There’s been lots of interesting discussion and debate recently regarding chatbots and their implications for thinking about artificial intelligence. I became aware of this issue from a post by economist Gary Smith pointing out an article by Google engineer Blaise Agüera y Arcas that made strong claims about computer language models. Agüera y Arcas claimed that, with these models, “statistics do amount to understanding,” but he presented no replicable evidence in support of this claim and avoided engaging with any questions about it. See here and here for followup. I don’t know whatssup with Google and nonreplicable research claims.
Anyway, the quick story of the debate is that Agüera y Arcas presented some purported examples of a chatbot sounding smart, and Smith presented some examples of a chatbot sounding clueless. Chatbots continue to get better, so it’s easier and easier to have a dialogue where the chatbot sounds smart, and harder and harder to have a dialogue where the chatbot sounds dumb.
The basic terms of the debate come down to: (1) when a chatbot usually sounds smart, can that be considered a form of intelligence?, and (2) when a chatbot sounds clueless, can that be taken as a refutation of the claim of intelligence, or does it just bound the domain of the chatbot’s intelligence?
You can find lots of recent discussion of this topic on the internet. For example, here’s the always interesting Thomas Basbøll:
Gary Marcus recently called the claim that LaMDA, or any other language model (like GPT-3), is sentient “nonsense on stilts.” Mark Coeckelbergh agreed, but with a twist. It is nonsense, he argued, not because of what we know about artificial intelligence, but because of what we don’t know about sentience. “The inconvenient truth,” he tells us . . . “is that we do not really know [whether LaMDA is sentient]. We do not really know because we do not know what sentience or consciousness is.” As he put it . . . in response to me, “we know how the language model works but we still don’t have a satisfactory definition of consciousness.” This strikes me [Basbøll] as a rather strange philosophy.
I quote this not because I want to engage with the specific arguments that these three people are offering but rather just to give a sense that lots of people are involved in this conversation, including some people I know and other people I’d never heard of. It’s a big conversation. Basbøll follows up here.
Flipping it around
The usual questions are arise when comparing computers to humans. How well can a chatbot mimic (or outperform) human conversation? How much is the chatbot like a human inside? Etc.
I want to ask it the other way: Why are human conversations so different from our conversations with computers, even in simple, computer-like settings?
I was thinking about this the other day when googling something. I don’t remember what it was, but let’s say I wanted to know the population of the United States in 1860. If that’s what I want, I’d google *US population 1860* or, if I’m really on the ball, *US pop 1860*. I wouldn’t google the full question, “What was the population of the United States in 1860?” That would be silly. I mean, sure, it would work, but there’s be no point. It would be like putting on a tux to go to McDonalds.
But here’s my question. Why, when we’re talking to each other, do we not use Google-speak? If I wanted to ask you the population of the United States in 1860″ and I thought you’d know the answer, I actually would just ask, “What was the population of the United States in 1860?” or some more polite version such as, “Hey, whassup, I was just wondering, do you happen to know the population of the United States in 1860?” I wouldn’t give you the telegraphic question, “US pop 1860?” or even its polite cousin, “Hey, whassup, I was just wondering, US pop 1860?” It’s not about politeness; it’s that the longer version works better for communication.
By asking why people don’t speak in telegraphic style, I’m not trying to claim that this is some sort of open question in science. I’m sure there are lots of linguists and information theorists who have worked this out in detail. My point is just that there are subtleties of human conversational interaction that we’re not usually thinking about when we look at a chatbot output or whatever. We talk to each other in full sentences, or some approximation to it, not because we’re demonstrating our humanity but because that’s how our conversations work.
Another way to see the challenges here is to consider some conversational environments that fail. Here I’m thinking of twitter, where people just throw out unsupported claims and there’s not a good place for following up. I guess other examples would be political ads and debates, where it’s all about the soundbite. In any case, one way we can understand what’s going on with conversation is to look at the places where it doesn’t work the way it seems that it should.
