On https://sensei-library-q-and-a.replit.app you can ask questions over Sensei’s Library ChatGPT-style.
7 Likes
Certified as accurate as chat gpt!!
What is the status of the bent four?
The status of the bent four depends on the context. In some cases, it is alive, while in other cases, it is dead.
What is the main difference between the Chinese and Japanese rules?
The main difference between the Chinese and Japanese rules is that traditional Japanese rules have a lot of peculiar situations that have special rules that go against these more clean rules, such as bent four in corner, triple ko, etc., while traditional Chinese rules discard the low-order bit of the score-difference, thus not allowing as fine control of komi as Japanese and Lasker-Maas rules. Additionally, Japanese rules emphasize harmony, while Chinese rules have frequent disputatious behaviour by Chinese pros.
“discard the low-order bit” sounds like total bulls hit but in a way, it’s true.
6 Likes
Add “in the corner”.
1 Like
Do you plan to share the source code?
4 Likes
Very cool idea. One thing I noticed, it is often references discussion pages rather than the pages itself. What is the cause of this, what do you think?
It would be interesting to parse different Go forums and Reddit to create a comprehensive database of text sources to reference.
Bonus mission: parse SGF reviews and chats to reference real games.
2 Likes
I indexed both the pages and the discussions.
Yep, I’ll share the source code. It’s pretty primitive. Web app code can be found at Sensei Library Q&A - Replit. I’ll add the indexing script to that same project later.
4 Likes
All doable if I can get access to the texts.
2 Likes
Cool project, but ChatGPT isn’t a sensei yet:
SenseiLibrary Q&A
Question:
When is it appropriate to monkey jump?
Ask
Answer
It is appropriate to monkey jump when connecting two groups or when responding to a monkey jump
The quoted source was the article on reverse monkey jump
2 Likes
That’s some nice games in view…
make longer input line, so we can paste screenshot that includes both input and output
1 Like
Done 
4 Likes
Indeed not, it can’t really reason. It just “speaks” really well.
1 Like
Actually, its not correct to say that it don’t understand anything and just generates nice looking text, it can correctly solve math problems, while no one on purpose tried to teach it to do it:
https://www.reddit.com/r/ChatGPT/comments/11remk0/gpt4_solves_this_2023_aime_math_contest_problem/
It does not understand what it is doing. It predicts the next token given the existing sequence. It’s essentially a next-word guesser. Sure, sometimes that can result in “solving maths problems”, othertimes it results in pure nonsense.
2 Likes
What we could do to make it smarter is integrate it with a go engine (like KataGo). You can then get it to invoke the go engine and do things with it (e.g. set up a board, place a stone, ask it to tell you where your opponent is most likely to play next).
to correctly predict next token in really complicated situation it had no choice but to actually learn some things. It doesn’t understand many things, but it do understand some things.
1 Like
I disagree. I think it gives the illusion that it does. Maybe there’s some emergent understanding somewhere, but I think it’s pretty esoteric.
1 Like