That’d be a really poor test. Part of the point is that the conversation is allowed to go anywhere at all. If the tester is asking all the same questions to multiple test-ees, then it’s impossible for the conversation to go anywhere.
The whole point of the discussion is that a system can understand. Every AI is a system of some sort. So is every human. And every human or AI can be deconstructed into individual parts each of which clearly does not “understand” (whatever that means).
Good question. Presumably it continues to be developed:
But I think your question is whether it has the ability to expand this knowledge base itself? Can it learn from its own dialogs? Can it annotate its own dialogs then train on those?
Yes, that’s my question. Humans learn through interaction with other intelligent beings (other humans) – we learn from conversations, media, whatever. We remember conversations we’ve had, corrections that the person we’re conversing with make, etc.
I haven’t found any articles stating one way or another. Lemoine claimed it ‘reads’ Twitter, but that doesn’t mean the process is without human intervention or even automatic.
It seems likely there could at least be some per-dialog weight boosting to help differentiate things like whether the word ‘buffalo’ refers to the animal or city. But at the same time, if that is not the intention of this research it may not have it.
“Don’t anthropomorphize X” is a weak, useless argument. It is just special pleading: humans have extra-special property X, which I take as axiomatic, and if anything else shows the same property it’s just fakery.
I’m almost certain you wouldn’t accept the argument elsewhere. If I claimed there’s no such thing as a happy puppy, and that what appears to be happiness in puppies is just mindless behavior selected by humans over time, and that only humans can experience the actual emotion of happiness, I’m pretty sure I know what your response would be.
The real lesson of the Turing test isn’t in the mechanism of evaluating a chatbot. It’s that there’s no difference between a sufficiently advanced fake and the real thing. I’d add the corollary that the real thing never existed in the first place; it’s fakes all the way down.
LaMDA isn’t a great fake at the moment. I think its advanced language skills are covering somewhat for a lack of emotional depth. But the mechanism of how it works is irrelevant to me. To display convincing emotion, that will have to be built into the system, as either an emergent phenomenon or something more explicit. That it’s all just zeroes and ones in the end is true for you as well.
No, I don’t think it is.
Consider this hypothetical situation, which I think is analogous to the Chinese Room thought experiment but which doesn’t involve translation:
Imagine a long list of all possible configurations of a chessboard that could arise in a game of chess, and with each one, one or more moves to be made. And, along with this, imagine an individual—could be a human being, could be a computer or robot—who “plays” chess by, on each turn, looking up the current state of the board on this master list and making the move indicated (or randomly choosing between the indicated moves, if the list has more than one for that particular board configuration).
Does the individual—the person or computer—who is looking up the board configurations and making the indicated moves—understand chess? No: they don’t necessarily understand the most basic things about chess, such as that the object of the game is to checkmate the opponent’s king, or that bishops can only move diagonally.
Does the big list of board configurations and associated moves understand chess? No: it goes against what I understand the word “understand” to mean to say that something like a printed book, or its electronic equivalent, can understand something.
Did whoever or whatever generated the list in the first place have to understand chess to do so? Maybe. But that’s a separate question.
Does the system consisting of the list of board configurations and associated moves (whether in the form of a huge book or scroll or electronic document or whatever) plus the individual looking up the board configurations and responding as directed (whether in the form of a human being, a computer, or a trained chicken) understand chess? That is the question at issue (if I understand correctly). And it doesn’t seem right to me to describe this situation by saying that the list plus the individual together understand chess. I’m not ruling out that I could be persuaded that it does, but I’d still deny that it does so clearly.
I consider myself sentient, and I didn’t get to choose my name.
You didn’t choose “Schnitte” when you registered for the SDMB?
Seems like the book is encoded with some understanding of how the chess pieces can be moved. I wasn’t sure if the book listed all possible moves or one or more ‘good’ moves. If it lists good moves, then the book has some understanding of how to move a chess game to an end state.
Obviously there is a bunch of information missing from the book: who goes first, what is win/lose/draw, has each player castled, threefold repetition, that this is a game, what a game is, why play a game, etc.
