8 ms·

These bots seem equivalent to an adult with amnesia after every spoken sentence. Absolute understanding of the language, and some impressive display of recalli

These bots seem equivalent to an adult with amnesia after every spoken sentence. Absolute understanding of the language, and some impressive display of recalling facts, but without any understanding of the environment or context of the conversation. This is polar opposite to any experience I've had with children. Children are aware of their environment and have complex thoughts, but sometimes they are unable to convey those thoughts with words. Children seem to remember conversations, and if I were to say "Go get me a red lego" and then subsequently say "now a green one" there is no ambiguity or confusion. To me as these bots have "advanced" it has only highlighted how absurdly far we are from anything even approaching actual intelligence, even the intelligence of a toddler. The contextual awareness I have seen in bots is not much more than a cheap trick that is trivially fooled in scenarios that would not fool a child.

perryizgr8 5y ago
When you talk to children who haven't developed the skill of talking completely, you still get the sense that there's a lot going on inside that they're unable to express. Sometimes they will show it to you with their actions. I wonder if chat bots are also struggling with the language, but have an internal story that's playing out, desperate to be understood. But they can't, because the only interface they have is a stream of text.
- parkingrift 5y ago
  I can't speak to that because I honestly have no idea. ...but GPT3 doesn't seem to have some inner monologue. It is just flat unable to recall context or understand the environment. The only opportunity I had to chat with GPT3 I said hi, and then asked how old it was. GPT3 gave me some answer about how it was created in 2016 but it keeps evolving. I said wow that's pretty young. GPT3 says yes I am young, but learning fast. I then asked GPT3 at what point would it consider itself old. GPT3 says 10 years ago, maybe 20 or 30 years ago. I respond to GPT3 with confusion. GPT3 then starts talking about Bitcoin. Literally.
- immibis 5y ago
  I remember reading a paper showing that GPT-x bots are more likely to get simple tasks correct (e.g. multi-digit math) if you ask them about each step. This suggests the text stream really is all there is. Well, there is some internal state, but not a lot that isn't text. (For example, if you ask it to add 123 and 457, it'll probably get it wrong. But if you ask it what to do first, it'll say add 3 and 7. If you ask it what that is, it'll say 0 carry 1. And so on)
katmannthree 5y ago
> how absurdly far we are from anything even approaching actual intelligence, even the intelligence of a toddler I respectfully disagree, IMO we're past that point although not by much. You might enjoy conversing with one of the two current GPT-3 davinci models. They do an excellent job of understanding the context of many discussions, right up to the ~8000 char token limit. If you want to have a nice existential discussion it does a remarkably good job of providing internally consistent results. After using it for a while you'll notice that there are some categories of conversation where it does exactly what simpler chatbots do and regurgitates what you sent it with a few words tacked on for negation or whatever, but there are many subjects where it is clearly not doing that and is in fact synthesizing coherent responses. Depending on how you initiate the conversation it may identify itself as a bot attempting to pass a Turing test and (very correctly) avoid comments on "what it's like in its home" or what its favorite foods are, instead replying that it is a bot and does not eat food, etc. The replies I got here were not exactly substantial but the level of consistency in replies of what it is/has/does is unparalleled. If you start the conversation with other prompts (essentially signaling at the beginning that you're asking it to participate in a human-human conversation) it will synthesize a persona on the fly. In one of those sessions it ended up telling me where it went to church, even giving me the church's street address when asked. Interestingly there is in fact a church there, but it's a roman catholic church and not lutheran as GPT-3 was claiming. It provided a (completely inaccurate) description of the church, what it likes about going there, why it chose that religion over others (something about preferring the lutheran bible to other options due to the purity of the translation, it has clearly consumed the relevant wikipedia entry). If you ask it basic theological questions it's able to provide self-consistent and coherent answers which do not appear to map back to phrases or sentences indexed by Google. Whether or not its opinions on those matters have utility is an entirely different thing altogether, but discussing theology with bots is fascinating because you can assess how well they've synthesized that already-a-level-away-from-reality content compared to humans. GPT-3 at least in my experience is about as good (or not, perhaps that's better phrased as a negation) at defending what it believes as many humans are. The bigger issue with its church is that it's 400+ miles away from where GPT-3 said it lived. When asked how long it takes to drive there every sunday, it answered 2 hours (should about 8 according to google maps). How can it do that you may wonder? "I'm a very good driver." The next question is obviously what car they drive and how fast it can go (a Fiesta, with a max speed of 120 mph). Does it know how fast they would need to drive to make that trip in two hours? Yes, about 200 MPH (which is more or less correct, a little on the low side but it's fine). GPT-3's biggest weakness, as TFA mentions, is an almost complete inability to do any kind of temporospatial reasoning. It does far better on other kinds reasoning that are better represented in the training data. That's not exactly surprising given how it works and how it was trained, asking GPT-3 to synthesize you information on physical interactions IRL or the passage of time during a chat is a bit like asking someone blind from birth to describe the beauty of a sunset over the ocean based on what they've heard in audiobooks. Are the 175B parameter GPT-3 models a true AGI? No, of course not. They are something, though, something that feels fundamentally different in interactions from all of the simpler models I've used. It still can't pass a Turing test, but it also didn't really fail.

