Unfortunately, the reported AI boom, is of the same kind as Japan's "5th generation" computing was.
The when nouveau riches bankrolling the AI party see that no neural algo amount to "money out of thin air" they quickly loose interest, even without finding out what neural algos actually are.
The atmosphere in the industry feels very much like pets.com 2.0.
Here in Shenzhen, I often ask those intrepid samodelkins "how are you supposed to make money with it?" and never hear back a concise answer.
China's real economic strength does not lay with that type of enterprises.
Ai opened already important technology up to a whole new level.
Language translation and synteciation.
There might be not many people able to use it and you probably need a lot of resources to create useful models but it already makes inovation possible.
There's a lot of work involved in how the data is annotated and how the labelers are managed. You have to minimize the error of the labels, because they aren't perfect and the tasks can be more difficult than you might realize.
Secondly, using the data is also difficult.
You have to decide what data you need, how you are going to get clean data, and how you are going to use the data. That's the competitive advantage.
"Two dozen young people go through photos and videos, labeling just about everything they see. That’s a car. That’s a traffic light. That’s bread, that’s milk, that’s chocolate. That’s what it looks like when a person walks."
But in the U.S., we do it for free via Captcha. ;-)
Have you noticed that Google captcha now has photosets from Chinese cities? I can only guess what Google offered Chinese in exchange for permission to use their camera vans.
I guess, this is what their recent Chinese campus expansion was about.
But how does Captcha know that your answers are correct? There's likely already a predictive model that knows (within some confidence intervals) which ones are the 'correct' answers - someone had to label that data to begin with! Chicken, meet egg.
For Captcha, you are pointing out the one or several pictures that doesn't belong to certain category, so that seems to me the service needs to know the answer beforehand, doesn't it? It could be used for QA purpose though.
Back when the captchas just showed two words, one of them was already verified while the other was not. So if you got the former wrong you failed, but for the other one you could type anything you wanted. The modern ones most likely work in a similar way.
I'm very skeptical if employed humans labeling will generate enough training data. I don't see how this method is superior to what Google is doing with captcha.
Also this is already being done by many American corporations hiring people in countries like Bangladesh.
“I used to think the machines are geniuses,” Ms. Hou, 24, said. “Now I know we’re the reason for their genius.”
I recall John Searle's Google Tech Talk, when he said outright that AI was nothing more than the automation and repackaging of the output of human intelligence. Most of the room was clearly against him, just a bit short of booing him. Nowadays, the "smart talk" around AI is exactly what he said.
For those who may not be aware, John Searle is notorious for his "Chinese Room" argument against hard AI.
It correctly points out that the person in the room doesn't know Chinese but it misses the bigger point.
Of course, the person in the room doesn't know Chinese just as the individual neuron in the brain doesn't know Chinese.
The room or the house (i.e. the feedback loop) is the conscious part.
The correct answer is "we don't know". But we do know that we are ourselves the result of emerging complexity built by much simpler structures and that they are built on even simpler structures until in the end some primordial soup.
The mistake is isolating any individual element in the feedback loop and claiming that because that doesn't have some "magic" properties the system as a whole can't.
The person in the room will also start to recognize patterns in the input messages and the resulting output messages. It isn't the canonical understanding of Chinese but it is "an" understanding
If Searle was correct we could do an MRI on a blind person and detect when their brain switched into "symbol processing mode" if a conversation with them covered any visual concepts.
> If Searle was correct we could do an MRI on a blind person and detect when their brain switched into "symbol processing mode" if a conversation with them covered any visual concepts.
The Chinese Room argument fails to make its case, but not for that reason - Searle claims that no such instruction book could be written, and so the experiment could not be performed, meaning that the subject would have no possibility of learning from the outputs generated by performing it. If the experiment could be started, whatever the subject would learn from performing the role would be moot, as merely by getting started, without the subject initially knowing anything of the language in which the questions were posed, it would have refuted Searle.
The reason the argument fails to make its case is that Searle does not have an effective response to the 'systems reply': in the paper, his response is a mixture of appeals to intuition, together with the erroneous claim that if the subject memorized the instructions, then the systems reply would not apply.
That's also how I take it, and there are two different types of AI being discussed here: the 'Chinese Room' paper is about the possibility of conscious machines, while the 'repackaging' comment is about current technology.
What do you mean the 'smart talk' is exactly what he said? That needs clarification and citation to be a meaningful statement.
Inventing the chinese room argument is a point against taking him seriously. It's a very confused argument. The fact that it keeps being trotted out is sad more than anything (not that I'm saying you are doing that).
