Sunday, June 16, 2013

GF 2045: First Sunday Session

Kurzweil is up...  Talking about exponentials and acceleration.  He's kind of sloppy - talking about social media and blogs taking only three years to spread (FB founded in 2004, blogs before 2000) .

Has a graph of microprocessor clock speed that doesn't seem to show the fact that clock speed has plateaud (we've moved to having many cores instead).

Kurzweil is explicit that the singularity metaphor is drawn from black holes (rather than mathematical singularities more abstractly).

He's putting up lots of log graphs.  I have a feeling Eroom's law isn't coming here - he's putting up exponentially declining DNA sequencing cost, but not the fact that overall drug discovery is slowing down.  No discussion of Flynn's law.  There's no discussion of any critic's points made yesterday (eg the fact that neuron's may be a lot more complex than realized).   In general, Kurzweil is a very clever pitchman, but he's highly selective in his choice of facts - watching him, I am not seeing an open-minded person (one characteristic of which is someone who addresses the strongest arguments against their own thesis, and shifts positions when their critics have a point).

Talking about 3D printing.  By 2020 we'll be able to print out clothing (that I'd be willing to countenance, at least for synthetic fabrics).   Will make the world of physical objects like IT (and thus improving exponentially).

Now talking about how the brain works.  Story about he wrote his first paper about how the brain worked at 14-15 and wrote a program based on it.  Don't think I'm going to be able to summarize his brain discussion effectively.

Argues that frontal cortex was enabler for human achievements (this would have to imply that Neanderthals didn't have a frontal cortex - is that right?  May be).

Questions.  Asked about the possibility that individuals neurons are much more complex and the quantum mechanics issues.  "That's wrong, no evidence at all for that".  Very dismissive tone of voice.  He's talking about Roger Penrose and dismissing him like he's obviously an idiot.

Now talking about metaphors (I missed the question).  Clearly gets the importance of metaphors in brain processing (something I learned from Lakoff).

Now a question about how the benefits of this stuff will go to the rich and powerful.  Argues that cellphones prove everything is fine - poor get cellphones .  "Only the rich have these technologies when they don't work".  That's a good line.  Doesn't address the fact that being poor still really sucks, even when they have cellphones.

Argues that open-source is the best way to democratize the availability of these technologies.

Overall, it was very helpful to me to see Kurzweil as I got a better feel for his personality structure.  He's clearly intellectually very brilliant, but I also see a dismissive personality structure where he ignores and dismisses anything that doesn't fit the story he's telling.  That's why his presentations don't address critics' points, or trends that might suggest other things going on, and he quickly turns obviously dismissive and reasons very weakly when questions challenge him.

Theodore Berger

Neuroscience is big science now.  Emphasizing that there are hundreds of people across more than a dozen labs that have contributed to the research he's presenting.

Presentation is about trying to developing brain prostheses.  Trying to develop an artificial partial- hippocampus - the part of the brain that transforms short term memories and turns them into long term memories.  Idea is to fix a partially damaged hippocampus (eg due to stroke) by bypassing it with a set of electrodes and a bunch of microprocessors.  Currently working with rats in an experimental set-up that involves remembering to press one lever rather than another to get water.

They have to do a lot of complex reasoning to tell what one neuron is saying - it's the temporal pattern of neuron's firing that carries information.  Seems to confirm the idea that there's a lot more than 100 ops per second going on in a single synapse.

Where able to demonstrate artificial long term memory formation in rats when a drug is used to knock out the hippocampus (they were able to reproduce the output firing of the hippocampus with electrodes, and show that the rat performance in the experimental setting was almost as good as with without the drug).

They can enhance the strength of memory formation in animals.

Now working in maqaque monkeys.  Hoping to move to humans before too long.

This was a really amazing presentation.

Note: I missed most of the next couple of presentations in order to go have some coffee and food - and there's only so much of this stuff I can swallow in one go - I tend to find it hard to take (eg I was pretty depressed for months after reading the Singularity Is Near).


dr2chase said...

Yeah, you can't just wave off quantum stuff as implausible -- plants do it.

More important than clock rate, is energy consumption per computation. Historically that's been a lovely ride just like Moore's law, and it has continued while clock rates have fallen, but to me it looks somewhat uncertain. Not impossible, but unlike the joy of turning the clock rate knob, hard at every level -- the hardware has to be designed with ECC deployed throughout, all the algorithms have to be redesigned to work in parallel, and most likely everything has to be designed for failure, in the form of checkpoint/restart, and everything has to be aware of energy allocation. I worked on a super-duper-computer project some years back, and one of the limits on the size of the computer (and its general-purpose use) was the ability to efficiently perform checkpoints and restarts; double the machine size, and you double its failure rate, but you also double the time needed to do a checkpoint, and may double the time needed to do a restore (I'm glossing over details here, but the rule generally holds for plain checkpoints or incremental checkpoints) . And there's an energy cost to checkpointing, too.

It would also be interesting to discover if Kurzweil has a plan for getting AI around the 25-year booting process necessary to form the human brain into something with reliable intelligence, judgement, and impulse control -- and no artificial mental illness, either. (Does he have kids? Have they passed through the teenage years?) The idea that we're we're going to replicate something we haven't yet fully debugged, and do it faster and better and without the buggy bits seems just a wee bit like hubris to me.

Greg T. Jeffers said...

Well, Stuart, I, for one, find your criticism and lack of high regard for Kurzweil somewhat comforting.

Keep it coming.

Mr. Sunshine said...

Dr2chase - good points. "25 year boot" especially.

Brian Bowman said...

Kurzweil is a modern day Cargo Cult clergy man preaching the gospel of technological optimism.

"The cults that won Cargoist adherents among the citizens of advanced nations were not always obviously religious. The Type II belief held that great technological breakthroughs would inevitably occur in the near future, and would enable man to continue indefinitely expanding the world's human carrying capacity. This was a mere faith in a faith...

"Technological optimism manifested itself in several pious hopes..."

Catton, William (1980) "Overshoot" University of Illinois Press, pp. 187-195.