From: Luke (wlgriffiths@gmail.com)
Date: Thu Nov 19 2009 - 20:49:26 MST
John K Clark said: "It's true that most programs that call themselves AI
are not very parallel, and that's probably why most AI programs suck."
--> Amen to you sir. I immensely enjoyed that, as well as other writings on
this list. Where do you work, and what do you work on, and how can I
transcend this web programming nonsense I do and work with you?
Mu In Taiwan said: "1. Asserts IBM simulated 1% of a human cortex this
year."
--> Sounds like hogwash to me. Perhaps they simulated a similar number of
neurons, but how could they know they were connected in the correct manner?
If we know one thing about the human brain, it's that it comes
pre-programmed with a lot of information. They also said they simulated a
rat's brain in 2007. Again, how could they know? Plug the computer into a
rat's spine and run it through mazes?
I know that dye techniques have improved for mapping connections in a slice,
but I'd really like to see how, or whether, they achieved a read of ALL the
connections in a brain.
The rest of your assumptions are not bad, though, and I think the final
conclusion can still stand even though point 1 is so weak. You predict
roughly 10 years for a real-time simulation. There's nothing holding
computing power back, so the real question is whether or not a complete
brain scan will be available in 2019 (allowing immensely expensive - i.e.
within $1B, not $1,000).
I give that 50%, with an increase to 95% 15 years out (2024). But ten years
- wow. Maybe I should say 80% by 2019, 99% by 2024.
I always write more than I intend. Yeesh. Good night Midwestern US.
- Luke
On Wed, Nov 18, 2009 at 9:20 PM, Mu In Taiwan <mu.in.taiwan@gmail.com>wrote:
> http://tech.yahoo.com/news/ap/20091118/ap_on_hi_te/us_tec_ibm_brain_mapping
>
> 1. Asserts IBM simulated 1% of a human cortex this year.
> 2. Not using fastest machine in the world.
> 3. Assuming that J.C. is correct with his observation of 'moore's law
> squared' (not improbable given that the graphics industry achieved moore's
> law CUBED for a good many years...)
> 4. That means about 5 years from now we can expect to have machines capable
> of simulating 100% of a human cortex.
> 5. If they're running at 1/100 speed like the other examples, add another 5
> years for real time.
> 6. Maybe sooner - I suspect as IBM gets closer to their goal, money will
> pour in.
>
> funny how these things creep up on you...
>
>
>
>
> On Wed, Nov 18, 2009 at 6:20 PM, John K Clark <johnkclark@fastmail.fm>wrote:
>
>>
>> On Tue, 17 Nov 2009 "J. Andrew Rogers"
>> <andrew@ceruleansystems.com> said:
>>
>>
>> >the Top 500 benchmark (LINPACK) doesn't measure much
>> > of anything useful.
>> > It is an embarrassingly parallel benchmark
>>
>> Nothing to be embarrassed about, most processes in the physical world
>> are parallel and predicting what will happen next is a key element of
>> intelligence. Vision recognition is parallel, and so is search, chess,
>> language translation, and the human brain itself. It's true that most
>> programs that call themselves AI are not very parallel, and that's
>> probably why most AI programs suck.
>>
>> And whatever LINPACK's shortcomings it must have had those same flaws 5
>> years ago, but machines have doubled it's speed every 9 months since
>> then. I find that fact significant.
>>
>> John K Clark
>>
>> --
>> John K Clark
>> johnkclark@fastmail.fm
>>
>> --
>> http://www.fastmail.fm - Choose from over 50 domains or use your own
>>
>>
>
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:01:05 MDT