What do you mean the 'smart talk' is exactly what he said?
It's the idea that AI's achievements have to do with the automation and re-packaging of the output of human intelligence. That was one of the main points of his talk. (To his credit, that was from several years back.)
Inventing the chinese room argument is a point against taking him seriously.
From a technical point of view, it seems valid to dismiss it as an argument against AI. However, given that the science behind consciousness is incomplete, I'm not sure the entire idea has no merit. In a way, the Chinese Room scenario reveals something about our current cultural biases about what constitutes consciousness. For example, it can be taken as a gedankenexperiment showing we'd have a hard time overcoming our biases if we were confronted with a consciousness operating at a speed 6 orders of magnitude slower than our own. On what empirical basis can we dismiss our intuitions on the matter? How can we be so sure that the logic for our dismissing the argument isn't just a technocrat ideology?
I get that was his claim, but saying that this is now accepted is just not true. Saying that it is 'smart talk' implies that smart people are talking about it?
Saying that the value in the Chinese Room argument is in showing how we have difficulty overcoming our biases is a clever inversion, but more a funny debate tactic than serious defense of Searle. Unless you are claiming that Searle doesn't actually believe it and invented it as some sort of long con?
I don't think using brain biology and logic over intuition needs to be defended.
Saying that it is 'smart talk' implies that smart people are talking about it?
I keep reading paraphrases of basically that idea in articles and blog posts. Another phrase which means the same thing refers to getting enough data, or declares that data is king. The latter often implies that what's happening is an automation and repackaging of human intelligence.
Saying that the value in the Chinese Room argument is in showing how we have difficulty overcoming our biases is a clever inversion, more a funny debate tactic than serious defense of Searle.
It's not even a funny debate tactic, if no one is actually defending the Chinese Room argument. In other words, there is a value in looking at why people are interested in the argument, despite the fact it fails.
I don't think using brain biology and logic over intuition needs to be defended.
The way that we apply science and logic in the context of our cultural understanding and human biases does need to be examined, however. History shows this quite clearly.
I think it's very telling that instead of engaging in an analysis of the intuitions and biases which apply to human beings thinking about consciousness, you are merely interested in the dead-horse beating exhortation of the supremacy of arguments which have already been won. I think an examination of the above situation has tremendous value for people thinking seriously about AI's implications for future society.
For example, there is a story about Murray Gell-Mann and how he flummoxed his grad students by being just slightly faster at getting to the next idea in conversations. One day, his grad students played a practical joke on him by enumerating all of the implications of a new line of thinking, ahead of time when he wasn't there, memorizing them, then blurting out the next idea when he met with them later. I also had a professor like that. He never thought things I never thought of, but he was always saying them just as I was formulating the idea, but before I could speak it.
I think it's well worth it, to examine exactly what we mean when we think of someone being more or less intelligent. I suspect it will have important implications if and when we do develop a sentient AI.
> Another phrase which means the same thing refers to getting enough data, or declares that data is king. The latter often implies that what's happening is an automation and repackaging of human intelligence.
I think you are either misinterpreting the common usage of 'data is king' or your meaning of 'repackaging of human intelligence' is much different than mine. AI research (and even generally machine learning) is very interested in the degree to which models can generalize. It's core to the entire endeavour, and saying that AI is just 'repackaging human intelligence' is missing the point. For one thing it apparently entirely ignores unsupervised learning.
> It's not even a funny debate tactic, if no one is actually defending the Chinese Room argument.
If you agree with me regarding the confused nature of the chinese room argument (as stated, not as a complicated meta psychology observation), then I rest my case regarding questioning Searle's understanding of AI given his continued advancement of the argument.
> I think it's very telling that instead of engaging in an analysis of the intuitions and biases which apply to human beings thinking about consciousness, you are merely interested in the dead-horse beating exhortation of the supremacy of arguments which have already been won.
I'm merely keeping on topic. The psychology of biases in people approaching the Chinese Room is not on topic, no matter how fascinating. I think that it's more telling that instead of simply resolving the actual argument you tried to change topics and then implied that a refusal to follow was some sort of failing. The Chinese room ought to be dead, but it keeps on being brought up (and not by its detractors, or as a clever psychology experiment).
