Adam Coates has done many more impressive things than that. The "cat detector" was an uninteresting piece of research and not very impressive at all. The image net results were bad and nets trained on one or two gpus were doing much more interesting things.

In Hinton's Coursera class, which was very good, he made a strong point about GPUs being the way to go (at least for now) in building deep learning networks.

Why are they still using GPUs? Maybe hardware companies should start making AI targeted processors instead. Could they boost performance 10 times with a dedicated processor?

GPGPU designs from NVidia are an AI targeted processor. It's the application they're designed for, and used everywhere for.

Until we know what is the best architecture for AI, it's just easier to play with it in software.

GPUs out perform distributed computing currently for deep learning training, but it will be interesting to see how MPI and CUDA will be combined in the future (there are already people working on that, but its still early).

In the presentation you linked by Baidu's Ren Wu on page 6 it says:

           Deep Learning:
    AI Brain   -   World's first RI

What is RI? I haven't seen the acronym before.

Might just be "Research Institute":

'The Institute of Deep Learning (IDL) is the first formally established research institution projected to be Baidu's “Artificial Intelligence Lab”.'

source: http://idl.baidu.com/en/IDL-about.html

Ah yes I think you're right - I guessed the I stood for intelligence and couldn't come up with anything.

I've seen the acronym used for Restricted Intelligence, though I'm not sure that its the intended meaning here.

Reinforcement Learning?

Says RI. Reinforced Intelligence?