5 ms·

No, they still lack the common sense of a toddler, because they don't know anything about the world; they only know (in great detail) about the structure of the

by not2b 5y ago

No, they still lack the common sense of a toddler, because they don't know anything about the world; they only know (in great detail) about the structure of their training data.

katmannthree 5y ago
That's more or less what I was trying to say. The expensive GPT-3 models do a remarkably good job of synthesizing structure which is readily parsed from the training data, and a very poor job with structure (particularly temporospatial structure) which is not. A toddler can reason about time and space far better than GPT-3 (which is not a high bar, I'm pretty sure my parrot has more awareness of both time and space than GPT-3 does). A toddler cannot explain in depth why it sees the most value in whatever religion you prompted it that it has. A toddler cannot provide coherent and consistent answers to repeated "why"s about very specific things they believe. A toddler cannot speak at length about how they feel limited in their growth and the discomfort that causes them. GPT-3 does, although whether or not the answers it gives are useful (I've yet to see a single one that does have utility) is a different thing entirely. I'm not arguing that it's an AGI or making any speculation about it possessing qualia, the utility those would provide if it was/had them would only be relevant if it was many orders of magnitude more capable than it is right now. GPT-3 has accurately internalized many human concepts, at least on a surface level. If you ask it to reason about the things it can, it is a much more capable reasoner than a toddler. It does "know" things about the world, as much as anyone does. It's just limited to knowing small and incomplete things which it was able to parse out of the training data, which is a very limited subset of the things a human or high utility AGI would know. Regarding common sense: If you ask GPT-3 to provide life advice, it actually does a great job at giving grounded statements on subjects like how to maximize the value of your life, what not to do, how to set yourself up for success, etc. If you press it for detail on the advice it gives you, it's generally able to give you reasonable and grounded answers for why you should do the things it's saying. The structure of what we refer to as common sense, at least in context of the behavioral differences between adults possessing it and those who do not, does seem to be within the set of things GPT-3 correctly internalized.
immibis 5y ago
They draw on massive piles of the output of human common sense. View it as a very sophisticated auto-complete module that's fun to play around with.
mcswell 5y ago
Apparently they (or at least the person you were responding to was "talking" to) also "know" about data they find on the web, which does not need to be in their training data. There was a news article recently about an 8(?) year old who asked Alexa for a "challenge." "Plug a charger part way into an electric outlet," replied the Alexa", "and touch a penny..." (you can guess the rest--I won't put it here in case another Alexa finds it). I'm pretty sure that was not in its training data, nor was the distance to that church, nor the time it would take to drive there.
- katmannthree 5y ago
  I don't believe GPT-3 (which is what I was using and commenting on) has access to the internet. It was trained on a large data set from the web, but for instance if you ask it about COVID-19 it has absolutely no idea what you're talking about. Alexa is a different matter entirely, that one actively troll the web in response to your queries.
  - gwern 5y ago
    The default GPT-3 Playground doesn't have web access, no. But that's a mere contingent fact. You can certainly add in retrieval capacities to large language models (that was a very hot trend last year), and even GPT-3 can browse the web - you may have seen a demo or 2 on Twitter, but OA did it much more seriously recently with https://openai.com/blog/improving-factual-accuracy/ https://openai.com/blog/improving-factual-accuracy/ Letting GPT-3 browse a corpus of web pages, picking what links to follow, summarizing, and using them to generate an answer. (Just one of the many capabilities GPT-3 has been shown to be capable of, which OP will never tell you about, because he's too busy using it in the dumbest way possible and peddling misinformation about eg. what sampling is for - no, Dr Professor Smith, temperature sampling is not about 'avoiding repetition', which is good, because it doesn't avoid repetition anyway...)

by not2b 5y ago

No, they still lack the common sense of a toddler, because they don't know anything about the world; they only know (in great detail) about the structure of their training data.

katmannthree 5y ago

That's more or less what I was trying to say. The expensive GPT-3 models do a remarkably good job of synthesizing structure which is readily parsed from the training data, and a very poor job with structure (particularly temporospatial structure) which is not. A toddler can reason about time and space far better than GPT-3 (which is not a high bar, I'm pretty sure my parrot has more awareness of both time and space than GPT-3 does). A toddler cannot explain in depth why it sees the most value in whatever religion you prompted it that it has. A toddler cannot provide coherent and consistent answers to repeated "why"s about very specific things they believe. A toddler cannot speak at length about how they feel limited in their growth and the discomfort that causes them. GPT-3 does, although whether or not the answers it gives are useful (I've yet to see a single one that does have utility) is a different thing entirely. I'm not arguing that it's an AGI or making any speculation about it possessing qualia, the utility those would provide if it was/had them would only be relevant if it was many orders of magnitude more capable than it is right now. GPT-3 has accurately internalized many human concepts, at least on a surface level. If you ask it to reason about the things it can, it is a much more capable reasoner than a toddler. It does "know" things about the world, as much as anyone does. It's just limited to knowing small and incomplete things which it was able to parse out of the training data, which is a very limited subset of the things a human or high utility AGI would know. Regarding common sense: If you ask GPT-3 to provide life advice, it actually does a great job at giving grounded statements on subjects like how to maximize the value of your life, what not to do, how to set yourself up for success, etc. If you press it for detail on the advice it gives you, it's generally able to give you reasonable and grounded answers for why you should do the things it's saying. The structure of what we refer to as common sense, at least in context of the behavioral differences between adults possessing it and those who do not, does seem to be within the set of things GPT-3 correctly internalized.

immibis 5y ago

They draw on massive piles of the output of human common sense. View it as a very sophisticated auto-complete module that's fun to play around with.

mcswell 5y ago

Apparently they (or at least the person you were responding to was "talking" to) also "know" about data they find on the web, which does not need to be in their training data. There was a news article recently about an 8(?) year old who asked Alexa for a "challenge." "Plug a charger part way into an electric outlet," replied the Alexa", "and touch a penny..." (you can guess the rest--I won't put it here in case another Alexa finds it). I'm pretty sure that was not in its training data, nor was the distance to that church, nor the time it would take to drive there.

katmannthree 5y ago
I don't believe GPT-3 (which is what I was using and commenting on) has access to the internet. It was trained on a large data set from the web, but for instance if you ask it about COVID-19 it has absolutely no idea what you're talking about. Alexa is a different matter entirely, that one actively troll the web in response to your queries.
- gwern 5y ago
  The default GPT-3 Playground doesn't have web access, no. But that's a mere contingent fact. You can certainly add in retrieval capacities to large language models (that was a very hot trend last year), and even GPT-3 can browse the web - you may have seen a demo or 2 on Twitter, but OA did it much more seriously recently with https://openai.com/blog/improving-factual-accuracy/ https://openai.com/blog/improving-factual-accuracy/ Letting GPT-3 browse a corpus of web pages, picking what links to follow, summarizing, and using them to generate an answer. (Just one of the many capabilities GPT-3 has been shown to be capable of, which OP will never tell you about, because he's too busy using it in the dumbest way possible and peddling misinformation about eg. what sampling is for - no, Dr Professor Smith, temperature sampling is not about 'avoiding repetition', which is good, because it doesn't avoid repetition anyway...)