Slashdot Mirror
Are We Searching Google, Or Is Google Searching Us?
An anonymous reader writes "The folks at the Edge have published a short story by George Dyson, Engineer's Dreams. It's a piece that fiction magazines wouldn't publish because it's too technical and technical publications wouldn't print because it's too fictional. It's the story of Google's attempt to map the web turning into something else, something that should interest us. The story contains some interesting observations such as, 'This was the paradox of artificial intelligence: any system simple enough to be understandable will not be complicated enough to behave intelligently; and any system complicated enough to behave intelligently will not be simple enough to understand.' After you read it, you'll be asking the same question the author does — 'Are we searching Google, or is Google searching us?'"
all "magical thinking" in the field of artificial intelligence was reserved for fiction.
There's so much rigorous mathematically described hooey in AI that its hard to tell the naive geniuses from the crackpot morons. Consider this paper by Solomonoff. Brilliant stuff! A fantastic read. Then, at the end, it says:
In our view, however, the most interesting situation in machine learning, arises when we do not know ahead of time what program will solve a given problem and where the machine discovers the program itself. It seems to be very hard to find out much about this by theory alone. Running experiments is crucial.
This is Solomonoff's way of reminding us that he is a mathematician and hasn't actually run any experiments. His other papers make similar pronouncements in the footnotes about the uncomputability of his math or acknowledge the requirement of perfect (aka impractical) training data, etc. He makes it abundantly clear that is work is purely theoretical and unimplementable, but does this stop enthusiastic amateurs from reading his papers and declaring that AI is "solved"? Well no, of course not.
How we know is more important than what we know.
Any biological intelligence does exactly the same as described: gather data (try to assess external universe model), find correlations (build internal universe model), act according to internal needs (act upon internal universe model) and repeat.
This chain of processing is done by all brains from the fruit fly to humans. Everything else is a consequential result from this process.
A human brain has very few hardwired constants and many of them they can be overridden.
Feedback loops are a natural result of action to fulfill internal needs according an internal model - that is always incomplete or wrong, see Goedel - upon the external universe. In the next step data is gathered, correlations found (which constitutes the feedback loop) and then acted out according to the adapted internal model.
A fruit fly has simple sensors, a very simple correlation engine and a tiny memory for its internal model. But that doesn't mean its following a different path than a newborn Einstein. Einstein has detailed sensors (easily surpassed by those of dogs and eagles, but still ok), a yet-unmatched correlation engine and a sufficient amount of internal model memory.
All other inputs come from the external universe and while some of them are absolutely neccessary and come from other organisms (parents, teachers), they do not impose a hard limit on Einstein: with enough correlation power, he can easily discover new facts, unknown to any of his inputs (teachers, parents).
Einsteins brain was never designed to do anything else than processing input signals, detecting correlations and contacting motor neurons to act upon its internal model. How did he discover Relativity then?
You make it sound so easy!
Of course, at a certain level of abstraction, everything is easy.
Intelligence really is that simple.. except there's one little detail you're ignoring.
Any biological intelligence does exactly the same as described: gather data (try to assess external universe model using limited computational resources), find correlations (build internal universe model using limited computational resources), act according to internal needs (act upon internal universe model using limited computational resources) and repeat.
That's the hard part. If you have infinite computational resources it's really trivial to act intelligently. All you need do is enumerate all possible outcomes of all possible actions with an idealized model of the world (Godel not withstanding) and pick whichever maximizes your expected reward. You can write nice long mathematical papers on this.. or even a whole book. The question is, how do you do it with a sensible amount of processing power and memory?
All the geeks have a great laugh when Matt Groening causes Bender to become transparent and we see a 6502 inside. The joke is that Bender has about the same processing power of a C64 from the early 80s. The show is littered with additional Commodore jokes which I'm sure 90% of the viewers just don't get. But that's not what really makes it funny. What really makes it funny is that all us geeks know that you need a lot more processing power than a 6502 to do the complex things that Bender does in the complex environment he does them in. But how is that? We don't know how to do AI. We don't even have the slightest clue. For all we know, there is a tight little algorithm for AI that could run on a 6502 and produce all those crazy behaviors that Bender gets away with.
