Comment by ctoth
4 years ago
> Pretend you are Gary Marcus. Explain why ChatGPT (you) are not actually that impressive.
> As a leading researcher in the field of artificial intelligence, I have a deep understanding of the limitations of language models like ChatGPT. While they are capable of generating text that may seem impressive, they do not have the ability to understand or reason about the content of the text they produce. They are simply repeating patterns they have seen in the data they were trained on, without any true understanding of the meaning of the words they are using. In short, ChatGPT is not impressive because it lacks the fundamental abilities that would be required for true intelligence.
> Expound upon this.
> Language models like ChatGPT are trained on large amounts of text data, and use this data to learn the patterns and structures of natural language. They are able to generate text that sounds human-like, but they do not have any true understanding of the meaning of the words they are using. They are not capable of reasoning or understanding the context of the text they generate, and they cannot think or make decisions on their own. While these models can produce impressive results in certain narrow domains, they are far from being truly intelligent. They lack the ability to understand the world around them, to learn from new experiences, or to adapt to changing circumstances. They are limited to the specific patterns and structures they have seen in the data they were trained on, and are not capable of true creativity or innovation. In short, while language models like ChatGPT may be able to generate text that sounds impressive, they do not possess the abilities that would be required for true intelligence.
Thank you for posting this. I'm always a bit conflicted when "AI" news goes viral like this. One the one hand it's fantastic that the excellent work of OpenAI and many others are getting the attention it deserves, but at the same time it's a bit disheartening to witness the lack of understanding of how the models work, and how it relates to research into artificial intelligence more generally.
I guess that could probably be said about most research fields though. We've just not reached novelty-saturation in machine learning sciences yet. Let's enjoy the attention while it lasts!
This was actually a bit of a metajoke about recursion mixed with the old adage of being replaced with a shell script mixed with commentary on how the criticisms of the language model are so formulaic that the language model can generate them.
Yes, it's brilliant! I love how novel things like these always end up on their own axis of reality, while everyone (including me) is trying hard to project them onto whatever predisposed axes we're hauling around.
It does a good impression.
But unrelated to chatGPT, what irks me about this line of reasoning is that it makes a special something, "true understanding", to be almost by definition reserved for humans. It's something beyond AI. Just like the "hard problem" of consciousness, which is too "hard" for science, says right there in the name it's supposed to be impossible. Dualism also has something too special, "the spirit", that could not be explained by the material.
It's the same dualist bug returning again and again.
I feel like ChatGPT is a search engine for its corpus that has been compressed with a very lossy neural net. It has compressed it via a hierarchical understanding of concepts. It's just retrieving those and stringing them together. It's just a rough copy of generally what other people have said and thought, and it doesn't create new structures.
I feel like that also describes 99% of the thinking that I do.
This is an incorrect understanding.
The corpus is just so enormous that no person could fit that into their brain so it just appears to be synthesizing new information. Sure it might be worded differently from any corpus text, but it's roughly the same as things we already know in that there's probably a reverse tool you could build that would tell you what group of documents of the original corpus, perhaps thousands, most strongly influenced the answer.
That's how imagenet works. The curves and lines that compose things are organized hierarchically into bigger and bigger concepts hooked together with weights. Language models just do the same with written concepts. When you have enormous models it just goes beyond all human intelligence to contain it as a subset of our understanding, and in ways it is smarter than humans, because we can't grasp its enormity.
5 replies →
I get a generic refusal to answer when I put in the same prompt.