Openai,anthropic should integrate symbolic methods into their neuro systems. its obvious that neurosymbilic ai can be much more powerful than pure neuro.
Prose(flashfill) framework from Microsoft for program synthesis could be used with llms. Program synthesis methods are extremely sample efficient - one example is often enough to infere program which makes this transformation/extraction/execution. And llm could generate such DSL or examples for prose. This would give a loop between neuro llm and symbolic program synthesis prose.
This is crazy dodo. Love it. But what's even more crazy is this pretty much sums up my mind at the moment...🤪. I am sending you a PM. Hope that's ok...K
Soon they will use weights and biases of multiple LLMs as input data to train an AI of a higher dimension. And then they will loop it onto itself. And this will be it.
The vague part for me, is the last step, "symbolic rule application". Not sure how to ask an LLM to do that. Let's say I have "Jane runs-faster-than John", and "John runs-faster-than Mark" in my database. If the user specifically asks, "Does Jane run faster than Mark", then I can see how the LLM could query the database, pull the relevant connections, and figure out the correct answer. BUT I don't know how I would ask the LLM to "please look for new connections that are suggested by existing connections" unless the entire graph database fits into one context window. Anyone else in the audience have any ideas about that last step?
@@code4AI right. i did some prolog in college during the ai winter; not saying i am any good at it, but i know enough to believe (perhaps mistakenly) that prolog was always able to take a graph of this-relates-to-that plus some rules, and find (through exhaustive measures) more this-relates-to-that connections, no LLM and no NN needed. are you trying to take me back to the 1990's? ;) Even if this is as good as it gets, it is still good because the LLM makes steps 1 and 2 relatively easy, compared to the 1990's when it was nearly impossible.
YVHW created everything with a divine code they created, which formed the highest order of this plane, symbols. With that in mind, and man's technological advancements, man still hasn't created anything close to a seed.