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.
(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)
...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.
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.
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.
Alexa is a different matter entirely, that one actively troll the web in response to your queries.
(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...)
Well... yes and no. Deployed models typically learn in a defined pattern, if at all. Various forms of data freshness, etc. to develop. But the chatbots don't have good history recall, typically, and know that what you mentioned 50 messages ago is relevant to message one prior and not current. Things like that. We don't program pareidolia very well, which is typically seen as a negative, but its a feature for finding useful patterns (not just lowest error patterns).
If so, I agree. I think that's our big flaw, in fact, because we instinctually apply patterns from birth, even when those patterns shouldn't be applied. So we see faces in the moon or on mars that aren't there. We see shapes moving in the dark that don't exist. And we seem to believe that ML models will develop over time as children or animals do, based on nothing more than our perceptions of similarity, our instinct to apply patterns even when we shouldn't.
Unlike a baby human, that ML model isn't going to develop increased complexity of thought over time. It's already maxed out. New models might up the complexity slightly, but that baby is going to vastly surpass any existing model in weeks or days.