Machine intelligence, part 1 • Hacker News Board

Re: Machine intelligence, part 1

#131

Post by upquark » Wed, Feb 25, 2015, 8:57 PM UTC

fchollet wrote: post #41

I have yet to meet a serious AI researcher who worries about AI ending the human race. At every AI conference I've been to lately, some guy would inevitably ask the same question to the speaker: "do you think we should be concerned by superhuman AI?". The answer was always the same, for instance from Andrew Ng at the Deep Learning Summit a few weeks ago: "dude, stop it. That's such a distraction". > One of my top 4 f…

> I have yet to meet a serious AI researcher who worries about AI ending the human race.

As an industry practitioner of machine learning / data science, I believe AGI poses a genuine risk to humanity.

Having said that, what I do for a living is of little relevance. People are pretty terrible at predicting the future (see late 19th / early 20th century predictions of year 2000, with food replaced by pills, etc). Unless someone has put enough thought and research into it, their predictions about the future of civilization are likely to be worthless regardless of their academic credentials.

Re: Machine intelligence, part 1

#132

Post by freyr » Wed, Feb 25, 2015, 9:01 PM UTC

It’s very hard to know how close we are to machine intelligence surpassing human intelligence.

I feel comfortable stating there is no evidence that the current trajectory of machine intelligence (e.g., developing a set of tools for optimization of mathematical functions on digital hardware) is bringing us any closer to sentience. So in that sense, Sam seems misinformed.

Of course, there's always a remote chance of a fundamental, revolutionary breakthrough in our understanding of AI or human intelligence (or anything), but it would represent a complete departure from our current state-of-the-art, not the mere evolution of our current progress. So in the sense that "it's possible," Sam is right -- in the same sense that it's possible teleportation or interstellar flight is right around the corner. When weighted by probability, it's not a pressing issue in my mind and should remain relegated to science fiction for now.

Re: Machine intelligence, part 1

#133

Post by Sakes » Wed, Feb 25, 2015, 9:03 PM UTC

fchollet wrote: post #41

I have yet to meet a serious AI researcher who worries about AI ending the human race. At every AI conference I've been to lately, some guy would inevitably ask the same question to the speaker: "do you think we should be concerned by superhuman AI?". The answer was always the same, for instance from Andrew Ng at the Deep Learning Summit a few weeks ago: "dude, stop it. That's such a distraction". > One of my top 4 f…

I don't understand.

> Since you are so knowledgeable about the unknowable

If knowing the unknown was a requirement to exploring it through playful thought, forming opinions and refining them over time through discussion, reflection and if possible testing, well... we'd still be chunking rocks. I think it is unfair to compare expressing concerns about the implications of AI to debating the gender of angels. No one is trying to build angels...

> for instance from Andrew Ng at the Deep Learning Summit a few weeks ago: "dude, stop it. That's such a distraction".

Ya, this makes sense in the context of the summit. But, applying it to a guy's blog? It is a blog...

> Yes, AI is something we should probably be worried about.

Agreed!

> "The history of technology advancing has been one of sigmoids that begin as exponentials."

Sure, but will we be irrelevant before the slope changes direction?

Anyways, I share your fear of advancements in technology being used to secure entrenched power structures. And yes, if it happened it would come before the singularity. But I just can't imagine how anyone wouldn't be afraid to be part of a species that had stopped evolving while a new competing "life" form was evolving at rates that known history has never seen.

Re: Machine intelligence, part 1

#134

Post by convexfunction » Wed, Feb 25, 2015, 9:04 PM UTC

datashovel wrote: post #130

This is a very interesting problem. A few things that I rarely see discussed, but I think are valid points. We, humans, are a byproduct of similar evolutionary processes that (at least in theory) the computers would go through. We as a species seem to have settled on a few interesting "rules of thumb". We value intelligent life, in all its forms. We value collaboration and have found that democratization is the best…

Or maybe that's anthropomorphizing and they won't! Let's just throw the lever when the time comes and find out, and urge people to not spend any time thinking about it too hard until then. This is a good and reasonable and safe strategy that isn't at all guided by an underlying desire to assert dominance over weird outgroup people who might threaten your own status.

(Not aimed at you, sorry. I do have a lot of trouble understanding the "maybe the universe isn't fundamentally unsafe" worldview though.)

Re: Machine intelligence, part 1

#135

Post by convexfunction » Wed, Feb 25, 2015, 9:06 PM UTC

freyr wrote: post #132

It’s very hard to know how close we are to machine intelligence surpassing human intelligence. I feel comfortable stating there is no evidence that the current trajectory of machine intelligence (e.g., developing a set of tools for optimization of mathematical functions on digital hardware) is bringing us any closer to sentience. So in that sense, Sam seems misinformed. Of course, there's always a remote chance of a…

I didn't see anything about sentience in the article. More generally, sentience has nothing to do with AI safety; whether or not the thing has qualia has very little to do with what the thing may or may not do to our civilization.

