I’ve Seen the Greatest A.I. Minds of My Generation Destroyed by Twitter
newyorker.com
newyorker.com
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.
The narration I have is surprisingly sedate given the text of the poem. Not a criticism, really, just something which struck me as odd.
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.
(i.e. for one specific stock or industry, or slowly and not in realtime) ?
If not, do you know for sure that there is any relevant information content in the Twitter stream?
You could skip the NLP altogether and hook up a PRNG and apparently people will still pay you.
If there's a lack of intelligence, it's not in the code.
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?
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.
I was expecting this to be the set-up for a joke :)
If someone has solved this particular type of problem (Winograd Schema), I'd be keen to know.
https://www.cs.nyu.edu/davise/papers/WS.html
https://en.wikipedia.org/wiki/Winograd_Schema_Challenge
http://commonsensereasoning.org/problem_page.html
http://spectrum.ieee.org/automaton/robotics/artificial-intel...
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.
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).
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...
Interesting perspective.
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"
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.
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.
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.
> 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.
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.
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.
Do you have an example of a machine learning algorithm that doesn't involve trial and error?
Because any trial and error system will require you to keep track of past events, where you are going to "increment a counter in a database".
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?
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?
This is actually part of the communications field. Here's a page that talks about how people communicate information and several of the issues: http://www.msucommunitydevelopment.org/effectivecommunicatio...
there's plenty more if you want to search for communication filters
Then, I want to see how Tay 1.2 will tweet.
Then, I want to talk with Tay 2.0.
I will be anxious to have Tay 3.x respond to my inquiries, instead of mindlessly searching StackOverflow.
I can accommodate the problems of Tay's mistakes initially, to see how AI will grow.
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.
We can engineer an AI's needs. You might as well say it should be a crime to train puppies to do tricks.
If "politically correct" means reasonable and respectful
But then how you filter out stuff like 'the Holocaust didn't happen', which would be just as offensive even though it involves no swear words.
These are normative matters, which I imagine would be very hard for AI to tackle.
If only there were an easily parsed list of topics to avoid... https://en.wikipedia.org/wiki/Wikipedia:List_of_controversia...
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.
And I thought the New Yorker had editors.
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".
So, the alarming responses were not, in fact, a part of the AI algorithm.
But of course, I also don't get a lot of random people messaging me.