P.S. After reading a bunch of commenters explain this to me, I think I have a better understanding here. In particular, I get why I don’t speak to someone by saying, “US pop 1860” or even “US pop 1860?” Here’s the deal. When communicating with Google search, I only ask questions. Google “knows” that a question is coming, so when I type *US pop 1860* it implicitly interprets it as such. But if I just say to someone “US pop 1860,” they don’t know coming into it that I’ll be asking a question. If I say, “US pop 1860?”, then, yes, this signals that I’m asking a question, but what sort of question isn’t so clear: the listener still has to figure out the context. The formulation, “Hey, whassup, I was just wondering, do you happen to know the population of the United States in 1860?”, is helpful because it tells you what’s coming, just like in the movies when we hear that threatening John Williams music we know the shark is coming (or might be).
I find that “normal” conversation becomes different when I am in a dialogue with someone who speaks a different language – my wording comes closer to my conversing with google or a chat bot. I’m not sure the distinction between human and machine is so clear, once we abandon the common language. And, even when we share a common language, conversing with students about statistics is more successful if I am more careful than casual conversation.
On your other point about the debates regarding computer “intelligence:” I find the issue of whether a computer can truly “understand” anything fairly uninteresting. I think some of the people that harp on how computers/AI can’t truly understand are guilty of the NHST trap: they reject the null hypothesis that chatbots exhibit understanding and then conclude that their favorite gripe against AI is correct. I’m perfectly willing to say that AI is incapable of understanding, but still believe that AI will increasingly take over decision making from humans. And I’m not saying that is a good thing, just that I believe it will continue to happen – and we will continue to be fooled by AI.
I think much of it is related to conventions of socio-pragmatics. Like what I talked about in this comment I just wrote
https://statmodeling.stat.columbia.edu/2022/11/17/note-that/#comment-2129058
Sometimes these are structural conventions that can help make communication clearer (as general practices even if it doesn’t always play out that way in specific situations). But sometimes they’re just an artifact of culture.
I think the use of the definite/indefinite article in English is a good example. If you ask people what are the rules for choosing between them, they’d be hard-pressed to give you a comprehensive, clear answer. And that’s in part because the proper use is often not needed for clear communication. But people know “intuitively” the conventions for their use.
That’s funny; about 10 years ago, I started talking to Google the same way I would ask a human a question. I found that it gave me better search results. My questions weren’t quite as easy as “What was the US population in 1960?”. I think in general my old search queries were trying to be too clever, just stringing together abstract keywords, as if it were a structured database query, instead of simply asking the exact question I had in natural language.
On the sentience of language models… my only thought is that I really liked a recent twitter comment (I forgot who said it) that we never should have called them “language models”, they should have been called something like “word sequence models”. The word “language” almost implies some sort of human intelligence, so that if a so-called “language model” produces human-like text, we jump to the conclusion that it’s “AI”. On the other hand, if a “word sequence” model produced human-like text, I think it would just imply that a computer can string together words in a realistic way, and the whole endeavor would seem less hype-driven, and easier to discuss/debate.
Also, lol at “It would be like putting on a tux to go to McDonald’s”.
As I’ve said before here, the “LLMs” are in principle no different from the Markov Chain models from years back. While technically accurate, this is sort of a cheap shot, since presumably they are actually parsing the text in their databases. But it is correct. But…
What crossed my mind recently, was that one of the first things computation linguists noticed when they pointed their natural language parsing programs at real text (e.g. NYT articles), was that real human language is wildly ambiguous; any reasonably long sentence has an enormous number of syntatically correct parses. Only one of which human readers ever notice. Basically, just knowing the part of speech (noun, verb, preposition) isn’t enough. Since the LLM folks claim they are not doing anything that could possibly reduce this ambiguity, it sounds to me that this game is actually much closer to the Markov Chain game than even I had thought. That is the “structures” that they claim make it so powerful are really only extremely short phrases.
Perhaps flip the question. When do we use telegraphic style in verbal communication with other people?
I find it happens quite a bit in high-tempo environments with lots of inter-personal experience. Closely knit teams in operational domains.
