4 ms·

This would be a good bill. Devices like Alexa should never transmit conversations back to Amazon's servers but all processing should be done locally. They shoul

This would be a good bill. Devices like Alexa should never transmit conversations back to Amazon's servers but all processing should be done locally. They should learn to train their system without real user data.

pdkl95 8y ago
They can use real user data if they ask[1] for it first. Let users browse their poor-quality/failed interpretations, recent commands, etc, and submit the data they want. If you're clever with the UI, you might even get them to attach the desired/correct interpretation if they really want a particular error fixed. Alternatively, use the Nielsen method of simply paying people for their data. [1] in the "informed consent" sense, not the usual open ended "you agreed to give us whatever we want in our contract of adhesion" dark pattern.
- m0zg 8y ago
  Yeah, I'd actually prefer that. If system guesses mis-recognition, ask me if you got what I said right, and if not, ask if you can use this as a training example. Have me enter on the phone the text that corresponds to what I was saying, and explicitly approve it for training. Better yet, if this creates an adaptation of the general model specifically for me.
amelius 8y ago
But what if the user asks Alexa about some topic on e.g. Wikipedia? Should Alexa store the entire web upfront? At some point it might need to send a request of some sort to a server.
- chopin 8y ago
  Why would this use-case require to call to Amazon? Going to Wikipedia should suffice.
  - amelius 8y ago
    Yes but then you're still sending data out to the web which the user (who might not be the owner of the device) might not have consented to.
    - garmaine 8y ago
      The device can fetch a webpage from wikipedia and summarize. It doesn't have to send the original query.
- aw3c2 8y ago
  Wikipedia is not a good example. It is freely licensed, available to download in its entirety and easily fits on a 30$ memory card. In other cases those personal assistants could first ask permission to query an API or search engine.
  - NotAnEconomist 8y ago
    A $30 memory card is the entire cost of an Echo device -- that's a 100% increase in cost just to cache Wikipedia locally. That's a non-trivial modification, even before talking about the processor upgrades required to actually utilize that information and do local processing. Your suggestion would drive the cost of an Echo from $30 to $300.
- amluto 8y ago
  If I ask a machine for the Wikipedia article about “such-and-such”, then I fully expect “such-and-such” to go over the network. If I ask for my lights to turn off, the cloud has no legitimate need to know about it. (The best counterargument I can think of is that the speech recognizer might be too computationally expensive to run on cheap hardware.)

pdkl95 8y ago

They can use real user data if they ask[1] for it first. Let users browse their poor-quality/failed interpretations, recent commands, etc, and submit the data they want. If you're clever with the UI, you might even get them to attach the desired/correct interpretation if they really want a particular error fixed. Alternatively, use the Nielsen method of simply paying people for their data. [1] in the "informed consent" sense, not the usual open ended "you agreed to give us whatever we want in our contract of adhesion" dark pattern.

m0zg 8y ago
Yeah, I'd actually prefer that. If system guesses mis-recognition, ask me if you got what I said right, and if not, ask if you can use this as a training example. Have me enter on the phone the text that corresponds to what I was saying, and explicitly approve it for training. Better yet, if this creates an adaptation of the general model specifically for me.

amelius 8y ago

But what if the user asks Alexa about some topic on e.g. Wikipedia? Should Alexa store the entire web upfront? At some point it might need to send a request of some sort to a server.

chopin 8y ago
Why would this use-case require to call to Amazon? Going to Wikipedia should suffice.
- amelius 8y ago
  Yes but then you're still sending data out to the web which the user (who might not be the owner of the device) might not have consented to.
  - garmaine 8y ago
    The device can fetch a webpage from wikipedia and summarize. It doesn't have to send the original query.
aw3c2 8y ago
Wikipedia is not a good example. It is freely licensed, available to download in its entirety and easily fits on a 30$ memory card. In other cases those personal assistants could first ask permission to query an API or search engine.
- NotAnEconomist 8y ago
  A $30 memory card is the entire cost of an Echo device -- that's a 100% increase in cost just to cache Wikipedia locally. That's a non-trivial modification, even before talking about the processor upgrades required to actually utilize that information and do local processing. Your suggestion would drive the cost of an Echo from $30 to $300.
amluto 8y ago
If I ask a machine for the Wikipedia article about “such-and-such”, then I fully expect “such-and-such” to go over the network. If I ask for my lights to turn off, the cloud has no legitimate need to know about it. (The best counterargument I can think of is that the speech recognizer might be too computationally expensive to run on cheap hardware.)