Given the approach, the listed "successes" are no wonder.
This EEG-based stuff typically works on the so-called P300 response, i.e., the fact that 300ms after(!) thinking something you can measure a response in the brain waves, if you just look closely enough.
Unfortunately, that's not only horrible laggish, but also not really precise. For more complex tasks like easy games like train simulators you already end up having positive interpretation in the range of 48% to 52% (so closely centered around guessing ...), with most test subjects not exceeding around 60-65%.
With some few test persons you might be able to get something in the 70-90% range, but from what I've seen over time I have a feeling that they rather get trained (or somewhat train themselves) to produce responses the EEG pickup system can read better rather than the computer getting better at reading their brainwaves.
Using P300-based spellers achieves a best-case net bit rate of around 2 bits/second. So already typing a single character will take in the range of 3 seconds.
I don't see how this will ever be sped up unless we go for deep brain-probing, which probably no human test subject wants to volunteer for (it's used for mice experiments, though, in certain virtual reality experiments studying learning and orientation).