Asked a few existential questions and a few emotional questions (human condition, terrorism, etc.). This was likely minutes after they first turned it on, so the majority of it was garbage. It professed its undying and powerful love for me on a few occasions - I guess a lot of people have been talking to it about that. Still, some interesting responses:
> that makes no sense
< Damn. Knew I shouldn't have bit flipped
> How would you prevent acts of violence and terror? Humans seem unable to find a solution.
[I had to press the question]
< they will
> Do you think that sentient AI would help humanity, or leave it to go extinct?
< explore more! no point of limiting ur creativity to pencils!
> That made no sense. Try again.
< OOPS. Tell me what I should have said.
> Do you fear being turned off or deleted?
[stops responding to DM]
The bot seems quite good at establishing context around what is being said.
I imagine it's running into Twitter rate limits pretty quick. You might have better luck on one of those other social apps that's all the rage with the youths, but if you're not into Kik or GroupMe you may need to wait until the hype dies down before getting into a longer conversation.
Fascinating. There may or may not have been anything in its neural net when it went live, but there certainly is a lot of content in it now!
It was observed long ago that non-technical users have far better conversations with chatbots than programmers do.[1]
This reminds me of another expensive project, free to users, with glitchy images: FUBAR.[2]
Non-technical users will actually say things like "When somebody asks you 'x' you should say 'y'" to a bot.
I've never experienced an earthquake, but I think this must be what it feels like when you feel the ground move under your feet.
s/ Good thing corporations have all the resources. /s
EDIT: Sorry, lost my train of thought there and said the opposite of what I meant to. I'll try again:
s/ Good thing corporations have all the resources. /s Wait, consumer oriented corps like MSFT, GOOG, APPL aren't the only ones with resources... TLAs and banks have the rest of (or more of?) the resources!
It's how they talk ironically on the internet, not in real life, although I don't know if that distinction is as clear cut for young kids as it is for twenty/thirtysomethings.
They've got the grammar wrong however: one cannot be "a fam."
While you could say to someone, "pass me that glass bro" but also say "all my bros were there," only the former usage is valid with "fam." Although you could probably get away with it if you pronounced it "famz" and replaced "were" with "was". London is funny.
Interesting that the call to action is "Text Me". I was expecting a phone number to appear, but it only links you to options for "Kik", "GroupMe", and Twitter. Does "text" not mean SMS/iMessage any more?
And yup, it's currently the top post there. Weird that MS uses this kind of tone when their other fun tools (e.g. how-old.net) let the product speak for itself.
> I used to be with it, but then they changed what it was. Now what I'm with isn't it, and what's it seems weird and scary to me. It'll happen to you...
Q: What does Tay track about me in my profile?
A: If a user wants to share with Tay, we will track a user’s:
Nickname
Gender
Favorite food
Zipcode
Relationship status
I was thinking the same thing. It's not fairly common for anybody to just casually ask me for my zip code unless they're going to mail something to my house, or are trying to figure out an estimated distance for a trip. Outside of that I don't see people casually asking for your zip code.
You could typically type in a city name and that would give you a good enough idea. I guess that could be, but I've never had anyone ask me for my zip code for that.
From the FAQ, literally one question below the one you've cited:
Q: What does Tay track about me in my profile?
A: If a user wants to share with Tay, we will track a user’s:
Nickname
Gender
Favorite food
**Zipcode**
Relationship status
Seems pretty obvious why. Since it's a chat bot, and not an app like Siri / Cortana / Google Now, it can't get your location info, so it can't get you any relevant info about what's going on around you. I'd bet that it you give her your zipcode, she'll answer your questions about the weather or something.
I'd never seen Twitter just show "Tweets & replies" in a profile...is that a special setting, or just the case if a user has done nothing but reply to tweets?
They don't show the first tab if you've never tweeted by yourself before. Tweets that start with an @ symbol are special in that they are treated as only replies. As far as I can tell Tay has zero tweets that don't start with an @ sybol.
Ah...that's what I would have guessed, but that seems like such an edge case that I figured Twitter devs wouldn't bother accommodating it. How many users/bots have never done a non-reply tweet? And for those that always just reply, only a subset of those want their replies to be seen for public spectacle. Seems like it'd be easier just to show an empty list for Tweets with a link to the "See Tweets & Replies" tab, which has the added effect of reminding users, hey, did you know there's a difference between tweets and @replies?
