Hacker News Reader

Comment by disgruntledphd2 | Hacker News Reader

Citations needed. You seem to be commenting like this on many topics (wherever its tangentially relevant). I would really like to know why you think this will happen, what you think are the most promising machine learning methods for achieving it, and who you think is going to program these machine learning methods.

I love machine learning don't get me wrong, and I'd like to see Skynet (hopefully friendly) in my lifetime, but the current state of the field does not reassure me on this matter.

View on HN

It's as easy to buy into the church of the singularity as it is to buy into the "We'll have a million customers and $10MM/month in subscriptions within a week of launching" belief after releasing your four-years-in-stealth-mode-with-no-customer-feedback startup.

Beliefs based on other beliefs generate even further out beliefs unhinged from reality. It's not that they aren't true -- they just aren't true yet, and imagination doesn't count as ground truth (because magic doesn't exist).

This is not about the singularity. This is about data vs programming.

And your attacks on the singularity are misguided. You think singularity people are retarded and your anti-singularity possie are holier, yet you don't have a coherent anti-singularity argument.

Any fool can tear down a barn, but it takes a carpenter to build one.

Oh, I certainly don't have an anti argument. We'll end up destroying ourselves, be obliterated by cosmic events, or eventually run civilization from computer simulations.

People have just been saying "real soon now" for 60 years. Obviously we are more "real soon now" than back then, but it's still unclear how soon we can be.

My favorite knock down is the "when computers get fast enough, we'll magically have AGI/hard AI." If that were the case, we could have really really slow versions now.

Theoretical vs. practical architectures matter. Nobody can imagine an iPad's software ecosystem if your model is vacuum tubes and single tape turing machines. But when you hit CPUs, shrinking processes, and GPUs with 6 teraflops on a card, you see how impossible your goals were with a 0.2 MHz room-sized computer with 500k of memory.

If you immediately know the candle light is fire, then the meal was cooked a long time ago.

2011 and 2012 have been watershed years. IBM watson, speech recognition using neural nets beating speech researchers, neural nets on imagenet beating computer vision researchers. The trend is that automated machine learning wins. You can be on the right side of history (automated machine learning) or take opposite side (and be wrong).

I think we're on different pages of the same book. The current stuff is all well and good, but it's the Penn & Teller phase of AI/ML/GMs. We'll dispense with the façade of ability and instill actual ability eventually. I'm just not overly rah-rah about the recent incremental developments. I don't mean to overly pooh-pooh your rah-rah, but my brain is telling my fingers someone is wrong on the internet!

Yeah, that wrong someone is you. You'll have to eat your words within 5 years at the most.

I'll take that wager. I mean that literally--put your money where your mouth is.

Email's in my profile.

You can read up presentations by Jeff Dean and also look at what those companies are doing. Listen to Craig Mundie from microsoft for example. See what's happening in competitions like kaggle or what are state of the art on difficult computer vision challenges like imagenet. Machine learning as used in android speech recognition and microsoft speech recognition has already decimated speech recognition researchers and computer vision researchers are next (in the sense that automated machine learning takes over feature engineering). Read all the big data hoopla. There's your citations. It's not all that difficult to figure this all out if you look into it.

Reply on news.ycombinator.com