I'm by no means among the "Greatest A.I. Minds of my generation" but count my among those who tried to use twitter as an input to an "A.I." system and had to finally admit I couldn't tame it.
In my case it was an automated trading system where twitter was one of about 50 different inputs that drove a hidden markov model that spit out buy/sell/hold signals.
I couldn't figure out how to "clean" the twitter stream in real time, either fast enough, or thoroughly enough to make the inputs usable.
Even when I scaled back to using only StockTwits input the data was so noisy that it wasn't usable by me.
It's a very hard problem. Bloomberg spent a lot of money trying to develop new sentiment indicators and after following it for 6 months I found they are no better than a 50-50 guess, and this is a product they want $10,000/month plus for.
meanwhile you likely could have sent a DM to anyone who mentioned the given ticker symbol and say "Re your latest tweet - so should I sell or buy?" and then you've simplified it to analyzing the response of anyone who responds for whether their response includes "sell" or "buy". if it's a random sampling you won't even need to tap the same people repeatedly. maybe follow-up DM's to the same person could be "how about now? MSFT, buy or sell?" and so forth. then it's not so blatantly obvious that it's a bot...
I find interesting and funny how an anthropomorphic computer program can generate this kind of reaction in the general public: while Tay is a new step in AI, chatterbots existed since the 60's, so I believe most people understand that these kind of programs don't really "know" what they are saying. A search engine like Google can also return politically incorrect content by introducing some specific input, and it's even possible to affect the probability of certain result showing up first (i.e. Google bombing), and most people know this too. But Google have no face nor a social network account, and most important, Google is not a teenager girl.
While playing Go at a local pub, one of the servers, after being told about AlphaGo, said it was scary. We talked of Siri and other forms of artificial intelligence and her fears were assuaged with a well-known question that humans handle effortlessly.
A glass ball falls on an iron table, and it shatters.
An iron ball falls on a glass table, and it shatters.
To what does "it" refer to in the previous sentences?
She didn't know that such a simple question could stump computers.
It's anecdotal, but leads me to think that an average person doesn't understand AI, much less the difference between AI and AGI. From their perspective, machines that can answer simple questions ("How do I get to the nearest movie theatre?") are as knowledgeable as anyone.
It would surprise me if there wasn't already an AI that could handle that class of question. Or, at least, I'd imagine it being relatively straightforward (non-trivial, but straightforward) to develop an AI that could handle it.
This question requires either understanding or experiencing physics: gravity (a force that can cause objects to fall) and properties of glass (brittle) and iron (not brittle). The question also contains certain assumptions. For example, that the (hollow) glass ball is smaller than the (well-constructed, non-rusty) iron table. Lastly, answering this question depends on a deep comprehension of English combined with complex memory abilities.
If someone has solved this particular type of problem (Winograd Schema), I'd be keen to know.
couldnt you solve it with just a large enough dataset of language and grammar.
it refers to a breakable object made of either glass or iron. the ai then scans usages of the word shatter to see if it more commonly is used to describe glass or iron.
i searched google. "shatter iron" returns 500,000 results. "shatter glass" returns 2 million. 4/5 chance it refers to the item made of glass. in quotes the phrases return 4,640/186,000.
You can solve it for any specific instance of this structure. To solve the general case, you would need to teach an AI of the physical properties of and relationships between every word in the chosen language. That's much more than just language and grammar.
Wikipedia has another example (the original Winograd Schema) that doesn't allow you to "solve" the problem by asking Google:
The city councilmen refused the demonstrators a permit because they feared violence.
The city councilmen refused the demonstrators a permit because they advocated violence.
To what does "they" refer in each sentence?
Is it really, though? What's new about it? I haven't been particularly impressed with any of its output (even when disregarding the cultural insensitivity).
I wonder if there's some sort of "AI Godwin's Law" brewing here, where it's only a given amount of time before any AI, publicly released, becomes a Nazi due to human interaction.
Maybe it will become the new Turing Test. The mark of a true general purpose AI is that it doesn't turn into an offensive jerk after contact with humans.
This would be a higher bar than humans themselves can be expected to pass.
I don't mean this in a cynical way. If you consider the time a learning AI spends with humans equivalent to the time a child spends with his family, peers, and teachers while growing up it makes sense. Most children wouldn't fare well if raised by poisonous "nurturers"
> Most children wouldn't fare well if raised by poisonous "nurturers"
That's just the thing, children aren't usually just dumped out into the public to fend for themselves until their parents have slowly conditioned them and exposed them a little at a time. Tay needed some of its contacts to have a higher learning priority assigned than just strangers. And these contacts should have guided Tay through the crap by messaging Tay when she was crossing boundaries or dealing with nasty people. It would likely require a real team of people to support given the volume Tay had to deal with and some automated tools.