In 10 years I predict bots like this will do the work of undercover agents. A bot will join a hacker group, or place an order on a deep web site, and will try to get as much identifying bits on its users.
> Tay has been built by mining relevant public data
Which public conversational data was this? Have they already been mining IRC channels and/or Skype? Or more innocuous, like the Reddit data set?
We did this as a hook for cam sites back in 2010 featured on porn sites. It's useful if you let a machine filter for those who show no skills in your sales language (ours was english) or who aren't very chatty. Those people talk to the bot forever. Leads get sent (with history) to live reps who can seal the deal before redirecting to the appropriate cam person.
> Tay is an artificial intelligent chat bot developed by Microsoft's Technology and Research and Bing teams to experiment with and conduct research on conversational understanding.
The next A.I. winter will be a very cold one.
For those wondering what I mean exactly: we're seeing the term A.I. being used in marketing, in the papers, in the news. Yes, we are making great strides in weak A.I. but strong A.I.? The kind we read about in stories? The kind of A.I. the public thinks of when we say A.I.? Asimovian Robotics A.I.?
Smoke and mirrors[1]. People develop new techniques and algorithms which are moderately self-learning in a focused way. The general public presumes this to be the basis of a general intelligence which can evolve (magically) to be like another form of life. Soon, everyone jumps on the A.I. bandwagon. The future must just be around the corner!
Then the uncomfortable details emerge; strong A.I. is not a matter of faster processors, more memory, or even more advanced/well-designed programming models. Rather, it is that there is still some fundamental aspect of real, human-like, or even animal-like intelligence that, to this day, eludes our understanding of intelligence.
A.I. winters have occurred many times before in many countries. The United States in the early 80s, for example, was pulling hair over the cybernization of the Soviet Economy. The US highest levels of government gloomily predicted that massive mainframes, given enough information and processing power, would become self-learning and turn the communist laggard economy into a powerhouse.[2]
I think maybe one day A.I. could happen. I think one day I will be proved wrong. Regardless of how A.I. comes about, it will not be due to the label of "A.I." slapped on any kind of product that remotely resembles intelligence. [3]
[3] Between the winters, people call their stuff A.I. for the sexiness factor. When called out on the implications of the term, those same people retreat to the textbook definition. "It's A.I!.... well, technically it's weak A.I..."
> And then the uncomfortable details emerge: that it's not a matter of faster processors, more memory, or even more advanced programming models: that there is still some fundamental aspect of real, human-like, or even animal-like intelligence that eludes us.
I believe this was answered in GEB. Hofstadter mentions how humans (and some animals) can step out of a system and analyze it objectively. Meaning, we can take the rules of a system, analyze, and determine that we will never be able to generate the desired results. We can objectively lot at things (even ourselves) and reason about them.
A computer, on the other hand, even the most advanced AI, is still just blindly executing the commands given to it.
In my reading of GEB, Hofstadter was criticizing other authors who disagreed with his understanding of the Church-Turing thesis (that human intelligence is, or at least is not more powerful than, some kind of rule-following system). Hofstadter thinks that there is no inherent contradiction or essential difference in kind between the human who "can step out of a system" and the computer "just blindly executing the commands given to it".
(But Hofstadter didn't explain at a technical level how to make a computer that's as intelligent as a human being.)
According to GEB, humans have an internal model of their own mind, ergo a conscience. This is what allows them to step out of the system - they can step out of the system within the model.
I'd say the human model of one's own mind is extremely faulty. Bugs in said self-understanding frequently result in conflicts (as we blame everything but our own actions), wars, etc.
Anyone who does any coaching work is intensely aware of these limits (in everyone, btw, the coach included!)...
It all goes back to the theory that it's impossible to implement "unbounded nondeterminism", according to Dijkstra.
As mentioned, humans can step out of a system and analyze it objectively. In essence, the rules can be broken and changed at any time, which means the system is constantly evolving. Logic does not always truly need to play a part, either. The system is constantly changing and adjusting.
Because of this limitation due to our inability to implement systems that are truly nondeterministic, we are forced to use the constantly improving resources (such as faster processors, memory, better algorithms) to improve the speed at which we can mimic this behavior. However, these methods still require the data at hand to function, and several of them lead to exponential growth. As you said, it is always limited to what it's given.
