3 ms·
Discussion of the actual paper here: https://news.ycombinator.com/item?id=25314830 https://news.ycombinator.com/item?id=25314830 It's a really neat one for pe
by 6gvONxR4sf7o 6y ago
Discussion of the actual paper here:
https://news.ycombinator.com/item?id=25314830 https://news.ycombinator.com/item?id=25314830
It's a really neat one for people who care about what's going on under the hood, but not immediately applicable to the more applied folks. I saw some good quotes at the time to the tune of "I can't wait to see the papers citing this one in a year or two."
- Fordec 6y agoI'm not in this subsection of the industry, so take this with a heaping grain of salt. I'm just using my own biases here. But I get the sense that this is one of a number of big pieces of the puzzle of how current cutting edge models can be cut down in operational size. I have a feeling that new methods of creating models will stick around, but any deep learning model can be reduced to a less resource intensive approximation through findings such as this. And if (IF!) so a lot of AI hype is going to hit a real post-hype cycle plateau of productivity because all this stuff will become easier and cost effective to deploy to the real world.