Update README.md

README.md

@@ -2,11 +2,7 @@
 
 ## RWKV v2: RNN with Transformer Performance
 
-RWKV v2 is a RNN which can also be directly trained like a GPT transformer.
-
-You only need x_t, a_t, b_t of position t to compute the vectors for position t+1. 
-
-Hence it can be 100x faster than GPT, and 100x more VRAM friendly.
+RWKV v2 is a RNN which can also be directly trained like a GPT transformer. You only need x_t, a_t, b_t of position t to compute the vectors for position t+1. Hence it can be 100x faster than GPT, and 100x more VRAM friendly.
 
 See the release for a **27M params model on enwik8 with 0.72 BPC(dev)**.
 
@@ -20,6 +16,8 @@ Write out the formulas for "token at pos 2" and "token at pos 3" and you will ge
 
 kv / k is the memory mechanism. The token with high k can be remembered for a long duration, if W is close to 1 in the channel.
 
+It's also using my SmallInitEmb trick https://github.com/BlinkDL/SmallInitEmb (applicable to all transformers).
+
 The pseudocode (execution from top to bottom):
 
 ![RWKV-v2-RNN](RWKV-v2-RNN.png)