Slashdot Mirror
A.I. Developer Challenges Pro-Human Bias
destinyland writes "After 13 years, the creator of the Noble Ape cognitive simulation says he's learned two things about artificial intelligence. 'Survival is a far better metric of intelligence than replicating human intelligence,' and "There are a number of examples of vastly more intelligent systems (in terms of survival) than human intelligence." Both Apple and Intel have used his simulation as a processor metric, but now Tom Barbalet argues its insights could be broadly applied to real life. His examples of durable non-human systems? The legal system, the health care system, and even the internet, where individual humans are simply the 'passive maintaining agents,' and the systems can't be conquered without a human onslaught that's several magnitudes larger."
The banking system is another example of a system much better than human intelligence for survival and resilience. Oh wait...
Survival is a terrible metric of intelligence. By that standard, lions and tigers and bears are the most intelligent species on the planet.
Check out my sci-fi/humor trilogy at PatriotsBooks.
By redefining intelligence to have nothing to do with what anybody means by intelligence, he can then claim that other systems exhibit more intelligence. Like a rock, presumably, since it survives far better than humans. I think this may be an example of somebody getting too interesting in specifics of tree-bark, and forgetting about the forest.
He essentially seems to be arguing that grey goo is the pinnacle of AI.
I much prefer the existing literature requiring that intelligence be an intelligence we can relate to as humans. Survivability is an interesting metric for creating more self-sustaining systems, but the goal of robotics should be fostering better knowledge and understanding of the universe. Searching for blind replication at the best rate possible just feels empty.
First of all, he doesn't actually say much about the survival as intelligence idea beyond the positing of the notion itself. It gives him a nice way to consider survival and intelligence as linked systems, with the "survival" of a system (that definition alone gets pretty abstract) being measured in terms of the logarithm of the number of humans required to shut it down.
/. article writers' need to get your attention, it suddenly becomes Human Intelligence Is Over.
He says you CAN consider the Internet, legal system, medical system, and others in terms of this notion, but doesn't get terrifically specific about it. He does, however, specifically state that road systems and the legal system are at least an order of magnitude more resilient than a human-level intelligence, which is nice, if you believe his examples are well-chosen. I'd be hard pressed to claim that they are.
In other words, he sets up an interesting research topic and then between his own poor choice of phrasing, the multiple Singularity references which surround the article, and the
Reading the article, it struck me as a good explanation of why AI is not getting anywhere.
By the authors criterion the solar system would take a huge number of people to shut down, and thus would be vastly more intelligent than any collection of surviving knives and forks used at AI conferences. I think that answers the other complaint of the author as well,
"There is a lack of scholarship in this area. This is, in large part, because most ideas about intelligence are deeply and fallaciously interconnected with an assumed understanding of human intelligence."
Oh well,
Stephan
http://stephan.sugarmotor.org
The available documentation for Noble Ape is fairly shallow and opaque, it describes a simple scripting language, and some high leve discussion about space, time, and so on... but that's about it. Where's the AI? How exactly does the model simulate an ape, what's the relationship of the model to ApeScript? Where, in short, is the FAQ?
Unsurprisingly, most of the people here haven't read, or perhaps not really absorbed, what TFA discusses, and are jumping to quick and irrelevant conclusions.
The author explains that Survival is a good metric of Intelligence, and he uses humans as an example. One human can definitely kill one lion, bear, mosquito, single bacteria, etc. if equipped with his intelligently designed tools such as a gun, or a mosquito zapper, antibacterial soap. He uses these tools, intelligently, to kill one bear, and hence, the human is more intelligent. However, if you take 10 bears, then sure, they may be able to kill the 1 human, but that means they are less intelligent, and take more numbers.
He simulates intelligence this way, and he defines a simulation as any environment with applied constraints, and that may include the internet, legal system, your neighbourhood community, etc.
So here's what he says: A system, such as the health care or legal system, will not be shutdown by one person. In fact, it probably won't even be shutdown by 10 people, maybe 100. And hence, the system is vastly more intelligent than a human, intrinsically since we worked in numbers to evolve this system.
I think it's a very interesting way of looking at intelligence. Again, this is all based Mr. Barbalet's assumptions.
