It's interesting that AlphaZero was finally applied to a different game, though. I wonder what architectural changes they had to make. I've read that pure MCTS isn't that good at playing Chess. How true is that?
2,750 karma · joined February 2, 2011
It's interesting that AlphaZero was finally applied to a different game, though. I wonder what architectural changes they had to make. I've read that pure MCTS isn't that good at playing Chess. How true is that?
We're seeing the reversal of PC revolution and most people don't seem to notice or care.
I'm not sure how we got to the point where something that worked without minification on year 2000 networks no longer works without per-processing on year 2017 networks.
Not sure why progressive JPG isn't the default. It's not any larger.
This is an interesting comment. I watch couple of streamers regularly, but not in real-time. I just watch their content on YouTube.
Why?
Some of them actually put good effort into creating interesting in-game situations and running entertaining commentary. It's less of a reality show and more of a comedy/talk show. Maybe.
Some of them are very good at the game, and watching the stream is an opportunity to learn something new. If the game is competitive, it sort of makes sense.
But overall, I agree that streaming and YouTube are becoming the new TV. Maybe the motivations I outline above are somewhat pathological in the sense that they don't justify the time spent on such simplistic entertainment.
A lot of videos are about competitive online games, and come to think of it the direct experience in those games isn't always great. For example, when you're watching a game on YouTube it's effectively per-selected. This way you can "skip" the games that are boring or have something broken or have annoying/rude players. You can cancel/pause at any time without loosing rank or annoying other people. In some ways it's a sad state of affairs. Games should be entertaining in their own right.
It seems today people write off those demos as showy, shallow and misleadingly optimistic. But if you think about it, the exact same thing can be said about modern ML demonstrations. A lot of them hint at nearly human-level intelligence, but that's often achieved through careful restriction on the problem domain and by choosing the most successful examples.
Considering that old-school AI was operating on laughable hardware with tiny hand-made datasets, with no crowd-sourcing options... I wonder, did attempts to generalize some of that research fail because the approach was fundamentally flawed, or was it because such efforts themselves weren't as good as initial projects?
I mean, the article on Wikipedia talks about "more realistic level of ambiguity and complexity", but the demonstrated domain is already more complex than 90% of all programming problems I solve on day-to-day basis.
* * *
I though Minsky also did something similar, but with physical blocks and a robotic arm, and without the NLP interface. Anyone knows about that? Or am I making it up? I was looking for info about this recently and couldn't find anything.
You can't base your reasoning about something like this solely on what's happening in some countries right now. That's how SSN mess started in the first place. They weren't intended as a global identifier, there were concerns that they will be used as such, but those concerns were ignored.
Also, I'm not convinced that spinning up a new ID system is necessary to solve the issue at hand.
This is a very confused statement. Fist, since we are talking about image recognition we are - by definition - talking about vision in the spectrum that can be captured by a digital camera and encoded in a typical image format. Second, there definitely is such a thing as essential property. It is a matter of correlation with reality, as well as internal consistency. For example, plants are green and have leaves because of the way they use sunlight. So permuting color of all leaves in the picture is fundamentally different from permuting luminosity of some random pixels.
One counter-point, though.
> I make no secret of hating the mouse.
If you look at the original uses of the mouse it was great. Especially in systems like Xerox Star. Star allowed people to perform complex tasks with almost no learning curve.
https://www.youtube.com/watch?v=Cn4vC80Pv6Q
(Note how they weren't shy of using keyboard either. There are dedicated hardware buttons for standard commands like copy, find, repeat, "properties" and even for common text editing actions. Meanwhile, our keyboards don't have dedicated keys for undo, redo, cut, copy and paste - operations that are used in almost every application today.)
Trouble is, we lost most of the driving ideas behind Xerox-style interfaces. Using a predefined set generic, powerful commands. Object-oriented UI. Uniformity of representations. Modern system have those things only as vestigial traits and in very limited contexts.
I don't think there were any quantum leaps in conceptual UI design since Xerox Park times. There were some minor improvements in very specialized apps and significant regression in software that's used universally. For examples, phones and tables almost completely lost drag-and-drop functionality and generic file UI.
For example, drag-and-drop is a very powerful concept, because it allows you to perform actions by combining things you already know about - and those "things" figure out how to interact in the best way possible. So, for example, instead of having N "Print" buttons in N applications you can have a singe drag-file-onto-printer-icon action that does different things based on the type of the file. [BTW, this is also the key idea behind the original notion of OOP.] Unfortunately, that's not how it works in modern UIs. They don't use either the keyboard or the mouse to their full potential.
Universality is the real hard problem of AI. In the long run, a mediocre AI that does a lot of different things is far more useful that most targeted "superhuman" AIs. Most domains simply don't require better-than-human performance, but could still reap tremendous benefits from automation.
Statements like these always make me wonder why certain obvious things weren't tried. If it's so generic, why wasn't it tried on Chess? Or was it tried, failed to impress and thus didn't make it into the press release?
This is a big problem with all these public discussion on AI. Almost no one speaks about algorithm failures. I haven't seen a single research paper that said "oh, and we also tried algorithm in X domain and it totally sucked".
How many people worked on Xerox Alto? They had to do everything from scratch, including bootstrapping their own tools and they had far worse computing capabilities/hardware available to them at the time.
All attempts in modern AI. Most interesting research from the 60s and 80s wasn't based on pattern recognition.
Funny that you mention it. Loosing form data on page refresh or clicking "back" button was a problem on the Web until browsers started keeping form state along with page history. Suddenly millions of websites got this new feature without actually doing anything differently. But that was only possible because page transfers were semantic. Frameworks of that time that tried to manage their own form state (like ASP.NET WebForms) didn't have real page transfers, so they didn't get this "upgrade" and currently are significantly worse off in terms of usability.
I think there is a lesson here.
This was fixed in browsers years ago when they started keeping form state along with page history. I can refresh this page, go back and forth in history, and this text box will keep my comment. Ironically, websites that tried to be clever about form state management and faked page transfers (like ASP.NET WebForms) didn't benefit from the fixes and currently are significantly worse off in terms of usability than websites that used vanilla forms.
I think there is a lesson here.
Programming with Agents: http://alumni.media.mit.edu/~mt/thesis/mt-thesis-Contents.ht...
Here is a short paper with a clear description of an ingenious idea.
Engineered Robustness by Controlled Hallucination: http://web.mit.edu/jakebeal/www/Publications/NIAI-2008.pdf
I like the simplicity of it. Most CS researches seem to be afraid of describing things that are simple, even if those things are non-obviosu and valuable.
Also, a lot of research in this field is done through working proof-of-concept systems, so there is no point in implying that those designs don't work "in practice".
And while recorders present some challenges, manually coding UI-level tests make the maintainability issue an order of magnitude worse.
The best way to make web UI tests stable is to involve engineers and make sure that the pages are structured in the way that make automation and validation easy. Usually the changes required for this are trivial. When they're not, it often points out something in the UI that doesn't make sense anyway, like having no confirmations for user actions.
And in any case, automated UI tests are still better than manual UI test, which is what a lot of people are stuck with.