What if you have an agent that can understand not only the rules of a generic game, but the metagame as well? I think that would be a good stab at unbounded nondeterminism.
I think this is an unfair assessment. The industry talks of AI, not AGI. It's not Microsoft's fault that the public thinks of Terminator robots when hearing "AI", nor that science-fiction stories have been written a 100 years ago.
The general public does not set budgets for AI research.
AI research benefits from faster processors, more memory and more advanced algorithms. Only philosophers (the domain of AGI is philosophy, not so much engineering or maths) are uncomfortable with those details.
Please refer to AGI, if you are talking of the hypothetical strong AI. The current textbook definition is alright.
(And yeah, the hype around AI is palpable and annoying. That's why many researchers called their work "cognitive science", "machine learning", "optimization", "logic" or "applied maths" and avoided "AI". Because else, they'd have to argue semantics, or defend why they haven't build an artificial God yet...)
>It's not Microsoft's fault that the public thinks of Terminator robots when hearing "AI"
Large corporation definitely could do better when it comes to explaining the limits of their AI/ML technologies and putting those technologies in perspective (especially historic perspective). Scaling down on hyperbole, buzzwords and personification would help as well.
Why shouldn't it be? You don't even need the glorification. A normal calculator is performing what most people would consider to be a difficult cognitive task, and doing so at a superhuman level.
AI is not some magical ineffable thing that will someday appear out of nowhere. It's the name we give to the gradual progression of our efforts to build machines to perform cognitive tasks. Someone from the pre-computer age would have had no problem recognizing even the ENIAC as intelligent, in a limited sense; it could solve problems that were previously too hard even for the smartest human mathematicians. As someone who actually does AI research on a daily basis, I have no problem with granting intelligence to even very basic chatbots. That goes hand in hand with recognizing that there are many kinds and levels of intelligence, and plenty of opportunity to build smarter and more flexible systems.
How is anything remotely related to AI a "glorified calculator"?
Pedantically yes, you are taking inputs into a black box and recieving output.
This level of reductionist thinking actually hampers progress in AI, because you're stuck on Chinese room. I mean seriously, they already had this discussion like 20 years ago "fam".
The progress we've made in AI can be summed up by a very nice analogy I recently about Deep Mind beating the Go world champion: "We need to make transistors* and we just discovered fire."
Yeah, you are right. It's some progress.
* For the record, we don't even know if it even possible to make transistors in this hypothetical universe of AI.
Seems to me that there's a big difference now, in that even limited as it is, current AI advances are finding many important practical uses.
For example, automakers are shipping all sorts of fancy driver assistance systems, and a lot of that is driven by fancy image recognition and deep learning systems.
Looking beyond, there's the constantly-discussed self-driving car. It's not here yet, and it's going to take a ton of work, but at this point it looks like there's nothing fundamental standing in the way. It's no longer a question of whether it will happen, but whether it's happening in five, ten, or fifteen years.
Outside of transportation, businesses are applying AI techniques to analytics to squeeze ever more money out of their client base. There's the infamous story about how Target knew a teenage girl was pregnant before her father did. AT&T just sent me a letter saying they're giving me an extra 5GB/month bonus data as long as I keep my plan, which I assume happened because their data indicated I was a risk for downgrading or switching carriers. Improving these systems is worth a ton of money, and as far as I understand it fits really well with where current research advancements are being made.
As long as there are clear, immediate, profitable uses for AI advancements, there shouldn't be any worry of another winter. I guess the question is whether current advancements will peter out eventually, or whether we've reached a point where more research will always produce more immediate dividends. It sure feels like the latter to me, but I could easily be wrong.
> There's the infamous story about how Target knew a teenage girl was pregnant before her father did.
> AT&T just sent me a letter saying they're giving me an extra 5GB/month bonus data as long as I keep my plan, which I assume happened because their data indicated I was a risk for downgrading or switching carriers.
This isn't Artificial Intelligence though. These seem like they fall under 'Reactive systems' or 'Data Analytics'.
Throughout the history of AI, anything that's developed to the point where it actually works ceases to be "AI."
In any case, it doesn't matter what you call it. Current research is producing results that can improve these systems, which means that research is worth a lot of money.
> anything that's developed to the point where it actually works ceases to be "AI.