One choice quote from the article:
The same reason you get the opinion "The primacy of human intelligence is one of the last and greatest myths of the anthropomorphic divide
Okay, human intelligence may be fuzzy and difficult to objectively measure. But that applies to many things such as CPU speed, Kolmogorov complexity, how complicated a shape is, or how much heat/sound insulation a particular material provides. Even how good a piece of music/art is.
They're tricky, but there's no doubt that exponentially low and high numbers can be given to each of those attributes.
Why OpalCalc is the best Windows calc
Survival is an instinct inherent in all living things, and survival is not measured by one's own ability but by the predators around them. Very few species (if any save humans) will willingly destroy themselves or lie down to die when the opportunity to survive presents itself.
There's something to be said for focusing on low-level survival issues, but one can do more than pontificate about it. As someone who's worked on legged locomotion control, I've made the point that a big part of life is about not falling down and not bumping into stuff. Unless you've got that working, not much else is going to work. Once you've got that working well, the higher level stuff can be added on gradually. This is bottom-up AI, as opposed to top-down AI.
Bottom-up AI is now mainstream, but it was once a radical concept. I went through Stanford at the height of the expert systems boom in the mid-1980s, when abstraction was king and the logicians were running the AI programs. The "AI Winter" followed after they failed.
Rod Brooks made bottom-up AI semi-respectable, but he went off onto a purely reactive AI tangent, with little insect robots. That works, but it doesn't lead to more than insect-level AI. My comment on this was that it was a matter of the planning horizon for movement planning. Purely reactive systems have a planning horizon of zero. That works for insects, because they are small and light, and can just bang feelers into obstacles without harm.
As creatures get bigger and faster, they need a longer planning horizon. The minimum planning horizon for survival is your stopping distance. (This is explicit in autonomous vehicle work.) Bigger and faster animals need better motion planners. This is probably what drove the evolution of the cerebellum, which is most of the brain in the mammals below the primates.
I've had horses for many years; I was on horseback an hour ago. Horses are interesting in this respect because they're big, fast, have very good short-term motion planning, but do little long-term planning. Horse brains have a big cerebellum and a small cortex, which is consistent with horse behavior. This gives a sense of what to aim for in bottom-up AI; good motion control, good vision, defer work on the higher level stuff until we have the low-level stuff nailed.
That's happening. The DARPA Grand Challenge, especially the 2006 season with driving in traffic, forced some work on the beginnings of short term situational awareness. BigDog has some of the low-level motion control working really well, but BigDog isn't yet very good at picking footholds. They're just getting started on situational awareness. There's some good stuff going on in the game community, especially in programs that can play decent football. This problem is starting to crack.
Short-term planning in these areas revolves around making predictions about what's going to happen next. The ability to ask "what-if" questions about movement before trying them improves survival enormously. This kind of planning isn't combinatoric, like traditional AI planning systems. It's more like inverting a simulation to run it as a planner.
I have no idea how we get to "consciousness", but if we can get to horse-level AI, we're well into the mammal range. I encourage people to work on that problem. There's enough compute power to do this stuff now without beating head against wall on CPU time issues. There wasn't in the 1990s when I worked on this.
That's "what" you "said", right?
Brain surgery - it's not rocket science!
The problem with the Loch Ness monster (as with Bigfoot and Yeti) is that, if they really resemble their known species colleagues (lizards and apes), they need at least a thirty or forty year cycle to reproduce. And they will have a lifecycle of around hundred years. And since they've been 'seen' for more than a hundred years, they must have had children, and since there were children, they must have had mates, and since they must have had mates, they must have had fathers, mothers, children, and by-and-large, represent at least one family of at least six members, but much more likely (to keep the gene pool a bit fresh), several tens of members, at the very least. Now you can hide one bigfoot in the hills, and one Nessy in a lake, but thirty ?
Religion is what happens when nature strikes and groupthink goes wrong.
Nothing like twisting the language to force a point. We have different words for "survival" and "intelligence" because they are different things. Redefining one to mean the other does not contribute to the discussion. It may be that Artificial Survival is a better goal for research than Artificial Intelligence - the point could be argued. But this semantic redefinition assumes that argument won, and claims victory in an Orwellian manner by redefining the language to state that victory has been won.
Consciousness is an illusion caused by an excess of self consciousness.