Situations where people can complete each other’s thoughts, communities where they reach out to each other – frequently.
Economy of communication often exposes the degree of shared knowledge and established context. Interpersonal is almost internal communication.
So maybe we are treating the computer as a part of our own internal cognition?
Right … I actually think ‘US pop 1860?’ works if asked over a cubicle wall, if it’s relevant information to what the office task is.
I think that would depend. Shouting “US pop 1860?” might be considered impolite in some contexts, and efficient in others. Socio-pragmatics is part of this issue.
In my area, the (incorrect) answer to that question would be Coca Cola (1880).
Not a reply to the main point here, but it seems Meta’s attempt at large language modelling has blown up in its face something fierce.
https://www.technologyreview.com/2022/11/18/1063487/meta-large-language-model-ai-only-survived-three-days-gpt-3-science
The author makes points that I’ve made here about the LLM game. Which is interesting, because it looks as though all his previous articles (Will Douglas Heaven, MIT’s Technology Review’s AI editor) have been completely non-critical of AI and acepting of the hype. It’s as though he suddenly noticed the emperor is completely lacking in threads of any sort whatsoever.
“Meta’s misstep—and its hubris—show once again that Big Tech has a blind spot about the severe limitations of large language models. There is a large body of research that highlights the flaws of this technology, including its tendencies to reproduce prejudice and assert falsehoods as facts.
However, Meta and other companies working on large language models, including Google, have failed to take it seriously.”
I will admit I only read part of the MIT article you cite – but I think I read enough to offer this comment. It is a perfect example of my comparison of AI critiques to NHST. The “failure” of Meta’s large language model is partially attributed to its “tendencies to reproduce prejudice and assert falsehoods as facts.” Hello World! It seems like a large portion of the human population is unable to distinguish between fact and fiction as well as exhibit prejudice. So, it is a failure of Meta’s model to achieve what they were aiming for – but it is also an example where AI and human performance increasingly converge. Put away the misplaced hype about how wonderful AI is – it is (unfortunately, in my mind) destined to be quite successful.
The problems with the argument that some vast subset of the human population are idiots are (a) that it ignores the fact that most of (well, some of) us aren’t idiots, and (b) that even the hate and inane conspiracy theory spewing idiots are actually functional humans who actually are able to think. You may not like what they think (much of the hate and conspiracy stuff is spouted exactly because it irritates us so much; the “owning the libs” game), but they know that a flying rock or baseball headed towards a window is going to result in problems, they have good jobs (and probably make more than you or I) and live in nice houses in the suburbs. The LLMs don’t have anything close to their reasoning abilities.
Also, you should have read more of the article. I agree that repeating the spout prejudice critique isn’t a technical critique, but, e.g. “Like all language models, Galactica is a mindless bot that cannot tell fact from fiction.”
This “cannot tell fact from fiction bit” is, of course, critical. There’s nothing in these programs that has any way of telling whether the stuff it glues together makes sense or not. It has no internal models of anything, just random words and phrases. It doesn’t know about rocks/baseballs and windows.
“A fundamental problem with Galactica is that it is not able to distinguish truth from falsehood, a basic requirement for a language model designed to generate scientific text. People found that it made up fake papers (sometimes attributing them to real authors), and generated wiki articles about the history of bears in space as readily as ones about protein complexes and the speed of light.”
You wrote:
“Put away the misplaced hype about how wonderful AI is – it is (unfortunately, in my mind) destined to be quite successful.”
I disagree. The increadible badness of the current approach (it’s based on an essentially religious belief that doing lots of stupid things really fast and in parallel will cause “intelligence” to magically appear) means it can’t not fail.
I strongly believe that “AI” someday will create programs that can actually reason about the world. But it’s way harder than anyone thinks, and the current round of AI isn’t doing anything that will make any progress.
“People found that it made up fake papers (sometimes attributing them to real authors)”
Yet another example of how parallel the actions of AI are to human actions. How many examples have we seen recently of faked data and false attributions? I agree that AI and human thinking are fundamentally different – but I don’t reach the conclusion you do that AI is destined to fail. It is destined to fail to achieve the hype, but I believe it is destined to increasingly replace human judgement, for better or (often) worse.
