Update README.md

README.md

@@ -18,9 +18,12 @@ MCTS is a decisive factor contributing to the world champion level performance.
 With the recent advancement of large language model in transformer[5] based decoder with a next token prediction objective[6], and it's application in Chess[7][8], how does a language model (the GoFormer here) perform in a Go game?  
 [9] finetunes 124M, 355M, and 744M GPT-2[10] on 56,638 Go game in SGF format. To the best of the knowledge, this is the first time a language model is trained from scratch with 1.36M Go games, with a specially designed tokenizer.
 
-Can GoFormer perform reasonably well just by next move (token) prediction, without MCTS[3][4]? Let's find out. The hope is that:
-- if language model can reason and plan, it can play Go very well. If it cannot, there is something worth investigating.
-- if GoFormer can perform reasonably well, it can be used as a baseline for future research in Go game, and even a baseline for heuristic search, without the use of tree search.
+Can GoFormer perform reasonably well just by next move (token) prediction, without MCTS[3][4]? Let's find out. 
+My research goals are that: 
+- if language model can reason and plan, it can play Go very well.
+- if GoFormer can perform reasonably well, it can be used as a baseline for future research in Go game, without the use of tree search.
+
+P.S: it is an intial release of model, and it is expected not to perform very well. But as we have more data, we will see if it can stand a battle with MCTS based engine like [Leela Zero](https://github.com/leela-zero/leela-zero).
 
 ## Data Preprocessing
 We take the leftmost variation of the game tree in SGF format and translate it into PGN.