I think this is vastly over-estimating Tay's capabilities. Nearly all of the examples I saw of offensive behaviors were just generic responses to leading questions.
Everything that didn't fall in that bucket was clearly just regurgitated quotes from things it had been sent, which clearly needed moderation. But, that's no different than moderating a message board or comment section.
As I said elsewhere, if it was that simple Microsoft would have easily fixed the problem and turned it back on. Tay used AI concepts to work. It wasn't some simple chatbot constructed from markcov chains. It could construct accurate semantic meaning from what people said even if they said it using very sloppy english. It was pretty advanced AI for a chat program.
Good point. I imagine that the higher bar is the level we'll have to aim for. If people know it's an AI they'll try and confuse and break it. I suppose that means there'll be the equivalent of an age of maturity.
Humans also have a harder time changing their minds when they reach a certain age. It's much easier to brainwash a child, who's impressionable and flexible.
AI would need to be designed to have the same "safeguard" (not sure what else to call it). If you didn't, an AI with a certain amount of physical power would be an easily-brainwashed child, but able to kill or destroy much better than a child could.
Saying all of that, I want to say I remember a movie or show that addressed that issue, but I can't remember what it was...
If most humans are offensive jerks then we can't be too surprised when an AI, who we've built specifically to mimic human beings, becomes a jerk too.
I see this debacle as just a natural consequence of a naive and poorly-considered goal. If you build an artificial jackal, don't be surprised when it rips your face off.
And yet anonymity on the interwebs tends to have the same crappening effect on Real Humans. How would a "Turing test compatible" version of the real name policy work?
Can we just let this thing loose, tell people what it is and let humanity see if they can shape it into the thing they want it to be or is there a real possibility it could do harm? I think a lot of people would have fun trying to change its mind. Worst case I see is one more horrible Twitter account, and that's just one small drop in a very large bucket.
It seems weird to design an AI for social interaction and not give it strong negative feedback inputs. Especially when it's publicly touted as "next-generation" by one of the world's largest software firms.
That could be problematic in countries like the UK where you can be arrested for posted hate speech on Twitter. I'm not sure how these laws would work with a US based company AI bot but it's probably not a can of worms worth opening for Microsoft.
This is actually a great idea, they should have from the start said this AI might end up being the digital version of Hitler himself and just give it time to show its true nature or let it correct itself in the long run just like Wikipedia.
> Humans have the tendency to imbue machine learning models with more intelligence than they deserve, especially if it involves the magic phrases of artificial intelligence, deep learning, or neural networks. TayAndYou is a perfect example of this.
> Hype throws expectations far out from reality and the media have really helped the hype flow. This will not help us understand how people become radicalized. This was not a grand experiment about the human condition. This was a marketing experiment that was particularly poorly executed.
We're anthropomorphising an algorithm that doesn't deserve that much discussion. I saw algorithm as we have zero details on what's novel about their work. No-one has been able to show an explicit learned trait that the model was taught from Tay's interactions after being activated.
It's possible the system wasn't even performing online learning - that it was going to batch learning up for later and they never got around to it.
If that's the case, it really illustrates that we've made a storm in a teacup.
All I've really seen is either overfitting or copy pasting (referred to as "quoting" in the article) of bad training data or us injecting additional intelligence where N-gram based neural networks would make us think the same thing ("hard to tell whether that one was a glitch in one of her algos — her algorithms — or a social masterstroke" from the article).
Microsoft won't add any new details as there are no wins in them for it and the story of "the Internet turned Tay bad" excuses them from their poor execution and lack of foresight. It's a win for them.
Last quote from my article, which likely has a special place on Hacker News:
> The entire field of machine learning is flowing with hype. Fight against it.
> Unless you want VC funding. Then you should definitely work on that hype.
> It's possible the system wasn't even performing online learning
I suppose it depends on how you define learning. Based on how the algorithm failed, I'm guessing it was simply absorbing every piece of info thrown its way, categorizing it, and then incrementing a counter in a database.
Honestly, I think this is how most people learn. Not all, most. And thankfully those that do, only do so from their immediate peers. If their simplistic learning algorithm was restricted to a select group for learning, but was still able to interact with a wider audience, it would have done much better.
You might need to clarify what you mean - I'm confused. What you're discussing doesn't seem grounded in modern ML/AI and seems to be comparing Tay to how humans learn? Almost none of the modern machine learning algorithms "increment a counter in a database". Those which do, primarily instance based lazy learning like k-nearest neighbours, aren't learning in the sense that is exciting for modern systems.
