In a round of Unreal Tournament 2004 played for 2012 BotPrize competition, two bots were mistaken for human 52% of the time. This is big. In fact, this is HUGE. One hundred years after Alan Turing invented the Turing Test, artificial intelligence was both smart enough and human-seeming enough to pass it.
Credit where credit is due: This achievement was the work of computer scientists at The University of Texas at Austin: Risto Miikkulainen, a computer science professor in the College of Natural Sciences, and doctoral students Jacob Schrum and Igor Karpov.
The human players identified the AI entities as human because they did not play perfectly. Their imperfection was by design. The way to trick a human into believing a robot is human, too, is to give it flaws.
Jacob Schrum explained the winning design in a press release:
“In the case of the BotPrize,” said Schrum, “a great deal of the challenge is in defining what ‘human-like’ is, and then setting constraints upon the neural networks so that they evolve toward that behavior.
“If we just set the goal as eliminating one’s enemies, a bot will evolve toward having perfect aim, which is not very human-like. So we impose constraints on the bot’s aim, such that rapid movements and long distances decrease accuracy. By evolving for good performance under such behavioral constraints, the bot’s skill is optimized within human limitations, resulting in behavior that is good but still human-like.”
A basic principle is clear: humans screw up, robots don’t.
Where do we go with that, I wonder? There are many situations in which perfection would be a positive boon: driving a car, for example; or performing any repetitive task. Bots do not get tired or bored, or wish they were somewhere else. (At least not yet.)
If your idea of perfection is to be able to aim at a video game target and never miss, go for it.
Then I thought of the great line from the miniseries Slings & Arrows: “Forget perfection! There’s nothing more boring than perfection.”
Consider the strange result from the same tournament where human players were correctly identified as human only 40% of the time. What is going on THERE?
Imperfect judgment such as that would drive a bot crazy, if it could be driven crazy and it probably could be.
Imperfection would keep bot-human relationships from being boring, though at some point the human might wonder why he/she is bothering. The bots would have to continually evolve, change as life changes us, to be interesting to be with.
I play Scrabble with a “bot” inside my Ipad. It beats me on the hard game every time and I wonder why I even bother to try again. Imperfection in this case would be a welcome change. Right now the skill set is uneven since the bot has acccess to every word in the dictionary and I have my brain and the words I’ve learned.
That’s a good example of the problem with perfection — at least as it is defined in games. There is no “game” when playing against such a bot.
>________________________________
I’m sorry, but I disagree with congratulating these technicians, because I think that it massively cheats the test that Turing envisaged or proposed to set out to mimic human characteristics, such as fallibility, with a deliberate aim of passing off one’s creation as human.
To use a stupid example, it’s like being asked to applaud someone for not stealing £1,000 from a chair in a shopping centre when the reason why he or she doesn’t do it (and would have done otherwise) is because of being tipped off that it is a trap.
Back at The Turing Test itself, human beings do whatever they do that is not machine-like repetitive perfection not to convince others that they are human, but simply because they *are* human, so it is a by-product or symptom, not a strategy to gain acceptance.
Subverting the test by making it one’s more or less sole goal to replicate such by-products or symptoms not only significantly devalues what the test is for, but also comes across as Alexander the Great slicing through The Gordian Knot because, truly, not big enough to admit that he could not untie it. Interpreting the Test as having been passed (and, even with this bogus introduction of human traits, it was only a marginal 52%) is about as valid as saying that Alexander undid the knot (meaning it in the sense of teasing it apart).
I doubt Turing intended to set the bar for the test as “human-like fallibility.” However, I think the victory has meaning. Rote, “robotic” behavior is inhuman, not only because no mistakes are allowed but because no learning is possible. Much research is under way, and progress is being made, in creating AI entities that learn from their experiences.
I still wonder about the statistic that indicated that humans misidentified other humans as robots most of the time.
>________________________________
I think that the way in which human memory works, which is not like a database or your PC’s RAM, because it is truly random access, will be the stumbling-block for the attempt to replicate thought-processes and the capacity to infer, recall pattern, and learn.
In a game called Pairs, where, either with two decks of cards, or a designated set of pieces where each and every image has its counterpart, a fully operational sensor would allow the machine-player to win every time, by, every time that a pair of cards is turned over, mapping which image is where – the same memory-trick that allows some card-players to know every card that has been played.
But compare that with this – I am trying to remember something, a word or an actress or the name of a film. I don’t have a list of every word / actress / film title that I know and which I go through one by one, checking whether it is the one that I am after. In fact, I don’t know how I ‘find’ this word / actress / title for which the one that I have in mind, and know is wrong, is a poor substitute.
With that list of words / actresses / titles, the computer / robot that knows that there is one that it is seeking could take a very small time to go through the lot and find the right one, because, although there would be labour involved, the computing time for each check would be negligible. But that, for me at least, is not how I think of what I’m seeking at all. First, I concentrate, and a first guess may soon be thrown up – I know that it’s wrong, but it’s on the trail, and I think why, and maybe soon after a second guess comes up.
I appeal to that experience that we all have where a friend says Johnny Depp was in that film, and we hear and respond to say that he wasn’t, but we can’t immediately place who it was. We start thinking, and maybe quite quickly the Christian name springs into one head or another, just like that – out of nowhere. It gets repeated, ‘Bruce —, Bruce —’, and maybe that aural trigger, the film in question, and someone like Depp brings up Willis as a match. But we both think and that’s not quite right, but, soon after, a slightly incorrect version of the surname comes from one of us, and the other corrects it, and we both agree.
An amazingly concise process in terms of the number of steps and the links that the thinking mind is making, with stimuli, to get to the right answer. I see no way of replicating that in a machine except by grafting on a human mind – which may or may not be the same thing as the brain – to its soulless self and give it creativity, intelligence and life.
I think you may be onto something.
>________________________________
The spider outside my kitchen window is a fraction of my size (I cannot guess what fraction), but it can do not only what I cannot do, but what I cannot imagine a man-made creature that size doing:
1. Hanging its web to fit the available points of support and first mapping out the size for its cathedral or stained glass in the air;
2. Producing the web to whatever tolerance without any lesson in geometry or trigonometry, and being able to negotiate its own trap;
3. Through the web, sourcing its alimentaton, and all to further the aim of mating and self-perpetuation.
Reductive science has ceased to see the marvel in any of this, though, as I challenge it, it cannot even reproduce what it explains away.
Excellent example. Beautiful.
>________________________________
Or think of something, as small as a fly. It’s said that there is no real intelligence – but, somehow, this claim is made about the memory-span of goldfish, but I do not see how it would ever have been established – and that the movements of the fly are a programmed strategy for exploring its environment, following an algorithm.
Yes, but something that small flies, and gets itself from one meal to the next to be able to do so. It’s true that, if you want a fly out of a room so that you can sleep without the noise of it, you can lead it by the nose, but that doesn’t detract from the wonder (rather than the nusiance or disgust) of it : turn on the light in the adjoining room, and, whilst the creature is in flight, switch off the light – it should head out of the door, following the light.
Not much of a pal to AI, am I?
AI is a heavy, clumsy thing compared with the least of life.
Sent from my iPad