obastani

Is going from recording to playback a trivial step? The article doesn't make this clear, but if this step is nontrivial, then Edison's contribution still seems like a big deal. The article even seems to imply that the format he recorded the sound in was not amenable to playback.

It's a classic case of prisoners' dilemma. Taking your kid out of daycare isn't going to help the planet very much, but could hurt your kid a lot.

I interpreted much of the article as being advice for the "hunter", who is not necessarily worth connecting to. For example, do your research beforehand, keep your requests short and reasonable, don't make false pretenses, don't oversell yourself, etc.

I think they are saying the volunteers performed no better than chance, i.e., volunteers are no better at guessing real vs. clone compared to randomly guessing. It's a really good outcome for the clone.

This doesn't really answer the question -- how come there are usually only a few clearly defined species that descended from this common ancestor? Why isn't there a spectrum?

I don't know anything about the phylogenetic tree of humans and our cousins, but I assume that at some point there was a spectrum (e.g., Neanderthals), but at some point the others got wiped out, without humans splitting any further. Perhaps this is due to the global mobility of humans? What do biologists think?

The point I was trying to make isn't that the NSA doesn't have secret programs (like spying on American citizens), it's that these programs didn't come about through the scheming of a few people. Maybe it did, but I think it is more likely that the organization organically pushed the limits of what it thought it could get away with (and I think the article would agree).

I'm not convinced that the NSA counts as a counterexample. From my very limited knowledge, it appears that the NSA's programs are known an approved by a reasonably large swath of government. I think the response of the article would be that, rather than being the result of scheming by a small group of people, the NSA programs arise from a combination of initial government consensus and subsequent expansion of powers. For example, Obama could have stopped programs like PRISM when he became president, but chose to continue them.

The point I got from the article isn't so much that secret and far-reaching programs can't exist. It's that these programs do not come into being based on the will of a small group of people, but rather based on a complex bureaucratic process in which many parties partake.

The most elementary explanation I ever got for why e^(i pi) = -1 is from the definition

e = lim_{n -> infinity} (1 + 1/n)^n

First, we can generalize this to

e^z = lim_{n -> infinity} (1 + 1/n)^(z n) = lim_{m -> infinity} (1 + z/m)^m

using the substitution m = zn. Therefore,

e^(i pi) = lim_{m -> infinity} (1 + i pi/m)^m

Now, converting the complex number (1 + i pi/m) in terms of polar coordinates (r, theta) yields

r = (1 + pi^2/m^2) ~ 1

theta = sin^{-1}(pi/m) ~ pi/m

Since the product of two complex numbers is

(r, theta) (r', theta') = (r r', theta + theta'),

we have

(1 + i pi/m)^m ~ (1^m, m * pi/m) = (1, pi) = -1.

I suppose you could take [0,1] \cup Z, and put a uniform measure on that set.

Nitpicking, but you can have a uniform distribution on the unit interval, which is infinite :). What you cannot have is a uniform distribution on an unbounded set.

Edit: As pointed out, the correct assumption is not unboundedness, but having infinite measure (since then no normalization constant exists). I thought they were equivalent, but a simple counterexample is [0,1] \cup Z.

How is the self driving car going to know it can't drive on that road? I imagine the argument is that the self driving car won't notice that the signs don't make sense (like a human would) and causes an accident.

What innovations do you have in mind? LSTMs and convnets were both invented in the 90s. Stochastic gradient descent, momentum, etc. have been around for decades. I feel like it's quite the opposite of what you are saying -- in the 2000s there was a lot of innovation, but it mostly turned out to be unnecessary.

What has improved is the variety of architectures, but I don't think there has been anything fundamentally new after LSTMs. The only exception I can think of is generative adversarial networks, and even ideas behind adversarial training have been around in other domains for many years now.

How did the proof you learned go? As far as I know, calculus is required to rigorously prove this fact.

The entire argument (both parent and the article) is misguided at best -- a large fraction of Asians coming to the US are here for higher education (College, Masters, or Ph.D.). Therefore, many Asian immigrants are the cream of the crop, which skews the income distribution of Asian Americans significantly upwards. The comparison is simply not meaningful.

"Spooky action at a distance" does not violate locality in the sense that information does not travel faster than the speed of light.

I'm curious -- does this kind of regulation apply to a site like WebMD as well?

This seems like a cool idea! Are there tools to automate the translation from C to SaferCPlusPlus?