The truth is that you are far more optimistic than I am. Yes, people are not idiots, but I think they increasingly are unaware of that. The ability to reason seems to be eroding rapidly – too rapidly in comparison with the development of AI. I don’t see the AI nightmares (e.g., Terminator, iRobot, etc.) as realistic. Instead I see a more subtle yielding of human judgement to algorithms. People may know that a flying rock headed towards a window will cause problems, but do they recognize that buying their lifestyle on credit will likewise cause problems?
I had to get AT&T to fix my phone the other day, and I had to go through a number of chat bots. Of course, they couldn’t help, and I finally got a person who could fix the problem. I imagine that the chat bots can solve the most basic problems, and in the future will be able to do more. So, “success” will be relative. They’re will be a market for these chat bots, and they will find a use making help desks more efficient. But, I agree. Anything that requires problem solving will require more than the current approach.
Blown up in its face? Galactica appears to be a useful tool.
Something like Wikipedia would never have come from a big company. Somebody would write an article about how easy it is to insert false information, and the company would withdraw it.
The wikipedia analogy makes no sense. People were worried that it might be dominated by false information, something which did not come to pass. Galactica actually does produce almost nothing but false information.
What is is useful for? What can you do with it?
You can use Galactica to find relevant papers and ideas, and then investigate them yourself. Just as how some people use Wikipedia.
“You can use Galactica to find relevant papers and ideas, and then investigate them yourself.”
If Galactica were just a search engine, then that would be true, but that’s not what it was; it recombined snippets from search-engine-returned papers in a manner that they hoped would be (and hyped as) interesting. It turns out that the “technology” can’t do that.
(Technology is in quotes here because the “technology” underlying these things is the random number generator.)
(It takes human understanding of the content to combine the content from multiple papers into something sensible. I’m perfectly happy with AI researchers thinking about what “human understanding” means and trying to implement that. That’s not what AI does nowadays.)
To reiterate: It turns out that the technology can’t do that.
It turns out that the technology can’t do that for principled reasons: it has no internal model of the content that it extracts and recombines, so the result of that extract-and-recombine process is exactly and only random garbage. Because the recombine operation is not based on a model of the underlying reality (or math or physics or whatever) that the various extracted pieces correspond to, the result is, in design principle, random garbage. It’s designed to be a random garbage generator. It’s game is to generate random meaningless garbage that, by coincidence, on occassion, appears to make sense.
Since the random meaningless garbage it generates consists of phrases that made sense in the context they were originally written in, it’s sometimes hard to recognize that the output is nothing other than random meaningless garbage, so you can spend some serious time figuring out where and how the logic of the connected snippets fails to connect.
There have been arguments that in the higher math world, the random recombination of logically unrelated random snippets can lead to insights that you might not have noticed. So it can be seen as a Rorsarch test generator. Meaningless images that one can read meaning into.
Seems like a totally inane total waste of time to me. YMMV.
Thanks; I hadn’t seen this META language model for articles, which does seem awful. (And pointless.)
Coincidentally, I learned from a colleague yesterday that META has a new LLM model for protein structure which (according to him) looks great:
https://ai.facebook.com/blog/protein-folding-esmfold-metagenomics/
Today, November 18, 2022 Andrew writes
“But here’s my question. Why, when we’re talking to each other, do we not use Google-speak? If I wanted to ask you the population of the United States in 1860″ and I thought you’d know the answer, I actually would just ask, “What was the population of the United States in 1860?” or some more polite version such as, “Hey, whassup, I was just wondering, do you happen to know the population of the United States in 1860?” I wouldn’t give you the telegraphic question, “US pop 1860?” or even its polite cousin, “Hey, whassup, I was just wondering, US pop 1860?” It’s not about politeness; it’s that the longer version works better for communication.”
