I worry that this could be even more succeptible for a 'bag of tricks' AI to win than a traditional Turing Test, after looking at the rules[1].
Chatbots that perform the best on a traditional Turing Test are all extremely specialized 'bag of tricks' approaches. They can be surprisingly convincing, because they often use huge databases of real human responses to construct their own -- if you haven't already, I'd recommend giving Cleverbot a try[2].
However these chatbots often fail when an important continuity of facts is being discussed in context, where their parrot-like nature is exposed. Therefore I think this "TED XPRIZE" will be much easier to fool in this sense, because there's only two questions from the judges. This essentially removes the most difficult AI challenges of a traditional turing test.
Of course, succeeding at the "TED XPRIZE" would without a doubt require legitimate and significant advancements to the field of artificial intelligence and machine learning. Synthesizing a convincing talk/essay with an engaging narrative and central theme is no easy task (though limited successes already exist [3]).
But even now computers can beat humans in very specialized areas (and specialization is a key requirement to this). For example, we already have human level-facial recognition[4] and superhuman character recognition[5]. Similarly, while it would indeed be a legitimate AI advancement to be able to construct such a speech, I do not believe this would be any more a holy grail of AGI than character recognition and facial recognition. Instead, it would be incremental progress disguising itself as having already arrived.
From these bags of tricks, we can learn a lot about, if not AI, at least humans. My favorite example in that category is MGonz [1]:
"In 1989 MGonz was able to fool a human subject for over an hour [...]
MGonz worked on a simple principle: don’t reply to statements. Merely
insult the person making the statement. Most of MGonz’s one-liners are
too profane to repeat – but “ah, type something interesting or shut-up”,
“what sort of idiot types something like that”, and “that’s it, I’m not
talking to you anymore!” are fairly representative of the style. The
programme simply pours a torrent of abuse on the unwitting human at the
other end of the internet connection."
> because they often use huge databases of real human responses to construct their own
Honestly, this often seems to be the conversational gambit of human beings.
Coming up with genuinely interesting conversation seems to be a task that takes too much effort for most everyday interactions.
Whether it be standard small chat, quoting cultural media, or reciting social group in-jokes, much of human interaction seems to also take a "bags of tricks" approach.
But how does yet another "bags of tricks" chatbot help to advance the field of AI? The Turing test and projects like this are a distraction on what should be our true goal of creating general AI. There is a large amount of evidence in neuroscience on the plasticity and uniformity of the neocortex, all pointing to a single algorithm for general intelligence.
I would much rather see a prize for finding more general purpose algorithms and using them to create general purpose AIs. Off the top of my head, show a picture a animal, building, logo, face, word, etc., and be able to pick out that same type from another group of pictures. Google has already demonstrated that we can teach a computer to recognize a cat with thousands of pictures and deep learning models, how about just one? The more general the test, the better, it should also be able to be trained to play new games and recognize sounds as well as pictures. Unlike another useless,specialized chatbot, this would be better aligned with Xprize's goals of "encouraging technological development that could benefit mankind."
That may be true, but at best you would just get animal level intelligence. We haven't refined our algorithms like evolution has over millions of years with natural selection, and we have far less computing power and far less training time than real brains.
Most animals can't learn language, or play games, and aren't very intelligent.
I recently watched a video with Stanford AI researchers talking about how the general case of searching for an object like a coffee cup in office is still an unsolved problem. At the same time, my German Shepherd with a pea sized mammalian brain has learned to pick up garbage when we go on nature hikes. After pointing out enough plastic bags, cups, cans, etc., she's somehow learned the concept of trash in the forest and will often find it on her own and bring it to me. Advancing AI to this level, would be a worthy goal.
The current rules are just meant as a high level baseline to start a discussion around designing the competition and the announcement was a call for people to do just that. An XPRIZE can take several months to years going through a design process before finalizing the rules. During that time domain experts are consulted to help define the competition.
I attempted to use the voice of Lieutenant Data from Startrek The Next Generation to interact with the Internet facing Cleverbot interface. I regret to inform, HN, that Cleverbot performed worse than anticipated.
I did, however, witness some potentially interesting if antiquated webdesign, not least a promotion for a potentially more human-like 'friend finder', located at the base of the page. It has 199k 'likes', 20.7k 'follows' and 22.2k 'shares', leading to a strong possibility Cleverbot is merely a pseudo-presence, and the true nature of this website will transpire only by clicking more deeply.
Did you read the same article as me? From what I could gather, it's more of a challenge in robotics. Nothing seemed to indicate that the speech should be machine generated. Rather it's a matter of making a robot charismatic/impressive enough to engage the audience.
This prize doesn't seem very useful to me. If someone actually was able to pull this off, they would be better off making billions off of that tech in the marketplace instead...
Even if the winner is a bag of tricks trick, it still would be something new in the way it is presented, with new tricks developed, and thereby surely would spark some innovation, would it not?
If so, then the XPrize served its purpose - to drive innovation.
