Comment by protocolture
1 year ago
I dunno, seems trivial to me.
Like if an arduino can reliably carry 1 wakeword, an iphone could carry what? 64 at minimum?
Instead of waking up recording the conversation and sending it to the stasi, it could just toggle a bool in your secret advertising profile, then sync that up at random intervals.
They dont need context, they just need "Biscuit" "Nappies" "Birthday".
Not to say it isn't all made up. It probably is. I just dont think its technically difficult to achieve. It is probably all psychology. But sitting there with a packet capture going as you talk about dogfood isnt necessarily a great test to confirm the negative.
But if you're listening for 64, or even 1024 wake words, you won't get to the level of granularity people are claiming to observe, like in the other comment here where they tested "pool fencing". If you were putting together a list of 64 words, would you really include "pool fencing"?
The wakewords could be simply "buy", "cheap", "sale", "price" etc - and then collect the exact words around it. I bet the pool fencing conversation followed with the exchange of ideas where to get equipment.
10 thousand-ish words make sense when one considers multiple languages.
Its an extreme low side estimate.
And the chance that Pool OR Fencing is one is pretty high.
A better keyword would be "buy."
There are certainly ways to make it more viable, but I hope that the systems up to billions of people use are sufficiently independently audited so that such features would quickly be found.