Moravec’s paradox is the hard problems are easy and the easy problems are hard. A computer can beat the world’s greatest chess player at chess, but it cannot beat a spider at getting around. If humans have been working on a problem for a thousand years, you can program a computer to do it. If evolution has been working on a problem for a hundred million years, not so easy.
It turns out that the vast majority of the functional human genome is information processing. A small proportion of the human genome codes for proteins, but most of the important genes, most of what matters, does not code for proteins. It is RNA world data processing, RNA genes, RNA generated primarily to process RNA.
Given that twelve to sixty percent of the human genome is data processing, is software, is programming, that is a lot of information processing – seven hundred megabytes to four gigabytes of software. A lot of this software is instructions on how to build a human being – where and when to express the proteins of which a human is made.
If, however, you have a massive system for processing data, seems likely that the brain is going to use it.
Particular RNA genes are expressed in particular kinds of neurones, often a particular RNA gene being expressed in few hundred or a few thousand very specific neurones in the entire brain, Protein expression is considerably less specific.
Most of the genetic complexity of the brain consists of very large numbers of very specific RNA genes being expressed in very specific neurons. Protein enzymes for editing RNAs are most highly expressed in the brain, and a disproportionate number of RNA genes are expressed only in the brain, and only in very specific neurons in the brain.
The human brain does thirty five times as much RNA editing per unit mass as the mouse brain. The smarter the animal, the more RNA data processing in neurons. Smarter animals not only have bigger brains with more neurons, they have substantially more RNA software expressed and running in each neuron. This is the missing complexity. Humans have about the same number of protein coding genes as a sponge or a flatworm. They have substantially more RNA genes, a large proportion of which are expressed only in quite specific neurons in the brain.
This suggests that neurons process data at the RNA level – that a large part of the evolution towards intelligence occurred in RNA world creating smarter individual free living cells, before cells got smart enough to gang up for attack and defense, and likely before they developed protein synthesis.
If brain data is processed in complex ways in RNA, there is no way that this can be emulated in silicon. Likely we have software that evolved over billions of years, which software is designed to run on RNA molecules in water solution and can only be efficiently run on RNA molecules in water solution.
So, if RNA world data processing, no possibility of emulating the human mind in silicon. Silicon consciousness would have to be built from scratch, rather than by copying existing software, which looks to me like a very hard project..
It appears to far more likely that we’ll hook brains grown in vats to run computers systems than to build brains in silicon. Which means in the future they’ll be 2 types of AIs: Heuristic based smart machines built through trial & error and copying systems humans already use, and systems run by organic brains grown just for the system.
Organic brains are really slow.
“It turns out that the vast majority of the functional human genome is information processing. ”
This is a classic Jimstatement – huge blanket assertion with absolutely no supporting evidence whatsoever. It might be true. I’m not saying it’s not true. But, given that at the very moment prior to my reading the sentence, my understanding of the situation (informed in part by my own reading, in part by several friends and family members who work and/or hold advanced degrees in this or closely related fields) was that the vast majority of the human genome is not understood in the least, I simply will not accept it as true based purely on your word without some sort of evidence to back it up. That is what hyperlinking is for, after all.
the fun thing is, Jim is often right about these Jim statements, because he says stuff that everyone is already thinking, but arranges it in ways that people haven’t really thought to yet.
In response to your complaint, I googled up some supporting evidence. If you read http://www.google.com/url?sa=t&rct=j&q=&esrc=s&source=web&cd=4&cad=rja&uact=8&ved=0CDYQFjAD&url=http%3A%2F%2Fwww.ansto.gov.au%2Fcs%2Fgroups%2Fcorporate%2Fdocuments%2Fstillimage%2Fmdaw%2Fmda4%2F~edisp%2Facs017309.pdf&ei=b537U-yqKYjh8AXR6oGoAg&usg=AFQjCNFMzce8fR8D4OOyQvNctJdhrpopOA&bvm=bv.73612305,d.dGc
You are likely to come around to the same conclusion, or at least the conclusion that the guy who collected this material collected it because he is thinking pretty much what I am thinking.
It’s been known for a while that most information processing takes place within cells in a way that’s probably impossible to simulate with a computer.
Brains grown in vats will not have the same capabilities as brains grown in humans who exist in a real environment. There’s a lot of non-linear learning that goes on that you can’t emulate purposely.
