Machines Beat Humans on a Reading Test
quantamagazine.org
quantamagazine.org
A human can sit down in front of any of these sorts of tests and do reasonably well given no prior training whatsoever. None of these systems are capable of that.
Here is a challenge for an AI researcher: create a system that can beat an arbitrary, unknown console video game considered easy[1] for humans. You are not allowed to train your system by having it play video games, since this test is supposed to measure the ability to learn to play video games for the first time. Additionally, your system only has access to the I/O that a human has access to (gamepad inputs, sound and video outputs), no direct access to console memory by any means.
[1] I consider a video game easy if humans finish it more than 80% of the time on their first attempt, having never played a video game before in their life.
tl;dr: Your brain is training 24/7 non-stop.
It's doubtful whether such data exists. Even if it did it's questionable whether such results have any bearing on human understanding or intelligence.
Like a human that never played a game, used a computer, used any physical buttons, learned a language, understands what a game is, knows what a sport like football is, etc...
There is lots of stuff that you'll need to feed your AI to be equivalent to a 6 year-old brain. That and a 6 year-old interaction with a game will be more random than logically driven.
If you are doing that you are already at the stage of AGI. It is not theoretically impossible but the impracticality of it right now is what makes scientist trying domain-specific AI.
That’s kind of my point. The article is asking the question “do they understand?” and my answer is yes, if you can do this. Read, to a 6-year-old, a kid’s book about a knight with sword and shield rescuing a princess and they will have no problem playing Zelda. Does the child understand the damsel in distress trope? I would argue yes.
This sort of test would be extremely difficult to get an AI to pass. Heck, it doesn’t even have to be video games. Just do any test that involves translating a concept from one medium to another (text and video). It’s extremely simple for humans to do and extremely hard for an AI.
As a kid I was able to gather the first 6 pieces of the Triforce without any outside help. I think an AI should be able to get at least one, given that it's playing a video game for the first time.
As for the "embedded knowledge", I'm happy to talk about context in terms of cultural references, heroes, the monomyth, etc. The Zelda series draws heavily on these concepts, though I am not sure how much I knew of them as a 6-year-old.
The difficulty could also be decreased, but in the end the challenge would be to grasp how the game works, not to beat highscores.
Does such a thing exist?
Here’s another task I’d like to see: have a kid who has never seen a film before read a children’s book and then watch the film adaptation. Test the child on their ability to identify all of the characters in the film. This is the sort of task any kid could do with high accuracy but an AI would struggle with unless it were specifically trained on a corpus of books and their film adaptations.
So maybe training on a corpus of books and film adaptations is cheating in this scenario, but the AI should still be allowed to train on a massive corpus of all of the things a child may be exposed to. Maybe instead of showing it films, show it thousands of hours of footage from head-mounted cameras that people have worn while going through daily life, and children's TV shows, books, and games, and have it talk to a bunch of humans.
Everything we would label without question as possessing intelligence has an instinctual drive and mechanical ability to improve itself, to beat itself at its own games. To move out of the way of danger if it has the opportunity to see it coming. To improve the probability of its survival.
There’s more to intelligence than building a machine of a predetermined spec to mechanically sort a predetermined set of data. These notions in comp sci seem to completely ignore physical reality. I can conceive of a planet called Krypton, can it exist? I can conceive things like tautology but can physical reality provide an example? I can conceive of a phone app giving me a sense it has intelligence but it doesn’t. It’s not self aware and cannot self replicate. It can’t save itself or beat the odds on its own.
This physical reality we occupy invented what we’d define as intelligence. Our definition ought to very specifically include physical capabilities of independent not programmed self awareness & capacity for self care. AFAIK, that’s what our basis for the concept of intelligence is based on. Why should we accept an artificial intelligence that’s less fully capable of that as somehow on par or better than is given its obvious limitations?
I think it's an incoherent premise to suggest that you could have a human with no contextual knowledge or experience that pertains to playing an arbitrary game. There's no blank-slate human that you could test like that. Even the act of finding food and eating, walking around and avoiding hazards, observing the movement and behavior of other people and animals..those experiences shape a person's ability to interpret an environment and model the agents within it. Those priors would be useful for any first-attempt at a game.
The point of my challenge is to demonstrate understanding by being able to translate cultural and contextual information into a new medium. So rather than just making connections between different texts, you need to be able to take a text and leverage it to understand what’s happening on the screen in a video game.
So while I get what you're saying and I'm sympathetic to your point of view, it's too late. The cloud bucks and 5x VC multiples that can be made on this stuff is too good. Even academics, whose whole advantage is that they can resist market forces, have very lucrative careers awaiting them if they publish the right stuff.
Then there's all the adjacent benefits. Like what good does it to be publicly skeptical of DeepMind? Google is everywhere in tech and science. Or even some random guy's paper, like where is he going to work? I can count the number of opportunities on my hand, and they all have a lot of influence in technology broadly. You don't even have to be doing AI for it to be rational to simply say, "Wow what an Amazing Discovery" and move along.
There's literally no forum that profits from skepticism. Nobody wins when "AI," whatever it is attached to, loses.
Usually we discover that the machine does not have consciousness and our definition for “consciousness” was wrong. Sometimes we discover that our definition of “consciousness” is so bad that we have to go back to the drawing board and come up with a completely new definition. Same thing happens with the word “intelligence”.
A venus fly trap is aware of it's environment and has a useful response. In many peoples books that's consciousness others want to limit things to just say it's aware but lacks consciousness.
IMO, that's the kind of divide that's prevents meaningful discussions.
Of course, there is a spectrum here, depending on the complexity and dynamicness of this internal representation of the world.
