bkcooper

Yea, it's boring for professor on the thesis committee to read, but this sort of stuff is very valuable for students and postdocs, who often struggle to reproduce poorly-documented results from other labs.

There's both absolute and relative poor documentation. The brevity of most scientific publishing (coupled with the lack of any sort of external incentive for clarity, as you point out) means that even papers that are good, by the standard of scientific papers, can be a real bear to turn into working examples. I personally feel that a lot of what makes various labs good is knowledge of the things that aren't going into the papers. A dissertation, with (potentially) a lot more room to go into details, can serve both as a source of this information for external readers, as well as a convenient internal reference for people in the lab who need to dig up some details of their prehistory.

edit: to clarify, my perspective on this is coming from physics. In other fields, there's more of an emphasis on explicit protocols and "methods sections" that presumably help with this sort of thing.

I would argue this is already essentially the dominant model. There are already more plausible projects than there is money to fund all of them, and the credibility of the person doing the proposing appears to be pretty decisive in terms of whether you actually get funding or not. It also isn't that uncommon for people to be a little dishonest about the funding and play games to shuffle it around to accomplish different ends. It would be nice to eliminate that layer, but I don't think it would be a revolutionary advance in quality or anything.

Yes, they are quite close to surface code thresholds. However, being close to the thresholds means that you need to be on the pessimistic side of how many physical qubits you need per logical qubit. As I mentioned in my other comment, even Martinis (whose group is, as you say, very good) is still several orders of magnitude short in terms of the number of physical qubits needed to implement one logical qubit, let alone the hundreds to thousands of logical qubits necessary to do interesting computations.

Packing more qubits in seems to me like it's going to be a very challenging problem. Looking at Martinis's recent Nature paper, 9 qubits are taking up something like 2 x 4 mm. Making these smaller would be nontrivial for many reasons: they each have their own microwave coupling line (which requires a certain amount of length); the coherence properties of superconducting qubits seem to care about how much surface you have relative to bulk metal, which is a loser for shrinking the devices; presumably jamming them together presents crosstalk issues; etc. Realistically, you also need to add a bunch of other types of electronics down there to handle multiplexing a la D-Wave. I don't know that these obstacles are insuperable, but I'm definitely a little skeptical that this road leads to useful technology.

I'm not sure what kind of scaling you mean in this context, but let me try to answer this.

I think the most important result in error correction is the threshold theorem, which (roughly) says that there exists an N such that if you have qubits that are basically error free for N gate operations, you can implement fault tolerant quantum computation. The particular threshold N depends on specifics of the architecture and error encoding that you use. What also depends on the encoding is the number M of logical qubits per physical qubit.

As far as I know, it is the case that once you've chosen an encoding, the number of physical qubits needed to realize Q logical qubits should just be MQ, so in that sense the scaling is linear. However, different encodings have wildly different thresholds. Shor's original code that could handle bit flip and phase errors had M=9, I believe, but I don't think the threshold was very good. For a long time, a popular figure quoted for the N you needed for error correction was ~10^4. I think there are state of the art codes ("surface codes") that can get the error threshold down to 10^2 or so, which is close to what's achievable in systems with small numbers of superconducting qubits (which is what IBM and Google are looking at.) But, the number M of physical qubits per logical qubit for a surface code is very, very high, I think on the order of 10^5. (It's been a while. EDIT: I got curious and looked up numbers from a 2012 paper; 10^4 is a better number here. That's still three orders of magnitude more qubits than is typical in state of the art circuits.) So in that sense, there is still a scaling problem.

Perhaps this will help clarify why it is different. In the classical case, adding one more spin is doubling the size of your set. In the quantum mechanical case, adding one more spin instead doubles the dimensionality of the space of possible states.

What you're thinking of is a a single basis state. The state of a 300 spin system can, in general, have any superposition of the 2^300 basis states available. These superpositions also mean that even in the single spin case, you're not representing states exactly, but truncating the representation at some point that is good enough.

A final word: where ever you go if you leave science you are likely to face the same fight, because it is in the nature of modern life and modern organizations.

I've thought before that much of the criticism of academia (which I've certainly engaged in) mostly reflects disappointment that it was ultimately not all that different from the rest of the world.

I think you'd need a very radical overhaul of how science works to replace the scorekeeping aspect. I don't think a Github CV is a good analogy to how this could work, because academic science is interested in hiring leaders (i.e. people who can get funded), not contributors. I think realistically you would need to change how science rewards/emphasizes certain activities first, and then publishing would follow. That would be a good outcome for science, but I think it'll be awfully hard to get out of this equilibrium.

edit: I'm curious what you think of the PubPeer model. That's obviously different from what you're envisioning, but thematically I think there are some similarities.

