dplavery92

Maybe Susskind. I'm not sure what definition of "monumental" we're going with, but he's surely "brand name."

Maldacena seems like a good consideration. Maybe Alan Guth.

And perhaps these are the exceptions that prove the rule, being that they are such remarkable cases.

How in the world do you figure $0.50 per mile?

Even if you assume a modest 20 mi/gallon at $3/gal, you're looking at about $0.15 a mile. A $40 oil change every 5000 miles, and a more expensive maintenance job every 10000 miles isn't going to push you past $0.25/mi.

Are you amortizing the cost of the vehicle?

That's an extremely unsatisfying answer.

This is something of an academic factoid that has nothing to do with the practice of training and using neural networks, or with the merits of deep networks that I was describing above.

Shallow feed-forward networks are "universal function approximators" [0] when the number of hidden neurons is finite but unbounded. Of course, the width of that layer grows exponentially in the depth of the deep network that you might wish to approximate [1].

The statement that "[i]t's just that deep neural networks are a lot more practical to train" (emphasis mine) sounds somewhat reductive; it's not only that depth is a nice trick or hack for training speed, but that depth makes the success of deep networks in the past decade at all possible. We live in a world with bounded computing resources and bounded training data. You cannot subsume all deep networks into shallow networks, and shallow networks into SVMs in the real world. So I am pretty sure of how important that distinction is.

And what's more, depth extracts a hierarchy of interpret-able features at multiple scales[2], and a decision surface embedded within that feature space, rather than a brittle decision surface in an extremely high dimensional space with little semantic meaning. One of these approaches generalizes better than the other to unseen data.

[0] https://en.wikipedia.org/wiki/Universal_approximation_theore... [1] https://pdfs.semanticscholar.org/f594/f693903e1507c33670b896... [2] https://distill.pub/2017/feature-visualization/

Significantly, it becomes a simple, 2-layer neural network. The power of the advances of neural networks in the past decade have largely relied on "deep" architectures with many layers. Very deep networks effectively learn the features from the data, rather than learn a decision surface over a set of hand-crafted features, as in learning with SVMs or shallow neural networks.

>The author of the paper is now in a position where they feel they have to defend themselves publicly

Due in large part to the fact that his paper was attached to a press release from Oxford--a point of contention for both Hossenfelder and Sean Carroll.

I get the gag here, but it's not a very apt comparison. The expected return from a roulette spin is really well understood: a loss of about 5% of your bet, albeit with high variance. The future of these investments is less understood. Again the variance is high, but its performance over time is also really good. Perhaps you have plenty of reasons to doubt the future performance of BTC, but it's not like it's a known bad bet.

> Those who read HN are likely aware of the differences and viable usecases.

I am unaware of any non-currency use case where blockchain brings any advantages to bear.

If your ankles connect to your knees, you might be in trouble, friend. :^)

$432k, actually!

https://www.nytimes.com/2018/10/25/arts/design/ai-art-sold-c...

There's something missing here.

How long ago did he finish his program? Is your friend limiting his search only to jobs in academia? Is he capable of getting a security clearance? What's your gauge of "respected"?

There are many government and industry jobs for fresh Physics PhDs right now.

I saw the most recent eclipse from Western Kentucky. I can't speak about all nocturnal creatures, but I can confirm that bats came out during totality. They also seemed to circle like they didn't have their normal sense of direction at dusk. It was pretty surreal.

That's simply incorrect.

https://www.snopes.com/fact-check/volcano-carbon-emissions/

https://www.climate.gov/news-features/climate-qa/which-emits...

https://www.skepticalscience.com/volcanoes-and-global-warmin...

https://www.forbes.com/sites/startswithabang/2017/06/06/how-...

https://www.scientificamerican.com/article/earthtalks-volcan...

What was the application process like, so many years out of school? Did you need old transcripts? Recommendation letters from supervisors?

>The beautiful part is that once home charging is no longer required, people will be happy to charge at home! We are an adorably irrational species.

The problem isn't that people are unhappy to charge at home now, it's that if and when I need to travel more than 200 miles without stopping at home, I need supporting infrastructure (and planning to find it) or to use another vehicle.

Research in carbon nanotubes only really matured from fundamental to applied in the past 15 years, so I don't think it's that surprising that the market isn't flooded with them yet.

We're already seeing some of the first commercially available CNT products, though. Take Vantablack[0] as an example. Or for non-commercial applications of CNT tech, see the CNT-reinforced bicycle frame that Floyd Landis used to win the Tour de France in 2006.[1]

When you consider that Bakelite was invented in 1907, but didn't see major manufacture until the 1940s, or that PVC was discovered in the 1870s, but didn't see commercial production until almost the same time, it doesn't seem like CNT are much behind the timeline for plastics. The entire chemical and plastics revolution didn't really take swing until the 1960s.

CNT probably aren't vaporware, and even if they were I don't know if that'd be too damning a story for Quantum Computers. Quantum Computing isn't "undiscovered" or "[not] proven to exist." Plenty of labs have produced physical quantum logic gates, and IBM even makes access to some of their real 5- and 16-qubit machines available online through the cloud.[2]

[0] https://en.wikipedia.org/wiki/Vantablack

[1] https://archive.fo/20120713061226/http://news.cnet.com/Carbo...

[2] https://quantumexperience.ng.bluemix.net/qx

Well, for certain classes of problems, it is certainly proven that quantum algorithms are faster than classical algorithms. And they're not totally theoretical pie-in-the-sky; you can implement a 5- or 16-qubit version of Shor's algorithm in IBM Q Experience[0] and run it on a real, physical quantum computer. The growing pains are in implementing quantum computers that are large enough to be of practical importance, and then building those systems several-fold larger for Quantum Fault Tolerance.[1]

Quantum computers are never going to replace or speed up every aspect of classical computation, but the idea of accessing them as a service for certain types of computation is probably not many decades off.

[0] https://quantumexperience.ng.bluemix.net/proxy/tutorial/full...

[1] https://en.wikipedia.org/wiki/Quantum_error_correction

Conway talks a little about this distinction in his talk at Google, at around 40 minutes: https://youtu.be/r1bDSlt1n9M?t=2401

The crux is that randomness is indistinguishable from a large collection of pre-determined hidden variables (imagine rolling all the dice for all of the outcomes of Physics ahead of time), but not all of the behaviors of quantum mechanics can be accounted for by (local) hidden variables (Bell's theorem.)