Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[Max™]

They're still going to serve you ads, but I think the idea is that you're saying "I'm going to go out of my way to avoid products if you try to serve ads based on my activities" or at least it is in my case. If I like your site I'll turn off adblock so you get revenue, but I'm keeping the userscript so the ads are just blank gray panels, and if your ads make it through that then I'm deliberately going to avoid that product and hunt down whatever loophole allowed it because fuck you, but that's also a reflex from growing up when pop-up blocking wasn't a thing and autoplaying audio with no mute button was totally acceptable.

[wierd]

You make me feel old. I remember when the internet was mostly email, newsgroups, and IRC.

I remember when advertisements on the internet were not a thing.

[Max™]

My first internet experience was compuserve back in 1992 or 1993, which made seeing AOL commercials with "you've got mail!" surreal later on.

[Starver]

I remember when advertisements on the internet were not a thing.

UN-altered REPRODUCTION and DISSEMINATION of  this IMPORTANT information is ENCOURAGED

[Max™]

INSEMINATION of this ENTOURAGE or DISCUSSION of ENTOURAGE or OBSERVATION of ENTOURAGE is, on the other hand, not CONDUCIVE to REPRODUCTION, but yeah, that's why I'm so virulently anti-advertising. No-ads+paying for a service to ads everywhere+paying for a service was a big fuck you.

[feelotraveller]

I remember when advertisements on the internet were not a thing.

Right wing commentators would now lead us to believe that advertising is responsible for the existence of the internet.  ROFL.

[Reelya]

Advertisements made the internet great again, after it had been usurped from it's free-market creators by the communists at university colleges.

Even now, communistic internet "standards bodies" are holding back the development of the internet. We need competing web standards, with internet sections gated according to which standards they use. This maximizes commercial opportunity (providing data translation services between the various standards and areas of the internet) and provides a fertile competition, which will benefit the consumer.

[Max™]

I feel like posting that America, Fuck Yeah! video again.

[McTraveller]

Where'd all the AI talk go? 

Regardless of method (neural net, genetic algorithm, etc.) the interesting limitation of modern AI is that it is still subject to the limitations imposed on it by the selection of the input set and the final "value" criterion that indicates if the mapping of inputs-to-outputs is "desirable" or "undesirable".

For instance, it's amusing to, say, take the Go neural net and feed it the binary representation of, oh, I dunno a game of DF and see what outputs it generates.  Can those AIs do things like "uh... that input makes no sense, yo, stop feeding me nonsense!" or will they just simply try to interpret it as a new in-context input?

[itisnotlogical]

From my understanding of machine learning, whatever input you get depends on how you've trained it. If you haven't trained the neural net, it might as well be random. The idea is that you give it a set of inputs with known outputs, and tell it "No, that's wrong" until it starts giving you the results you expect.

[itisnotlogical]

There actually was an AI that did very well at several NES games. Of course the difference is that the AI isn't "playing" the game as we understand it by looking at the screen, it's reading values out of memory every frame and directly activating controller inputs. It actually doesn't do very well at Mario, but is god-tier at Gradius.

[helmacon]

Just because it isnt incredibly useful in the way you imagine dosnt mean it's not incredibly useful in some other way. Google's project magenta has done some pretty cool stuff with classifying and ordering sounds in ways that make sense to us, but would be difficult or near impossible to do ourselves.

[Reelya]

Ipsil, that's just some guy's single home-brew program, yet it's able to go in blind into a wide variety of game types and do ok at them. e.g. it kicks the ass of a side-scrolling space shooter or pinball without needing to be told anything about how those genres work. That's far more universal than you're giving it credit for.

Notice that he mentions that for Mario he needed to tell it extra information such a "prioritize extra lives" etc, but that then worked for every mario level he threw it at. That also implies he didn't need to do such things for the other games.

The video represents a single guys homemade AI program being thrown up against games it's never been told about. You cannot infer anything about the limits of AI/machine learning as a whole from that.

[Reelya]

I'm not treating it like magic, I proposed using it for an extremely limited use case. Then you made a bunch of objections which looked a lot like changing the goalposts constantly, like you were gunning to prove it wrong no matter what, even if there's no logical connection between the varied points you're making.

Remember what I original proposed, which was using it to tune the jump physics for an arcade game by co-evolving the challenge parameters against the jumper AI, so that there was a controllable rate of failure/victory based on your accuracy with the controls. None of that is treating machine leaning like "magic", you're just plainly strawmanning now. Adding insult to the other fallacious arguments you've already deployed.

[Reelya]

The other one that was brought up was Starver's vehicle design challenge, with the trillions of possible loadouts. I pointed at Eurisko as a type of solution that can be applied to that sort of thing. You mutate a set of heuristics that are used in a car-designed engine, instead of directly mutating car designs themselves. That's not calling it "magic" either, it's just realizing that you can mutate car design rules instead of directly mutating the cars. If you mutate the thing (car) directly then your system is too low-level, and it doesn't learn anything about "what makes a good car", whereas if you mutate sets of design criteria then that system can then build good cars, even with parts it's never heard of.