Death of Giraffe, but hopefully not ML in chess!

[matthewlai]

It is with great sadness that I am announcing the discontinuation of the Giraffe project. I am not sure why I am making a post about it instead of just having it silently fade away... I suppose it is to give myself some closure. It's something that I had a crazy amount of fun with in the past few months.

I am now an employee of Google DeepMind, and I felt that Giraffe is a bit too close for comfort to what I do at work.

I can only legally use what's public knowledge in Giraffe, but the real world is not so clear-cut. I have been exposed to a lot of really cool machine learning techniques due to the nature of my work, and a lot of times it's difficult to discern what's public knowledge and what isn't. It's no fun knowing there are several techniques I can try that will probably drastically improve Giraffe, when I can't try them because they are still trade secret.

I am sure there are ways to make this work if I really wanted to, but I have many other passions outside of machine learning as well, and at this point it's just much easier to just pick something totally unrelated to work on - where I can go 100% and not have to worry about spilling trade secrets.

Plus, I already do 8 hours of ML a day, 5 days a week. I would like to take a break from ML in my free time, and unfortunately, computer chess is not really interesting to me anymore without ML.

I may pick up Giraffe again if I ever leave DeepMind. That's not likely to be any time soon, though. It's an awesome company with amazing people, amazing research, and amazing free food. What more can you ask for in a job?

Since Giraffe was released under the GPL, anyone is obviously free to pick it up and continue working on it. I hope someone will at least continue exploring using ML in chess, even if not through Giraffe. I achieved all I did in about 4 months of work. It was very much just an early proof of concept. There were many things that I was almost certain would help, but didn't have time to try. Imagine what YOU can achieve if you put in more time than I did. There is still much potential in ML + chess that is begging to be discovered.

Quit the rat race. Just leave Stockfish and Komodo to it. Start taking great risks and work on the next big thing . Machine learning has defeated hand-crafted systems in just about every other field. That will happen in chess, too, and the only question is when.

Keep in mind that Stockfish running at 5 nps (same speed as humans) is still weaker than anyone who has been playing chess for more than 2 weeks.

Disclaimer: This post is all personal opinion, and do not reflect the views of my employer. They have never contacted me about Giraffe (well, besides telling me they like it and to set up the first interview).

[Steve Maughan]

Congrats Matthew!

Can I ask - was the Giraffe project the reason why you got the job? Did they headhunt you?

Cheers - Steve

[matthewlai]

Congrats Matthew!

Can I ask - was the Giraffe project the reason why you got the job? Did they headhunt you?

Cheers - Steve

Thanks Steve!

Yes, most likely. They contacted me after seeing my thesis on arXiv.

[Uri Blass]

It is with great sadness that I am announcing the discontinuation of the Giraffe project. I am not sure why I am making a post about it instead of just having it silently fade away... I suppose it is to give myself some closure. It's something that I had a crazy amount of fun with in the past few months.

I am now an employee of Google DeepMind, and I felt that Giraffe is a bit too close for comfort to what I do at work.

I can only legally use what's public knowledge in Giraffe, but the real world is not so clear-cut. I have been exposed to a lot of really cool machine learning techniques due to the nature of my work, and a lot of times it's difficult to discern what's public knowledge and what isn't. It's no fun knowing there are several techniques I can try that will probably drastically improve Giraffe, when I can't try them because they are still trade secret.

I am sure there are ways to make this work if I really wanted to, but I have many other passions outside of machine learning as well, and at this point it's just much easier to just pick something totally unrelated to work on - where I can go 100% and not have to worry about spilling trade secrets.

Plus, I already do 8 hours of ML a day, 5 days a week. I would like to take a break from ML in my free time, and unfortunately, computer chess is not really interesting to me anymore without ML.

I may pick up Giraffe again if I ever leave DeepMind. That's not likely to be any time soon, though. It's an awesome company with amazing people, amazing research, and amazing free food. What more can you ask for in a job?

Since Giraffe was released under the GPL, anyone is obviously free to pick it up and continue working on it. I hope someone will at least continue exploring using ML in chess, even if not through Giraffe. I achieved all I did in about 4 months of work. It was very much just an early proof of concept. There were many things that I was almost certain would help, but didn't have time to try. Imagine what YOU can achieve if you put in more time than I did. There is still much potential in ML + chess that is begging to be discovered.

Quit the rat race. Just leave Stockfish and Komodo to it. Start taking great risks and work on the next big thing . Machine learning has defeated hand-crafted systems in just about every other field. That will happen in chess, too, and the only question is when.

Keep in mind that Stockfish running at 5 nps (same speed as humans) is still weaker than anyone who has been playing chess for more than 2 weeks.

Disclaimer: This post is all personal opinion, and do not reflect the views of my employer. They have never contacted me about Giraffe (well, besides telling me they like it and to set up the first interview).

