"GPT-3 is the first AI system that has obvious, immediate, transformative economic value."
Uh, what? This is not obvious to me at all. It looks like a ginormous potential legal liability for anything you mission-critical you put it into. Having it go haywire in your beta-test of search is just embarrassing. Using it in something that is supposed to have "immediate, transformative economic value" also sounds like using it in something where you get in big trouble if it goes really badly. And, unlike more conventional software, this kind of thing is very hard to test in a reliable way.
Not saying it will never happen, but that it is "obvious", is not obvious to me at all.
The author thinks that ChatGPT and similar systems can easily be scaled up because they only cost about 5 million dollars to train. However, the author seems to ignore the cost of inference in his analysis. I've seen figures saying it costs OpenAI $0.02 per query to run ChatGPT. Until they solve that bottleneck, nobody's gonna be massively scaling up LLMs by a factor of 100 anytime soon.
3 comments
[ 3.6 ms ] story [ 16.0 ms ] threadUh, what? This is not obvious to me at all. It looks like a ginormous potential legal liability for anything you mission-critical you put it into. Having it go haywire in your beta-test of search is just embarrassing. Using it in something that is supposed to have "immediate, transformative economic value" also sounds like using it in something where you get in big trouble if it goes really badly. And, unlike more conventional software, this kind of thing is very hard to test in a reliable way.
Not saying it will never happen, but that it is "obvious", is not obvious to me at all.