obastani

> He had received an offer at a start-up, and he would have loved to take it, but it paid half as much, and he felt locked into a lifestyle that made this pay cut impossible. “My wife laughed when I told her about it,” he said.

It sounds like part of the problem is his family (or at least with his wife)...

Perhaps they are earning $1.2M/yr because they never found anything they enjoyed to do with their time outside of work?

This is far from a new idea, and is called the "hidden-variable theory" [1]. In fact, it has been mathematically proven that locality and hidden variables are fundamentally incompatible in the context of quantum mechanics.

[1] https://en.wikipedia.org/wiki/Hidden-variable_theory

They do two things. First, you get right of way (so highest priority). Second, the only cars they do not stop are right turns. As other commenters have mentioned, the point is to improve efficiency, both since cars can turn right once there are no pedestrians, but also since pedestrians can cross for a longer period of time. On the other hand, it can also be annoying for cars, since they often never get a dedicated right turn without pedestrians.

Honestly I think stop lights aren't a big danger for pedestrians. Crossing at intersections with stop signs is far, far worse -- drivers have to pay attention to multiple directions of traffic as well as pedestrians.

To add to your list, a big one for me is the ritualization of coffee. Can't focus until you've had a cup!

Google's original rational for exiting China was to protest suppression of information. Has it done anyone any good? China simply developed their own search technology in house to replace Google, but with censorship. By doing so, they strengthened their local tech talent, and now they might outsource the technology they developed to other regimes. If Google had developed this tech from the beginning, they may have been able to avoid letting this tech get into the hands of regimes even worse than China.

Of course, we will never know the alternative outcome. And I appreciate the value of a company standing for moral principals even when it goes against practical considerations. But I think it is highly reductionist to suggest that yours is the only or "obvious" moral view on the situation.

Thanks for the response! I think you're right that it's something along these lines. I didn't find any further information about Euclidea, but for anyone interested, I did find some information about open source engines that achieve similar purposes [1,2]. Need to do some more digging...

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

[2] https://dev.geogebra.org/trac/wiki/TheoremProving

Does anyone know what are the algorithms behind this game, i.e., how does the game decide whether a given construction is valid? I'd appreciate any pointers!

It's probably easier to give an example of a number that I can't represent symbolically. However, I can't give you a specific example, since if I uniquely define a number then that definition is itself a symbolic representation of that number. But here is an example: a uniformly random real number on [0, 1]. Or, computationally speaking, uniformly sample an infinite number of Bernoulli random variables, and write down the binary representation. Of course, this process takes infinite time, which suggests that they are not representable.

For a proof, consider the fact that every representable number, by definition (I know I haven't given a formal definition), is represented by some finite string. But the number of finite strings is countable. So it must be a measure zero subset of [0, 1], so a random sample is representable with probability zero.

This is an issue if the non-AI company using humans is not profitable (supported by VC funding) with the goal of "eventually replacing humans with AI". If they are not transparent to investors about this point, then investors may have preferred to invest in a company that already has AI that works well enough (but not as well as the humans in the first company). Perhaps the first company will never be able to build the AI, leading to a market failure.

But... aren't they? It's one share one vote, not one employee one vote, but I think that still qualifies as democratic. Then, there are elected board members (the representatives) who hire the CEO (the prime minister?), etc.

I think "dynamic" is not such a bad term either -- the original application of DP was to optimization in dynamical systems.

Though the other comments are correct, I want to point out that you can get some nontrivial behavior with only linear functions. For example, low-rank matrix factorization is kind of like a neural network

f(x) = U * V * x,

where U is an n by k matrix and V is a k by m matrix, where k is much smaller than n and m. Basically, we are constraining the set of allowed linear transformations, which is a form of regularization. Convolutional layers in neural networks similarly restrict the allowed linear transformations.

Nevertheless, the power of linear neural networks is far less than that off nonlinear networks.

But a huge part of the problem is that many users of facial recognition software will not be experts on how to properly use the software. If the ACLU improperly used it, do you really trust law enforcement agencies to do a better job?

Perhaps we are looking at a different graph, but in the one I am looking at, blue is "sparse" (plateaus at 70) and orange is "dense" (very quickly hits 80). I believe "dense" means they are doing reward engineering.

That makes sense, thanks for clarifying!

I think this is quite impressive. I'm a bit confused about the section saying that "binary rewards can give good performance". Is it saying that binary rewards (instead of continuous rewards) work fine, but end-of-rollout rewards (instead of intermediate rewards such as kills) work poorly?

I think the "simple random search" algorithm in the paper you linked is not so simple -- it's basically using numerical gradient descent with a few bells and whistles invented by the reinforcement learning community in the past few decades. So perhaps it would be more fair to say that gradient descent (not random search) has proven to be a pretty solid foundation for model-free reinforcement learning.

Can you clarify why variance only scales linearly in the duration you are optimizing over? I would have expected it to be exponential, since the size of the space you are searching is exponential in the duration.

Very little of what goes into a current-generation self-driving car is based on machine learning [1]. The reason is exactly your point -- algorithmic approaches to self-driving are much safer and more predictable than a machine learning algorithms.

Instead, LIDAR should exactly identify potential obstacles to the self-driving car on the road. The extent to which machine learning is used is to classify whether each obstacle is a pedestrian, bicyclist, another car, or something else. By doing so, the self-driving car can improve its ability to plan, e.g., if it predicts that an obstacle is a pedestrian, it can plan for the event that the pedestrian is considering crossing the road, and can reduce speed accordingly.

