"Neurallambda".
"The Problem: My premise all comes down to 'reasoning', and the lack thereof, in current AI models. I'll provide my working definition of 'reasoning', but for a moment, please bear with a couple examples of reasoning failures."
"Transformer models cannot reason."
"Diffusion models cannot reason."
"AI is currently like a living textbook."
"What is Reasoning? Reasoning is the ability to know true things without having learned them. It is building knowledge/predictions/retrodictions/actions atop principles, instead of evidence."
"What are programs?" "Turing Machines are machines capable of executing programs which can calculate anything calculatable. Your computer is a Turing machine (in the limit of infinite memory and time). I'd also suggest that your conscious mind is Turing Complete."
"Tiers of 'Programming Ability' / 'Reasoning Ability'":
"1. An AI can execute programs"
"2. An AI can verify traits of a program"
"3. An AI can generate novel programs during training"
"4. An AI can generate novel programs post training"
"So far, this library provides an existence proof up to Level 1. It contains code which can execute arbitrary programs, written in a custom lisp dialect, in a fully differentiable setting. (Some tantalizing tests have proven up to Level 3, that AI can learn novel programs to solve toy problems via SGD, but, there are still frontiers of research here)."
By "this library", he is talking about his creation "Neurallambda". What is Neurallambda? It's a dialect of lisp designed to be generated by AI systems but at the same time be human-readable. It also has the important attribute that all the code generated using it is "differentiable". That means the code itself can be incorporated into a stochastic gradient descent model. That's what the "SGD" stands for above. In this form, Neurallambda code can be deterministically translated into tensors, compiled or interpreted, and then the resulting tensors be read back out, and presented in human-readable form.
What do y'all think? Is this a path to computer reasoning?
