What if you take the speculative execution strategy and have multiple interpreters translating every possible semantic branch and then throwing out the recordings of the interpretations that were incorrect? 🙃
I’ve been thinking about the suggestion over and over and I just don’t see any way to make it worthwhile, whether or not you’re in a scenario where you need on-the-fly translations.
Let’s says you have 20 translations immediately ready by the time the sentence ends and only 1 correctly predicted the sentence. Okay you already have the translation while a single translator would be only partly through it on their own. But now you’ve got to submit this translation (I’m assuming it’s being shown digitally?? How tf else would this work, you can’t have 20 interpreters all saying random different shit at once).
So you need 19 people to not submit their sentence and the one correct one to do so. The incorrect predictions can drop out as the sentence continues and hopefully two people don’t have equally correct interpretations—at this point it’s clear this won’t work with people. Especially since you have to pay them.
So AI right? That requires computing a shit ton of branches as the sentence progresses and having them cease their branch as the sentences invalidates their predictions. So once the sentence is mostly complete the branch count should be low and whichever branch computes correctly first is submitted, once it knows the sentence is over, which isn’t instantaneous. There must be a delay for the computer to know the sentence is over so as not to submit preemptively. And in that short time, a modern computer can already translate your sentence correctly and find the best interpretation instead of the first available one, which may not be ideal.
What if you take the speculative execution strategy and have multiple interpreters translating every possible semantic branch and then throwing out the recordings of the interpretations that were incorrect? 🙃
Can’t wait for the Spectre/Meltdown exploit.
AI does this with MLM. I rest my case.
Why would you do that?
You have to wait for the last work in the sentence anyway so doing extra calculations to throw out later isn’t making it faster.
I’ve been thinking about the suggestion over and over and I just don’t see any way to make it worthwhile, whether or not you’re in a scenario where you need on-the-fly translations.
Let’s says you have 20 translations immediately ready by the time the sentence ends and only 1 correctly predicted the sentence. Okay you already have the translation while a single translator would be only partly through it on their own. But now you’ve got to submit this translation (I’m assuming it’s being shown digitally?? How tf else would this work, you can’t have 20 interpreters all saying random different shit at once).
So you need 19 people to not submit their sentence and the one correct one to do so. The incorrect predictions can drop out as the sentence continues and hopefully two people don’t have equally correct interpretations—at this point it’s clear this won’t work with people. Especially since you have to pay them.
So AI right? That requires computing a shit ton of branches as the sentence progresses and having them cease their branch as the sentences invalidates their predictions. So once the sentence is mostly complete the branch count should be low and whichever branch computes correctly first is submitted, once it knows the sentence is over, which isn’t instantaneous. There must be a delay for the computer to know the sentence is over so as not to submit preemptively. And in that short time, a modern computer can already translate your sentence correctly and find the best interpretation instead of the first available one, which may not be ideal.
Unrelated but is your username based on the O gauge model train manufacturer?
Nope