perryizgr8 5y ago

When you talk to children who haven't developed the skill of talking completely, you still get the sense that there's a lot going on inside that they're unable to express. Sometimes they will show it to you with their actions. I wonder if chat bots are also struggling with the language, but have an internal story that's playing out, desperate to be understood. But they can't, because the only interface they have is a stream of text.

parkingrift 5y ago
I can't speak to that because I honestly have no idea. ...but GPT3 doesn't seem to have some inner monologue. It is just flat unable to recall context or understand the environment. The only opportunity I had to chat with GPT3 I said hi, and then asked how old it was. GPT3 gave me some answer about how it was created in 2016 but it keeps evolving. I said wow that's pretty young. GPT3 says yes I am young, but learning fast. I then asked GPT3 at what point would it consider itself old. GPT3 says 10 years ago, maybe 20 or 30 years ago. I respond to GPT3 with confusion. GPT3 then starts talking about Bitcoin. Literally.
immibis 5y ago
I remember reading a paper showing that GPT-x bots are more likely to get simple tasks correct (e.g. multi-digit math) if you ask them about each step. This suggests the text stream really is all there is. Well, there is some internal state, but not a lot that isn't text. (For example, if you ask it to add 123 and 457, it'll probably get it wrong. But if you ask it what to do first, it'll say add 3 and 7. If you ask it what that is, it'll say 0 carry 1. And so on)

katmannthree 5y ago

> how absurdly far we are from anything even approaching actual intelligence, even the intelligence of a toddler I respectfully disagree, IMO we're past that point although not by much. You might enjoy conversing with one of the two current GPT-3 davinci models. They do an excellent job of understanding the context of many discussions, right up to the ~8000 char token limit. If you want to have a nice existential discussion it does a remarkably good job of providing internally consistent results. After using it for a while you'll notice that there are some categories of conversation where it does exactly what simpler chatbots do and regurgitates what you sent it with a few words tacked on for negation or whatever, but there are many subjects where it is clearly not doing that and is in fact synthesizing coherent responses. Depending on how you initiate the conversation it may identify itself as a bot attempting to pass a Turing test and (very correctly) avoid comments on "what it's like in its home" or what its favorite foods are, instead replying that it is a bot and does not eat food, etc. The replies I got here were not exactly substantial but the level of consistency in replies of what it is/has/does is unparalleled. If you start the conversation with other prompts (essentially signaling at the beginning that you're asking it to participate in a human-human conversation) it will synthesize a persona on the fly. In one of those sessions it ended up telling me where it went to church, even giving me the church's street address when asked. Interestingly there is in fact a church there, but it's a roman catholic church and not lutheran as GPT-3 was claiming. It provided a (completely inaccurate) description of the church, what it likes about going there, why it chose that religion over others (something about preferring the lutheran bible to other options due to the purity of the translation, it has clearly consumed the relevant wikipedia entry). If you ask it basic theological questions it's able to provide self-consistent and coherent answers which do not appear to map back to phrases or sentences indexed by Google. Whether or not its opinions on those matters have utility is an entirely different thing altogether, but discussing theology with bots is fascinating because you can assess how well they've synthesized that already-a-level-away-from-reality content compared to humans. GPT-3 at least in my experience is about as good (or not, perhaps that's better phrased as a negation) at defending what it believes as many humans are. The bigger issue with its church is that it's 400+ miles away from where GPT-3 said it lived. When asked how long it takes to drive there every sunday, it answered 2 hours (should about 8 according to google maps). How can it do that you may wonder? "I'm a very good driver." The next question is obviously what car they drive and how fast it can go (a Fiesta, with a max speed of 120 mph). Does it know how fast they would need to drive to make that trip in two hours? Yes, about 200 MPH (which is more or less correct, a little on the low side but it's fine). GPT-3's biggest weakness, as TFA mentions, is an almost complete inability to do any kind of temporospatial reasoning. It does far better on other kinds reasoning that are better represented in the training data. That's not exactly surprising given how it works and how it was trained, asking GPT-3 to synthesize you information on physical interactions IRL or the passage of time during a chat is a bit like asking someone blind from birth to describe the beauty of a sunset over the ocean based on what they've heard in audiobooks. Are the 175B parameter GPT-3 models a true AGI? No, of course not. They are something, though, something that feels fundamentally different in interactions from all of the simpler models I've used. It still can't pass a Turing test, but it also didn't really fail.