I still end up doing this in LaTeX, only now there's a time lag.

True, but conda has its own virtual environments that work quite nicely. You can make a new one with conda create and switch between them with activate/deactivate. You even get the same sort of interaction with pip (for the packages you want that aren't in conda or binstar), and I've been able to use e.g. Emacs packages that expect to work with virtualenv and have (mostly) had success using conda's envs instead.

Man, no kidding. It feels like a very inelegant info dump. While I realize those tend to show up in Stephenson books at some point or another, it's a rough way to start.

I liked a lot about Anathem but found Reamde unmemorable. I'd love to see a return to form.

A good example from this article would be M-q for auto-fill. I was super excited when I discovered that command after spending some time redoing line breaks manually before thinking "I'm sure somebody has solved this problem in Emacs before." I'm sure there are tons of useful text editing hacks that would be really useful that just haven't even occurred to me yet.

Also, with smaller labs you can afford to hire more PIs and give a career path to more of the postdocs that you are "training."

That would probably be better for the postdocs, but I think it would have a negative impact on the overall production of new science. A PI who runs a big lab with a proven track record of success is going to be a better investment most of the time. Splitting the labs up also means that even more people are exposed to the type of grantwriting inefficiencies that you describe above.

What I like about the superdoc idea is that right now (modulo some pretty rare positions) science feels like "PI or bust." I think there are lots of people who can make useful scientific contributions but probably shouldn't be PIs. Figuring out something for these people would be good for them and (IMO) good for the system.

I'm with the GP here, your experience sounds very weird to me. Tuition remission is a pretty standard benefit, often handled by the department at the beginning, and then eventually by your advisor (but often by that point you're not taking much in the way of classes anyway.)

Sure, but the argument is that we're training many more people for those sorts of careers than can realistically have them. Having the filter come at the beginning rather than the end is way better for people who are going to get screened out one way or the other.

If the need for these positions to be filled is sufficiently low that your group can be that picky for almost a year, it would seem demand is pretty low, thus poor evidence for a shortage.

If postdocs are so prized, then one obvious solution is to reward them. Both the 2014 National Academies report and earlier reports urged US lab heads to consider creating senior staff scientist, or ‘superdoc’, positions. These would be higher-paid, permanent jobs for talented postdocs who have no desire to start their own labs.

This type of thing becoming more common would fix a lot of academic career problems. I think it would also be a very good first step in actually reducing the number of incoming PhD students, since the productivity of such a researcher will usually exceed the amount of graduate student you could replace them with.

The warning about choking on large data literals (which are probably one of the places where prettifying would be most useful, at least to me) also seems ominous.

edit: for my personal use, I tend to use flycheck with flake8 in emacs. This keeps me honest. Is the primary use case for something like this cleaning your own code or other people's?

What the author says is: In any case, the effect certainly isn’t that the more people know, the less they think they know. And he's right; the data he shows does not demonstrate that. The more talented people in those plots, at worst, think they know as much as the less talented.

What the graphs do demonstrate is that the less talented people consistently overestimate their ability, whereas the most talented people underestimate. So what's actually negatively correlated isn't actual knowledge and perceived knowledge, it's actual knowledge and the gap between actual knowledge and perceived knowledge.

In conclusion,

Personally, I'm getting sick of this kind of uber-skeptic mentality that's popular here and on reddit. I think there's something ego gratifying for the INTJ male about being this loud-mouth contrarian who tells everyone they're wrong.

Um...

I don't think that's true, at least in areas where physicists or engineers are actually going to be trying to use the relevant math. I think the difference is mostly in norms of written presentation. In physics, the concrete motivating example is often seen as an important thing to present; in math, it seems more typical to polish your definitions until they're as minimal and general as can be and start the presentation there. This often obscures the more concrete examples, but I've found plenty of times that if you read or talk to mathematicians in informal settings they can be pretty good about the examples that motivate the ideas.

You've heard the expression "let's get busy"? Well this is a textbook that gets biz-zay. Consistently and thoroughly.

You actually can. multiprocessing defines a few classes (e.g. Array) that store their data in a mmap kept in the module. Data there will be visible to all processes, and you can even have a numpy array as a view to this data.

That said, my attempts at working with this recently have been pretty painful. multiprocessing tries, but at least in Python 2.7 does not succeed in abstracting away the differences between Unix forks and Windows spawns. This results in a lot of weird issues: things running differently in command line vs. console IPython vs. IPython notebook; the need to structure your code to avoid pickling errors; etc. This is probably the biggest portability issue I've encountered with Python thus far (using it mostly for scientific problems.)

I'm aware that there are solutions to some of these issues out there, but it's too bad that the standard library implementation has these issues.

There's also a cost associated with having the most visible discussion of human travel to Mars be a scam.