Y
HN Search
Hacker News Search
new
|
comments
|
top
|
jobs
zoe_mnode
searching Neon…
1.
▲
2.
▲
3.
▲
4.
▲
5.
▲
6.
▲
5 ms
·
1.
▲
by
zoe_mnode
8mo ago
This direction is clearly not unique to DeepSeek. Similar ideas have shown up before, like Gemma’s per-layer embeddings, RWKV’s deepembed, and ByteDance’s UltraMem. From an algorithm perspective, using hash-based n-gram lookup is far from i
2.
▲
by
zoe_mnode
8mo ago
This is impressive work on quantization. It really validates the hypothesis that software optimization (and proper memory management) matters more than just raw FLOPs.