3 ms·
Have you actually used ChatGPT? The output is way beyond being grammatically correct. It does make sense and is correct on all levels - vast majority of the tim
by p1esk 4y ago
Have you actually used ChatGPT? The output is way beyond being grammatically correct. It does make sense and is correct on all levels - vast majority of the time. I'm talking about multi-paragraph long semantic coherence, to the extent where it's impossible to tell it was not human produced. Who cares whether it "means" anything in a philosophical sense?
What makes you think our brains do anything other than "blind statistics"? What is a human brain if not a multi-trillion parameter statistical machine, trained on a vast amount of data, implicitly over the billions of years of evolution, and explicitly over decades of individual lifetime? It's actually amazing that a completely randomly initialized model like GPT can be so rapidly trained to communicate effectively - using just text as training material!
GPT still fails on many occasions, such as ignoring context or simply not knowing enough about the world, but these are not fundamental limitations. It can and will be trained to pay attention to context, to query external data sources, and to construct a better world map. It will be trained using multiple data types, like audio and video inputs in addition to text, that way it will learn a lot more about how the world works, physics, etc.
People are constantly redefining "true intelligence" as whatever these ML models can't do yet. That definition has changed a lot in the last ~20 years.
- gamblor956 4y agoI have used ChatGPT. To say that is is correct on all levels, or even most of the time is like saying that the sky is neon green. ChatGPT does not communicate effectively; it communicates quite poorly. It simply does a good job of hiding the lack of substance in its output. ChatGPT is syntactically correct, but it's rare to get something that is semantically correct, or meaningful, for anything but the simplest prompt. To put it bluntly: AI isn't intelligent until it can reason, and that's not something that can be learned from statistics. ChatGPT can say the sky is blue, and regurgitate a Wikipedia article about it, but it will never understand why the sky is blue, or be able to use that knowledge to answer a question that for which the answer isn't already in its dataset. That's not intelligence. That's just a very good search program paired to a wiki.
- p1esk 4y agoI don’t know why the sky is blue. Someone once tried to explain it to me and I didn’t quite understand it. What does it make me? Many of my answers - including most of this one - are regurgitation of information I learned previously. Many, if not most people just follow established patterns when communicating, with no deep understanding involved. Even that deep understanding is also pattern matching, where patterns are simply more complex. You can easily find examples of an ML model performing multi step reasoning to answer a complex question, just like a human would. I can google those examples for you if you like.
- gamblor956 4y agoYou're changing the goal posts from what I said. I merely said understanding; you've decided this means "deep understanding," whatever that is. I don't appreciate gaslighting, so I'll be withdrawing from this conversation and you're free to continue your fantasy about ML taking over the world, like VR, blockchain, and big data before it.