anarmorarm

8 ms·

1. 1.▲
   by anarmorarm 5mo ago

   Edit: GPT-2, not GPT-2 Medium. The 2nd paragraph should read: "With 23.8 PPL on WikiText-103, WaveletLM beats both GPT-2, which was trained on 80× more data, and Transformer-XL Standard, which..."
2. 2.▲
   Show HN: WaveletLM – wavelet-based, attention-free model with O(n log n) scaling(github.com)

   7 points by anarmorarm 5mo ago|1 comments
3. 3.▲
   by anarmorarm 4y ago

   This deserves to be on the front page. Fantastic read.