A mere one year ago...

[Jim Glass]

Kingscrusher on Leela v Stockfish #1.

https://youtu.be/T85mELZeqso

A lot can happen Ina year.

[Paul Berger]

Sure but on the other hand I'm wondering if there is any real progress since 6 months.

All in all, this year AI somewhat stalled, IMHO. 

I sincerely hope to be wrong otherwise beware yet another AI winter in 2020!

[Shuo Xiang]

LC0 could not be defeated by ANY ENGINE in either premier league or the cup.

If that's not perfection, I don't know what is.

The only engine capable of defeating LC0 is stockfish, and only during the SuFi, where it itself was subsequently defeated by a margin of 10.

--
You received this message because you are subscribed to the Google Groups "LCZero" group.
To unsubscribe from this group and stop receiving emails from it, send an email to lczero+un...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/54f8188e-5fa1-4872-a994-7bea115d63da%40googlegroups.com.

[Cary Knoop]

Leela is far from perfection, she just blundered against Komodo at the TCEC.

$fen 1rb2r2/4Nkpp/5p1P/8/p7/2Q5/PPB5/1K1R3q b - - 1 35

Leela played Qh2 and getting checkmated.

On Sunday, October 27, 2019 at 9:42:37 AM UTC-7, Shuo Xiang wrote:

LC0 could not be defeated by ANY ENGINE in either premier league or the cup.

If that's not perfection, I don't know what is.

The only engine capable of defeating LC0 is stockfish, and only during the SuFi, where it itself was subsequently defeated by a margin of 10.

On Sun., Oct. 27, 2019, 10:34 a.m. Jim Glass, <jimgl...@gmail.com> wrote:

Kingscrusher on Leela v Stockfish #1.

https://youtu.be/T85mELZeqso

A lot can happen Ina year.

--
You received this message because you are subscribed to the Google Groups "LCZero" group.

To unsubscribe from this group and stop receiving emails from it, send an email to lcz...@googlegroups.com.

[Shuo Xiang]

Yeah I just saw that. Like literally the minute after I sent the prev post. (Cuz I realize I'd forgotten to check on the progress of the Cup today after what appears to be an eternal battle btw alliestein and Houdini yesterday)

Bummer!!!

To unsubscribe from this group and stop receiving emails from it, send an email to lczero+un...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/f2b97a5f-119b-4b10-828e-e4142c7c2006%40googlegroups.com.

[Paul Berger]

So, as I said and not only from the idealized environment of computer chess but everywhere, AI IS STALLING once again. 

We got sophisticated pattern recognition but not more. Still struggling on transfer learning, deep reasoning and overall unsupervised progress.

We all be dead before being served at home by a "Rosie", sad but true :-(

To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/f2b97a5f-119b-4b10-828e-e4142c7c2006%40googlegroups.com.

[Shuo Xiang]

I myself am still confident about the technological singularity...

To unsubscribe from this group and stop receiving emails from it, send an email to lczero+un...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/36df2ed6-8410-48d9-b26c-a350c9a557e0%40googlegroups.com.

[Gokhan M]

AI isn't stalling. It probably never did, just people thought so vs their projections. Proof of today's progress? See glue leaderboard and stanford squad data sets.

To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/CAFJZE7jKwNfqW1N9Deay9MO-2GtDGXzHRUvg5e8QRdb%3D90DzJA%40mail.gmail.com.

[Gokhan M]

You might argue Leela is stalling. Yeah that's kinda true, given it takes 6 months+ to develop a new network from scratch. My suggestion to the developers would be trying to be incremental (ie build on top of existing networks). Vs learning everything from scratch. (That is unless you have deepmind's resources :)). I really doubt if the current process (of building two new networks every ~6 months) will yield anything. 

[Cary Knoop]

You can't really do that, it's the "you can't teach an old dog new tricks" or even worse it starts forgetting what it learned before!

On Sunday, October 27, 2019 at 11:33:45 AM UTC-7, Gokhan M wrote:

You might argue Leela is stalling. Yeah that's kinda true, given it takes 6 months+ to develop a new network from scratch. My suggestion to the developers would be trying to be incremental (ie build on top of existing networks). Vs learning everything from scratch. (That is unless you have deepmind's resources :)). I really doubt if the current process (of building two new networks every ~6 months) will yield anything. 

On Sun, Oct 27, 2019, 11:16 AM Gokhan M <gokhan...@gmail.com> wrote:

AI isn't stalling. It probably never did, just people thought so vs their projections. Proof of today's progress? See glue leaderboard and stanford squad data sets.

On Sun, Oct 27, 2019, 10:15 AM Shuo Xiang <shuo...@gmail.com> wrote:

I myself am still confident about the technological singularity...

To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/36df2ed6-8410-48d9-b26c-a350c9a557e0%40googlegroups.com.

--
You received this message because you are subscribed to the Google Groups "LCZero" group.

To unsubscribe from this group and stop receiving emails from it, send an email to lcz...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/CAFJZE7jKwNfqW1N9Deay9MO-2GtDGXzHRUvg5e8QRdb%3D90DzJA%40mail.gmail.com.

[Shuo Xiang]

Wow I worship you Google ML God Gokhan M! 🙇🙇🙇🙇🙇🙇🙇🙇🙇🙇🙇🙇🙇🙇🙇🙇🙇

On Sun, Oct 27, 2019 at 2:45 PM Gokhan M <gokhan...@gmail.com> wrote:

That's actually not true. There's something called warm start (also sometimes called pretraining) where one net is used as an initial state to another. In my opinion, this is by and large superior to starting from scratch. Especially if your goal is to try out new ideas fast.

Basically, init from previous solution. Raise the temperature to encourage exploration. Raise step size. (Hopefully) try out new ideas faster.

PS: I build ML models for a living (in Google of all places).

