Isn’t this simply a byproduct of the training data?
For conversations in US English or more generally sourced from USians that end up in training data, there will be supporters and deniers, as with any issue. Conversations about it outside of the US are likely to have far less conspiratorial support.
The model isn’t doing any critical thinking here… what mechanism would even support that under the hood? Feels like another thing that is only weird if you assume AI actually works like a human.
Even if they’re trained on the same data, and I suspect they are, why maintain 2 unfathomably large and probably very similar sets of data
It still might be doing it on its own based on its own internal logic
Like maybe it’s worked out that users who use miles expect the conversation to proceed in a different way than users who use kilometers and is tailoring its responses to fit those expectations.
Not that I’m ruling out that it’s something deliberately coded into it by a longshot, just kind of pointing out another possibility
But both options are terrifying when you think about them.
Isn’t this simply a byproduct of the training data?
For conversations in US English or more generally sourced from USians that end up in training data, there will be supporters and deniers, as with any issue. Conversations about it outside of the US are likely to have far less conspiratorial support.
The model isn’t doing any critical thinking here… what mechanism would even support that under the hood? Feels like another thing that is only weird if you assume AI actually works like a human.
The training data is the same whether you are in the US or Canada.
Even if they’re trained on the same data, and I suspect they are, why maintain 2 unfathomably large and probably very similar sets of data
It still might be doing it on its own based on its own internal logic
Like maybe it’s worked out that users who use miles expect the conversation to proceed in a different way than users who use kilometers and is tailoring its responses to fit those expectations.
Not that I’m ruling out that it’s something deliberately coded into it by a longshot, just kind of pointing out another possibility
But both options are terrifying when you think about them.
Do we know for a fact that the models being used for US requests are the same as those used for foreign requests? If so, how do we know this?