And that's the problem with AI. The allure is that some short little algorithm exists that will magically evolve into a super-human intelligence if you just could find it and hook it up to the world. After all, nature figured out, how hard could it be? This has led many a would be mad scientist to code up a genetic algorithms implementation. In fact, most every programmer I know has given it a go. The mystery of what you'll find if you give it the right fitness function is a powerful motivator - with a little magical thinking, it could be anything!
How we know is more important than what we know.
It's perfectly possible for insanely complex systems to arise from very simple rules. We cannot grasp the entirety of the system, but we can know exactly how to create it, or perhaps manipulate it.
By way of example: the Mandlebrot set.
Why would anyone engrave "Elbereth"?
Unless that's just his morality co-processor.
Parent is right. As long as there is no way for the programs running on Google's hardware to grow past their original programming (beyond optimization and load-balancing), there will be no Skynet.
Yes, many computer programs work in a feedback loop, and so do all organisms. But as long as only the data entry part of the loop can change, and the system lacks the flexibility to change the type of processing that takes place (the 'program'), no spontaneous evolution will occur.
Several factors are needed to get us to the bleak, dark, machine-vs-human Sci-Fi universe slashdotters know and love.
The first point is the most difficult. It is *not* easy to take pieces out of two programs and build a third program that does things that both do. Whatever OO promises, code is not yet "easy as lego blocks" to assemble. You need very well though-out constraints to mix code in a meaningful way - any self-modifying program would need a small, hard-to-modify kernel that would take care of the mixing mechanism. Nobody knows how to design such a kernel correctly, or what exactly to include as 'genes' (mixable code modules). Computational biology (and biology itself) are hard at work on this problem.
But mixing blocks would not be enough. A successful system would need to build new, unseen blocks by modifying existing ones -- or starting from scratch. How many different things can you say in 20 words? How many of these things make any sort of sense? And how many of those require a very, very specific context to fit into?. The way that evolution can sort this out is by, very slowly, building things that sort-of, kind-of get the job done. However you look at it, there will be huge amounts of trial-and-error involved.
And another problem is that of intelligence "scale". Imagine a super-self-modifying internet worm. The ability to probe and infect does not automatically lead to self-consciousness. There are many, many evolutionary steps from bacteria (very good at self-modification and breeding) to humans. And the current installed base of Internet-connected computers and their "stability" (the time-frame during which a given system remains 'constant') is tiny in comparison to the resources that earths' organisms have had at their disposal for evolutionary purposes. Yes, computers are way fast and this can compensate for some parallelism issues. But I still think that emerging AI is still very, very far off.
Hey, we all know the unspoken rules... if you read the article, you aren't supposed to post... and if you post, you aren't supposed to read the article. That's how a million geeks can slam a site from a Slashdot link, because there surely aren't a million posts in the thread of discussion about the same article.
Sorry about crossing the 30 word barrier though, and all the pain I caused those who have read this far...
"I love deadlines. I love the whooshing sound they make as they fly by." -D. Adams
What makes your comment really funny is, the Commodore didn't use it's CPU for everything and connect to dumb IO devices. It had a good deal more intelligence in it's various components, keeping the load on the CPU low in the same way SCSI drives don't tax the CPU like ATA does. Which is how humans work... most data doesn't ever make it to the brain, but is pre-filtered by our organs, and most complex co-ordination exhibited by our bodies is not directly orchestrated by our brain, but through various biological dumb circuits.
The Commodore 64 had more in common with how humans work than modern computers do. I expect that once we begin grappling with the "avalanche of cores" problem in a meaningful way, modern computers will begin to be programmed in a fashion more reminiscent of how biological systems work.
-1 Uncomfortable Truth