Talk:AlphaGo versus Ke Jie

Merge with Future of Go Summit?

Wrong question, as this article is currently only a stub, with nothing to merge. The key question is whether the detailed coverage of the AlphaGo – Ke Jie match should have its own article. See the discussion I have started at Talk:Future of Go Summit#Should the detailed coverage of the Ke Jie match be in its own article? ~ J. Johnson (JJ) (talk) 19:06, 24 May 2017 (UTC)

Not a wrong question. A merger proposal is just to discuss whether a page, even a stub, should be an independent article or be redirected to another one. --Neo-Jay (talk) 16:26, 25 May 2017 (UTC)

Whether this should be an independent article is indeed the essence of the question. But, as I have already said: there is nothing to merge. So it is nonsensical to speak of "merging" specifically. The question is effectively whether material should be taken from the Summit article (more in the nature of a split), and depends entirely on how that article is handled. The result here will derive from the result there, so we ought to have the discussion there. ~ J. Johnson (JJ) (talk) 20:17, 25 May 2017 (UTC)

When a page is a stub and has noting to merge, merging it to another article means simply redirecting it to that article. Wikipedia does not have a special procedure to discuss redirecting a page to another article. The relevant procedure is just Wikipedia:Merging. Therefore it is correct to add Template:Merge to a stub that even has nothing to merge if an editor proposes redirecting it to another article. --Neo-Jay (talk) 00:32, 26 May 2017 (UTC)

No, it's not correct. A redirect usually follows a merger, but (is this too subtle for you to understand?) there is nothing to merge; your fanciful interpretation is idiosyncratic at best. In fact, what I have proposed is a moving content from another article, rather the reverse of a merge. Just for the record, I oppose your this cockamamy proposal for a "merge" that you are so adamantly set on. ~ J. Johnson (JJ) (talk) 21:06, 26 May 2017 (UTC)

Yes, it is correct. If an editor thinks that a stub, which may even have nothing to merge, should not be an independent article and should be redirected to another article, then Wikipedia:Merging is the the right procedure to follow. What else procedure can we follow to discuss this issue? You unfortunately misunderstand the nature and essence of Wikipedia:Merging. And just for the record, it is not I that propose this merge. It is 128.112.16.122 who added Template:Merge to AlphaGo versus Ke Jie at 18:12, 24 May 2017. What I did was just doing him/her a favor to add Template:Merge from to Future of Go Summit at 12:14, 25 May 2017, which is required by Wikipedia:Merging. I have not expressed any opinion on whether AlphaGo versus Ke Jie should be merged (redirected) to Future of Go Summit. You may propose moving content from Future of Go Summit to AlphaGo versus Ke Jie. 128.112.16.122 may also propose merging (redirecting) AlphaGo versus Ke Jie to Future of Go Summit. What you and him/her need to discuss is actually the same issue, and should follow the procedure of Wikipedia:Merging. --Neo-Jay (talk) 02:26, 27 May 2017 (UTC)

• Not merge. Keep Future of Go Summit and AlphaGo versus Ke Jie as separated articles, and move relevant content from Future of Go Summit to AlphaGo versus Ke Jie.--Neo-Jay (talk) 02:18, 28 May 2017 (UTC)

SGF game records

To many readers, SGF records of the games would be more useful than the diagrams currently provided. I'd provide them myself, if I knew where to find them all. Maproom (talk) 08:45, 27 May 2017 (UTC)

SGF records are not readable at Wikipedia pages. I don't think that it is appropriate to add SGF records to Wikipedia articles. --Neo-Jay (talk) 10:18, 27 May 2017 (UTC)

Not the records themselves, but links to the records. (E.g., see AlphaGo_versus_Lee_Sedol#SGF_files.) But note the discussion above: this article may be replaced. ~ J. Johnson (JJ) (talk) 21:38, 27 May 2017 (UTC)

Move discussion in progress

There is a move discussion in progress on Talk:Future of Go Summit which affects this page. Please participate on that page and not in this talk page section. Thank you. —RMCD bot 21:03, 27 May 2017 (UTC)

No cited source states DeepMind had a stronger version on April 7

"AlphaGo Master was actually the second best version that DeepMind had at the time, for it was already in possession of AlphaGo Zero, a version much stronger than the Master version; this can be known by the fact that Nature received their paper on AlphaGo Zero on April 7, before the games with Ke Jie.[7] DeepMind did not reveal the existence of Alphago Zero until the paper was published in Nature in October 2017."

