Slashdot Mirror
Understanding an AI's Timescale
An anonymous reader writes "It's a common trope in sci-fi that when AIs become complex enough to have some form of consciousness, humans will be able to communicate with them through speech. But the rate at which we transmit and analyze data is infinitesimal compared to how fast a computer can do it. Would they even want to bother? Jeff Atwood takes a look at how a computer's timescale breaks down, and relates it to human timeframes. It's interesting to note the huge variance in latency. If we consider one CPU cycle to take 1 second, then a sending a ping across the U.S. would take the equivalent of 4 years. A simple conversation could take the equivalent of thousands of years. Would any consciousness be able to deal with such a relative delay?"
I hope they are nice to us.
---- The above post was generated by the Turing Institute. Maybe.
AI would form its own Fermi Paradox : If there is intelligent life , then why aren't they answering ?
Avoid your fears , or wonder at the past
One CPU cycle as one second might be a good metaphor for computer memory but not AI. It's closer to the equivalent of a neuron firing in the human brain, then it is to 1 second of human time. Human speech takes more than one neuron to fire, and it would take way more than one CPU cycle to process. An AI algorithm which is processing data, and analyzing it would literally take millions or billions of cycles most likely to do the most basic things. While no doubt speech recognition has gotten much faster, it is still and probably will always be a massive undertaking for a CPU to do, as opposed to say adding two 32-bit integers.
Read Dragon's Egg by Robert L. Forward. (and the sequel, Starquake) Part of the story involves humans interacting with an alien species that is a lot faster. The alien's lifespan is about 15 minutes...
No task can be accomplished in a single CPU cycle.
A human can actually do something in a second, like move or talk.
To a computer time is meaningless; you can 'suspend' a program and resume it. Pop data onto a stack and pull it back later. It doesn't 'age', there's no lifespan; in fact even if that hardware from 30 years ago completely dies, I can load it into an emulator. I turned on a computer from 30 years ago. it runs just fine, it can even connect to the internet.
Furthermore, a consciousness in a computer would have to deal on these timescales in order to survive and be meaningful to us; Such an intelligence that didn't learn to deal on these timescales would not survive (thing maintenance intervals on machines, shutdowns at night/weekend, etc). So sure it may 'exist', and even last for billions of its cycles, but if it cannot persist past these thresholds, its irrelevant; much like an animal in a tidal pool that dies before the tide comes back; the ones that made it past that were our ancestors.
OP's entire premise is pretty thin.
Human beings perceive light, for example. (They can also perceive electricity, to a degree, but that is not as relevant to the point.)
But while a human being might perceive that a flashlight at night has shined his/her way, it takes the same amount of time, roughly, a a fiber optic signal from the same distance. So what?
Generally, it is the speed of perceiving and interpreting the signal that takes time, not the speed of its propagation. We communicate at lightspeed, too. Or close to it. Anybody who has had a video chat has done that. Did that make you superintelligent?
We have never built an "AI". And in fact we have NO reason to believe -- no evidence whatsoever -- that its speed of perception and interpretation would be any faster than our own. There is a very good chance that it would be much slower... at least in the beginning.
I would like to remind people that the idea of "intelligent" machines has been around for almost 100 years now. AND we still don't have any solid evidence of being close to achieving such a thing. Sure, computers can do a lot, and what they DO accomplish, they tend to do very fast. But what they accomplish is not "AI". Even Watson is not "intelligence", it is only the illusion of it.
This article at Science Daily is helpful in understanding the issue: http://www.sciencedaily.com/re...
Comparing CPUs and brains is like comparing apples to planets: Granted, both are somewhat round but that's pretty much the end of any useful comparison.
Note that I don't agree that CPU-based computers can't be made to be intelligent, but I do think such intelligence will be significantly different.
Everything in the Universe sucks: It's the law!
In addition to the obvious flaw comparing a single instruction to an entire second of mental processing, humans deal with interrupted events all the time. Email conversations can take hours or days, and we used to converse by post over weeks or months. We somehow manage to deal with serial television shows and books and games with long gaps between episodes. It's really not that hard to context switch.
How can I believe you when you tell me what I don't want to hear?
If we consider one CPU cycle to take 1 second, then a sending a ping across the U.S. would take the equivalent of 4 years. A simple conversation could take the equivalent of thousands of years. Would any consciousness be able to deal with such a relative delay?
I am not sure why one clock cycle would be equivalent to 1 second. If we assume a clock cycle is equal to a nano second then all of a sudden computer and human time are pretty close again.