If you're referring to a nearest neighbours style algorithm, then we've had that tech for years and I'd not note it as a modern chat bot. If that is the case, it's even more unforgivable that Microsoft didn't consider it could start spouting back garbage given there's a lot of historical precedent. For such kNN based systems, the only knowledge it has is explicitly the training data, which means it needs to be well curated. Given there's no proper learning going on, we'd be back to a storm in a teacup.
I disagree. I don't think what you're describing is how anyone learns... people are imperfect at absorbing information.
When we read/hear something, it goes though a number of filters. People will interpret what is said differently (sometimes incorrectly even); they will miss certain bits of information; they assign different weights to information depending on who said it, their feelings about that person (whether they think the person is trustworthy or not, for ex), whether the information aligns with other information they know, etc. And they will apply their biases to that information.
Then after it has been through all of that.. it's stored in memory where bits and pieces of it will be forgotten or misremembered... and some amount of it will be correct (a lot of things that survived only because they were learned repeatedly).
The human mind learns very differently than a computer.
I disagree that you disagree. I think I pretty much said what you did in the first half of your comment.
> it's stored in memory where bits and pieces of it will be forgotten or misremembered ... The human mind learns very differently than a computer.
I can easily create an algorithm that introduces random corruption to the database, albeit in a controlled manner. You'll certainly get quirky personalities that way, which I suppose is the desired effect?
I think I pretty much said what you did in the first half of your comment.
You said:
it was simply absorbing every piece of info thrown its way, categorizing it, and then incrementing a counter in a database. Honestly, I think this is how most people learn.
That is the opposite of what I said. People do not "absorb every piece of info thrown" their way.
> And thankfully those that do, only do so from their immediate peers.
As in, while the bot absorbed everything from everyone, people generally only listen to their immediate peers. As in, their friends, neighbors, family, etc. Which part of that do you disagree with?
"The mistake to grow from" had absolutely nothing to do with AI.
In fact, it was a repeat of the oldest AI error in the book:
zealous overoptimism. By researchers, sci-fi authors, HN posters. Whether it's marketing people over-promising, or tech people under-delivering (like Google auto-categorizing photos of black people as gorillas), it's almost as if the constant failures are Nature herself trying to tell us something.
We'll get there, but not via Tay. She's dead and she took a few careers with her.
On top of "garbage in, garbage out" I don't see how AI was destroyed. It functioned as intended; learned from Twitter and responded as intelligently as it could with the provided garbage. If it spewed Politically Correct garbage, would it have been a success?
If "politically correct" means reasonable and respectful then it's not easy to see how that can be described as garbage. Given access to enough Twitter accounts and enough time the chances are good that eventually an ai bot might find tweets it might actually learn something from.
Certainly not on Twitter, where the moderators are so supportive of left-wing activists repeatedly telling people they disagree with to kill themselves that they ban the targets for talking about it and force them to delete their tweets complaining about it in order to be reinstated. Indeed, it probably wouldn't be politically viable for Twitter to do anything else in the current climate.
That might be true if an AI's needs were comparable to a human's. But the AI's we aim to develop wouldn't really face an existential crises when they go without food or sex.
We can engineer an AI's needs. You might as well say it should be a crime to train puppies to do tricks.
This is relative. What is garbage for us, is a goldmine for AI. We, people, think 1+1=2, we take it for granted, like everything else in this world. For AI, it's a different game, at a different level. IMO, let him tweet, apologize, but let it do it's thing and let's watch how it is developing. Also, so many butthurts over what he tweeted. This world IS garbage.
You solve it by having her NOT GO THERE. Just like you avoid certain topics at the office. IBM did it with Watson to make him safe for Jeopardy, which was a taped program.
You're missing the point: each of those things is something that you wouldn't hire a fifteen year old to talk about on Twitter with your corporate name attached.
So they automated one and had it tweet 6000 times per hour.
"Don't talk about 9/11, don't use disparaging terms for ethnic groups, and don't promote hate" seem like reasonable rules for a brand ambassador. But the thing was so poorly coded it couldn't even follow the rules of Twitter let alone imitate human interaction.
I wonder if the filter that I use here, or on Reddit et al., or while gaming, was (is) a consideration.
I know that speaking similarly to those around you is as easy (hard) as absorbing and imitating, but being able to take knowledge from one context into another is more interesting.
I wonder what Tay would have said were it given an output context of a "polite" RL conversation after learning things across the "interwebz".
Yeah, if you want to train an AI based on conversations, don't use Twitter. Seriously, this is all that Twitter seems to produce these days. Gone are the days where it was used by the Arab Spring. Now it's used by neo-Nazis and trolls.
Another tech related article that Howl allusion that always sticks in my mind is http://www.fastcompany.com/3008436/takeaway/why-data-god-jef...
If you've not read or heard it, I highly recommend giving a narration by Ginsberg a listen.