From 8e2a779c4c8b496a3e0f343dba66821c7545ee92 Mon Sep 17 00:00:00 2001
From: PENG Bo <33809201+BlinkDL@users.noreply.github.com>
Date: Wed, 18 May 2022 13:06:28 +0800
Subject: [PATCH] Update README.md

---
 README.md | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/README.md b/README.md
index 05d7e01..09fa4ad 100644
--- a/README.md
+++ b/README.md
@@ -52,6 +52,8 @@ The **simplified formula** for GPT:
 
 ![F[\mathrm{t}+1]=\frac{\sum_{\mathrm{i}=0}^{\mathrm{t}} \exp (\mathbf{Q}x[\mathrm{t}] * \mathbf{K}F[\mathrm{i}]) \cdot(\mathbf{V}F[\mathrm{i}])}{\sum_{\mathrm{i}=0}^{\mathrm{t}} \exp (\mathbf{Q}x[\mathrm{t}] * \mathbf{K}F[\mathrm{i}])}](https://render.githubusercontent.com/render/math?math=%5Ccolor%7Bblack%7D%5Cdisplaystyle+F%5B%5Cmathrm%7Bt%7D%2B1%5D%3D%5Cfrac%7B%5Csum_%7B%5Cmathrm%7Bi%7D%3D0%7D%5E%7B%5Cmathrm%7Bt%7D%7D+%5Cexp+%28%5Cmathbf%7BQ%7Dx%5B%5Cmathrm%7Bt%7D%5D+%2A+%5Cmathbf%7BK%7DF%5B%5Cmathrm%7Bi%7D%5D%29+%5Ccdot%28%5Cmathbf%7BV%7DF%5B%5Cmathrm%7Bi%7D%5D%29%7D%7B%5Csum_%7B%5Cmathrm%7Bi%7D%3D0%7D%5E%7B%5Cmathrm%7Bt%7D%7D+%5Cexp+%28%5Cmathbf%7BQ%7Dx%5B%5Cmathrm%7Bt%7D%5D+%2A+%5Cmathbf%7BK%7DF%5B%5Cmathrm%7Bi%7D%5D%29%7D)
 
+It's very capable in theory, however that **does not mean we can fully utilize its capability with usual optimizers**. I suspect the loss landscape is too difficult for our current methods.
+
 Compare with the **simplified formula** for RWKV-2 (the parallel mode, looks similar to Apple's AFT):
 
 ![F[\mathrm{t}+1]=\sigma(\mathbf{R}x[\mathrm{t}]) \cdot \frac{\sum_{\mathrm{i}=0}^{\mathrm{t}} \exp (\mathbf{W} \cdot(\mathrm{t}-\mathrm{i})) \cdot \exp (\mathbf{K}F[\mathrm{i}]) \cdot(\mathbf{V}F[\mathrm{i}])}{\sum_{\mathrm{i}=0}^{\mathrm{t}} \exp (\mathbf{W} \cdot(\mathrm{t}-\mathrm{i})) \cdot \exp (\mathbf{K }F[\mathrm{i}])}](https://render.githubusercontent.com/render/math?math=%5Ccolor%7Bblack%7D%5Cdisplaystyle+F%5B%5Cmathrm%7Bt%7D%2B1%5D%3D%5Csigma%28%5Cmathbf%7BR%7Dx%5B%5Cmathrm%7Bt%7D%5D%29+%5Ccdot+%5Cfrac%7B%5Csum_%7B%5Cmathrm%7Bi%7D%3D0%7D%5E%7B%5Cmathrm%7Bt%7D%7D+%5Cexp+%28%5Cmathbf%7BW%7D+%5Ccdot%28%5Cmathrm%7Bt%7D-%5Cmathrm%7Bi%7D%29%29+%5Ccdot+%5Cexp+%28%5Cmathbf%7BK%7DF%5B%5Cmathrm%7Bi%7D%5D%29+%5Ccdot%28%5Cmathbf%7BV%7DF%5B%5Cmathrm%7Bi%7D%5D%29%7D%7B%5Csum_%7B%5Cmathrm%7Bi%7D%3D0%7D%5E%7B%5Cmathrm%7Bt%7D%7D+%5Cexp+%28%5Cmathbf%7BW%7D+%5Ccdot%28%5Cmathrm%7Bt%7D-%5Cmathrm%7Bi%7D%29%29+%5Ccdot+%5Cexp+%28%5Cmathbf%7BK+%7DF%5B%5Cmathrm%7Bi%7D%5D%29%7D)