Perhaps it goes something like: absolutely zero awareness (e.g. a rock), to barely aware (a Venus fly trap), to highly aware (elephants, humans, etc). I'm sure the scale doesn't stop there, but we haven't had a chance to observe anything other than the highly aware animals here on Earth.
You could break this down further, looking at things like episodic memory, the different senses of self, and so on. But humanity has only scratched the surface when it comes to understanding these things.
It doesn't prevent discussions for the sake of it. It's simply that we don't know.
Any position you take in this matter is nothing more than unfounded speculation.
This philosophical question keeps on being asked . From my understanding right now I think it should be approached the same way Feynman approached quantum physics.
Just train your models.
(When you read a novel, the experience of reading the novel is different for everyone. It’s like being on a rollercoaster or amusement park ride. Everyone has roughly the same stimulus but a different mental landscape and a different reaction. Our response to the book is largely a response to our own experience of reading the book, and only partly a response to the content of the book itself. This distinction isn’t just sophistry. Are we prepared for machines that react to books in ways that are completely alien to our own reactions?)
That question has been asked by philosophers and psychologists for a long time. There's a branch of philosophy called epistemology -- the study of how we know what we know -- where questions about whether we even know what we claim to know (or understand what we claim to understand) belong. In addition, nearly a hundred years ago behaviorist psychologists were positing that all we can see is other people's behavior, so making claims about their supposed mind states, understanding, etc, were unwarranted and effectively off-limits... and then psychology, cognitive science and linguistics had a backlash against behaviorists, rejecting their approach and assumptions, with mental states, thought, and understanding being admissible in the field for many decades now. So none of this is really new, and mountains of books and papers have been written discussing and debating such issues for at least a hundred years.
The Chinese Room thought experiment discussed in the article has been a staple of many Philosophy 101 courses, where the obvious followup question to ask (even for students just getting their first exposure to philosophy) is whether humans understand what they read to begin with. Pick up any related philosophy book and they'll discuss such questions ad nauseum.
Taking a much further step back in to history, more than 2000 years ago, the Skeptics were questioning whether we can know anything, and some were denying that anything can be known. Philosophers in the West have been debating such questions ever since. In Eastern philosophy and religion, such concerns are also very old, with a variety of people asserting the illusory nature of worldly knowledge, and positing that true reality is unknowable, or at least not accessible or limited by rational thought. Similar themes are echoed in Western mysticism.
Jumping forward yet again (forgive the rambling nature of this response), we have Descartes Duck[1], and the popular Enlightenment and Post-Enlightenment view that "man is just a machine". Questions about whether we really understand or are really doing no more than a sophisticated version of what the Chinese Room or "dumb" neural networks do could be seen as modern restatements of that view.
Because the logic underlying the Chinese room argument is unsound, it naturally leads to some bizarre conclusions. I think it’s a bit of a tragedy that it’s a staple of philosophy 101 courses, unless you use it as an example for teaching logic by dissecting it and showing how it is unsound.
Personally my view is that as much as we need our scientists and ML researchers to study epistemology, we need our epistemologists to study science and mathematics. Any coherent and comprehensive epistemological system in the future will have to somehow synthesize things like substance monism, post-Fisherian / post-Bayesian statistics, post-Popper science, and computational complexity theory. Just as too many ML researchers are reluctant to dip into philosophy and reinvent a poorer version of old philosohpical arguments, there are too many philosophers that would rather e.g. rehash some ideas of Hegel rather than step outside their own field long enough to synthesize new ideas.
The Chinese Room is a refutation of the Turing Test and similar claims on AGI. It has nothing to do with the person in the room, but with the outside observer.
You, the observer, could be convinced by the evidence that a Chinese speaker is in the room despite that not being the case.
So you can't use outcomes to infer intelligence, and formal computaion of symbols is not thought. A neural network is still just an extremely fancy, well tuned lookup table.
It will produce many images that do contain the object in question. It is amazing how well it works.
But then it will also produce photos that have nothing even similar to what I or anyone else would consider "a chair", "a bike" etc.
If the question is now: does Google AI understand what a chair is?
The answer is: obviously it does not. It does something completely different than "understanding" what a chair is. That method does work well, but it is not the "chair" we all know about.
Understanding will start with questions like "cat sitting on the chair".
I mean, I was so thoroughly trained to take standardized tests that I really didn't understand what I was reading. Through speed-skimming, selective underlining and recognizing the most probable answers, I was able to speed through reading tests without doing any normal reading.
Oh, you're talking about machines?
I can't link directly to the result, but if you select "co-reference" from the dropdown, and enter the above text, this demo correctly links "they" to "Machines".
A: "Answer that question."
Mark Miller pointed out, "A Computer's Perspective on Moore's Law: Humans are getting more expensive at an exponential rate."
https://web.archive.org/web/20160316014939/http://www.caplet...
If there's nothing we can do that they can't..?
The ability to think like a human. The ability to see and recognize the world around you. The ability to change your behavior. The ability to communicate. The ability to be free in a society driven by fear and self interest. The ability to look for love and trust in others. The ability to learn how to love yourself. The ability to be independent. The ability to be successful by your own choice.
A human can make you laugh, cry, scream, and cry again.
A human can design, analyze, execute, innovate, understand and apply new technology. A machine cannot do that.
- GPT-2 774M (cherry-picked samples)
If you want to know the difference between a machine and a human you need a human to compare to. (That's not quite it either, I think, but perhaps is a little less obscure?)
Rather than thinking "this machine can't do it as good as me", why not think of ways you could use the machine to enhance your own abilities? I think it would be really interesting to have a system at my disposal that can "read" a book and generate book reports/digests. This could allow me to spend less time reading speculatively and focus more of my (finite) reading time on books I am already reasonably confident contain what I'm after.
(Edit: maybe it's because the headline frames it as a competition, setting people to think of themselves competing with the machine, rather than collaborating with it.)