Yet, yesterday(!), November 17, 2022
https://statmodeling.stat.columbia.edu/2022/11/17/note-that/
he advocated dropping phrases from written communication that are entirely human such as
“Of course,” “very,” “quite,” “Importantly,” “Interestingly,” and a few more), and delete them—or, I should say, decide by default to delete all of them, but keep them where you decide they are absolutely necessary.”
Hence, the concept of “longer version” and relevance to human interaction is a bit inconsistent. That also goes for all those quotation marks within quotation marks which I am ignoring/violating.
And, before someone else chimes in about “consistency,” and the famous quotation of Emerson, I will do it myself:
“A foolish consistency is the hobgoblin of little minds, adored by little statesmen and philosophers and divines. With consistency a great soul has simply nothing to do.”
Ditto “Consistency is the hallmark of the unimaginative.” ― Oscar Wilde.
The problem with “US pop 1860?” is that it equally encodes pretty much all these questions:
“Do you know the population of the US in 1860?”
“Was there a population in the US in 1860?”
“What was the population of the US in 1860?”
“What was popular in the US in 1860?”
“What popped in the US in the 1860?”
“Do you know what popped in the US in 1860?”
“How many things popped in the US in 1860?”
“How many pops were sold in the US in 1860?”
…
Etc. It only works because context allows us to eliminate most of those from prior knowledge. If the Hindenburg disaster had been in 1860 we’d have a lot more ambiguity.
Some things are irrelevant “You know, like, that’s a really nice dress” conveys no more information than “that’s a really nice dress” and that conveys only slightly more information than “that’s a nice dress”.
But “what was the population of the US in 1860?” conveys a lot more than “US pop 1860?”
Nice!
This is an interesting point but I don’t think it’s the ‘gotcha’ that you think it is.
When you’re writing a paper…well, actually I have no idea what happens when you’re writing a paper. When _I’m_ writing a paper, I sometimes find myself writing something like “…so we end up with y = log(x-z). Note that (x-z) is always positive because of the way we derived z.” Andrew would suggest killing “Note that”, or at least that the default should be to remove that phrase since nothing is gained by including it. But we would still be left with non-telegraphic writing. He isn’t necessarily advocating writing “Thus y = log(x-z) where (x-z) > 0 by def z”, which is the way Lev Landau would have written it….and maybe is the way I should write it too.
Saying you should cut all the garbage is not the same as saying you should cut all the words. It’s not ‘inconsistent’ to say you should cut filler words but keep the good stuff.
As a math major I always appreciated more words than the absolute minimum needed. Too many math books are approaching the modem line-noise level of symbology. It can help a lot to have a few “note thats” and “in addition to”s it’s even better when the author explains the “why” even though the why is actually not part of the formal theory. Or maybe especially because the why is not part of the formal theory.
Even in maths the words can be a problem. A colleague of mine once filled in teaching a complex numbers class for another (sick) colleague and was met with bewildered looks when he started talking about the “roots of unity”. Turns out his colleague had never used the word “unity” for the number 1.
Do mathematicians still use “unity” when talking about the number 1?
Howard:
One of the advantages of teaching in a foreign language that you don’t know very well is you can’t use elegant variation, you have to just go with brute force every time.
I think written mathematics need lots more words instead of symbols, not just for “note that” and not just for “why”. Quantifiers should usually be written out in English (“for every” , “there is”) rather than with the symbols from logic. “So” and “therefore” and “it follows that” are much friendlier than the symbol for implication. All too often students think that “writing math” means replacing words by symbols.
Howard: Mathematicians talk about a ring with unity, but refer to the unity in the ring of complex numbers simply as “1”, except when using the whole phrase “roots of unity”.
Ethan:
I’m with you on that, but I guess that tastes differ. I once was working on a paper with a colleague and he used all those goofy symbols that you and I hate. I was talking with some else about this once, and I expected him to agree with me, but he said that, for him, it was much easier for him to picture what was going on with the symbols than the words. I guess one issue is that words are collections of letters, whereas these symbols are unique, and for some people, the visual images of the symbols make the math more vivid.