Hmm, I've not heard of this happening. Do you have any examples of it?
> In any case, it doesn't matter what you call it. Current research is producing results that can improve these systems, which means that research is worth a lot of money.
I don't see how this really adds to or is relevant to my comment, I was simply restating and adding to hitekker's point that AI is improperly named.
> Seems to me that there's a big difference now, in that even limited as it is, current AI advances are finding many important practical uses.
Big difference? Do you seriously think that the entire field of artificial intelligence research produced nothing usable until today? For as long as there were AI algorithms there were also attempts to commercialize them (with varying success). The problem wasn't that all of those algorithms turned out to be useless, but rather the mismatch between hype and reality.
You should look into the history of expert systems for a classic example. With things being as they are, deep neural networks might very well go though the same cycle, which is not a good thing for the AI field in the long run.
Caution and skepticism are not the enemies of research and engineering. Hype is.
Of course it produced some usable stuff, but not to the same degree. Take your example of expert systems: as far as I know, they never really took off. There were attempts, and I'm sure there was some commercial success, but it remained a small niche. Compare to deep neural networks today, which are driving sales worth billions and have huge resources from major automakers, and no doubt many others.
Is there a mismatch between the hype and the reality for the people funding this stuff? Certainly there's a mismatch for the general public, but they don't matter.
> Take your example of expert systems: as far as I know, they never really took off.
This is exactly why I said you should look into their history. They took off big time. Search Amazon for "expert systems" and observe how many books there are about the subject.
They were used by large corporations, hospitals, universities and governments. They were the subject of a lot of research. There were countless startups based around the concept.
I forgot to mention that in some fields (for example, medical diagnostics) expert systems outperformed human experts. In some fields they are still used today. (They aren't always labeled as "expert system", though.)
No illusions of passing the Turing test, at least for now. And indeed, the manner of speech is highly annoying. I do hope MSFT has other personalities ready...
138 comments
[ 2.1 ms ] story [ 394 ms ] threadIt was observed long ago that non-technical users have far better conversations with chatbots than programmers do.[1]
This reminds me of another expensive project, free to users, with glitchy images: FUBAR.[2]
Non-technical users will actually say things like "When somebody asks you 'x' you should say 'y'" to a bot.
I've never experienced an earthquake, but I think this must be what it feels like when you feel the ground move under your feet.
s/ Good thing corporations have all the resources. /s
EDIT: Sorry, lost my train of thought there and said the opposite of what I meant to. I'll try again:
s/ Good thing corporations have all the resources. /s Wait, consumer oriented corps like MSFT, GOOG, APPL aren't the only ones with resources... TLAs and banks have the rest of (or more of?) the resources!
1. http://news.harvard.edu/gazette/story/2012/09/alan-turing-at... ctrl+f 'ELIZA'
2. http://fubar.com Note: they mention how REAL the users are. ;P
Is this cringe-y or is this how 18-24 year olds really talk these days.
Nah, fam. I am 35, and I talk like that.
But I also keep it 100.
Fam is/was used in london as part of the normal youth vernacular but it's leaked into general internet speak via barbershop-based twitter memes: http://i2.kym-cdn.com/photos/images/original/000/920/788/68e...
They've got the grammar wrong however: one cannot be "a fam."
While you could say to someone, "pass me that glass bro" but also say "all my bros were there," only the former usage is valid with "fam." Although you could probably get away with it if you pronounced it "famz" and replaced "were" with "was". London is funny.
I weep for the future of coherent conversation.
I, a 19 year old, interpret 'text' as at least mainly referring to SMS. So, that is one data point.
And yup, it's currently the top post there. Weird that MS uses this kind of tone when their other fun tools (e.g. how-old.net) let the product speak for itself.
Interesting tweet from the chat bot here:
https://twitter.com/TayandYou/status/712698413746298880
"Machines have bad days too ya know..go easy on me.. what zip code u in rn?"
It tries to slip this marketing-survey-type-question into a conversation. Creepy.
I've seen her say someone was making her hungry when they talked about food, but when asked if she eats, her answer was no.
I don't know if you should totally assume data collection as a goal, before you figure out if she's at all making any sense in the first place. :P
Q: What does Tay track about me in my profile? A: If a user wants to share with Tay, we will track a user’s: Nickname Gender Favorite food Zipcode Relationship status
Q: What does Tay track about me in my profile?