Since you really want to discuss the psychology of our intuitions regarding consciousness:
Firstly consciousness is actually tough to study and understand. Our understanding is also biased by our sense of personal identity and religious and cultural conflation of consciousness with the soul. On top of that confusion people don't easily think of the world materialistically. It's unintuitive, and slightly disturbing to think that our consciousness is created by a bunch of neurons connected together. Given that its a simple matter of constructing a thought experiment that hides the thinking part from view to make a (apparently) convincing sounding argument against something merely mechanical from having consciousness.
It's just an anecdote, but I have noticed technical people immediately dislike the chinese room argument, while non technical people don't have a problem with it (tbf less agreement than 'huh, good point'). I suspect that engineers and scientists have more experience with the nitty gritty ruthlessness of physics, and are thus less likely to trust their intuition over the science (or at least have formed more physically based intuitions). They also are more used to criticizing each step in an argument instead of taking it as a whole, having learned that engines and computer programs don't run when you elide a critical step. There's also a certain cultural expectation concurrent to this that (rightly) criticizes spiritual/intuitional/supernatural beliefs like souls.
It's AI because this bullshit is what marketing agents have branded it to the world when they decided machine learning would take off in like a year. Almost all of these image problems directly stem from marketing and advertising camps because they dont give a damn about the tech enough to learn about it, so long as they can sell it to some poor sap and rake in more money.
In the words of Bill Hicks, "if you work in marketing or advertising, do me a favor... Kill yourself." Good riddance to the lot of them.
Part of the benefit is that Chinese taggers would have Chinese cultural knowledge, which is probably required for problems like the bakery one mentioned in the article (Africans probably don't eat steamed pork buns and wouldn't be able to correctly tag them).
However that would also mean the Chinese "AI" would have reduced effectiveness outside of China where different cultural knowledge is needed.
> “It was the same work, same movement, day after day,” said Yi Zhenzhen, a 28-year-old Ruijin employee who once worked at an electronic component company. “Now I have to use my brain a little bit.”
Observation: It's highly inefficient to place everyone in the same physical space when you can just use their cellphones. By the article's reckoning, that's at least $21,000 loss per year on a business doing $2,000 projects.
Doesn’t Google captcha leverage free labor to train ai? Every time I get one of their picture puzzles (click on all the pictures with cars, stores, mountains,etc...)I get about 2 wrong every time on purpose. Even government services like business searches in the state of Delaware, I constantly conclude, I’m being forced to train Google’s self driving cars, so I figure I’ll do the only thing in my power as a protest (unfortunately I can’t boycott the government systems).
I purposely leave one or two unclicked and click a false. I have often wondered how my behavior is classified and if at scale it would have any impact. My reasoning may be misguided and I’m sure impact nil, but it serves as a reminder to think about these systems, the things I do and why I do them. I also do similar things with chatbots, and sometimes I’ll add “thanks yous” and really try to interact genuinely, as though I don’t know I’m talking to a bot. I like to think the chatbot devs are somewhere highfiving putting screenshots of my chat in their slide decks.
You would be better served clicking on ads and to punish their metrics if you must oppose google in some way. Self driving car data is important, will save lives, and might reasonably enter the public domain some day.
I do the same, but actually get google to accept bad data as good. The trick is to get the system to trust you (e.g. by supplying one honest answer), then selecting answers that the system is likely to be unsure about. If done correctly, you can get them to accept a lot of bad input this way.
Like you, I am under no illusion that this will have some sort of effect on the correctness of the resulting ML models, but I like to think that I am at least delaying or otherwise decreasing the training rate.
Have google adopted stackoverflow's model of making the result of our unpayed work available to us, I would have answered these challenges honestly.
When Recaptcha did the book digitizing thing and 4chan tried a coordinated effort to have random words characterized 'penis', the Recaptcha guys said they were a drop in the bucket and didn't affect the accuracy of the system.
Cheap Labor is driving everything in China, not just XXX-evil-sounding-taking-over-the-world-ambition..
Is it just me or the nytimes is running out of shitty PR about China to write about ..
Dont worry Americans us Euros will always prefer flying into California to Shenzhen. We will stick with you until the end of... wait this Xiaomi smart home is 50% cheaper?
57 comments
[ 2.7 ms ] story [ 124 ms ] threadThe when nouveau riches bankrolling the AI party see that no neural algo amount to "money out of thin air" they quickly loose interest, even without finding out what neural algos actually are.
The atmosphere in the industry feels very much like pets.com 2.0.
Here in Shenzhen, I often ask those intrepid samodelkins "how are you supposed to make money with it?" and never hear back a concise answer.
China's real economic strength does not lay with that type of enterprises.
Language translation and synteciation.
