Hacker News Reader

Comment by siegecraft | Hacker News Reader

Hopefully Mycroft (or similar open-source alternatives) can provide the benefits without all the privacy / behavioral modification downsides. But I fear open source products like this will always languish behind commercial offerings. Makes me think that the democratization of AI is very important in this regard (you shouldn't have to sell your soul to get good speech recognition, etc).

View on HN

There's never going to be a democratization of AI without a democratization of the data. But the data itself comes from privacy violating surveillance so I don't think we really want democratization of data.

You could share the pre-trained models that take tons and tons of data and compute time to generate? The models themselves are abstract enough to be considered anonymous, by their very nature.. I think.

Take a look at the literature on de-anonymization. It's easier than you might think.

De-anonymization of data or of trained neural-nets? I've read somewhere that you can sort of 'squash' a trained net down into a simpler version for performance benefits - the caveat being that it can't do any more learning. I would think that'd make it harder to de-anonymize.

If you simplify the function, you throw away accuracy (or maybe gain, if you over-simplified). The less accurate, the harder to de-anonymize. However, you only need 32 bits of information to uniquely identify someone.

s/over-simplified/over-fit/

I'm confident that in 20 years we'll have a very different conception of privacy. Certainly, one of the worst privacy outcomes would be if government had a monopoly on ubiquitous surveillance.

Already I don't expect to be anonymous in any public location. Given a public enough lifestyle and enough acquaintances, the likelihood of someone recognizing you anywhere on the globe or in a random pub is non-negligible.

I agree that we will have a different conception of privacy. I'm hopeful (but not particularly optimistic) that we can have something like David Brin's "Transparent Society" (https://en.wikipedia.org/wiki/The_Transparent_Society).

I like that term "sousveillance". It's tough to communicate the concept with phrases when folks tend to skim rather than read carefully.

Completely this. The companies talking about "democratization of AI" means "yeah, come use the models we built" but the data is the real source a democratization of AI should be focusing on. That's where the source of power is.

Mycroft seems to rely on Google for text-to-speech.

Open-Source TtS works ok (though inferior), but StT does not really exist.

Best (only) open source solution available seems to be pocketsphinx. On a quick try, it was awful.

Reply on news.ycombinator.com