I like this article a lot, but I would ask one or two things.
First, can we get a bit more quantitative? This article is mainly scientifically-minded anecdotes because of the lack of quantification or specificity. How many decks use the 8/8 split, e.g.? You say 'most', but this isn't duplicatable.
Second, does this apply to Limited? Limited, by its nature, has quite a bit of quantification of knowledge. E.g., play 17-18 lands, you want more than 12-13 creatures, removal is key, etc. How does the act of drafting change the knowledge gained from this article? Is it completely different, a modified form, or true in full force?
Inquiring minds want to know.
I'm curious as to the extent people think that the seeded pack has distorted the view of the environment.
If Tutelage never won a game against you, I have difficulty understanding how it 'did some work'.
The reason I ask is because the above makes me think that the network 'learned' that Lightning Helix is a 'red' effect, when it is actually a 'white/red' effect. Given the realities of Magic design, this might point to the need for a tighter encoding of mana costs. E.g., you may need to code RB as a single character, rather than two separate mana symbols. This would allow the network to more easily learn that certain effects belong to interactions of colors rather than just one. I am afraid of the lack of learning it might make overall, but Lightning Helix's text should never appear on a monored card.
I dunno... The issue is one of space, I guess.
Mustard: Well, from a philosophical perspective, you might want to ask why it is that humans do have consciousness to the degree that we can design things such as, e.g., the forum we're using, and yet no other lifeform, to our knowledge, can do this.
It can't be a question of just the sophistication of our brains, though that certainly assists. Even with my limited knowledge of NNs, I can get the 'equivalent' of very robust 'thinking' out of a NN that is very simple.
What I think is missing, really, is the self-spurring nature of things. It is fairly straightforward to use a NN to do what we have done, here. But a NN doesn't 'learn' in the sense that we do. There is no consciousness internal to the NN directing it.
Let us use your example of language acquisition. Assuming that most people primarily use words to think with, the question is not how did you learn so much as why. My own anecdotal evidence (from having a large extended family and thus many nieces and nephews) is that children acquire language to communicate. It is something they want to do, and having made the leap that the sounds they're hearing are a way to communicate. Once they decide to acquire language, then yes, the method of learning can be approximated with a NN-type approach.
The only exception given for singular nouns are ancient proper names that end is -es or -is, or Jesus. Quite literally.
Now, if they had decided to use this exception, Erebos would probably slide in fairly well.
I have some knowledge of statistics. First question, what is the distribution of?
Also, should we not be able to give it a historical design skeleton (say, for Magic 2014) and see how close it can come to reproducing the set?
I like how Delusion only really showed up in UBR (with one White card with a Black outlier). Is this representative, or did you cherrypick?
Now, another question. Can the NN be used for hole-filling? Let's say that you have a set made (with the NN or not), and you have a certain low number of holes in the set. Could you take your NN, feed it your set, and have it hole-fill for you?
Also, could you have the network make cards with the following ability?
Delusion X - When ~ enters the battlefield, gain X life. When ~ leaves the battlefield, lose X life.
Just wanted to say that this thread is awesome, and I would like to point out that the above card Dragon's Revelation is a completely reasonable card that could actually be in Origins.
Sword of the Pierced Totem is actually completely feasible as well. Well, except for the minor point that I don't think Spell Mastery shows up on non-Sorcery/Instant spells.
I am really looking forward to a created set of cards with the network.
Talcos, I am correct in believing that this will be able to (at some point) take a design skeleton and mechanics for a set and create the set, correct?
But sure, I'll take this Kodama's Reach variant.