BlindType has a new Demo video available that looks simply jaw-dropping! Ray, I'd like to see you beat the person's speed in THIS video! There is also a new hands-on article as well:
Exclusive: I Used BlindType, Virtual Keyboard of the Future | Singularity Hub
YouTube - BlindType - Demo2
I want to think out loud for a second.
I kind of work in AI--research, not product development--and I have a basic idea of the problem being solved here. Like Google's speech recognition engine, BlindType is using contextual information like statistical associations between words, and between letters within words (e.g., how often does the letter sequence "xqa" occur in English...how often does "pra" occur, etc.) to assign probabilities (or confidence estimates) to letters (e.g., hidden Markov Models, Bayesian inference, etc.).
In principle, this is how autospell works: you type "blant" and the computer says "'blant' not found in dictionary...let's assume it's off by one letter...which 1-letter substitutions create a real word?...'bland'? 'blunt'?"
But a good AI designer will think up better heuristics, and a good developer will collect mountains of user-data to feed the statistical estimates. For example, which letters are easy to reach? which ones are easy to miss?
Bottom line? I have two different reactions at this point:
(1) This program is amazing, and will win dozens of awards for solving the touch-type recognition problem.
Or
(2) What we're seeing is a very carefully crafted demonstration that does not reflect real-world use.
Call me a skeptic. ATM I'm leaning toward (2).
-Matt