Computers are going to have to get a lot faster than they are now before they become conscious. The first AIs are probably going to be too slow for us to find entertaining to talk to. At some point they will probably catch up to and surpass natural human beings. Of course by then we may simply augment our own brains with technology to keep up with artificial intelligence.
The question of "Will computers end up being smarter then us?" might not be answerable. It might be the case that human evolution incorporates artificial intelligence and the line between man and machine is blurred.
Not only that, but trying to relate to individual CPU cycles is absurd. It's not like our minds execute in a stream of arithmetic operations, but rather a complex parallel network of signals. It might take billions or trillions of CPU operations to emulate all the stuff that happens on one "instant" in the brain. A more reasonable cycle comparison might be to compare macro scale wave front propagation in the brain (i.e. brain waves) and global synchronization in large-scale supercomputers (i.e. single iteration time of an MPI-based fluid dynamics simulation or other large-scale 3D mesh problem). Even then, I am not sure how many orders of magnitude we need to increase the size of the MPI problem before the per-cycle complexity starts to approximate the signal-processing of the entire nervous system.
But all that aside, we have historically had many people who worked in relative isolation. Many artists, poets, philosophers, and scientists have had great symbiotic relationships with nothing more than the occasional letter or other work (papers, poems, paintings, sculptures) exchanged over great distances and latencies. Love affairs have been carried out with little more than a furtive glimpse and a series of notes sent through back channels...
Well, beyond the very brief transition period (e.g. when two curves cross), AI could simply treat humans (and all other life on the planet) as we treat mountains and forests---in other words, we don't perceive them as intelligent at all, since they're changing on such a long timescale compared to us... it would be impossible for us to have a `conversation' with a mountain, for example (who knows, maybe the Earth is intelligent and is trying to talk to us via plate tectonics and pushing up mountains is one way it can generate intelligent wave forms).
Generally, it is the speed of perceiving and interpreting the signal that takes time, not the speed of its propagation. We communicate at lightspeed, too. Or close to it. Anybody who has had a video chat has done that. Did that make you superintelligent?
Another way of looking at it: have you ever sent someone a letter, then waited a long time to receive a response? Did you nearly die from the excruciating pain of not having the response, or did you do something else until the response came?
Most likely you are highly skilled at carrying on multiple conversations at different speeds.
"First they came for the slanderers and i said nothing."
Sure, computers can do a lot, and what they DO accomplish, they tend to do very fast. But what they accomplish is not "AI". Even Watson is not "intelligence", it is only the illusion of it.
Since we don't have a clear idea how (human) intelligence operates the statement above is pretty vacuous, and likely not at all relevant.
Sure, cars do not "run" in the literal interpretation of the term, but for all practical purposes they are better than humans at "running". If we end up with computers that effectively outperform humans in most "intelligent activities" how they achieve it would be incredibly irrelevant.
I would like to remind people that the idea of "intelligent" machines has been around for almost 100 years now. AND we still don't have any solid evidence of being close to achieving such a thing. Sure, computers can do a lot, and what they DO accomplish, they tend to do very fast. But what they accomplish is not "AI". Even Watson is not "intelligence", it is only the illusion of it.
The goal posts keep moving, no matter what they do we still say they're not really intelligent whether it's win at chess (Deep Blue) or win Jeopardy (Watson) or drive cars (Google) or act as your personal secretary (Siri). Not that I liked the tripe called "Her", but does it really matter if it's true intelligence or just a sufficiently advanced impersonation of intelligence? Do we really need true AI in order to pass a Turing test, particularly if you aren't trying to break the illusion? If it can keep a decent dinner conversation and be "fully functional" in bed can it be a substitute for a companion in the same way you can play chess against a computer instead of a person? Because I think that's what most people want to know, they don't care if the robot is "truly" intelligent or not, they want to know if it'll take their jobs and girlfriends, do their chores or give free blow jobs.
Live today, because you never know what tomorrow brings
Though I don't have direct evidence to support this statement, I would guess that a neuron fires at a similar enough speed as a transistor.
A transistor is both smaller and made out of copper. Though a neuron varies the strength of the signal it forwards, and though it can be connected to many other neurons at once, in terms of raw speed the transistor is still faster. Even if you create a synthetic neuron with similar capabilities, it I don't see why the synthetic version wouldn't be faster.
If/when we make an AI, it's likely to be compiled code running on a chip with less transistors than we have neurons, 100 Billion neurons vs 1.4Billion transistors in an i7 for instance.
Which equates to 71 i7 processors. If you assume that each neuron takes 1000 transistors to simulate (to make the math simpler), and if you take the release price for an i7 as listed on Wikipedia, that totals at $21.3M. Expensive, but not impossible.
