Hacker News Reader

Comment by swombat | Hacker News Reader

Be sure to read Kurzweil's well-argued response too: http://www.technologyreview.com/view/425818/kurzweil-respond...

(imho, more convincing than Allen's).

One of Kurzweil's arguments in his response is "...the design of the brain (like the rest of the body) is contained in the genome. And while the translation of the genome into a brain is not straightforward, the brain cannot have more design information than the genome"

However the brain's design information is the genome, the laws of physics and a universe.

You only have to look at protein folding, and the complexity of the resulting molecule and how that interacts with the rest of the world, so see how much complexity is there outside of the genome - and it's that complex end result that you have to simulate or replicate.

He goes onto say that "..the amount of design information in the genome is about 50 million bytes, roughly half of which pertains to the brain. That’s not simple, but it is a level of complexity we can deal with and represents less complexity than many software systems in the modern world."

As above, this is greatly simplifying the end result that you're trying to replicate - the genome is just the software that's running on the OS/hardware of the universe. To replicate that virtually we'd also need a virtual universe.

The virtual brain could use the real world, just as the real brain is. Just like the Google cars and Watson are doing.

The main complication of protein folding is how the amino acid chain interacts with itself, not its interactions with an external environment.

That proves the point. To be able to fold a simple protein, you have to be capable of simulating the very challenging simulation of "real world physics at an atomic scale".

To simulate a brain, you get to do that for a few hundred trillion proteins that make up a brain...and all their interactions with all the other trillion proteins.

The complexity doesn't come from the proteins so much as the physics that govern the proteins.

It's certainly longer and more flowery, but a lot of it upon even the shortest reflection rings hollow:

  And while the translation of the genome into a brain is   
  not straightforward, the brain cannot have more design 
  information than the genome

How could a Game of Life grid possibly have more state or require more storage than the ruleset for generating it! Gasp! This is just sloppy on his part.

  We do need to understand in detail how individual types 
  of neurons work, and then gather information about how 
  functional modules are connected. The functional methods 
  that are derived from this type of analysis can then 
  guide the development of intelligent systems.

I suspect this reflection would be about as useful in developing new intelligent systems as meditation on the semantics of various MIPS instructions would be for developing new processor implementations. He's looking at the wrong level.

  The Google self-driving cars (which have driven over 
  140,000 miles through California cities and towns) learn 
  from their own driving experience as well as from Google 
  cars driven by human drivers.

A vacuous statement--to what degree does a Kalman filter "learn" from its state estimates of speed or similar? His statement may well be technically accurate, but its explanatory power is wanting.

All the buzzwords and nerd feelgood talking points are there, but I'm very put off by the random anecdotes and clumsy assertions.

How could a Game of Life grid possibly have more state or require more storage than the ruleset for generating it! Gasp!

It can't have more state. The initial conditions have just as much state as any later generation. The initial state is {off, off, off, off, off, off, ... }. That's just as many bits as any later stage.

What you really mean is that the initial state is easily compressible. (Run-length encoding can express an entire empty grid in two numbers as length, value.) And that's pretty much what Kurzweil means, actually the Game of Life is a great analogy here. The genome is a compressed data set, which doesn't express or exhibit anything useful in itself, but can be decompressed using external inputs of organic matter and energy to procedurally generate the neurons that will form the intelligence.

That encoding is only viable if you restrict the game to a finite grid.

While true, the infinite case is of no interest here: a brain is finite. (Many infinite grids require an infinite amount of information to store, and clearly many of our assumptions break down there.)

The initial conditions in the game of life isn't normally all-off, because then every generation would be all-off.

In any case, the amount of information in the game of life is merely (initial state, number of iterations) which is certainly finite for a finite board.

There's a lot of good in his response, but I do take issue with two of the concepts:

    > How do we get on the order of 100 trillion connections 
    > in the brain from only tens of millions of bytes of 
    > design information? Obviously, the answer is through 
    > redundancy.

No. That's called algorithmically generated complexity. It's where the function generates data that's more complex than the function itself, because the function (DNA in this case) is fitted explicitly to match that output. This isn't a complicated idea; stop avoiding the issue. The brain is complicated.

    > Combining human-level pattern recognition with the inherent 
    > speed and accuracy of computers will be very powerful.

No kidding. If we had 'human level pattern recognition' this whole thing would be a lot easier wouldn't it? ...but that doesn't just magically appear from no where.

The reality is that most progress follows a log normal curve (http://en.wikipedia.org/wiki/File:Lognormal_distribution_CDF...); it starts linear, and ends flat. Some times it starts flat, looks like its going exponential and then goes flat... but it always ends up flat.

The trick is, does the singularity happen after we hit the flat curve (which means it's going to be some nebulous time in the future that may never arrive) or does it happen while we're still on the 'exponential acceleration' curve (which means it may happen in-this-lifetime).

I don't see us on an exponential increase path for cpu power, cost reduction or algorithmic strength at the moment.

If anything I see us as passed the main hump of growth and into the 'slowdown' phase for all of these.

So, I'm dubious of this rebuttal, which is basically endorsing the stupid 'growth for ever! exponential growth! yay! growth!' idea. -_-

If you use a Kolmogorov measure of complexity, the complexity of a function and/or its output is defined to be the description of that function using some universal language. So a function cannot, by definition, generate data that is more complex than the function itself.

Pi can be described by a relatively small formula/function. It can generate very long sequences of seemingly random digits. The complexity in all of those infinity digits is still just that of the simplest formula that generates Pi.

Poor phrasing on my part.

Pi is a great example.

The full length of Pi cannot be compressed using traditional means into the same number of bytes that the algorithm that generates Pi can exist in.

The point is that a function can generate a dataset that is complex, without redundant information in the data. A dataset that cannot be reduced to a smaller dataset.

The idea that a complex function can only generate a simple data set with 'high redundancy' is obviously nonsense.

My point being that the OP is suggesting the DNA that is the functional generator for the brain generates highly redundant physical structures because there is a limited amount of information in the functional generator (DNA).

While the DNA may generate redundant or repeated structures, it does so (if it does) because of the function. Not because of the information density of the DNA.

As your example of Pi shows, a trivial information density can generate complex datasets without repetition.

I don't think your argument about "traditional means of compression" holds.

The existence of redundant data can be proven by the ability to store the same information with less data.

Pi's information density is in fact no larger than the small function that generates Pi's digits -- even if traditional compressors fail to identify it.

  A classical example is the laws of thermodynamics [...]

I'm not sure what is argument is here. The laws of thermodynamics are scientific laws because we can test them and repeat them in a laboratory. This is untrue of Kurzweil's LOAR as far as I understand the requirements to deem something a scientific law [1].

[1] http://en.wikipedia.org/wiki/Scientific_law

Reply on news.ycombinator.com