[smoldesu]

How do smartphones address the privacy concerns? I'm still surprised people are buddy-buddy with commodity smartphone manufacturers.

> What is stopping them from collecting and storing these samples for "training" models?

From the top of my head, I'd wager the cost of transmitting that data is not worth it for how low-quality the input would be.

[prollyjethi]

> From the top of my head, I'd wager the cost of transmitting that data is not worth it for how low-quality the input would be.

Why is that? If the data is good enough to make an inference, will it not be good enough for transcription?

I am curious what makes it low quality?

[smoldesu]

For starters, the integrated microphone in devices like Humane's pin don't face the user. Recording things without the user knowing would be difficult, if you could even manage a good recording. Then you have to transmit the data, which would almost certainly add another layer of compression. Even if you manage to get a lossless file back on the server-side, it's going to be noisy and imperfect data that wouldn't make for clear training material. Unless you're training a denoiser (at which point there are much better approaches), that sort of noisy data probably isn't good for much. Nevermind the cost of human-assisted labeling...

I Am Not Georgi Gerganov, I cannot denounce entire AI concepts with a single refutation. But I think logically, stealing that data would be kinda pointless. Not to say it's impossible, but at-scale I'm not sure why you'd implement it.