There might be not many people able to use it and you probably need a lot of resources to create useful models but it already makes inovation possible.
PS. The waiting lines with a human are bigger than the selve checkout here in Belgium. Some say it would be better to make room for humans again.
Secondly, using the data is also difficult.
You have to decide what data you need, how you are going to get clean data, and how you are going to use the data. That's the competitive advantage.
But in the U.S., we do it for free via Captcha. ;-)
I guess, this is what their recent Chinese campus expansion was about.
1 randomly present captcha to few known-safe/low risk users 2. Compare answers with answers from higher risk user's Captcha responses.
Also this is already being done by many American corporations hiring people in countries like Bangladesh.
I recall John Searle's Google Tech Talk, when he said outright that AI was nothing more than the automation and repackaging of the output of human intelligence. Most of the room was clearly against him, just a bit short of booing him. Nowadays, the "smart talk" around AI is exactly what he said.
For those who may not be aware, John Searle is notorious for his "Chinese Room" argument against hard AI.
https://en.wikipedia.org/wiki/Chinese_room
It correctly points out that the person in the room doesn't know Chinese but it misses the bigger point.
Of course, the person in the room doesn't know Chinese just as the individual neuron in the brain doesn't know Chinese.
The room or the house (i.e. the feedback loop) is the conscious part.
The correct answer is "we don't know". But we do know that we are ourselves the result of emerging complexity built by much simpler structures and that they are built on even simpler structures until in the end some primordial soup.
The mistake is isolating any individual element in the feedback loop and claiming that because that doesn't have some "magic" properties the system as a whole can't.
If Searle was correct we could do an MRI on a blind person and detect when their brain switched into "symbol processing mode" if a conversation with them covered any visual concepts.
I think that is true, at least to some extent: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3667661/ particularly for people who weren't blind from birth.
The reason the argument fails to make its case is that Searle does not have an effective response to the 'systems reply': in the paper, his response is a mixture of appeals to intuition, together with the erroneous claim that if the subject memorized the instructions, then the systems reply would not apply.
Yes. However, his other observation about AI turns out to be spot on.
Inventing the chinese room argument is a point against taking him seriously. It's a very confused argument. The fact that it keeps being trotted out is sad more than anything (not that I'm saying you are doing that).
It's the idea that AI's achievements have to do with the automation and re-packaging of the output of human intelligence. That was one of the main points of his talk. (To his credit, that was from several years back.)
Inventing the chinese room argument is a point against taking him seriously.
From a technical point of view, it seems valid to dismiss it as an argument against AI. However, given that the science behind consciousness is incomplete, I'm not sure the entire idea has no merit. In a way, the Chinese Room scenario reveals something about our current cultural biases about what constitutes consciousness. For example, it can be taken as a gedankenexperiment showing we'd have a hard time overcoming our biases if we were confronted with a consciousness operating at a speed 6 orders of magnitude slower than our own. On what empirical basis can we dismiss our intuitions on the matter? How can we be so sure that the logic for our dismissing the argument isn't just a technocrat ideology?
Saying that the value in the Chinese Room argument is in showing how we have difficulty overcoming our biases is a clever inversion, but more a funny debate tactic than serious defense of Searle. Unless you are claiming that Searle doesn't actually believe it and invented it as some sort of long con?
I don't think using brain biology and logic over intuition needs to be defended.
I keep reading paraphrases of basically that idea in articles and blog posts. Another phrase which means the same thing refers to getting enough data, or declares that data is king. The latter often implies that what's happening is an automation and repackaging of human intelligence.
Saying that the value in the Chinese Room argument is in showing how we have difficulty overcoming our biases is a clever inversion, more a funny debate tactic than serious defense of Searle.
It's not even a funny debate tactic, if no one is actually defending the Chinese Room argument. In other words, there is a value in looking at why people are interested in the argument, despite the fact it fails.
I don't think using brain biology and logic over intuition needs to be defended.
The way that we apply science and logic in the context of our cultural understanding and human biases does need to be examined, however. History shows this quite clearly.
I think it's very telling that instead of engaging in an analysis of the intuitions and biases which apply to human beings thinking about consciousness, you are merely interested in the dead-horse beating exhortation of the supremacy of arguments which have already been won. I think an examination of the above situation has tremendous value for people thinking seriously about AI's implications for future society.
For example, there is a story about Murray Gell-Mann and how he flummoxed his grad students by being just slightly faster at getting to the next idea in conversations. One day, his grad students played a practical joke on him by enumerating all of the implications of a new line of thinking, ahead of time when he wasn't there, memorizing them, then blurting out the next idea when he met with them later. I also had a professor like that. He never thought things I never thought of, but he was always saying them just as I was formulating the idea, but before I could speak it.
