3 ms·

What is the "deep reasoning" that leads one to the conclusion that a language model is capable of any kind of insight? How does training something like this to

What is the "deep reasoning" that leads one to the conclusion that a language model is capable of any kind of insight? How does training something like this to be better and better at mimicking somehow indicate something about general AI? Does reasoning ability lie somewhere on the path of better mimicry? What's the evidence for that? When is one of these going to solve the Riemann hypothesis?

atleastoptimal 4y ago
> What is the "deep reasoning" that leads one to the conclusion that a language model is capable of any kind of insight? The ability to notice that language models make the same fundamental abstract connections that humans do, which is the essential building block of human reasoning. > How does training something like this to be better and better at mimicking somehow indicate something about general AI? Does reasoning ability lie somewhere on the path of better mimicry? What's the evidence for that? Claiming that the fact that LLMs are trained to "mimic" means they don't reason implies that human beings aren't sentient because they spend the first 20+ years of their life learning through imitation of their peers, mentors and media. Does the fact that Einstein was once a helpless baby who needed his mother to feed him and teach him his first words mean that he is just a parrot with no insight? Does the fact that most artists in human history went to school and mimicked the old masters means they never developed creative insight? > When is one of these going to solve the Riemann hypothesis? When are the tens of thousands of fine-tuned 150+IQ intellects going to solve the Riemann hypothesis?
- alienicecream 4y ago
  Imitation is not the same as reasoning, imagination, or abstraction. You're arguing that Einstein babbled his way to general relativity. There is a fundamental difference in what Einstein did as a baby and what he did when he was doing science, it's not just 20 years of more of the same but with a bigger data set. You're going to say that that what you mean is that we are at the beginning of AI development, but that is not an argument unless you can put forward a plausible theory of how you get from babbling to reason and abstraction. In a certain sense we don't even know why humans haven't solved the Reimann hypothesis. There is a creative leap somewhere that we are missing. And we don't how to make it, or how to make it occur to us. But if reasoning is just the idiocy of mimicry then GPT should have solved it already.
pessimizer 4y ago
> What is the "deep reasoning" that leads one to the conclusion that a language model is capable of any kind of insight? That almost all of the failures one notices with them are failures to drag along and maintain/update state? To me, that's just a lack of scaffolding. If LLMs can't add, give them calculators. I feel outdated even saying this. Certainly people are quickly coming up with dozens of elaborate ways for LLMs to build and update state. They're just not releasing them.
- alienicecream 4y ago
  It doesn't know how to store and transform the state, it knows symbols but it doesn't know their meaning so how is it going to transform the state? How is it going to know what the state and limits of objects are or the state and limits of concepts are? It's there a way around that without ushering in complete nonsense? These models are an attempt to circumvent the issue. It's a gorilla in a suit with a team of researchers that are trying to make it look like a person.

atleastoptimal 4y ago

> What is the "deep reasoning" that leads one to the conclusion that a language model is capable of any kind of insight? The ability to notice that language models make the same fundamental abstract connections that humans do, which is the essential building block of human reasoning. > How does training something like this to be better and better at mimicking somehow indicate something about general AI? Does reasoning ability lie somewhere on the path of better mimicry? What's the evidence for that? Claiming that the fact that LLMs are trained to "mimic" means they don't reason implies that human beings aren't sentient because they spend the first 20+ years of their life learning through imitation of their peers, mentors and media. Does the fact that Einstein was once a helpless baby who needed his mother to feed him and teach him his first words mean that he is just a parrot with no insight? Does the fact that most artists in human history went to school and mimicked the old masters means they never developed creative insight? > When is one of these going to solve the Riemann hypothesis? When are the tens of thousands of fine-tuned 150+IQ intellects going to solve the Riemann hypothesis?

alienicecream 4y ago
Imitation is not the same as reasoning, imagination, or abstraction. You're arguing that Einstein babbled his way to general relativity. There is a fundamental difference in what Einstein did as a baby and what he did when he was doing science, it's not just 20 years of more of the same but with a bigger data set. You're going to say that that what you mean is that we are at the beginning of AI development, but that is not an argument unless you can put forward a plausible theory of how you get from babbling to reason and abstraction. In a certain sense we don't even know why humans haven't solved the Reimann hypothesis. There is a creative leap somewhere that we are missing. And we don't how to make it, or how to make it occur to us. But if reasoning is just the idiocy of mimicry then GPT should have solved it already.

pessimizer 4y ago

> What is the "deep reasoning" that leads one to the conclusion that a language model is capable of any kind of insight? That almost all of the failures one notices with them are failures to drag along and maintain/update state? To me, that's just a lack of scaffolding. If LLMs can't add, give them calculators. I feel outdated even saying this. Certainly people are quickly coming up with dozens of elaborate ways for LLMs to build and update state. They're just not releasing them.

alienicecream 4y ago
It doesn't know how to store and transform the state, it knows symbols but it doesn't know their meaning so how is it going to transform the state? How is it going to know what the state and limits of objects are or the state and limits of concepts are? It's there a way around that without ushering in complete nonsense? These models are an attempt to circumvent the issue. It's a gorilla in a suit with a team of researchers that are trying to make it look like a person.