Hi Steven, I'm worried the anthropomorphic ("a breed of intellect that humans have not seen before") & hyperbolic ("supreme insight") statements mislead. Also that the article overstates the ease of generalizing from function approximation to broader categories of problem solving
-
-
-
Thanks for raising these very good counterpoints!
End of conversation
New conversation -
-
-
Beautiful piece! As a pretty strong chess player myself (FIDE 2300 USCF 2400), the stiff AlphaZero comes up with is astonishing. Some of its games with Stockfish are the most elegant have ever seen.
-
Thanks Ron. And “pretty strong” is quite an understatement on your part!
End of conversation
New conversation -
-
-
Very nice essay. It’s rare reading an article in popular media that refers to chess and actually has intelligent things to say. The black box problem is annoying. It would be nice to know if AlphaZero has discovered some as yet unknown principles, or if it’s just Kasparov^n
- End of conversation
New conversation -
-
-
A great read, Steven.
-
Thanks Paul!
End of conversation
New conversation -
-
-
Excellent piece.
-
Thanks
(BTW, I enjoy your tweets about chess!)
End of conversation
New conversation -
-
-
for those interested,
@agadmator analyzes (if that word can still be used) some of AlphaZero's games against Stockfish on YouTube -
Here's one game where AlphaZero's "insight" was quite astonishing - a game
@agadmator calls an "immortal Zugzwang" game (Zugzwang, German for "forced to move", is where a player wishes they could "pass" because any move they make hurts their position)https://youtu.be/lFXJWPhDsSY -
and another where AlphaZero seems to violate opening principles, moving one piece several times in the opening, only to obtain lead in development & play according to "romantic" principles of initiative, attack, movement of pieceshttps://youtu.be/NaMs2dBouoQ
End of conversation
New conversation -
-
-
One key difference between AlphaZero and “AlphaInfinity” is the lack of a clear objective function to maximize.
-
Alphainfinity would be a true General Intelligence, as I understand it. Rather than a system that merely optimizes narrowly, but considers millions of variables and predicts outcomes reliably and interprets those outcomes with 'wisdom'.
-
Right, my point is that none of the current approaches even begin to point to how we’d get to such a system, as they’re based on having an objective function and ground truth to work with.
-
1/n The generalized obj fn is energy efficiency. Under specific thermodynamic energy eff conditions, a system has to learn+predict its environment. Brains are not energy eff because they learn. The learning is a manifestation of energy efficient dynamics.
-
2/n So the goal is how do we design/fabricate systems that satisfy required thermodynamic cond as opposed to computational ones? I think these guys at UCLA are a perfect test-bed to build uponhttps://journals.plos.org/plosone/article?id=10.1371/journal.pone.0042772 …
End of conversation
New conversation -
-
-
Nicely written! But I don't quite agree that problems in cancer or consciousness would lend themselves to the kind of impressive but still deterministic problem-solving that chess benefits from. ML is still not generating models or hypotheses which are hallmarks of creativity.
- 1 more reply
New conversation -
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.