However, the only purpose of this reliance on the machine learning classification should be to improve the comfort of the drive (e.g., avoid abrupt braking). I believe we can reasonably expect that within reason, the self-driving car nevertheless maintains an absolute safety guarantee (i.e., it doesn't run into an obstacle). I say "within reason", because of course if a person jumps in front of a fast moving car, there is no way the car can react. I think it is highly unlikely that this is what happened in the accident -- pedestrians typically exercise reasonable precautions when causing the road.

[1] https://www.cs.cmu.edu/~zkolter/pubs/levinson-iv2011.pdf

I thought it was pretty clear. They specifically say "certain items" were to be blocked, not the entire website.

D-Wave is not a "real" quantum computer in the sense that it is not asymptotically faster than a classical computer. Quantum computers with very few Quibits have existed for a while -- I believe the current record is 17 quibits [1]. I believe that to factor an n-bit integer, you need roughly n bits. So right now, we can factor 17 bit integers, which isn't very useful yet. However, I have read that for quantum simulations, it 30 quibit quantum computers might already be useful.

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

From that article, I got a very different impression (despite its clearly negative tone). The quote that really stood out to me was

"How long does it take to build such a factory? From that other car company, you may have heard that car factories could be brought from no plans to mass production in two years, no time at all, or whatever is less. I ask Mr. Demeunyck how long the Luqiao plant took from start to finish, and he proudly tells me that it “took only five years. They do things fast in China”... "In Europe, or in other places with tight permitting, it would take a few years longer,” Demeunyck says. If anyone tells you he has picked no location, and secured no permits, but wants to have a new car plant up and running in two years, call an ambulance."

Apparently, what they are really arguing is not that Tesla is bad at production, but that Tesla is trying to do something nearly impossible. In this light, the problems Tesla faces are not because they lack the skills or expertise to deliver, but problems that any other car company would face if they were to do things on the schedule Tesla is trying to achieve.

Thus, Tesla is really just taking a huge gamble -- they have a chance to pull off something very impressive, even if they fall 6 months or a year behind schedule. And if they fail, they failed because they took on a challenge that was simply too hard for any company to achieve. To me, it really resembles the gamble that SpaceX took, which paid off very well in the long run. So while I think there's a real chance that Tesla will fail like all of their detractors believe, I think there is also a real chance that they will be wildly successful.

Couldn't this finding be explained by the fact that the hospitals with lower readmissions tend to be poorer (since poorer patients are less likely to be able to afford care)? I only took a quick look, but I didn't see how the paper accounts for this and other possible causality issues. On the other hand, it's certainly plausible to me that lower readmissions can cause higher mortality.

Why would disolving the FDA fix the problem? If you don't do clinical trials to establish efficacy, then how will doctors know which treatments are effective and which ones are bogus? Randomized trials are important here due to the causality issues inherent in establishing efficacy in hindsight.

As many other commenters have pointed out, it is possible for net global wealth to increase (e.g., the increase in total real GDP). As far as I understand, the main reason this is possible is the development of new "technology" -- in the very general, economic sense of the word technology [0], which includes anything that increases the productivity of a single worker; this can include everything from new technology in the traditional sense to new processes (e.g., agile vs. waterfall for software development, or more optimal supply chains) and investment in infrastructure (e.g., better road networks).

For example, a single farmer today can not only produce many times more crop compared to a farmer 100 years ago, but also safely and efficiently distribute this produce to consumers across the globe. In general, assuming everyone in the economy works equally hard, the total production (i.e., the global net wealth) can increase as long as technology is increasing.

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

To be precise, negative energy prices are good for consumers with flexible energy usage (and bad for other consumers). For example, if you have a battery at home, you could charge it when energy prices are negative and use the battery when energy prices are high. Ideally, such a system of incentives would reduce peak consumption, making the energy market overall more efficient.

It's a bit difficult to explain the visualization without pictures, but I'll give it a shot.

The transpose is really about converting the matrix to operate on a different vector space, namely, the dual space. In particular, the dual space of a vector space V is the vector space of "linear functionals", which are linear functions

\phi: V -> R

A linear functional on R^2 looks like a gradient (the "gradient fill" gradient, not a calculus gradient). These gradients are in one-to-one correspondence to vectors in R^2. In particular, given a vector w \in R^2, the direction of the gradient is along the direction of v, and the speed with which the gradient is changing corresponds to the magnitude of w.

The precise mathematical correspondence is that (i) given a vector w \in R^2, the function

f_w(v) = <w, v>

is a linear function (here, <,> is the inner/dot product), and (ii) every linear function has this form. Now, note that f_w is exactly multiplication by the transpose w^T of w! In particular,

f_w(v) = w^T v

More generally, for any linear map A : V -> W, the adjoint A* of A is defined to be the linear map from the dual space of W to the dual space of V that satisfies

<w, A v> = <A* w, v>

The transpose A^T is the adjoint of A when V is a finite-dimensional real vector space:

<A^T w, v> = (A^T w)^T v = w^T A v = <w, A v>

In summary, you can try to visualize A^T as a linear map "acting" on dual vectors. For example, let v \in R^2 and let w be a dual vector (i.e., a gradient), and suppose that A rotates v clockwise by 90 degrees. To preserve the inner product <w, A v>, A^T rotates w counter-clockwise by 90 degrees.

I agree that having possible examples in mind is a great way to learn mathematics. There are whole books on useful counterexamples, e.g.

https://www.amazon.com/Counterexamples-Analysis-Dover-Books-...

These counterexamples are sometimes a bit involved, but I find they are often useful for understanding the purpose of the technical assumptions that accompany many theorems.

I'm not supporting the GP's viewpoint, but I think their point is that we don't have enough information to attribute warming specifically in the Arctic to human causes. I don't know very much about the subject, but given that the Arctic is a fairly large region, I'd be inclined to guess that this viewpoint is still wrong.