Discrepancy between AlphaGo Teach and Leela Zero

I started studying a blog post shared by @mark5000 on AlphaGo josekis.

To my surprise, AlphaGo Teach (left)and Leela Zero (right) have a very different view on the large avalanche:

Any explanation on that? Am I reading it wrong?


That IS interesting. A couple of things:

  1. LeelaZero is zero-based and AlphaGo Master is not, so it’s understandable there is a difference of opinion.

  2. Try forcing those two moves for LZ individually and see if the output changes. Maybe the output would change if you followed the rabbit trail even further. I don’t know.


Thanks for the reply!

What’s also interesting is that the zero-based one agrees with humans.

I pushed a couple more moves and both stick to their evaluation:


Knowing discrepencies between Leela Zero and Pangafu’s Leela Master. This does not suprise me. LZ tends to think very lightly of value of influence. For what it’S worth I would guess AG master guess is more relevant for human players.

1 Like

I thought AlphaGo Teach was based on AlphaGo Zero rather than AlphaGo Master. Anyone got some more information on that?

1 Like

That’s what I was hoping, but the Tool uses AlphaGo Master version.

See https://www.facebook.com/aja.huang/posts/1761304863904077

Machine translation:


But then in my example it’s LZ that agrees with humans…

Regardless of humans, I was not expecting such a gap between the two on a local sequence. The sequence is joseki for LZ, a black failure for AlphaGo Master.

1 Like

This might be a redundant post for anyone reading this (or anyone capable of using a search engine) but here is the link for AlphaGo Teach:

I only post as I was about to ask where I could find this tool and then remembered that I am (just about) capable of working a search engine.


And if you follow the link in my original post you’ll find it too :slight_smile: (and much more)

1 Like

First (and obvious) question: is the whole-board position the same in both cases?

It is, and as mentioned in my post it comes from a link shared by @mark5000.

If you’d like to read more click here. At the time LZ analysis was not available.

1 Like

A small number of professional games do omit the exchange, but it’s not like this variation became any more popular after the release of AlphaGo Teach. In fact, it’s only appeared a minimum of three times since and including 2016, and only once since this thread got posted.

OGSJ doesn’t include the idea. Josekipedia has it marked as a “Good Move”, and the main line as “Ideal”.

(PS. Diego’s link is dead.)

aka OJE :wink: … because compound acronyms FTW.

1 Like

KGS is KGS Go Server, they already got us beat ;)

Kinda like the B. in Benoît B. Mandelbrot stands for Benoît B. Mandelbrot.


Technically, those are a special set of compound acronyms that are recursive. No wonder KGS is in trouble, if they set up a stack overflow the first time you look at them…

At OGS we like to make sure we’re functional :wink: