Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

DeepMind has open-sourced the heart of AlphaGo and AlphaZero

twitter.com

Post reply on HN
Print view View original

31–40 of 91 posts

Re: DeepMind has open-sourced the heart of AlphaGo and AlphaZero

#31

Post by bluecalm »

mtlmtlmtlmtl wrote: post #20
throwaway81523 wrote: post #8

Earlier quoted context omitted.

These particular programs were bleeding edge when they were announced, but have been recreated and improved on outside of DeepMind since then. For example, LCZero today is stronger at chess than any of the DeepMind chess programs that were shown to the public. Of course we don't know if they got further behind the scenes. Of course LCZero and everything else relied on the published papers of DeepMind. That is, the pa…

There's an interesting bias I've noticed on HN where a lot of people still believe AlphaZero is the state of the art in computer chess, when that hasn't been true except for a short while after the release. AZ still gets posted and upvoted today while newer improvements in other engines are discussed far less. I guess it's because AZ came out of Google?

What is interesting is that current state of the art comes from another place: https://en.m.wikipedia.org/wiki/Efficiently_updatable_neural...

Lc0 is based on AlphaZero ideas and is significantly weaker than NNUE based modern Stockfish.

Re: DeepMind has open-sourced the heart of AlphaGo and AlphaZero

#32

Post by Closi »

robertlagrant wrote: post #2

I can't help but wonder when this stuff happens that the devs (or the leads) are worried about being let go, so they open source their stuff to be able to take it elsewhere. Anyone here from DeepMind?

> I can't help but wonder when this stuff happens that the devs (or the leads) are worried about being let go, so they open source their stuff to be able to take it elsewhere.

Open-sourcing isn't a dev-level decision - it is a business leadership decision.

Re: DeepMind has open-sourced the heart of AlphaGo and AlphaZero

#33

Post by Yuioup »

londons_explore wrote: post #16

Worth noting that while AlphaGo and AlphaZero are incredible achievements, the amount of actual code to implement them isn't very much. If you have the research paper, someone in the field could reimplement them in a few days. Then there is the large compute cost for training them to produce the trained weights. So, opensourcing these bits of work without the weights isn't as major a thing as you might imagine.

It's already been done:

https://github.com/lightvector/KataGo

Re: DeepMind has open-sourced the heart of AlphaGo and AlphaZero

#34

Post by tyler33 »

est wrote: post #6
kubb wrote: post #3

Must feel pretty bad being a rank and file dev at DeepMind and realizing that your leads had all the talent and funding in the world that they could have used for building language model products and instead they heavily invested in solving board games. Then you realize they get paid >x10 of what you are and they're fine, but the next layoff is likely gonna get you.

They have their own LLMs such as PaLM

why they dont make public apps?(at least)

Re: DeepMind has open-sourced the heart of AlphaGo and AlphaZero

#35

Post by ta_tunestub »

londons_explore wrote: post #16

Worth noting that while AlphaGo and AlphaZero are incredible achievements, the amount of actual code to implement them isn't very much. If you have the research paper, someone in the field could reimplement them in a few days. Then there is the large compute cost for training them to produce the trained weights. So, opensourcing these bits of work without the weights isn't as major a thing as you might imagine.

Yep, there are many reimplementations. Here is a reimplementation that swaps out a neural net with a GBDT to address compute costs:

https://github.com/cgreer/alpha-zero-boosted

Re: DeepMind has open-sourced the heart of AlphaGo and AlphaZero

#38

Post by woah »

ta_tunestub wrote: post #35
londons_explore wrote: post #16

Worth noting that while AlphaGo and AlphaZero are incredible achievements, the amount of actual code to implement them isn't very much. If you have the research paper, someone in the field could reimplement them in a few days. Then there is the large compute cost for training them to produce the trained weights. So, opensourcing these bits of work without the weights isn't as major a thing as you might imagine.

Yep, there are many reimplementations. Here is a reimplementation that swaps out a neural net with a GBDT to address compute costs: https://github.com/cgreer/alpha-zero-boosted

How does the performance of this version compare?

Re: DeepMind has open-sourced the heart of AlphaGo and AlphaZero

#39

Post by rhdunn »

londons_explore wrote: post #16

Worth noting that while AlphaGo and AlphaZero are incredible achievements, the amount of actual code to implement them isn't very much. If you have the research paper, someone in the field could reimplement them in a few days. Then there is the large compute cost for training them to produce the trained weights. So, opensourcing these bits of work without the weights isn't as major a thing as you might imagine.

There is Leela Zero (https://github.com/leela-zero/leela-zero) for Go and lc0/Leela Chess (https://github.com/orgs/LeelaChessZero/repositories) for Chess, where both provide trained weights. The Leela Chess project specifically have been working for a long time on training and refining the weights for Chess, as well as providing the code -- they allow you to see the history and performance over time for the various trained models.
Post reply on HN