But it depends on the definition of ‘understand’. If ‘understand’ must be more than a collection of facts, but also some agency to use that data, then the book doesn’t understand anything.
Chess is a closed system, with clearly defined rules and a specific (if mind-boggling) set of moves. There’s nothing in principal that says that chess couldn’t be solved, every possible board and move figured out (except for limiting things like the number of particles in the universe, time before heat death, etc.).
Language and conversation isn’t a closed system with a specific set of rules and allowable responses. I’m not sure that you could, even in principle, lay out all the possible conversations, as I tried to demonstrate above.
If you could construct a Chinese language room that really worked (I’m not sure it can be, even in principle), I would say that it has as much understanding as your neurons and hormones do. Otherwise, what else is there?
If it’s possible for that chess book to even exist in the Universe, and for it to be used by any entity in the Universe, then it must be written in a format that requires a great deal of understanding of chess to be able to write or read it. A chess book with notation simple enough to be used by an entity without understanding would need to be many orders of magnitude larger than the entire Universe. And I think it silly to claim that a system many orders of magnitude larger than the entire Universe couldn’t possibly “understand” anything.
Sure you could. Every language can be encoded to a binary string. Every statement is finite in length, so every statement can be a distinct integer. Just map that integer to a page in the book with the response.
You can handle full conversations this way, too, if you want to have some kind of “memory”. Just encode the entire conversion so far to an integer and map to the response.
If you want some element of randomness, that’s not hard to add; just have a large number of possible responses for each conversation and pick a random one.
Not really. “Hey, what is one divided by 3?” “Zero point three three three three three three three three…three three…etc.” could go on indefinitely.
The set of all finite strings is not finite.
I wouldn’t consider that a valid statement. How could there ever be a followup response? I have to wait forever to decide if it was correct or not, because the person might stick a seven in there at some point.
Of course, I can say “Zero point three, repeating”, which is a valid answer and finite in length.
True. Though the set of strings expressible by a human within a normal lifetime is finite.
Or, maybe each page in my book is 1% thinner than the previous.
Right, but just because any particular human only has time to express a finite number of strings, the number of strings that the Chinese Room has to account for would be infinite, right? It has to respond “intelligently” to any of the infinite strings that could be put to it, not just those expressible by a particular human.
If the Chinese room just has to be as good as a human, it can be finite in size. If it needs to handle all possible strings, no matter how long–then sure, you’ve got some problems with infinity there.
Personally, I think the fact that it’s physically unrealizable, even in the simple case of chess, makes it a somewhat uninteresting example in the standard presentation. What we need is some kind of “data compression” that allows things like knowledge of the rules of grammar, the meaning of words, and so on. Then it doesn’t need a response to all possible input strings; it can extrapolate the meaning of novel strings from ones it already knows.
But taken to an extreme, we just have a brain. Data compression is synonymous with understanding. It’s exactly the thing that makes language possible in our universe. Another entity that can do the same thing should also be considered to have an understanding of language.
Part of the appeal of the thought experiment is that the room seems very mechanical–it just takes an input and maps to the output. But ultimately, so is the universe. Instead of doing a straight lookup, we could have our room keeper do a series of calculations–take these two numbers, multiply them, add to this third number, etc. But if the keeper can do that, they can simulate an entire universe–a universe that contains our brains already.
If Google can make chatbots like LaMDA that can pass the Turing test most of the time, then the Chinese can do it too. If the Chinese included a (non-chinese-speaking) human being and a comprehensive rulebook inside the workings of their chatbot, then they would have a Chinese Room that would fulfil nearly all of Searle’s constraints. The human is just doing some of the processing, but the rulebook is doing the donkey work. You could probably still detect the difference if you talked to it long enough.
Give it ten or twenty years and there will be chatbots that can’t be distinguished from humans by the best judges. Will they be sentient? Probably not, because they won’t feel real happiness or fear, or have the ability to formulate their own goals. It will take a lot longer before that is possible, and I’m glad to say I won’t see it happen.
Fighting the hypothetical? What about if we replace chess with a game with fewer possible configurations, like tic-tac-toe?