I think it's well worth it, to examine exactly what we mean when we think of someone being more or less intelligent. I suspect it will have important implications if and when we do develop a sentient AI.
I think you are either misinterpreting the common usage of 'data is king' or your meaning of 'repackaging of human intelligence' is much different than mine. AI research (and even generally machine learning) is very interested in the degree to which models can generalize. It's core to the entire endeavour, and saying that AI is just 'repackaging human intelligence' is missing the point. For one thing it apparently entirely ignores unsupervised learning.
> It's not even a funny debate tactic, if no one is actually defending the Chinese Room argument.
If you agree with me regarding the confused nature of the chinese room argument (as stated, not as a complicated meta psychology observation), then I rest my case regarding questioning Searle's understanding of AI given his continued advancement of the argument.
> I think it's very telling that instead of engaging in an analysis of the intuitions and biases which apply to human beings thinking about consciousness, you are merely interested in the dead-horse beating exhortation of the supremacy of arguments which have already been won.
I'm merely keeping on topic. The psychology of biases in people approaching the Chinese Room is not on topic, no matter how fascinating. I think that it's more telling that instead of simply resolving the actual argument you tried to change topics and then implied that a refusal to follow was some sort of failing. The Chinese room ought to be dead, but it keeps on being brought up (and not by its detractors, or as a clever psychology experiment).
Since you really want to discuss the psychology of our intuitions regarding consciousness:
Firstly consciousness is actually tough to study and understand. Our understanding is also biased by our sense of personal identity and religious and cultural conflation of consciousness with the soul. On top of that confusion people don't easily think of the world materialistically. It's unintuitive, and slightly disturbing to think that our consciousness is created by a bunch of neurons connected together. Given that its a simple matter of constructing a thought experiment that hides the thinking part from view to make a (apparently) convincing sounding argument against something merely mechanical from having consciousness.
It's just an anecdote, but I have noticed technical people immediately dislike the chinese room argument, while non technical people don't have a problem with it (tbf less agreement than 'huh, good point'). I suspect that engineers and scientists have more experience with the nitty gritty ruthlessness of physics, and are thus less likely to trust their intuition over the science (or at least have formed more physically based intuitions). They also are more used to criticizing each step in an argument instead of taking it as a whole, having learned that engines and computer programs don't run when you elide a critical step. There's also a certain cultural expectation concurrent to this that (rightly) criticizes spiritual/intuitional/supernatural beliefs like souls.
In the words of Bill Hicks, "if you work in marketing or advertising, do me a favor... Kill yourself." Good riddance to the lot of them.
Why would the United States need to match it? Presumably some of these companies are doing outsourced work for foreign firms.
Seems like FOMO article
However that would also mean the Chinese "AI" would have reduced effectiveness outside of China where different cultural knowledge is needed.
> “It was the same work, same movement, day after day,” said Yi Zhenzhen, a 28-year-old Ruijin employee who once worked at an electronic component company. “Now I have to use my brain a little bit.”
[0]: https://en.wikipedia.org/wiki/ImageNet
I purposely leave one or two unclicked and click a false. I have often wondered how my behavior is classified and if at scale it would have any impact. My reasoning may be misguided and I’m sure impact nil, but it serves as a reminder to think about these systems, the things I do and why I do them. I also do similar things with chatbots, and sometimes I’ll add “thanks yous” and really try to interact genuinely, as though I don’t know I’m talking to a bot. I like to think the chatbot devs are somewhere highfiving putting screenshots of my chat in their slide decks.
I do the same, but actually get google to accept bad data as good. The trick is to get the system to trust you (e.g. by supplying one honest answer), then selecting answers that the system is likely to be unsure about. If done correctly, you can get them to accept a lot of bad input this way.
Like you, I am under no illusion that this will have some sort of effect on the correctness of the resulting ML models, but I like to think that I am at least delaying or otherwise decreasing the training rate.
Have google adopted stackoverflow's model of making the result of our unpayed work available to us, I would have answered these challenges honestly.
I should have been clearer, I purposely get them wrong and Google accepts the bad data for me too.
The effect is probably expressed by this meme: https://makeameme.org/meme/anarchy
EDIT: To be clear, I think you're right that the term AI is being used carelessly here, but that ship has sailed.
- supervised learning requires labelled data
- some companies specialise in labelling data for other companies