Regarding my original point of “Note that,” etc.: I think the analogy to “um,” “uh,” and throat-clearing in speech is a good one. Sometimes it can be helpful to say “um” or “uh,” but usually I think it’s distracting, just something we do when trying to figure out what to say next. That’s why I say that it’s just fine to use phrases such as “Note that” when writing: there will be plenty of time to remove most of these later in the revision stage.
“As a math major I always appreciated more words than the absolute minimum needed.”
Why? Needless words tell people to tune out. They say: “this writing is about me making myself sound important. There’s no content here. Take a nap.” The more useless words are in your writing, the more your writing starts to sound like Charlie Brown’s teacher “Wa wa wa wa, wa wa WA WA wa”.
Because the minimum number of words in math would be like posting this
http://content.lessonplanet.com/resources/thumbnails/322081/large/bwluav9tywdpy2symde0mduxmi0xnjg5ni0xmgxkmg0uanbn.jpg?1399890687
as an entire set theory textbook since it contains the entirety of ZFC.
Daniel, as another math major, I concur. Using more words helps students, in lectures and written format. Sheldon Ross’s probability books were decent about that.
Ethan says using more words is friendlier. Makes sense: Mostly formulas and thickets of logic symbols can intimidate students. It might even seem pretentious.
For practitioners in industry, who know the concepts, there ARE use cases for what Andrew describes as ‘all those goofy symbols’. Sometimes I prototype math, statistics, and probability solutions for business or regulatory purposes. Then I explain (methodology, input data, and expected output) to a lead software developer, to be recoded and run in production. Developers were often terse men whose native languages were Chinese or Russian. Showing them my SAS code while writing the formulas it captured, using Greek letter math notation and logic symbols, was best. They did NOT want to listen to me explain in English sentences, nor did I!
Ellie, the pure logic symbols can be useful absolutely, if you’re writing code, it can be much easier to translate math symbols to code than a bunch of words. I also think it’d be better if we had a pseudocode that was widely accepted in math. In my opinion computability and code are as fundamental as set theory. Math notation can be shockingly ambiguous once you try to convert it to something executable on a computer.
For example, is dy/dx a function of x, or some particular value of the derivative (evaluated at a point) or an operator? What is an operator anyway? The correct definition would be something like a function that takes a function and returns a new function.
“conversational repair” is a good starting point. imagine multi-level ECC added to your words, that’s why we have expected sentence structures and redundancy and more than two phonemes.
We don’t use “normal speech” with Google because typing is harder and slower than speaking. And we don’t use Googlespeak with people because we’re trained from day one to use complete sentences.
We don’t use normal speech patterns with google because typing is harder and slower than speaking. We don’t use “googlespeak” with people because we’re trained relentlessly for our first 20-25 years use complete sentences when speaking.
That’s a good point: it’s partially cultural. Part of why we don’t usually speak in short choppy sentences is because there is a cultural preference for complete sentences and some polite waffling. We associate shorter, choppier sentences with not having a good command of language or rudeness. We associate longer, more elaborate sentences with being smart, polite, well educated, high class, etc. (Then we judge you based on what you said and how you said it.)
(Also: shorter, more abrupt sentences are used to convey either urgency or that the person you’re talking to isn’t worthy of the polite waffling we typically use. e.g. barking orders at a servant.)
On the other hand, context matters. If I’m making breakfast, I might turn to my partner and say “Bagels?” and they will generally be able to work out what I mean from the context (where are the bagels, do we have bagels, do you want bagels, I am offering you bagels, etc.). Or I might say the same to a customer service person at a buffet (using body language or tone of voice to convey the polite waffling of “excuse me, sorry to bother you, do you happen to know…”). I wouldn’t just yell “BAGELS???” at someone, though – it’s not enough information for a query and the delivery will likely result in a negative interaction.
I’d also like to throw in that google has no emotions/no social status. I can type “US Pop 1860” and if it confuses google, I can just type in “United States population 1860” or get more specific as needed, repeatedly, without worrying that my conversational partner will be annoyed or upset by my behavior. I would not act like this if I was talking to a research librarian, or at least, I would be including words/tone/body language that communicates “I respect you and your time, thank you for helping me with this weird query” in a way that google does not need.