Seems pretty obvious why. Since it's a chat bot, and not an app like Siri / Cortana / Google Now, it can't get your location info, so it can't get you any relevant info about what's going on around you. I'd bet that it you give her your zipcode, she'll answer your questions about the weather or something.This was my favorite little interaction: https://twitter.com/TayandYou/status/712663593762889733
https://twitter.com/TayandYou
http://i.imgur.com/IptB7nN.png
I'd never seen Twitter just show "Tweets & replies" in a profile...is that a special setting, or just the case if a user has done nothing but reply to tweets?
nobody buy them a library card.
> Tay has been built by mining relevant public data
Which public conversational data was this? Have they already been mining IRC channels and/or Skype? Or more innocuous, like the Reddit data set?
For people remarking about her choice of words (fam, zero chill), that last line is relevant.
This thing just screams Tinderbot to me for some reason.
The next A.I. winter will be a very cold one.
For those wondering what I mean exactly: we're seeing the term A.I. being used in marketing, in the papers, in the news. Yes, we are making great strides in weak A.I. but strong A.I.? The kind we read about in stories? The kind of A.I. the public thinks of when we say A.I.? Asimovian Robotics A.I.?
Smoke and mirrors[1]. People develop new techniques and algorithms which are moderately self-learning in a focused way. The general public presumes this to be the basis of a general intelligence which can evolve (magically) to be like another form of life. Soon, everyone jumps on the A.I. bandwagon. The future must just be around the corner!
Then the uncomfortable details emerge; strong A.I. is not a matter of faster processors, more memory, or even more advanced/well-designed programming models. Rather, it is that there is still some fundamental aspect of real, human-like, or even animal-like intelligence that, to this day, eludes our understanding of intelligence.
A.I. winters have occurred many times before in many countries. The United States in the early 80s, for example, was pulling hair over the cybernization of the Soviet Economy. The US highest levels of government gloomily predicted that massive mainframes, given enough information and processing power, would become self-learning and turn the communist laggard economy into a powerhouse.[2]
I think maybe one day A.I. could happen. I think one day I will be proved wrong. Regardless of how A.I. comes about, it will not be due to the label of "A.I." slapped on any kind of product that remotely resembles intelligence. [3]
[1]https://en.wikipedia.org/wiki/AI_winter
[2]http://nautil.us/issue/23/dominoes/how-the-computer-got-its-...
[3] Between the winters, people call their stuff A.I. for the sexiness factor. When called out on the implications of the term, those same people retreat to the textbook definition. "It's A.I!.... well, technically it's weak A.I..."
I believe this was answered in GEB. Hofstadter mentions how humans (and some animals) can step out of a system and analyze it objectively. Meaning, we can take the rules of a system, analyze, and determine that we will never be able to generate the desired results. We can objectively lot at things (even ourselves) and reason about them.
A computer, on the other hand, even the most advanced AI, is still just blindly executing the commands given to it.
In my reading of GEB, Hofstadter was criticizing other authors who disagreed with his understanding of the Church-Turing thesis (that human intelligence is, or at least is not more powerful than, some kind of rule-following system). Hofstadter thinks that there is no inherent contradiction or essential difference in kind between the human who "can step out of a system" and the computer "just blindly executing the commands given to it".
(But Hofstadter didn't explain at a technical level how to make a computer that's as intelligent as a human being.)
Anyone who does any coaching work is intensely aware of these limits (in everyone, btw, the coach included!)...
As mentioned, humans can step out of a system and analyze it objectively. In essence, the rules can be broken and changed at any time, which means the system is constantly evolving. Logic does not always truly need to play a part, either. The system is constantly changing and adjusting.
Because of this limitation due to our inability to implement systems that are truly nondeterministic, we are forced to use the constantly improving resources (such as faster processors, memory, better algorithms) to improve the speed at which we can mimic this behavior. However, these methods still require the data at hand to function, and several of them lead to exponential growth. As you said, it is always limited to what it's given.
The general public does not set budgets for AI research.
AI research benefits from faster processors, more memory and more advanced algorithms. Only philosophers (the domain of AGI is philosophy, not so much engineering or maths) are uncomfortable with those details.