It may not even be human brains when we get around to making them. Lots of animals solve problems better than humans who don’t have the training in solving a particular problem.
I doubt that people will be able to properly emulate the nervous system of a c. elegans in our lifetime. There is too much data processing going on on a cellular level, and in ways we can’t really simulate from first principles.
What Is Thought? (2004) said that the hard part was the software, which evolution had thrown vastly more computational resources at designing than we will ever have available (If memory serves 10^34 creatures have lived and died), and that the discoveries are coded for in the genome for detailed expression in the brain.
http://www.amazon.com/What-Thought-Bradford-Books-Eric/dp/0262524570/ref=tmm_pap_title_0
If the RNA computation operates with a predictable and measurable input-output mapping, it can be approximated by computer. Period.
RNA is infeasible to simulate directly, but it is exceedingly unlikely that the full detail is required. The actual information-theoretic function implemented is likely to be much higher level and much simpler.
We cannot simulate quantum gravity wave function directly, but we can still predict where a projectile will go, because it has predictable high-level behaviour. The same will apply to neurons.
Neurons are different, because they DON’T have predictable high-level behavior. For instance, they can perceive individual quantum-level events.
At what cost?
Quantum effects matter in RNA oxidation. And even if they don’t matter in the sense that we can use a quasi classical approximation, what is the computational cost?
My point is that neurons are unlikely to do complex information processing at the quantum level. They are likely to have a relatively simple probabilistic IO mapping with maybe a Kb of memory at the top level, even if there is all kinds of quantum nonsense going on beneath that.
As an example, if neurons are sensitive to individual quantum events, that is likely to be just thermal noise and high-level rules, rather than actual computing. If you replaced that whole mechanism with a random number generator with similar distribution, would the brain stop working? Unlikely.
To actually construct a device finely enough to do computing at that scale, you need milliKelvins and nanoTeslas, without which you get just noise and high-level rules.
Now it might be the the high level rules follow from some computationally difficult quantum dynamics, and while being the same every time, are not easily packed into a table lookup or approximation. In this case, it still seems that one can replace neurons with cheaper devices that organized the same way or in an easily computable permutation, do the same thing.
So I’m changing my mind here; it is plausible that the RNA function is effectively cryptographically semisecure by being big, expensive and random enough that there does not exist a cheap approximation. However it still seems exceedingly unlikely that the overall system can’t be either hacked or redesigned along the same lines to be much cheaper.
So the RNA thing, under some possibilities, might slow development down until it can be cracked.
They can detect an individual photon.
I suspect that if there was a cheap simulation of neuron functions, nature would have already developed it and be using it.
There was a Russian Jewish gentleman by the name of Efim Liberman who invented the field of bioinformatics. His theory was that most information processing takes place within cells using a cytoskeletal 3d lattice using sound waves for transmission of information.
I don’t know whether they do complex information processing at the quantum level or not, but what makes you think they are unlikely to do information processing at the quantum level?
We have increasing reason to believe the output bears a non trivial relationship to the inputs, in which case there is an unknown information processing mechanism inside the cell.
Redesigned implies a rewrite. Rewriting hundreds of megabytes of the sort of code that natural selection produces sounds hard.
Hard but tractable without understanding intelligence.
DUH!
Hardly any processing at all, at that level.
Test comment (please delete) – my other comments are hitting a “403 – forbidden” wall.
Bug: Any html results in a 403 forbidden.
Not clear why.
I reinstalled the blog at a new location. It worked.
Reinstalled it at this location. Did not work.
Thanks Jim.
Posted without html and it went through.
testing with html
testing with
Sometimes html goes through and sometimes it does not. Cannot figure out what makes the difference.
[…] Moravec’s paradox. […]
Would it be so fucking bad if we were just farmers?
No. Work with a few engineers turned farmers now (and visa versa).
Greatest men I ever knew. Ever.
To bad they’re almost 70.
…and too bad I can’t spell.
fuxorz it
Article in Kurzweilai today regarding alternative shuffling of identical exons critical in species brain differentiation.
http://www.kurzweilai.net/why-youre-smarter-than-a-chicken
Video of white blood cells attacking a parasitic worm. These individual cells seem to appear on the same order of intelligence as an ant or wasp.
https://pbs.twimg.com/tweet_video/B-bAbPxIMAA5atN.mp4
I think this recent article may be of interest.
http://www.nextbigfuture.com/2017/03/computing-exponentially-faster-using-dna.html