Such a thought provoking question!
I think our terse Google queries reflect how poorly we think of Google’s ability to answer interesting questions.
“US population 1860?” only has one answer when posed to a dummy.
We expect humans to be capable of much more interesting responses so we need more words to narrow how we want our question interpreted. Otherwise, it is possible you would you like to know about the demographic composition of the population? What about how the population was changing—perhaps by region? Or how about who was counted as part of the population and who was excluded?
I don’t think that these types of questions are especially common in everyday human conversation but our belief in their potential (compared to talking with a Google) is what demands the extra words.
I just tried googling *US pop 1860*
Google answered:
31,443,321
POP culture: 1860
If it was intelligent, I could ask it what the heck it means.
Hi Andrew and everyone:
I really enjoyed this post (which I can honestly say about nearly all of them). I think what they meant by “consciousness” goes something like this. With these huge AI models
(called Foundation models), there is an enormous number of parameters to fit. And they literally spend millions of dollars just fitting them (and not over-fitting). So “consciousness” is what they get when the algorithm does something smart that it wasn’t trained on. For example, AlphaGo made moves that human grandmasters had never done in the training set. This is what they are seeing from the Foundation models. Anyways, the word “consciousness” is not quite right. These actions are something like a Turing test. Do we need to coin a new phrase? D.E.M. for Deus ex machina?
Andrew says:
“When communicating with Google search, I only ask questions.”
You’re still missing the difference.
Whether you’re asking a question or not isn’t the the issue. If you said “US Pop 1860”, even without the question mark, most people would figure out that you mean to ask the question “what was the US population in 1860?”. What would be strange about the question is why the hell you would expect any human to be able to answer it. Only very small number of people interested in population trends could answer that question even roughly without looking it up, and even most children would intuitively understand that. Like Kaz pointed out above, you can request information in short form when there is a reason for you to expect someone to know the answer. You can say to your partner “Bagels” when you’re really asking “Are the bagels ready?”. OTOH, if you’re getting breakfast ready and you say to your partner: “Mean Jan Temp Berlin”, they understand your question. What they don’t understand is why you would expect them to know the answer.
You’re mistaken when you say google is “answering” your question or that you’re “asking” it a question. That’s not what’s happening. What’s happening is that you’re issuing a request for information and google does the only thing it “knows” how to do, which is find the best match to your request. It’s sophisticated enough that you can phrase your request in many different ways and it still fulfills your request, but regardless of how you think your phrasing it, Google “understands” you to say “find the population of the US in the year 1860”. It’s not really a question. It’s an order, and Google fulfills it with the best match it can find.
That’s all google does. It just looks up matches.
Want to ask google a question? Try this one:
“Is cherry pie better than steamed oysters?”
Google does a poor job of “answering” this question. That’s because it doesn’t “answer” questions, it just finds matches to the words in your request for information.
Here’s another way to think of this:
If you say to a human: “Hey bro, hey do you happen to know what the population of the US was in 1860?”, what you’re really doing is rephrasing a demand for information as a question to be polite. In effect what your doing is issuing the command that Google actually responds to, regardless of how you phrase it: “Please provide the population of the US in the year 1860”
My concern (expressed here, for example) is that mimicking human conversation is very different from making reliable recommendations about buying stocks, approving loans, hiring people, setting prison sentences, and so on.
As for human conversation, I think that we often communicate in “efficient” ways, leaving out words because of our understanding of context; for example, “Dakota likes pineapple pizza” instead of, “A person named Dakota likes eating pizza with a pineapple topping;” or “Brooklyn likes pizza with his kids” instead of “A person named Brooklyn has children and likes eating pizza with them.”
Gary,
Sure, but it’s not like anyone is saying you could type a text into a language model it would “generate a paper with relevant references, formulas, and everything” . . . ummmm, whoops!
Link was lost: https://medium.com/@garynsmith47/turing-tests-are-terribly-misleading-cc1165c5b09e