Recursive Recursion in the R Chain Blockchain

Speaker 2
Let's talk about recursion because I haven't heard anything
Speaker 1
yet about that actually so right right so Greg so Greg has fully recursive, reflexive, Turing complete version of the distributed process calculus. So we're doing programming language for smart contracts in the R chain blockchain. Okay, so what we're doing now. We're not working with Greg to modify his rolling interpreter to be basically part of the plumbing for our, our language so we can, when we want to compile meta for rapid execution, we compile meta to rolling, we run on the new version of Greg's rolling interpreter. Now, the rolling interpreter, during the properties of the row calculus, it allows a great variety of algorithms to run a linear speed up on a GPU or other parallel processing infrastructure. Yeah, so we can, we can take most algorithms we want to deal with implement them in meta in a very abstract way, compile meta to rolling, then the rolling calculus interpreter, it will run these on say a GPU with a thousand processors and the way that gets roughly a thousand times
Speaker 2
speed up. So then, so then
Speaker 1
this gets back to your question like, yeah, we can make this very abstract general stuff, including logics in all sorts of different logic We can make this really fast using the row calculus formalism and this, this gets into the whole algorithm against infrastructure of big tech like every big tech company now is based on map reduce, map reduce being a certain way of boiling down algorithms to run fast on on a bunch of GPUs, right, and however, if you look at the algorithms you need for a GI, these don't work well for map reduce. So you need a more sophisticated way to boil your algorithms down to run officially on on on the computer use which is most of the machine computer architecture we have now, and Greg's row calculus does that, right, which is quite interesting now in the neural net world. They're doing fine with map reduce because they're not doing recurrent neural
Speaker 2
nets too much forward, yeah, transformers
Speaker 1
have a limited amount of recurrence but if you if you got into more richly recurrent neural networks, which I think you do need for really abstract creative cognition, if you go to recurrent neural networks, you're going to get things that are irritating to parallelize using using map reduce, okay, and you're going to want something more like Greg's row calculus also so I mean this, so we're so yeah in implementing the plumbing of open conge hyper on, we're diving quite deep into you know, how do you use mathematics computer science advanced programming language implementation to create a really efficient way to run neural nets logic systems evolutionary learning other sorts of AI, all together on large knowledge graphs that run officially in modern GPU infrastructures were in the middle of that I'd say by the end of this year beginning of next year, we should be far enough through that that we shift more of our effort to building a AI on this infrastructure, okay, and I mean we're playing with that now we're doing logical inference in open conge hyper on me so we're already and we're, we're doing some episodic memory stuff behind the grace Sophia and Desdemona humanoid robots and open the background so on this sort of experimental basis we're playing with parts of the system and some real stuff now, but it's looking like probably early next year before it's really ready for for primetime once we get there, there's really two separate projects that I'm putting a lot of energy into one is, one is coming up with the next thing beyond chat GPT by adding logical reasoning on to chat GPT and then we're on to LLM, so
Speaker 2
would that have recursion in it? And
Speaker 1
without recursion on the logic side, okay, not on the neural side, right, yeah, so that's a bit of a, that system in a way is a bit of a hack job and if you can take a logic engine and hybridize it with a large language model in a way that makes a large language model more capable of multi-step reasoning and more able to ground its other answers in reality, yeah, I mean this will be incredibly huge from an economic and practical and look
Speaker 2
more

Ep. 237 - Artificial General Intelligence and Philosophy w/Dr. Ben Goertzel

Parker's Pensées

Remember Everything You Learn from Podcasts