/You would probably need a very powerful desktop (or a laptop with a desktop processor, >$4,000) with the best imaginable specifications to pull this off. So, it is of extremely limited use for the disabled individual./
Not necessarily. They have an extremely small-dimensional input, and have a pretty straightforward problem. The fact that they can train a good model with just a few hundred sentences suggests that it's an easy problem...
For other context, I do neural speech generation on phones for my day job, using an RNN that makes 4000 inferences per second. It works fine on a single thread with most phones produced in the last few years. Another helpful point of context might be 'swipe'-style phone keyboards, which are often RNN based, and turn paths into words.
The focus on giant-model work I think hides how effective small models can be, and how much progress has been made on making models run faster in limited resource environments.
(Do you have a reference on the divorce rate? Not sure I understand the causal link there...)
My work in AI revolves around biophysical signals. I do not use AI generally in this case but I use a 19 channel EEG (using an EEG cap) for experiments in closed-loop controls. It requires a lot of RAM (ideally 32 GB if not more) to prevent latency.
Thanks for the article; it was a super interesting read. My takeaway was that poking brains can lead to major personality changes, which can lead to divorce and other bad outcomes. (Just hearing 'leads to divorce' made me wonder if it was due to previously non-communicative people expressing themselves in ways they couldn't before... but sounds more like a 'sometimes you get subtle kinds of brain damage' problem.)
Not necessarily. They have an extremely small-dimensional input, and have a pretty straightforward problem. The fact that they can train a good model with just a few hundred sentences suggests that it's an easy problem...
For other context, I do neural speech generation on phones for my day job, using an RNN that makes 4000 inferences per second. It works fine on a single thread with most phones produced in the last few years. Another helpful point of context might be 'swipe'-style phone keyboards, which are often RNN based, and turn paths into words.
The focus on giant-model work I think hides how effective small models can be, and how much progress has been made on making models run faster in limited resource environments.
(Do you have a reference on the divorce rate? Not sure I understand the causal link there...)