I do not think that stockfish running at 5 nodes per second is weaker than anyone who has been playing chess for more than 2 weeks.

I know that there are many children who play chess for more than 2 weeks and they do many one ply mistakes and do not use their time.

Stockfish at 5 nodes per seconds can have enough time to get depth 2 that is probably good enough to beat most children who started 2 weeks ago and even month ago.

Edit
Here is a game to show you the level of beginners who started to play
more than 2 weeks ago(age of the loser less than 8).

I am sure stockfish with 5 nodes per second beat both of these players

[pgn][Event "Computer chess game"]
[Site "URIBLASS-THINK"]
[Date "2016.01.22"]
[Round "?"]
[White "UriBlass"]
[Black "UriBlass"]
[Result "0-1"]
[BlackElo "2400"]
[ECO "A06"]
[Opening "Reti Opening"]
[Time "00:41:50"]
[Variation "Tennison (Zukertort-Lemberg) Gambit Accepted"]
[WhiteElo "2400"]
[TimeControl "120+6"]
[Termination "normal"]
[PlyCount "18"]
[WhiteType "human"]
[BlackType "human"]

1. e4 d5 2. Nf3 dxe4 3. Ne5 f6 4. Nd3 e6 5. Nc5 Bxc5 6. Bb5+ c6 7. Bf1 f5
8. a3 Qd4 9. a4 Qxf2# 0-1[/pgn]

[JoshPettus]

You absolutely sure google wouldn't be interested in having a google chess app with resources in the cloud? Could be a nice demo, just saying.

It's a shame, it's such a great engine. I'll be sure to make a compile of the last source code for OSX when I get a chance.

[Vinvin]

A new discussion rise a couple a days ago on Lichess where the Crazyhouse variant was introduced ... Does Giraffe can learn Crazyhouse variant ???

[supersharp77]

It is with great sadness that I am announcing the discontinuation of the Giraffe project. I am not sure why I am making a post about it instead of just having it silently fade away... I suppose it is to give myself some closure. It's something that I had a crazy amount of fun with in the past few months.

I am now an employee of Google DeepMind, and I felt that Giraffe is a bit too close for comfort to what I do at work.

I can only legally use what's public knowledge in Giraffe, but the real world is not so clear-cut. I have been exposed to a lot of really cool machine learning techniques due to the nature of my work, and a lot of times it's difficult to discern what's public knowledge and what isn't. It's no fun knowing there are several techniques I can try that will probably drastically improve Giraffe, when I can't try them because they are still trade secret.

I am sure there are ways to make this work if I really wanted to, but I have many other passions outside of machine learning as well, and at this point it's just much easier to just pick something totally unrelated to work on - where I can go 100% and not have to worry about spilling trade secrets.........Disclaimer: This post is all personal opinion, and do not reflect the views of my employer. They have never contacted me about Giraffe (well, besides telling me they like it and to set up the first interview)

Congrats on your new job with the Google "Braintrust"...well of couse
the Giraffe project will be shelved for good and used to further to interests of Google Inc. (getting paid) thats what happens to all the other
chess based artificial intelligence projects...they all end up working for national defense or some "black project" or a high powered corporation
most of what we see is the initial efforts..then ZAP!! they disappear
gone...gone..gone..and the chess engine created remains..the way of the world...AR




ps... that w32 version still not working in XP

[melajara]

I have mixed feelings about that.

First, big congratulations to you Matthew, DeepMind is probably the dream team inside a dream company, very big achievement to have been recruited to join them

Btw, I just saw this (somewhat unusually bombastic) presentation from a man I respect very much and who is usually much more reserved, Jurgen Schmidhuber, see https://www.youtube.com/watch?v=Ya9YfYveFXA

It happens you should have 2 colleagues who got their PhD under his supervision and times are indeed very exciting in the field of deep learning right now.

Now, I'm very sad that potential conflict of interests (and burning out when staying in ML for spare time too) is actually threatening to kill Giraffe, very very sad news.

On the other hand, Google has made some strides to open its AI research, e.g. with TensorFlow and Alphabet, officially is supporting the Open AI initiative pioneered by Elon Musk.

I'm very curious about the limit here of what is trade secrets and what is on the verge to become public, I envy you to work in such an exciting environment, again big congratulations to you,
and, please, stay around to shepherd (like e.g Marco Costalba is doing now for Stockfish) whoever has the capability to push Giraffe forward

[matthewlai]

I have mixed feelings about that.

First, big congratulations to you Matthew, DeepMind is probably the dream team inside a dream company, very big achievement to have been recruited to join them

Btw, I just saw this (somewhat unusually bombastic) presentation from a man I respect very much and who is usually much more reserved, Jurgen Schmidhuber, see https://www.youtube.com/watch?v=Ya9YfYveFXA

It happens you should have 2 colleagues who got their PhD under his supervision and times are indeed very exciting in the field of deep learning right now.

