On Wednesday, November 22, 2023, at 3:21 PM, Matt Mahoney wrote: > Symbolic AI is dead. Expert systems, knowledge representation, Cyc, and > OpenCog/RelEx showed it doesn't work. It would have died sooner if we had > petaflop GPUs and petabytes of training data in the 1980's. We had neural > networks then but couldn't scale it to AI. We knew the brain's computing > power and how much input data we receive over a lifetime.
It is just that ANNs got clued up sooner, I believe. On Wednesday, November 22, 2023, at 3:21 PM, Matt Mahoney wrote: > Symbolic language modeling doesn't work because it put grammar before > semantics, like a programming language. You can't parse English without > knowing what the words mean. Putting semantics before grammar, the way > children learn language, let's you train a neural network one layer at a > time. Our brains didn't just evolve to understand language. Language also > evolved to make learning efficient. The way I see it, ANNs are black box program synthesis machines. You can do program synthesis by copying brain in XYZ petabytes. It would work. But you can do the same synthesis in other, more mathematical ways. It usually leads to combinatorial explosion, but the results are 100% guaranteed to reproduce correctly the learned material if you are willing to wait for the response. With ANNs, we just have to rely on hardware progress, and we will eventually be there. Given the brute force, even to surpass the human intelligence sooner or later. But with symbolic AI... things are getting tricky and puzzling in trying to optimize mentioned combinatorial explosion. And I'm very interested in those puzzles. The solution to those show the real creativity of the human kind. Imagine some extraterrestrials come down to the Earth tomorrow, and ask about how we did the AI. Which would be cooler to say: (1) sure, we did it, we just copied the Nature and inherited all the virtues and flaws, no big deal, or (2) oh, we did it our own way, it was a bit itchy, but we finally clued it up and tweaked it so it runs just the way we want, no more, no less, wait to hear the whole story... ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T3c8c80a465a14daa-Mca6a607fb15a361c7f5617a7 Delivery options: https://agi.topicbox.com/groups/agi/subscription