Do other car companies allow cars to be controlled via an Android app? Some people might see that as a feature, but to me that seems like a pretty bad product decision precisely because Android is so vulnerable.

Maybe it wouldn't keep a cure, once discovered off the market, but perhaps it could reduce investment in trying to find a cure to begin with? I can imagine that this wouldn't require much collusion, and there's nothing illegal about not investing in finding a cure. Even from a moral standpoint, it's a tougher argument to make.

I think this is the link you want:

https://en.wikipedia.org/wiki/List_of_aircraft_carriers_in_s...

Something just struck me -- one day, there's probably going to be a U.S.S. Donald J. Trump.

I think the grandparent's point is that it doesn't matter whether it's technically possible for us to prevent climate change, it's politically impossible even with a Democrat as president. Which I hope is not true, but with Trump in office it seems all the more probable.

The fact that the electoral college exists probably distorts the popular vote. People in non-swing states may not vote in the same proportions if the president were decided by popular vote. Of course, this trend could go either way.

Maybe cars sit idle 90-95% of the time, but isn't it the same 90-95% of the time? I would imagine the peak utilization is much higher. Furthermore, if self-driving cars become a reality, then I imagine utilization would dramatically increase. Now that I can just send my car to pick up the grocery, I'm much more likely to send it on random errands.

It's impossible to guarantee that your theorem prover is bug free -- for example, if it has a bug that causes it to always return "correct", then it proves its own correctness, even though it is clearly wrong.

This is not to say that self-hosting isn't useful, since most bugs are not likely to be of this form.

It's not at all the same -- self-driving cars can risk the lives not only of the passengers, but of innocent bystanders. Several commenters have pointed out problems such as rain, snow, darkness, etc., but there are more fundamental problems. For example, self-driving cars based on machine learning may be susceptible to adversarial attacks that cause the car to behave unpredictably [1,2]. How will self-driving cars know how to react when things go wrong, e.g., the stop light is broken? What about construction?

I'm not saying humans react perfectly in these situations, but self-driving cars are rule-based, and have zero ability to adapt to unanticipated situations. I'm not arguing that Tesla shouldn't push forward. However, I believe that so far, Tesla has demonstrated a lack of concern for user safety. Companies like Google, Mercedes-Benz, etc. have tech at least as good as Tesla's, but have been much more cautious about deploying it.

[1] https://arxiv.org/abs/1606.04435

[2] https://arxiv.org/abs/1312.6199

AI/ML algos cannot magically guarantee safety, for example, deep nets are known to be very un-robust: https://arxiv.org/abs/1312.6199. We are not yet at the point where AI/ML based algos can be guaranteed to be safe.

Isn't Hoffman's argument pretty mundane?

The conclusion seems pretty obvious; we only see light in a certain (relevant) spectrum, we only taste certain (relevant) chemicals, we only hear sounds in certain (relevant) frequencies, etc., because it doesn't help us (and in fact distracts us) to perceive more information.

Also, it seems like a big aspect he's missing is that within each relevant spectra, we've gotten pretty well tuned to reality. And that's simply because oftentimes, fitness is highly correlated with perceiving reality. For example, seeing all physical objects realistically is important for survival, since if you can't see perceive certain physical objects, then predators/prey may take advantage of the failure (e.g., zebras blending in with tall grass).

Is there some greater point I'm missing?

I don't think that's as surprising/chilling as people make it out to be. Let's say I run a compression algorithm (e.g., Huffman coding) on a billion strings, and a very common substring is "hello, world". I wouldn't be surprised if the compression algorithm "learned" to compress "hello, world" to a single bit 0. There are a few big jumps to get to cat faces (e.g., the representation has to be approximate), but I don't see why the idea is fundamentally different.

RL can in principle recover necessary subsystems, but you can't tell if that has actually happened. A human designed system may not be as comprehensive asymptotically as a learned system, but the designer can ensure that critical portions of the state space are covered.

I might be misunderstanding the article you linked, but it seems very misleading to me (though I agree with the conclusion that the ProPublica article is misleading). They claim to conclude:

"The predictor is probably not biased against any particular race - the race_factorAfrican-American:score_factorHigh term is not statistically significant. Or, as ProPublica puts it, it's "almost statistically significant"."

I don't think you can conclude that the predictor is probably not biased against any particular race -- you can only use significance to reject the null hypothesis, not prove the null hypothesis (especially since the significance is still somewhat high). Am I misunderstanding the claim in the article?