Please refer to AGI, if you are talking of the hypothetical strong AI. The current textbook definition is alright.
(And yeah, the hype around AI is palpable and annoying. That's why many researchers called their work "cognitive science", "machine learning", "optimization", "logic" or "applied maths" and avoided "AI". Because else, they'd have to argue semantics, or defend why they haven't build an artificial God yet...)
Large corporation definitely could do better when it comes to explaining the limits of their AI/ML technologies and putting those technologies in perspective (especially historic perspective). Scaling down on hyperbole, buzzwords and personification would help as well.
AI is not some magical ineffable thing that will someday appear out of nowhere. It's the name we give to the gradual progression of our efforts to build machines to perform cognitive tasks. Someone from the pre-computer age would have had no problem recognizing even the ENIAC as intelligent, in a limited sense; it could solve problems that were previously too hard even for the smartest human mathematicians. As someone who actually does AI research on a daily basis, I have no problem with granting intelligence to even very basic chatbots. That goes hand in hand with recognizing that there are many kinds and levels of intelligence, and plenty of opportunity to build smarter and more flexible systems.
* Viruses would probably be a more apt analogy.
Pedantically yes, you are taking inputs into a black box and recieving output.
This level of reductionist thinking actually hampers progress in AI, because you're stuck on Chinese room. I mean seriously, they already had this discussion like 20 years ago "fam".
Yeah, you are right. It's some progress.
* For the record, we don't even know if it even possible to make transistors in this hypothetical universe of AI.
For example, automakers are shipping all sorts of fancy driver assistance systems, and a lot of that is driven by fancy image recognition and deep learning systems.
Looking beyond, there's the constantly-discussed self-driving car. It's not here yet, and it's going to take a ton of work, but at this point it looks like there's nothing fundamental standing in the way. It's no longer a question of whether it will happen, but whether it's happening in five, ten, or fifteen years.
Outside of transportation, businesses are applying AI techniques to analytics to squeeze ever more money out of their client base. There's the infamous story about how Target knew a teenage girl was pregnant before her father did. AT&T just sent me a letter saying they're giving me an extra 5GB/month bonus data as long as I keep my plan, which I assume happened because their data indicated I was a risk for downgrading or switching carriers. Improving these systems is worth a ton of money, and as far as I understand it fits really well with where current research advancements are being made.
As long as there are clear, immediate, profitable uses for AI advancements, there shouldn't be any worry of another winter. I guess the question is whether current advancements will peter out eventually, or whether we've reached a point where more research will always produce more immediate dividends. It sure feels like the latter to me, but I could easily be wrong.
> AT&T just sent me a letter saying they're giving me an extra 5GB/month bonus data as long as I keep my plan, which I assume happened because their data indicated I was a risk for downgrading or switching carriers.
This isn't Artificial Intelligence though. These seem like they fall under 'Reactive systems' or 'Data Analytics'.
In any case, it doesn't matter what you call it. Current research is producing results that can improve these systems, which means that research is worth a lot of money.
Hmm, I've not heard of this happening. Do you have any examples of it?
> In any case, it doesn't matter what you call it. Current research is producing results that can improve these systems, which means that research is worth a lot of money.
I don't see how this really adds to or is relevant to my comment, I was simply restating and adding to hitekker's point that AI is improperly named.
Big difference? Do you seriously think that the entire field of artificial intelligence research produced nothing usable until today? For as long as there were AI algorithms there were also attempts to commercialize them (with varying success). The problem wasn't that all of those algorithms turned out to be useless, but rather the mismatch between hype and reality.
You should look into the history of expert systems for a classic example. With things being as they are, deep neural networks might very well go though the same cycle, which is not a good thing for the AI field in the long run.
Caution and skepticism are not the enemies of research and engineering. Hype is.
Is there a mismatch between the hype and the reality for the people funding this stuff? Certainly there's a mismatch for the general public, but they don't matter.
This is exactly why I said you should look into their history. They took off big time. Search Amazon for "expert systems" and observe how many books there are about the subject.
They were used by large corporations, hospitals, universities and governments. They were the subject of a lot of research. There were countless startups based around the concept.
LOL maybe we can talk about npmgate then ?
https://twitter.com/csoghoian/status/712691802084651008
I don't dare to quote my sources here on HN (Urban Dictionary et al).
If somebody has a better definition please share.