On Mon, Dec 16, 2024 at 6:42 AM Matt Mahoney <mattmahone...@gmail.com> wrote:
> YKY, what is the title of your thesis? What is the central idea you are > trying to prove? What unknown knowledge do you hope to discover? Are you > going to do any experimental work to prove your ideas, or is this purely a > math paper? > Working title is: Algebraic Logic for Deep Learning. The goal is to build logic-based AI from deep learning. Yes, I will do experiments on TicTacToe. > I know you have been working on a symbolic approach to AGI for a couple of > decades now. The reason everyone was doing this in the 1980's was because > we had no choice. We didn't have either the hardware or the training data > to run LLMs. It wasn't until we used neural networks that we solved > problems like brittleness, ambiguity, and the lack of a learning algorithm > for language models. We have known about these problems since the late > 1950's. Given the amount of work that has gone into solving them, do you > have any new insights? > Deep learning is revolutionary for the single reason: it learns faster, making it possible to learn from huge datasets, previously impossible with symbolic methods. I remake symbolic AI with neural networks to gain the same learning efficiency. So it'd be different this time. Also this approach enables to use the intuition we had about logic-based systems to build AGI. Right now, the Transformer is state-of-the-art, but everyone seems beguiled by how to interpret its mechanism. ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T8685950780e86bd5-Mcb61cef24e02991d69b3e2f6 Delivery options: https://agi.topicbox.com/groups/agi/subscription
