In machine learning terms, it seems that they can manage supervised learning (given a known task, they can figure out how signals correspond to movements) but not unsupervised learning (given a new user's signals, they can't decode them without knowing what the user is trying to do).
So I can imagine this working if new users first had to complete a short initialization sequence (maybe gamified in some way) but could then use it without errors more common than typos when using a keyboard.
so the setup might be shit, but if the system itself works well, and the expected value of it working well actually bears fruit, then theres still a chance it'll all work out
Think of it like this. If you have the choice between a regular mouse with 3 buttons, and a mouse with 22 buttons that randomly flies off to some direction for 3 seconds every 2 minutes or so, which would you chose as a power user?
Obviously the first, because reliably having 3 buttons is just so much better than a wonky mouse even if it does have 19 more buttons (that you likely wont use anyway).
It's not really a case of them needing to find a new paradigm, the technology needs to get to a place where someone (like the interviewer) cant at the very very least just pick it up and use it. And after that they need to get to a point where it's reliable and consistent. EEG headsets have been trying to get there for some 10-15 years now without much luck, and these guys might have a good funding pitch claiming that "it's different, we are using muscles!" (EEG headsets pick up your facial muscles as well, but lets forget that for a second) and of cause "We're using Machine learning to build a model!" (machine learning has already been employed on EEG headsets naturally, but it's not a problem of analyzing the signals, it's about getting consistent signals from hardware that freaks out if a metal cart is nearby).
I'm not saying these guys can't do this, I'm just saying that nothing they've shown or done puts them apart from the other companies who have spend 10 years failing to provide something other than cool looking demos in the field.
Your many mouse example fails in that learning it doesnt actually lead to significant benefit, so ofc it'll never see converts; but if it did, if it was actually better than the three button mouse, then the initialization cost is overriden by the long-term benefit, and it remains possible for it to succeed.
My point is primarily about the initialization; if this tech's only substantial negative is that it has to be trained to function correctly, and its benefit is substantial, then there remains a path to success. That is, the training step (alone) does not necessarily kill the technology.
I can't think of any good examples off the top of my head
The promise was lofty, but the device was near unusable in practice.
In the demo application you could learn to move the ball around a little bit but it was ridiculously hard.
Nothing like Macross Plus...
This section: "Then he strapped the bracelet on my arm. I had worse luck — the thumb on the computerized hand reflected the motions of my thumb, but the index and pinkie finger didn’t — they remained stiff. Berenzweig had me recalibrate the system by angling my wrist slightly, but to no avail." and this: “It acts like an antenna,” he said, “so it’s susceptible to interference.”
Sounds a lot like my experience. If you can't make a system people can just pick up an use right away, I predict it will just be too much hassle for anyone to want to use. And throwing buzzwords like ML at the problem doesn't seem to change that.