Re: Machine intelligence, part 1

#136

Post by choppaface » Wed, Feb 25, 2015, 9:11 PM UTC

neuralk wrote: post #116

sama wrote: post #105

Earlier quoted context omitted.

though i dropped out, i studied AI in college. i also worked in andrew ng's lab. as a current grad student, why do you believe whatever makes us smart cannot be replicated by a computer program?

I never said that. I think karpathy (also an AI researcher) summed up my feelings, particular the Ryan Adams quote: https://news.ycombinator.com/item?id=9109140 edit: apologies about the 'no background' part

Nice link. I also did AI in grad school, and I firmly agree that posts like sama's are, as Ng says, "a distraction from the conversation about... serious issues." The OP is much much more aimed at marketing a plausible future of AI than producing any sort of rigorous prediction. It doesn't even matter if the OP predicts correctly; the post doesn't contribute anything substantially meaningful. I'm sad to see Sam spend so much of his precious time and energy on this post.

Re: Machine intelligence, part 1

#137

Post by AndrewKemendo » Wed, Feb 25, 2015, 9:15 PM UTC

Tossrock wrote: post #77

fchollet wrote: post #41

I have yet to meet a serious AI researcher who worries about AI ending the human race. At every AI conference I've been to lately, some guy would inevitably ask the same question to the speaker: "do you think we should be concerned by superhuman AI?". The answer was always the same, for instance from Andrew Ng at the Deep Learning Summit a few weeks ago: "dude, stop it. That's such a distraction". > One of my top 4 f…

I have a feeling this is a leakage of Eliezer Yudkowsky's opinions into tech entrepreneur circles via his friendship with Peter Thiel.

This is exactly what it is, which is so terribly terribly disheartening as Yudkowski is at best a marginal sci-fi writer with and outsized ego and some math chops. He has an almost cult like following over at less wrong.

His writings don't do anything for fundamental AI research but just handwave a lot of philosophical arguments.

Re: Machine intelligence, part 1

#138

Post by bcheung » Wed, Feb 25, 2015, 9:17 PM UTC

One possibility that I think is often overlooked is that as humans develop technology they use it to increase their leverage. It is my belief that as we develop neuroscience and AI then we will augment our own brains.

In this case, humans will gain intelligence linear to non-human intelligence.

The bigger problem from SMI, and it relates to all forms of leverage, even education, capital, the Internet, and machinery; is that it creates inequality and concentrates power. This is not necessarily a problem in and of itself because leverage usually means improving the standard of living but it can cause social and political problems.

I see increasing class warfare as the biggest threat of SMI.

Silicon Valley is going to become the new capital of power in the world and it will create an us vs them scenario.

Re: Machine intelligence, part 1

#139

Post by fchollet » Wed, Feb 25, 2015, 9:17 PM UTC

Tossrock wrote: post #77

fchollet wrote: post #41

I have yet to meet a serious AI researcher who worries about AI ending the human race. At every AI conference I've been to lately, some guy would inevitably ask the same question to the speaker: "do you think we should be concerned by superhuman AI?". The answer was always the same, for instance from Andrew Ng at the Deep Learning Summit a few weeks ago: "dude, stop it. That's such a distraction". > One of my top 4 f…

I have a feeling this is a leakage of Eliezer Yudkowsky's opinions into tech entrepreneur circles via his friendship with Peter Thiel.

I think you nailed it.

"AI doom" is a powerful psychological trope for people in technology, and much like the hypothetical rogue super-AIs, it has gone out of control.

Re: Machine intelligence, part 1

#140

Post by datashovel » Wed, Feb 25, 2015, 9:19 PM UTC

convexfunction wrote: post #134

datashovel wrote: post #130

This is a very interesting problem. A few things that I rarely see discussed, but I think are valid points. We, humans, are a byproduct of similar evolutionary processes that (at least in theory) the computers would go through. We as a species seem to have settled on a few interesting "rules of thumb". We value intelligent life, in all its forms. We value collaboration and have found that democratization is the best…

Or maybe that's anthropomorphizing and they won't! Let's just throw the lever when the time comes and find out, and urge people to not spend any time thinking about it too hard until then. This is a good and reasonable and safe strategy that isn't at all guided by an underlying desire to assert dominance over weird outgroup people who might threaten your own status. (Not aimed at you, sorry. I do have a lot of troubl…

No, I completely agree. I did mention in the comment that I would of course be cautious. And I think it's very plausible (at least there's no reason I can think of to disagree) that the universe is a fundamentally unsafe place.

Even though it may be anthropomorphizing to an extent, there are alot of attributes humans have that are not so desirable, such as mortality. As far as we know computers (sufficiently replenished) have a theoretically infinite lifespan. I completely agree that there's no way to determine what kinds of effects those differences will have on how machine intelligence would evolve.