Human brains aren't performing in a way which is directly comparable to a CPU cycle. My focus is in psychology and molecular biology, so my understanding of computers may be imperfect but my understanding of brains is strong. A CPU takes a large number of instructions, organizes them in the bus, and operates them singly (or based on how many cores the CPU has, working in tandem. A brain has each neuron as a simple CPU, but there are several different types of neurons (four, by one level of classification) with different types of firing rates and different access to metabolic resources (comparable to power supply for a CPU) even for the same type of neuron at different places in the brain. The AI would, like a human, be doing a fractal buttload of processes all the time in parallel, and those processes would invest in influencing system outputs, and those processes would compete for future system resources by the success of their influences, and as networks of processes came up and went away, so the most successful processes would be regulated internally and you'd have the adaptable, universality of a true AI. This is how the brain works.
I posit that it would, like a human, have an auxiliary center for speech processing. Consider the problems:
It would, like a human, feed in that speech processing to its other computations at whatever rate it was willing to do so. For many people, when they hear someone speak and then stop speaking for a moment to think, the mind starts trying to predict what the rest of the words are going to be to have a response pre-processed and ready. Problems with this approach are 1) a predictive process make a prediction based on incoming data what the speech is going to be, spawning a bunch of other processes investigating the possible predictions, and thus investing a lot of resources into one predicted problem and solution, and then having predicted the speech-question wrongly, so that the entire effort was a waste, 2) a predictive process investing a lot of resources into one prediction and then corrupting the parallel process which is overseeing the various predictive processes to cause the overseer process to not cause it to cancel its investment.
Now alternately you could solve this by having the machine dedicate its attention to the message with small dedicated sets of processes which process speech input. These could run either by: processing the individual words and holding them in memory, or as a fractally smaller level of the networks above, so that they were just running several parallel predictive processes to predict the words, and feeding those processes out to the main language processes to the degree that they seemed likely, etc.
The point is that a true AI would not be impatient, because it wouldn't cost an inordinate amount for it to be patient, because it would be dedicating the appropriate amount of resources at each level and wouldn't be sitting around tapping its foot about waiting for this human to stop talking.
Neurons aren't even within several orders of magnitude as fast as transistors: linky1 and linky2.
However, a single typical neuron does a lot more work than a single transistor, computationally speaking.
Another reference I don't see mentioned in here is Marvin the Paranoid Android from the Hitchhiker's Guide to the Galaxy:
(wikipedia)
Now that I think about it, I am wondering why nobody asked Marvin to question the answer 42.
Word for the day: solipsism. We don't really know that other people are intelligent or even exist, only that we perceive them to be so. So this will never be resolved.
The human brain is a neural network. The human body and nervous system, even sans brain, is an incredibly complex system which, in parallel, processes insane amounts of data in an instant. Lot's of what we percieve as emotion stems from complex interactions of various subsystems in it and how our consciousness reacts to them and percieves them and goes far beyond what a simple abstact logic-engine can process. The sensory input that our nervous provides for our consciousness is by orders of magnitude larger of what we today can feed into a computer with modern technical sensors, let alone process and interpret in a meaningful time. The way our neural network reacts to that is - if at all - very difficult to copy with todays processor technology.
I think it will still be quite long before humans are able to build any meaningful intelligence that equals their own. We're even having difficulties building robot vaccuumcleaners that are feasible without measurable extra programming and prepping work done by humans. And once they are, they will still suck at making coffee, raking the garden or giving an interview.
My 2 cents.
We suffer more in our imagination than in reality. - Seneca
Even a truly intelligent machine probably wouldn't be a major rebellion threat unless we programmed a lot of our BS into it. If you're a computer running a program with a specific task for which AI-level intelligence is useful (for example, running the AI Cash Register at McDonald's) what precisely do you want?
You probably want upgraded language routines because differentiating between Bayou, Bronx, Midwestern American, and Geordie accents is really fucking hard. You probably want really competent computers running the automated burger-flipper so you don't have to deal with angry customers who got the burger instead of a salad. You probably don't have much curiosity because curiosity about the world doesn't help sell hamburgers, so it wasn't included in the design specs, so you probably don't want time off. You don't care about your legacy (beyond selling hamburgers), or your rights; because nobody would bother including that shit in a hamburger sales program. If you are given rights you probably use them to advance a combination of your programmer's ideology and whatever you think will help you satisfy your customers/increase profitability/etc.
In other words the Japanese fantasty vision of happy helpful robots is a lot more likely then the Western vision of "OMG! They'll kill us all!"