To unsubscribe from this group and stop receiving emails from it, send an email to lczero+un...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/43f0a007-ce64-4e07-bd59-e0494dd6c060%40googlegroups.com.

--
You received this message because you are subscribed to the Google Groups "LCZero" group.

To unsubscribe from this group and stop receiving emails from it, send an email to lczero+un...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/CAEfYzwkvgNnNGthua3viVB8m0N6FtDr-m-u0rO5kTDWqdNUrwA%40mail.gmail.com.

[Gokhan M]

Peace ✌️

https://xkcd.com/1838/

[Deep Blender]

Not sure what you mean, because I am seeing continuous progress. There has been a lot of progress in language modeling. There has been a lot of progress in video understanding and unsupervised learning is being researched in more and more domains. Just to give one example: It is being used to find the motion flow between frames, while also figuring out the camera properties and the depth, just based on pairs of images during the training.

[Anil Dharan]

Hi

    Maybe just a Conspiracy Theory, but I was recently in a Chat Room where someone was laughing and saying that around 6 months ago, when the second set of AlphaZero-Stockfish Matches were released, Google's AlphaZero Programmers also deliberately released false Information/Data to misguide and mislead the top Leela Programmers and lead them in a wrong Direction ! He said that Google was worried about the rapid Progress that the Free Lc0 was making and wanted to stop it, without appearing obvious about it. Hence this Strategy.

I have no way of proving whether this is Right, but it could explain the lack of Progress in the recent months.

I'm no Programmer, but I would humbly suggest that maybe the Top Leela people should spare a Look at the grassroots or at least from 6 months ago to see whether any subtle but crucial mistake was made in the Progress of Leela, around that time.

Regards

[Deep Blender]

It is not maybe a conspiracy theory, it is definitely a conspiracy theory!

The initial paper was missing some critical details which were later presented in a technical report. If you let people chew long enough on that and you mix in with some hear say, you end up with tons of conspiracy theories. There is no reason why Google or DeepMind should be worried by the rapid progress of others when it comes to Chess. They have pretty much no commercial interest in Chess. It is more or less a side project from them to test their reinforcement learning approaches. This claim makes absolutely no sense and has no basis!

[Jim Glass]

"He said that Google was worried about the rapid Progress that the Free Lc0 was making and wanted to stop it"

Did he also explain why Google cared?

Did you ask?

[Anil Dharan]

Quite probably you are right and Google has no Commercial interest in Chess in itself. But perhaps, just pehaps, they were worried that if lc0 was allowed to progress unfettered and one day became as strong as AlphaZero and making mincemeat of all AB Engines including Stockfish in the process, non-Google, non-chess people would take an interest in the free-for-all lc0, if only to learn and understand the underlying Concepts/Algorithms, so that they could utilize them in Fields unrelated to Chess, Fields that Google WOULD HAVE a Commercial interest in !

                    Remember, we are just seeing things from the Computer Chess perspective, while Google has to see from a much broader Perspective.

[OmenhoteppIV]

Well not quite.. leela itself prove that deepmind is right. Demis hassabis mention leela many times

[Shuo Xiang]

Noooooooooooooooooooooo Leeeeeeeeeeeeeeeeeeela loooooooooooooooooost the CUP!!!!!!!

😱😱😱😱😱😱😱😱😱😱😱😱😱😱😱😱😱😱

On Tue, Oct 29, 2019 at 12:40 AM OmenhoteppIV <khaki...@gmail.com> wrote:

Well not quite..  leela itself prove that deepmind is right. Demis hassabis mention leela many times

--
You received this message because you are subscribed to the Google Groups "LCZero" group.

To unsubscribe from this group and stop receiving emails from it, send an email to lczero+un...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/688fa949-6680-4445-b5c5-49e7d772a5c7%40googlegroups.com.

[Jim Glass]

"But perhaps, just pehaps, they were worried that if lc0 was allowed to progress unfettered and one day became as strong as AlphaZero and making mincemeat of all AB Engines including Stockfish in the process, non-Google, non-chess people would take an interest in the free-for-all lc0, if only to learn and understand the underlying Concepts/Algorithms, so that they could utilize them in Fields unrelated to Chess, Fields that Google WOULD HAVE a Commercial interest in !"

And which of Google's competitors aren't doing that already?

Which really need to think "Oh my God, a good open source chess program!!" to go no-holds-barred on AI development?? Apple? IBM? ...?

Or perhaps Alpha0 embodies the magic key to AI, that other AI researchers can't imagine, and the paper has been rewritten to hide it?

[Stephen Timothy McHenry]

Not to worry, they are playing old Leelas, and she did not lose by much. T60 will fix it soon enough, look at her progress.

On Tuesday, October 29, 2019 at 7:54:03 AM UTC-5, Shuo Xiang wrote:

Noooooooooooooooooooooo Leeeeeeeeeeeeeeeeeeela loooooooooooooooooost the CUP!!!!!!!

😱😱😱😱😱😱😱😱😱😱😱😱😱😱😱😱😱😱

On Tue, Oct 29, 2019 at 12:40 AM OmenhoteppIV <khaki...@gmail.com> wrote:

Well not quite..  leela itself prove that deepmind is right. Demis hassabis mention leela many times

--
You received this message because you are subscribed to the Google Groups "LCZero" group.

To unsubscribe from this group and stop receiving emails from it, send an email to lcz...@googlegroups.com.

[Shuo Xiang]

👍👍👍👍👍👍👍👍👍👍👍👍👍👍👍👍👍👍👍👍👍👍👍👍👍👍👍👍👍👍👍👍

To unsubscribe from this group and stop receiving emails from it, send an email to lczero+un...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/70df9519-791a-43f2-9baf-5a7847d3be4e%40googlegroups.com.