Now, I'm very sad that potential conflict of interests (and burning out when staying in ML for spare time too) is actually threatening to kill Giraffe, very very sad news.

On the other hand, Google has made some strides to open its AI research, e.g. with TensorFlow and Alphabet, officially is supporting the Open AI initiative pioneered by Elon Musk.

I'm very curious about the limit here of what is trade secrets and what is on the verge to become public, I envy you to work in such an exciting environment, again big congratulations to you,
and, please, stay around to shepherd (like e.g Marco Costalba is doing now for Stockfish) whoever has the capability to push Giraffe forward

Thanks Mel. It is indeed a very cool group to join, and I never imagined I would be able to join such a famous AI lab so early in my career, and being able to work with some of the (if not the) brightest minds in the field.

It's a very interesting presentation! There has certainly been a lot of discussion surrounding those topics in and outside of Google, and I think it's good that we are talking about those things early (a few years to a few decades before they become reality).

Google does have a fairly open stance towards AI, and there is always tons of really cool AI research going on at Google, so stayed tuned for more cool stuff!

[matthewlai]

It is with great sadness that I am announcing the discontinuation of the Giraffe project. I am not sure why I am making a post about it instead of just having it silently fade away... I suppose it is to give myself some closure. It's something that I had a crazy amount of fun with in the past few months.

I am now an employee of Google DeepMind, and I felt that Giraffe is a bit too close for comfort to what I do at work.

I can only legally use what's public knowledge in Giraffe, but the real world is not so clear-cut. I have been exposed to a lot of really cool machine learning techniques due to the nature of my work, and a lot of times it's difficult to discern what's public knowledge and what isn't. It's no fun knowing there are several techniques I can try that will probably drastically improve Giraffe, when I can't try them because they are still trade secret.

I am sure there are ways to make this work if I really wanted to, but I have many other passions outside of machine learning as well, and at this point it's just much easier to just pick something totally unrelated to work on - where I can go 100% and not have to worry about spilling trade secrets.

Plus, I already do 8 hours of ML a day, 5 days a week. I would like to take a break from ML in my free time, and unfortunately, computer chess is not really interesting to me anymore without ML.

I may pick up Giraffe again if I ever leave DeepMind. That's not likely to be any time soon, though. It's an awesome company with amazing people, amazing research, and amazing free food. What more can you ask for in a job?

Since Giraffe was released under the GPL, anyone is obviously free to pick it up and continue working on it. I hope someone will at least continue exploring using ML in chess, even if not through Giraffe. I achieved all I did in about 4 months of work. It was very much just an early proof of concept. There were many things that I was almost certain would help, but didn't have time to try. Imagine what YOU can achieve if you put in more time than I did. There is still much potential in ML + chess that is begging to be discovered.

Quit the rat race. Just leave Stockfish and Komodo to it. Start taking great risks and work on the next big thing . Machine learning has defeated hand-crafted systems in just about every other field. That will happen in chess, too, and the only question is when.

Keep in mind that Stockfish running at 5 nps (same speed as humans) is still weaker than anyone who has been playing chess for more than 2 weeks.

Disclaimer: This post is all personal opinion, and do not reflect the views of my employer. They have never contacted me about Giraffe (well, besides telling me they like it and to set up the first interview).

I do not think that stockfish running at 5 nodes per second is weaker than anyone who has been playing chess for more than 2 weeks.

I know that there are many children who play chess for more than 2 weeks and they do many one ply mistakes and do not use their time.

Stockfish at 5 nodes per seconds can have enough time to get depth 2 that is probably good enough to beat most children who started 2 weeks ago and even month ago.

Edit
Here is a game to show you the level of beginners who started to play
more than 2 weeks ago(age of the loser less than 8).

I am sure stockfish with 5 nodes per second beat both of these players

[pgn][Event "Computer chess game"]
[Site "URIBLASS-THINK"]
[Date "2016.01.22"]
[Round "?"]
[White "UriBlass"]
[Black "UriBlass"]
[Result "0-1"]
[BlackElo "2400"]
[ECO "A06"]
[Opening "Reti Opening"]
[Time "00:41:50"]
[Variation "Tennison (Zukertort-Lemberg) Gambit Accepted"]
[WhiteElo "2400"]
[TimeControl "120+6"]
[Termination "normal"]
[PlyCount "18"]
[WhiteType "human"]
[BlackType "human"]

1. e4 d5 2. Nf3 dxe4 3. Ne5 f6 4. Nd3 e6 5. Nc5 Bxc5 6. Bb5+ c6 7. Bf1 f5
8. a3 Qd4 9. a4 Qxf2# 0-1[/pgn]

Ha! That's cool! I have never actually seen children play. I suppose computers still have the advantage of not blundering (much) within the search horizon, no matter how close that is.