I am not under the impression that XPrizes are intended to be impossible goals or "holy grails" - just goals that drive progress.
I respectfully disagree. As far as I understand, it is not about Q&A/jeopardy kind of context. It's about generating a TED speech that would involve deep domain expertise but more importantly tone and speech technics to make it great to hear...
I'm surprised that for a mere $10M they were willing to simultaneously torpedo the xprize brand and forever enshrine their irrelevance to actual innovation.
That video is the perfect response for why this is a silly idea. A clever chat bot can beat the Turing test, but in a TED talk you don't even have to deal with probing questions. All you have to do is babble on about something vaguely inspiring.
Yeah, I was already wondering how they were going to reconcile that particular speech with whatever nuances of the prize rules would be required to keep people from giving a speech like that one.
It gets weirder when you consider that a) an AI could have plausibly (though not likely) have generated such a speech in answer to a real question (for example, "discuss a flaw in the TED program", and b) no one rejected Watts' talk as unacceptable, though they probably would have if an AI had generated the same talk.
1. Pick two keywords that describe this event. "TED" and "XPrize" seem like good candidates since they (together) identify this.
2. Go to Google (in a private browsing window, if you wish non-bubble results) and type in these keywords together. Here is an example URL: https://www.google.com/search?q=TED+Xprize . Order is irrelevant so long as phrases are maintained as phrases.
4. Observe that the page says that the rules are not yet set in stone but that a 'sample' exists: http://www.xprize.org/ted-rules
5. Follow that link to reach a page which says that "...the competing A.I. is given 30 minutes to prepare a compelling 3 min TED Talk" once it receives a subject.
The conclusion is the following:
The AI must actually construct the TED Talk, i.e. the words of the TED Talk must be programmatically constructed given a topic.
This being the case, I doubt your TI-99/4A would deliver a performance compelling enough to stand and applaud. While I'm sure you searched before asking the question, I hope that this guide helped you construct your search for information better.
Ha ha, certainly, that would have been nice. I'm disappointed that I didn't even draw a chuckle. After all, I'm a human posting a mechanical sounding comment on a thread about machines speaking like humans.
That's just a really simple markov chain generator. If someone could get a bunch of transcripts of TED talks in plain text, this would be pretty straightforward.
> All of this must be accomplished with no human interaction with the mechanical being.
Well that makes things much easier. Answering questions without giving the game away would be impossible, after all. Generating an expert system to just walk around and generate a passing talk without any interaction is pretty easy, though.
How many humans could get a standing ovation for a TED talk on an arbitrary subject after 30 minutes preparation? I know I couldn't. It seems bizarre to create an AI challenge that most people couldn't meet.
I wonder how far you could get, though, with an AI that strung together phrases like a struggling beauty pageant contestant or politician: "I think world peace is important because uh education is the key to success and uh what really matters is the children and uh that's what makes America great!"
> I wonder how far you could get, though, with an AI that strung together phrases like a struggling beauty pageant contestant or politician: "I think world peace is important because uh education is the key to success and uh what really matters is the children and uh that's what makes America great!"
Hmm -- you clearly haven't listened to many TED talks, because lots of them sound just like that. Having listened to about 100 of them over the years, I eventually came to the conclusion that the most popular ones acquired their acclaim by being as formulaic as possible -- say positive things, don't mystify your audience, build your talk out of a series of verbal bumperstickers.
Yes, this should work, because the organizers have a realistic grasp of (a) the present state of AI, and (b) the low intellectual level required for a typical TED talk.
Yes that's true, but I think that we are far away from being able to create artificial charisma, it is pretty hard to write a machine learning algorithm that can mimic emotions.
Chatbots that perform the best on a traditional Turing Test are all extremely specialized 'bag of tricks' approaches. They can be surprisingly convincing, because they often use huge databases of real human responses to construct their own -- if you haven't already, I'd recommend giving Cleverbot a try[2].
However these chatbots often fail when an important continuity of facts is being discussed in context, where their parrot-like nature is exposed. Therefore I think this "TED XPRIZE" will be much easier to fool in this sense, because there's only two questions from the judges. This essentially removes the most difficult AI challenges of a traditional turing test.
Of course, succeeding at the "TED XPRIZE" would without a doubt require legitimate and significant advancements to the field of artificial intelligence and machine learning. Synthesizing a convincing talk/essay with an engaging narrative and central theme is no easy task (though limited successes already exist [3]).
But even now computers can beat humans in very specialized areas (and specialization is a key requirement to this). For example, we already have human level-facial recognition[4] and superhuman character recognition[5]. Similarly, while it would indeed be a legitimate AI advancement to be able to construct such a speech, I do not believe this would be any more a holy grail of AGI than character recognition and facial recognition. Instead, it would be incremental progress disguising itself as having already arrived.
[1] http://www.xprize.org/ted-rules
[2] http://cleverbot.com
[3] http://www.geekosystem.com/journal-accepts-nonsense-paper/
[4] https://www.facebook.com/publications/546316888800776/
[5] http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.84.1...