I can key morse code at 40wpm with two muscles. With one hand I can chord at 120wpm. On a stenowriter I can transcribe about as quickly as most people can read - 250wpm.

I've invested an extraordinary amount of effort into improving the speed at which I can interface with a computer; I think the practical limit is about 300 baud, half-duplex.

Of course, we're trying to establish an interface with a bafflingly complex lump of grey meat, but are we really daunted by the idea of outpacing a V.21 modem?

Your judgment that present technology is inadequate is based on the assumption that computers need to learn to read the human thoughts.

What about the inverse, that the humans learn how to think in a way that a computer understands? That will be much easier, as humans learn much better than computers, and also much safer - I will have complete control over which of my thoughts the computer can detect and interpret.

The human learning to adapt to the machine has been the way EEG-based brain computer interfaces have been made for a couple of decades. Using machine learning to adapt the machine to the human is a much more recent development.

It is possible today to make EEG controlled devices. They typically differentiate between a small number of real or imagined movements in the user. This is awesome, because it can allow severely paralyzed people to communicate, control a wheelchair. etc. Nevertheless, the algorithms used to do this are perfectly useless when it comes to distinguishing whatever words the user is internally vocalizing.

The keyboard is not very good at determining which words I'm internally vocalizing either, still seems to work. The point I'm trying to convey is that maybe we can learn to transmit words using some form of brain reader, but that measures something else than vocalizing.

Doesn't have to be "brainwaves". The brain has a few outputs that can be highjacked (e.g. a computer with a neural interface that appears to be another muscle in the body). I don't know whether the bandwidth of these outputs is sufficient for interesting communication; we've evolved to take in far more data than we produce.

Edit It seems that more direct methods of neural interface are already plausible: http://www.technologyreview.com/biomedicine/37873/

It doesn't actually need to be noninvasive. If an invasive procedure is useful enough and can be made safe, eventually it will be ubiquitous.

The problem with invasive is upgrading.