Got a Roomba.

It was bound to be a disappointment – It would get stuck, the battery would be insufficient and it would do a crummy job – but I bought one anyway… and it is incredible.
We have had it for 2 months now and it does just exactly what it says on the tin. To be fair we do have a flat that is basically designed with the robot in mind. Open plan, wood floors, a small amount of furniture with decent floor clearance – it can even get under the sofa.

(more after the jump…)

It is like a dishwasher: not really a necessity but it would be a huge pain to give up on now we are used to it. Imagine shutting the door on a grubby flat and coming back to find the robot charging in its hutch, a spotless floor testament to the hard work it just put in. Dust motes, re-dried corn-flakes from the kids breakfasts (and Lego) are no problem (remember to fish the Lego out when emptying the hopper).
It seems to mostly sweep up the dirt, the sucking is pretty week, but it maximises its vacuum power by sucking through a thin slit. This works fine on our wood floor, but I can’t imagine it sucking dust-mites out of a deep shag carpet.

It has a low IQ, it ricochets off the furniture to get the job done. This is the equivalent of my generation’s grandparents’ first TV. It is a milestone, but will look pretty dim-witted pretty quickly.

Eric the half a mouse.

1/2 a mouse a philosophically…

When I first read about a simulation of a mouse brain running at 10% C, I dismissed it immediately as too far fetched – We Can’t Do That Yet. Reading past the headlines the actual simulation turned out to be significantly less far fetched, but only slightly less portentous. According to the article, standard Moore’s Law improvements will allow hardware of the same cost to run a human mind in real-time in 20 years. Or 1000 human minds in 30 years, 10,000,000 human minds in 50 years.

(more after the jump…)

Sure at the moment this is just a statistical model with no behaviour. The regions of this simulated brain have the neural topology and firing patterns of a real mouse brain, but the software it is running is just noise.
True AI equivalence will require the software too, and this is a non trivial problem. But even if we do it by brute force – start with a population of 1,000 agents each with a human brain’s allocation of processing resources, run them at 1000 times real-time and let various fitness functions optimize the various brain regions – I can’t imagine it taking more than a couple of decades to crack.

I personally think we will have some very clever software ready to run on those simulated brains by the time they arrive on our lab benches. Neuronal implementations of Jeff Hawkins HTM hierarchies, e.g. My MSc thesis, would already look pretty smart utilising all those synapses.

Mouse brain simulated on computer