find one of countless really good minima. The search space is not one needle in one very big haystack. A meaningful percentage of the hay is needles.
Well...
One approach to symbolic AI is that it is about putting together logic and databases and another is that it is about SMT solving and constraint satisfaction and such. Lisp was a revolutionary language not because of the parens or even macros or that people used it to implement "expert systems" but because it had a garbage collector, hashtables and such. (e.g. most of the examples in famous Lisp programming books by Graham and Norvig can be coded up just fine in Python for pedagogical purposes) You can make a superhuman chess engine by alpha-beta search, etc.
One of many great applications of neurosymbolics right now is using an A.I. to translate a program or proof or informal description of such a thing into inputs for a theorem prover. If math learns anything about A.I. it will be that they won't let you submit a proof to a journal if you can't write it up formally. Like it turns out that the ability of A.I.(s) to do "translation" (e.g. something that sorta requires understanding but doesn't quite) and handle tedious tasks unlocks the ability to use formal methods.
If we can plug in a boolean circuit, I wonder what else we can do. What about plugging in something that can serve as dynamic long term memory?
Okay but the difficulty here is getting the LLM to use the neural circuit effectively without training just wiping out the circuit. You'd need to have the circuit be a differentiable black box where gradients cannot affect the black box
Why would that be a problem? If you have a differentialable program in the circuit it should be compatible with backpropagation. It just doesn't have weights to update.
Okay, so when do we get FPGAs in LLMs