3 ms·

An AI-driven assistant doesn't have to necessarily run on someone else's servers and be under their control to be useful. In fact, if I teach my assistant patt

An AI-driven assistant doesn't have to necessarily run on someone else's servers and be under their control to be useful. In fact, if I teach my assistant patterns like how when I say "David" I mean David O, then I don't see a) why this type of learning can't be fully localized, or b) how phoning home with this kind of data is going to be of any particular use to anybody else, anyway. To put it another way, what's wrong with letting my Siri and your Siri just be two different "people" instead of feeding back into a single, networked hivemind?

icebraining 10y ago
I'm not involved in AI projects, but my understanding is that current state-of-the-art work, including in digital assistance components like speech recognition, the approach involves gathering large amounts of data and throwing it into deep learning algorithms. This is why Google had free services like Voice, which served as sources for their modelling data.
- flukus 10y ago
  IMO this is a bunch of BS by companies that want to control and centralize computing. Years ago we had products like Dragon naturally speaking that could do full voice dictation reasonably accurately. They had limited digital assistants too (which is a subset of full dictation, much less valid inputs), though not much more limited than today. These ran on Pentium 1's. The downside to this was the training time, but in practice I found this to be a strength. It was trained to my voice, not whatever generic profile Google stocks me into. I got better results nearly two decades ago (got a Dragon demo in a carnival bag) than I get with Google voice today.
tmsldd 10y ago
> what's wrong with letting my Siri and your Siri just be two different "people" instead of feeding back into a single, networked hivemind? Absolutely nothing wrong.. that would be great, actually. Question is how to do that without data and GPUs.
- rwallace 10y ago
  It's perfectly okay to say, if you want to get the most out of this program, you need a big GPU rig. The price of such should come down over time. That leaves the question of how to get the data. If you make it optional, will enough people opt in? Is it an acceptable compromise to make it opt-out?
  - tmsldd 10y ago
    Your idea of opt-{in, out} is a good one. It would integrate enough data with time, I think. "Open Data Foundation" + "World Wide Distributed Processing" would really make a revolution here. We need break urgently the monopoly over data.

icebraining 10y ago

I'm not involved in AI projects, but my understanding is that current state-of-the-art work, including in digital assistance components like speech recognition, the approach involves gathering large amounts of data and throwing it into deep learning algorithms. This is why Google had free services like Voice, which served as sources for their modelling data.

flukus 10y ago
IMO this is a bunch of BS by companies that want to control and centralize computing. Years ago we had products like Dragon naturally speaking that could do full voice dictation reasonably accurately. They had limited digital assistants too (which is a subset of full dictation, much less valid inputs), though not much more limited than today. These ran on Pentium 1's. The downside to this was the training time, but in practice I found this to be a strength. It was trained to my voice, not whatever generic profile Google stocks me into. I got better results nearly two decades ago (got a Dragon demo in a carnival bag) than I get with Google voice today.

tmsldd 10y ago

> what's wrong with letting my Siri and your Siri just be two different "people" instead of feeding back into a single, networked hivemind? Absolutely nothing wrong.. that would be great, actually. Question is how to do that without data and GPUs.

rwallace 10y ago
It's perfectly okay to say, if you want to get the most out of this program, you need a big GPU rig. The price of such should come down over time. That leaves the question of how to get the data. If you make it optional, will enough people opt in? Is it an acceptable compromise to make it opt-out?
- tmsldd 10y ago
  Your idea of opt-{in, out} is a good one. It would integrate enough data with time, I think. "Open Data Foundation" + "World Wide Distributed Processing" would really make a revolution here. We need break urgently the monopoly over data.