User:Maproom ‎The source does not state that DeepMind's AlphaGo Zero was "much stronger than the Master version" on April 7, as the Wikipedia page currently claims. We don't know what the draft version submitted on April 7 stated, and Zero would obviously have time to progress significantly in power since April 7. Therefore this line should be removed unless it can be appropriately sourced. Rolf H Nelson (talk) 19:44, 24 November 2017 (UTC)

The source (which you can read without payment at https://deepmind.com/documents/119/agz_unformatted_nature.pdf) shows a graph of AlphaGo Zero ("Reinforcement Learning") overtaking AlphaGo Master ("AlphaGo Lee") after about 25 hours of self-training. The source is described in Nature as "Received: 07 April 2017". Admittedly, I am inferring that something described in the source must have occurred before the source was submitted for publication, which might be regarded as Original Research. Maproom (talk) 20:16, 24 November 2017 (UTC)

I'm thinking more that there could have been improvements to AlphaGo between the draft in April and final publication in October. The published version may be a revised version of AlphaGo Zero. If the tweaks were minor, they wouldn't even have to update the methods part of the publication, just the results. Perhaps the initial version of Zero only beat the previously-published AlphaGo, but was worth publishing in Nature due to its ability to start from scratch. Of course, it is also plausible that everyone at DeepMind took their hands completely off the project after April; IMHO we just don't know without a source. Rolf H Nelson (talk) 19:52, 25 November 2017 (UTC)

Why does it matter whether Deepmind "was already in possession of ... a version much stronger" when AlphaGo Master defeated Ke Jie? ~ J. Johnson (JJ) (talk) 23:33, 25 November 2017 (UTC)

To those (like me) who have been following the impressive achievements of the AlphaGo team, it was astonishing to learn that when AlphaGo made headlines by convincingly beat the world's strongest player, the team had already developed a significantly stronger program. Maproom (talk) 08:59, 26 November 2017 (UTC)

So you are "astonished" that development on a newer model did not stop just because the previous model was entered in a tournament? Why are you astonished? And back to my original question: why does that matter? ~ J. Johnson (JJ) (talk) 20:20, 26 November 2017 (UTC)

What astonished me was the rapid progress in the strength of AlphaGo. It advanced from beating Fan Hui to beating Ke Jie in 15 months – that's impressive. Then I learned that when it had beaten Ke Jie, a much stronger version had already existed. That was astonishing.

Why does it matter? It doesn't, really. None of it matters. Editing articles like this is a waste of time. Maproom (talk) 20:41, 26 November 2017 (UTC)

The rate of progress is, indeed, astonishing. Also the reduction of hardware, going from 48 TPUs to only four, and even simplifications in the software. And the results? We are still assessing that. That Deepmind was going so fast they had the new model in the wings before the older one had even strutted its stuff is impressive, but I think rather anticlimatic. And not as suggestive as some others seem to think; just a result of how fast they were going. ~ J. Johnson (JJ) (talk) 22:22, 26 November 2017 (UTC)

IMHO adding a timeline section or article for AlphaGo or Computer Go would be a better way of neutrally conveying what is actually known. The paper submission and the Ke match could both be included on the timeline alongside other events, without making any unsourced speculations about what the timing means. In theory you could ask DeepMind to tweet their development timelines, or scrounge around in interviews in hopes of finding clues, but that seems a poor cost/benefit ratio to me. BTW I've linked to this discussion from the Talk:AlphaGo Zero since they have a similar discussion. Rolf H Nelson (talk) 02:48, 27 November 2017 (UTC)

The only reason (seen so far) for having the acceptance date is in support of this crazy conspiracy theory. Lacking any other basis, inclusion of the acceptance date is at best a trivial, insignificant, irrelevant detail, that verges on non-neutrality. ~ J. Johnson (JJ) (talk) 21:50, 27 November 2017 (UTC)